Datasets:
ccdv
/
arxiv-classification

Dataset card Data Studio Files Community
4
text
stringlengths
2.85k
2.55M
label
class label
11 classes
Submitted to the Annals of Statistics PREDICTION ERROR AFTER MODEL SEARCH By Xiaoying Tian Harris, arXiv:1610.06107v2 [math.ST] 9 Feb 2017 Department of Statistics, Stanford University Estimation of the prediction error of a linear estimation rule is difficult if the data analyst also use data to select a set of variables and construct the estimation rule using only the selected variables. In this work, we propose an asymptotically unbiased estimator for the prediction error after model search. Under some additional mild assumptions, we show that our estimator converges to the true prediction error in L2 at the rate of O(n−1/2 ), with n being the number of data points. Our estimator applies to general selection procedures, not requiring analytical forms for the selection. The number of variables to select from can grow as an exponential factor of n, allowing applications in high-dimensional data. It also allows model misspecifications, not requiring linear underlying models. One application of our method is that it provides an estimator for the degrees of freedom for many discontinuous estimation rules like best subset selection or relaxed Lasso. Connection to Stein’s Unbiased Risk Estimator is discussed. We consider in-sample prediction errors in this work, with some extension to out-of-sample errors in low dimensional, linear models. Examples such as best subset selection and relaxed Lasso are considered in simulations, where our estimator outperforms both Cp and cross validation in various settings. 1. Introduction. In this paper, we consider a homoscedastic model with Gaussian errors. In particular, (1.1)  ∼ N (0, σ 2 I), y = µ(X) + , where the feature matrix X ∈ Rn×p is considered fixed, y ∈ Rn is the response, and the noise level σ 2 is considered known and fixed. Note the mean function µ : Rn×p → Rn is not necessarily linear in X. Prediction problems involve finding an estimator µ̂ which fits the data well. We naturally are interested in its performance in predicting a future response vector that is generated from the same mechanism as y. Mallows (1973) provided an unbiased estimator for the prediction error when the estimator is linear µ̂ = Hy, MSC 2010 subject classifications: Primary 62H12, 62F12; secondary 62J07, 62F07 Keywords and phrases: prediction error, model search, degrees of freedom, SURE 1 2 XIAOYING TIAN HARRIS where H is an n × n matrix independent of the data y. H is often referred to as the hat matrix. But in recent context, it is more and more unrealistic that the data analyst will not use the data to build a linear estimation rule. H, in other words, depends on y. In this case, is there still hope to get an unbiased estimator for the prediction error? In this article, we seek to address this problem. Some examples that our theory will apply to are the following. In the context of model selection, the data analyst might use some techniques to select a subset of the predictors M to build the linear estimation rules. Such techniques can include the more principled methods like LASSO (Tibshirani 1996), best subset selection, forward stepwise regression and Least Angle Regression (Efron et al. 2004) or some heuristics or even the combination of both. After the selection step, we simply project the data onto the column space of XM , the submatrix of X that consists of M columns, and use that as our estimation rule. Specifically, (1.2) µ̂(y; X) = HM · y, T T HM = PM = XM (XM XM )−1 XM , M = M̂ (y) where M̂ can be any selection rule and PM is the projection matrix onto the column space of XM . In the case when M is selected by the LASSO, µ̂ = XM β̄M (y), and β̄M (y) is known as the relaxed LASSO solution (Meinshausen 2007). In general, the range of M̂ does not have to be 2{1,...,p} , the collection of all subsets of {1, 2, . . . , p}, but we assume the hat matrix HM̂ depends on the data y only through M̂ . In this sense, M̂ is the abstraction of the data-driven part in H. This paper will study the prediction error of µ̂ = HM̂ · y. In this paper, we want to estimate the prediction error for µ̂,   (1.3) Err = E kynew − HM̂ · yk22 , ynew ∼ N (µ(X), σ 2 I) ⊥ y. There are several major methods for estimating (1.3) (Efron 2012). Penalty methods such as Cp or Akaike’s information criterion (AIC) add a penalty term to the loss in training data. The penalty is usually twice the degrees of freedom times σ 2 . Stein’s Unbiased Risk Estimator (Stein 1981) provides an unbiased estimator for any estimator that is smooth in the data. For non-smooth estimation rules, Ye (1998) use perturbation techniques to approximate the covariance term for general estimators. PREDICTION ERROR AFTER MODEL SEARCH 3 Nonparametric methods like cross validation or related bootstrap techniques provide risk estimators without any model assumption. Methods like Cp assume a fixed model. Or specifically, the degrees of freedom is defined as df = tr(H) for fixed H. Stein’s Unbiased Risk Estimator (SURE) only allows risk estimation for almost differentiable estimators. In addition, computing the SURE estimate usually involves calculating the divergence of µ̂(y). This is difficult when µ̂(y) does not have an explicit form. Some special cases have been considered. Works by Zou et al. (2007), Tibshirani & Taylor (2012) have computed the “degrees of freedom” for the LASSO estimator, which is Lipschitz. But for general estimators of the form µ̂ = HM̂ y, where HM̂ depends on y, µ̂ might not even be continuous in the data y. Thus analytical forms of the prediction error are very difficult to derive Tibshirani (2014), Mikkelsen & Hansen (2016). Nonparametric methods like cross validation are probably the most ubiquitous in practice. Cross validation has the advantage of assuming almost no model assumptions. However, Klement et al. (2008) shows that cross validation is inconsistent for estimating prediction error in high dimensional scenarios. Moreover, cross validation also includes extra variation from having a different X for the validation set, which is different from the fixed X setup of this work. Efron (2012) also points out that the model-based methods like Cp , AIC, SURE offer substantially better accuracy compared with cross validation, given the model is believable. In this work, we introduce a method for estimating prediction errors that is applicable to general model selection procedures. Examples include best subset selection for which prediction errors are difficult to estimate beyond X being orthogonal matrices (Tibshirani 2014). In general, we do not require HM̂ to have any analytical forms. The main approach is to apply the selection algorithm M̂ to a slightly randomized response vector y ∗ . This is similar to holding out the validation set in cross validation, with the distinction that we do not have to split the feature matrix X. We can then construct an unbiased estimator for the prediction error using the holdout information that is analogous to the validation set in cross validation. Note that since y ∗ would select a different model from y, this estimator will not be unbiased for the prediction error of µ̂. However, If the perturbation in y ∗ is small and we repeat this process multiple times so the randomization averages out, we will get an asymptotically unbiased and consistent estimator for the prediction error of µ̂ = HM̂ (y) y, which is the original target of our estimation. Moreover, since our estimator is model based, it also enjoys more efficiency than cross validation. In fact, we prove that under mild conditions on the selection procedure, 4 XIAOYING TIAN HARRIS our estimator converges to the true prediction error as in (1.3) in L2 at the 1 rate of n− 2 . This automatically implies consistency of our estimator. The Cp estimator, on the other hand, converges in L2 at the rate of n−1 for fixed 1 hat matrix H. So compared with Cp , our estimator pays a small price of n 2 for the protection against any “data-driven” manipulations in choosing the hat matrix HM̂ for the linear estimation rules. 1.1. Organization. The rest of the paper is organized as follows. In Section 2, we introduce our procedure for unbiased estimation for a slightly different target. This is achieved by first randomizing the data and then constructing an unbiased estimator for the prediction error of this slightly different estimation rule. We then address the question of how randomization affects the accuracy of our estimator for estimating the true prediction error. There is a clear bias-variance trade-off with respect to the amount of randomization. We derive upper bounds for the bias and the variance in Section 3 and propose an “optimal” scale of randomization that would make our estimator converge to the true prediction error in L2 . Since the unbiased estimator constructed in Section 2 only uses one instance of randomization. We can further reduce the variance of our estimator by averaging over different randomizations. In Section 4, we propose a simple algorithm to compute the estimator after averaging over different randomizations. We also discuss the condition under which our estimator is equal to the SURE estimator. While SURE is difficult to compute both in terms of analytical formula and simulation, our estimator is easy to compute. Using the relationship between prediction error and degrees of freedom, we also discuss how to compute the “search degrees of freedom”, a term used in Tibshirani (2014) to refer to the degrees of freedom of estimators after model search. Finally, we include some simulation results in Section 5 and conclude with some discussions in Section 6. 2. Method of estimation. First, we assume the homoscedastic Gaussian model in (1.1), y ∼ N (µ(X), σ 2 I), and we have a model selection algorithm M̂ , M̂ : Rn × Rn×p → M, (y, X) 7→ M. As we assume X is fixed, we often use the shorthand M̂ (y), and assume M̂ : Rn → M, y 7→ M, where M is a finite collection of models we are potentially interested in. The definition of models here is quite general. It can refer to any information we extract from the data. A common model as described in the introduction PREDICTION ERROR AFTER MODEL SEARCH can be a subset of predictors of particular interest. value of the observation y and maps it to a set of also the inverse image of M̂ −1 induces a partition will discuss this partition further in Section 3. However, instead of using the original response we use its randomized version y ∗ , (2.1) y ∗ = y + ω, 5 In such case, M̂ takes a selected variables. Note on the space of Rn . We variable y for selection, ω ∼ N (0, ασ 2 I) ⊥ y. For a fixed α > 0, after using y ∗ to select a model M , we can define prediction errors analogous to that defined in (1.3), (2.2) h i ynew ∼ N (µ(X), σ 2 I) ⊥ (y, ω). Errα = E kynew − HM̂ (y+ω) yk22 , The subscript α denotes the amount of randomization added to y. Note that although randomization noise ω is added to selection, Errα integrates over such randomization and thus are not random. The prediction error Err as defined in (1.3) corresponds to the case where we set α = 0. In this section, we show that can get an unbiased estimator for Errα for any α > 0. Before we introduce the unbiased estimator, we first introduce some background on randomization. 2.1. Randomized selection. It might seem unusual to use y ∗ for model selection. But actually, using randomization for model selection and fitting is quite common – the common practice of splitting the data into a training set and a test set is a form of randomization. Although not stressed, the split is usually random and thus we are using a random subset of the data instead of the data itself for model selection and training. The idea of randomization for model selection is not new. The field of differential privacy uses randomized data for database queries to preserve information Dwork (2008). This particular additive randomization scheme, y ∗ = y + ω is discussed in Tian & Taylor (2015). In this work, we discover that the additive randomization in (2.1) allows us to construct a vector independent of the model selection. This independent vector is analogous to the validation set in data splitting. To address the question of the effect of randomization, we prove that Err and Errα are close for small α > 0 under mild conditions on the selection procedures. In other words, since we have an unbiased estimator for Errα for any α > 0, when α goes to 0, its bias for Err will diminish as well. For details, see Section 3. In addition, Section 5 also provides some evidence in simulations. 6 XIAOYING TIAN HARRIS 2.2. Unbiased estimation. To construct an unbiased estimator for Errα , we first construct the following vector that is independent of y ∗ , y− = y − (2.3) 1 ω. α Note this construction is also mentioned in Tian & Taylor (2015). Using the property of Gaussian distributions and calculating the covariance between y − and y ∗ = y + ω, it is easy to see y − is independent of y ∗ , and thus independent of the selection event {M̂ (y ∗ ) = M }. Now we state our first result that constructs an unbiased estimator for Errα for any α > 0. Theorem 2.1 (Unbiased Estimator). Suppose y ∼ N (µ(X), σ 2 I) is from the homoscedastic Gaussian model (1.1), then (2.4) dα = y − − H Err M̂ (y ∗ ) y 2 2 + 2 tr(HM̂ (y∗ ) )σ 2 − 1 2 nσ α is unbiased for Errα for any α > 0. Proof. First notice y= 1 α y∗ + y−, 1+α 1+α y = µ(X) + , if we let ∗ = y ∗ − µ(X) and − = y − − µ(X), then (2.5) = 1 ∗ α −  +  . 1+α 1+α 2 Note ∗ ⊥ − and ∗ ∼ N (0, (1 + α)σ 2 I), and − ∼ N (0, 1+α α σ I). With this, we first define the following estimator for any α > 0 and any M ∈ M, (2.6) err c α (M ) = y − − HM y 2 2 + 2 tr(HM )σ 2 − 1 2 nσ . α We claim that err c α (M ) is unbiased for the prediction error conditional on ∗ {M̂ (y ) = M } for any M ∈ M and any α > 0. Formally, we prove h i h i (2.7) E err c α (M ) | M̂ (y ∗ ) = M = E kynew − HM · yk2 | M̂ (y ∗ ) = M . To see (2.7), we first rewrite (2.8) h i h i E kynew − HM yk22 | M̂ (y ∗ ) = M = E kµ − HM yk2 | M̂ (y ∗ ) = M + nσ 2 . PREDICTION ERROR AFTER MODEL SEARCH Now we consider the conditional h i 2 E y − − HM y 2 | M̂ (y ∗ ) = M h i =E kµ − HM yk2 | M̂ (y ∗ ) = M + h i =E kµ − HM yk2 | M̂ (y ∗ ) = M + h i =E kµ − HM yk2 | M̂ (y ∗ ) = M + 7 expectation of err c α (M ). Note h i 1+α 2 nσ − 2E (− )T HM y | M̂ (y ∗ ) = M α    1+α 2 2α nσ − tr HM E y − (− )T α 1+α 1+α 2 nσ − 2 tr(HM )σ 2 α The equalities use the decomposition (2.5) as well as the fact that y ∗ ⊥ − . Comparing this with (2.8), it is easy to see (2.7). Moreover, marginalizing dα in (2.4) is unbiased for Errα . over M̂ (y ∗ ), it is easy to see Err In fact, using the proof for Theorem 2.1, we have a even stronger result dα . than the unbiasedness of Err dα is not only unbiased for the prediction error marginally, Remark 2.2. Err dα is also unbiased but conditional on any selected event {M̂ (y ∗ ) = M }, Err for the prediction error. Formally, h i h i dα | M̂ (y ∗ ) = M = E kynew − HM · yk2 | M̂ (y ∗ ) = M . E Err This is easy to see with (2.7) and h i h i dα | M̂ (y ∗ ) = M = E err E Err c α (M ) | M̂ (y ∗ ) = M . dα in (2.6) has some resemblance to the usual Cp The simple form of Err formula for prediction error estimation, with 2 tr(HM̂ )σ 2 being the usual correction term for degrees of freedom in the Cp estimator. The additional term nσ 2 /α helps offset the larger variance in y − . 3. Randomization and the bias-variance trade-off. We investigate the effect of randomization in this section. In particular, we are interested in the difference Errα − Err for small α > 0 and the variance of our dα as a function of α. estimator Err There is a simple intuition for the effects of randomization on estimation of prediction error. Since y ∗ = y + ω, ω ∼ N (0, ασ 2 ), the randomized response vector y ∗ which we use for model selection will be close to y when the randomization scale α is small. Intuitively, Errα should be closer to Err when α decreases. On the other hand, the independent vector y − = y − ω/α 8 XIAOYING TIAN HARRIS which we use to construct the estimator for the prediction error is more variant when α is small. Thus, there is a clear bias-variance trade-off in the choice of α. We seek to find the optimal scale of α for this trade-off. The unbiased estimation introduced in Section 2.2 does not place any assumptions on the hat matrix HM̂ or the selection procedure M̂ . However, in this section we restrict the hat matrices to be those constructed with the selected columns of the design matrix X. This is not much of a restriction, since these hat matrices are probably of most interest. Formally, we restrict M̂ to be a selection procedure that selects an “important” subset of variables. That is M̂ : Rn → M ⊆ 2{1,...,p} . Without loss of generality, we assume M̂ is surjective. Thus the number of potential models to choose from is |M| which is finite. Moreover, the map M̂ induces a partition of the space of Rn . In particular, we assume (3.1) Ui = M̂ −1 (Mi ) ⊆ Rn , i = 1, . . . , |M|, where M = {M1 , . . . , M|M| } are different models to choose from. It is easy to see that |M| a Ui = R n , i=1 and we further assume int(Ui ) 6= ∅ and ∂Ui has measure 0 under the Lebesgue measure on Rn . Now we assume the hat matrix is a constant matrix in each of the partition Ui . In particular, (3.2) HM̂ (y) = |M| X HMi I (y ∈ Ui ) . i=1 The most common matrix is probably the projection matrix onto the column space spanned by a subset of variables. Formally, we assume T X )−1 X T , Assumption 3.1. For any M ∈ M, we assume HM = XM (XM M M where XM is the submatrix of X with M as the selected columns. It is easy to see, HM is symmetric and 2 HM = HM , ∀ M ∈ M. Moreover, we also assume that M̂ does not select too many variables to include in a model, and the rank of the matrix rank(HM ) is less than number of variables in M . Specifically, 9 PREDICTION ERROR AFTER MODEL SEARCH Assumption 3.2. For any M ∈ M, rank(HM ) ≤ |M | ≤ K. Furthermore, we assume K grows with n at the following rate, √ (3.3) K 2 log p = O( n), where p is the number of columns in X. Assumption 3.2 requires that none of the models M ∈ M is too large. However, its size can grow with n at some polynomial rate. For penalized regression problems, choices of the penalty parameter λ such that the solution is sparse has been studied in Negahban et al. (2009). Also Assumption 3.2 allows p to grow as an exponential factor of the number of data points n. Thus it allows applications in the high dimensional setting where p  n. We also assume the model selection procedure M̂ to have reasonable accuracy in identifying the underlying model. Assumption 3.3. Suppose M̂ satisfies: "  # 2 2 (3.4) E = O(n), µ − HM̂ (y0 ) · µ y 0 ∼ N (µ, τ ), 2 where σ 2 ≤ τ ≤ (1 + δ)σ 2 for some small constant δ > 0. Assumption 3.3 assumes that the subspace spanned by XM̂ is a good representation of the underlying mean µ. But it does not require µ to be in this subspace. Namely, we allow model misspecifications. Remark 3.4. In the context of sparse linear models µ = Xβ 0 , we assume β 0 has support Γ. Then µ − HM̂ · µ 2 0 = (I − HM̂ )XΓ\M̂ βΓ\ M̂ 0 ≤ XΓ\M̂ βΓ\ M̂ 2 2 Assuming X is normalized to have column length √ n and that for any j ∈ Γ, we have that µ − HM̂ · µ 2 2 2 = Γ\M̂ log p. βj0 =O q log p n  10 XIAOYING TIAN HARRIS 2 Thus Assumption 3.3 is close to placing a condition like Γ\M̂ √ Op ( n), which is analogous to that in Equation (3.3). log p = With these conditions above, we show in the following that the bias of d Errα is O(α) (Theorem 3.5) and its variance is O((nα2 )−1 ) (Theorem 3.10). There is a clear bias variance trade-off with regard to the choice of α, which we will discuss in more detail in Section 3.3. The proofs of the theorems uses some well known results in extreme value theories. dα is introduced by the fact that selection is 3.1. Bias. The bias of Err ∗ performed with y , the randomized version of y. In fact, this is the only source for the bias. However, for small perturbations, the resulting bias will be small as well. Formally, we have the following theorem. Theorem 3.5. Suppose Assumptions 3.1, 3.2 and 3.3 are satisfied, then dα is bounded by: the bias of Err 1 hd i 1 E Errα − Err = |Errα − Err| ≤ Cα, n n for α < δ. where C is a universal constant and δ > 0 is a small constant defined in Assumption 3.3. Essential to the proof of Theorem 3.5 is that the performance of the estimation rule µ̂(y) = HM̂ (y) y is resistant to small perturbations on y. This is true under the assumptions introduced at the beginning of Section 3. Formally, we have Lemma 3.6. Suppose our hat matrix is of the form in (3.2) and Assumptions 3.1-3.3 are satisfied. Then for α < δ and ω ∼ N (0, ασ 2 I), we have i h i 1 h E kµ̂(y + ω) − µk2 − E kµ̂(y) − µk2 ≤ C1 · α, n where C1 is a universal constant and δ > 0 is a small constant defined in Assumption 3.3. The first expectation is taken over (y, ω) and the second expectation is taken over y. With Lemma 3.6, it is easy to prove Theorem 3.5. 11 PREDICTION ERROR AFTER MODEL SEARCH Proof. First notice that HM̂ (y∗ ) y = HM̂ (y∗ ) y ∗ − HM̂ (y∗ ) ω = µ̂(y + ω) − HM̂ (y+ω) ω. Thus we have 1 |Errα − Err| n     2 2 1 = E HM̂ (y+ω) y − µ − E HM̂ (y) y − µ n  i h i 1 1 h ≤ E kµ̂(y + ω) − µk2 − E kµ̂(y) − µk2 + E HM̂ (y+ω) ω n n h i h i i 1 1 h ≤ E kµ̂(y + ω) − µk2 − E kµ̂(y) − µk2 + E kωk2 n n ≤(C1 + σ 2 )α 2  The proof of Lemma 3.6 relies on the following lemma which will also be dα in Section 3.2. The proofs of used for the bound on the variance of Err both Lemma 3.6 and Lemma 3.7 are deferred to Section 7. Lemma 3.7. value in Suppose Z ∼ N (0, In×n ), and HM̂ is a hat matrix that takes {HMi , i = 1, . . . , |M|} , where |M| is the total number of potential models to choose from. The value of HM̂ may depend on Z. If we further assume HM̂ satisfies Assumptions 3.1 and 3.2, then   2   2 2 E HM̂ Z 2 ≤ 18 K log(K|M|) . 3.2. Variance. In this section, we discuss the variance of our estimator d Errα . As previously discussed at the beginning of the section, it is intuitive dα will increase as α decreases. Before we establish that the variance of Err dα with respect to α, we first quantitative results about the variances of Err state a result on the variances of Cp estimators. This will provide a baseline of comparison for the increase of variance due to the model selection procedure. Formally, suppose y ∼ N (µ, σ 2 I) and the hat matrix HM̂ is constant independent of the data y, that is HM̂ = H. Then the Cp estimator Cp = ky − Hyk22 + tr(H)σ 2 is unbiased for the prediction error. Moreover 12 XIAOYING TIAN HARRIS Lemma 3.8. jection matrix, 1 Suppose y ∼ N (µ, σ 2 I) and HM̂ = H is a constant pro- Var [Cp ] = 2 tr [(I − H)] σ 4 + 4 k(I − H)µk22 σ 2 . Furthermore, if we assume kµk22 = O(n),     Cp 1 (3.5) Var =O n n The proof of Lemma 3.8 uses the following lemma whose proof we defer to Section 7: Lemma 3.9. Let Z ∼ N (0, In×n ) and A ∈ Rn×n be any fixed matrix, h i  Var kAZk22 = 2 tr A4 Now we prove Lemma 3.8. Proof. h i h i Var ky − Hyk2 = Var k(I − H)k2 + k(I − H)µk2 + 2µT (I − H)2  h i = Var k(I − H)k2 + 2µT (I − H)2  h i   = Var k(I − H)k2 + 4Var µT (I − H)2    2 = 2 tr (I − H)4 σ 4 + 4 (I − H)2 µ σ 2 = 2 tr [(I − H)] σ 4 + 4 k(I − H)µk2 σ 2 The last equality is per Assumption 3.1, as it is easy to see that I − H is also a projection matrix. Finally, since tr(I − H) ≤ n, k(I − H)µk22 ≤ kµk22 , it is easy to deduce the second conclusion. C 1 Lemma 3.8 states that the variation in np is of order Op (n− 2 ). In the dα is compared following, we seek to establish how inflated the variance of Err dα to Cp . Theorem 3.10 gives an explicit upper bound on the variance of Err with respect to α. In fact, it is simply of order O((nα2 )−1 ). Formally, we have 1 Lemma 3.8 is inspired by a talk given by Professor Lawrence Brown, although the author has not been able to find formal proof for reference. 13 PREDICTION ERROR AFTER MODEL SEARCH Theorem 3.10. Suppose Assumptions 3.1, 3.2 and 3.3 are satisfied, then # "   dα 1 Err =O . Var n nα2 Compared with the variance of the Cp /n in (3.5), we pay a price of α−2 but allows our hat matrix to be dependent on the data y. Particularly, if we dα will have choose α−1 = o(n1/2 ), our estimator for the prediction error Err diminishing variance and be consistent. To prove Theorem 3.10, we again use Lemma 3.7 which is stated in Section 3.1. Proof. First notice that y = y − − α1 ω, thus y − − HM̂ (y∗ ) · y 2 = y − − HM̂ (y∗ ) · y − 2 + 1 HM̂ (y∗ ) ω α2 2 . Therefore, we have,   h i 2 2 dα = Var y − − H Var Err · y + 2 tr(H )σ M̂ (y ∗ ) M̂ (y ∗ )    i h 2 1 2 − − ≤2Var y − HM̂ (y∗ ) y + 2 tr HM̂ (y∗ ) σ + 2Var 2 HM̂ (y∗ ) ω α 2  First using the decomposition for conditional variance, we have  i  h 2 − − Var y − HM̂ (y∗ ) y + 2 tr HM̂ (y∗ ) σ 2    − − 2 2 ∗ + 2 tr (HM ) σ M̂ (y ) = M =E Var y − HM y   2 + Var (I − HM̂ (y∗ ) )µ 2      2 1 2 4 1 ≤2n 1 + 1+ σ + 4E (I − HM̂ (y∗ ) )µ σ2 α α 2   2 + Var (I − HM̂ (y∗ ) )µ 2 Note for the last inequality, we used both Lemma 3.8 as well as the independence relationships: y ∗ ⊥ y − . Furthermore, with Assumption 3.3, we have  h i  2 − − Var y − HM̂ (y∗ ) y + 2 tr HM̂ (y∗ ) σ 2 ≤ O(nα−2 ). 14 XIAOYING TIAN HARRIS Moreover, per Lemma 3.7 we have "   2 1 1 Var 2 HM̂ (y∗ ) ω ≤ 4E HM̂ (y∗ ) ω α α  2 2 # ≤ 18 [K log(K|M|)]2 σ 4 , α2 where |M| is the number of potential models in M. Finally, since we would only choose models of size less than K (Assumption 3.2), |M| ≤ pK and with the rate specified in Assumption 3.2 we have the conclusion of the theorem. 3.3. Bias-variance trade-off and the choice of α. After establishing Theorem 3.5 and Theorem 3.10, we combine the results and summarize the bias-variance trade-off in the following corollary. Corollary 3.11. then we have Suppose Assumptions 3.1, 3.2 and 3.3 are satisfied,   1d Bias Errα = O(α) n     1 1d Errα = O Var n nα2 Furthermore, if we choose  1 E 2 n 1 α = n− 4 , h i2   1 d Errα − Err = O n− 2 . Proof. The first part of the corollary is a straightforward combination 1 of Theorem 3.5 and Theorem 3.10. Moreover, if we choose α = n− 4 ,  h   i2   1 1 d 1 2 E 2 Errα − Err = O(α ) + O = O n− 2 . n nα2 It is easy to see the optimal rate of α that strikes a balance between bias 1 and variance is exactly α = n− 4 . This should offer some guidance about the choice of α in practice. 4. Further properties and applications. In Section 3.3, we show dα will have diminishing variances if α is chosen properly. However, that Err dα is computed using only one instance of the randomization ω, since Err its variance can be further reduced if we aggregate over different randomizations ω. Furthermore, in the following section, we will show that after dα is Uniform Minimum Variance Unbiased such marginalization over ω, Err (UMVU) estimators for the prediction error Errα under some conditions. PREDICTION ERROR AFTER MODEL SEARCH 15 4.1. Variance reduction techniques and UMVU estimators. We first indα can be reduced troduce the following lemma that shows the variance of Err at no further assumption. Lemma 4.1. (4.1) The following estimator is unbiased for Errα , i h d d(I) Err α = Eω Errα | y , Furthermore, it has smaller variance,   h i (I) d dα , Var Errα ≤ Var Err The lemma can be easily proved using basic properties of conditional expectation. In practice, we approximate the integration over ω by repeatedly sampling ω and taking the averages. Specifically, Algorithm 1 provides an d(I) for any α > 0. algorithm for computing Err α d(I) Algorithm 1 Algorithm for computing Err α for any α > 0. 1: Input: X, y (I) dα 2: Initialize: Err ← 0, N ∈ Z+ 3: for i in 1 : N do 4: Draw ω (i) ∼ N (0, ασ 2 I) 5: Compute y ∗ = y + ω (i) , y − = y − α1 ω (i) 6: Compute M̂ ∗ = M̂ (y ∗ ) − ∗ d(i) 7: Use Equation (2.4) to compute Err α from y, y , M̂ . (I) (i) dα + = Err dα /N 8: Err (I) dα return Err d(I) has the same expectation as Err dα with smaller variances, it Since Err α 2 d(I) is easy to deduce from Corollary 3.11 that Err α also converges to Err in L 1 at a rate of at least O(n− 2 ) (after a proper scaling of n−1 ). Furthermore, we show that such estimators are UMVU estimators for any α > 0 when the parameter space µ(X) contains a ball in Rn . Lemma 4.2. If parameter space of µ(X) contains a ball in Rn , then d(I) Err α are UMVU estimators for Errα for any α > 0. Proof. Without loss of generality, assume ω has density g with respect to the Lebesgue measure on Rn , then the density of (y, ω) with respect to 16 XIAOYING TIAN HARRIS the Lebesgue measure on Rn × Rn is proportional to   µ(X)T y (4.2) exp g(ω). σ2 However, since (4.2) is an exponential family with sufficient statistics y. Moreover, when the parameter space of µ(X) contains a ball in Rn , then we have y is sufficient and complete. Thus taking an unbiased estimator dα and integrating over ω conditional on y, the complete and sufficient Err statistics, we have the UMVU estimators. 4.2. Relation to the SURE estimator. In this section, we reveal that our dα is equal to the SURE estimator for the prediction error Errα estimator Err if the parameter space of µ(X) contains a ball in Rn . First, we notice that for any α > 0, Errα is the prediction error for h i µ̂α (y) = E HM̂ (y+ω) y | y , ω ∼ N (0, ασ 2 I). Although µ̂(y) might be discontinuous in y, µ̂α (y) is actually smooth in the data. To see that, note (4.3) µ̂α (y) = |M| X i=1 Z Hi y φα (z + y)dz, Ui where φα is the p.d.f for N (0, ασ 2 I). Due to the smoothness of φα and the summation being a finite sum, we have µ̂α (y) is smooth in y. Therefore, in theory we can use Stein’s formula to compute an estimate for the prediction error of µ̂α (y). Note such estimator would only depend on y, the complete and sufficient statistics for the exponential family in (4.2) when the parameter space of µ(X) contains a ball in Rn . Thus it is also the UMVU estimator for Errα . By Lemma 4.2 and the uniqueness of UMVU estimators, dα is the same as the SURE estimator. we conclude Err However, the SURE estimator is quite difficult to compute as the regions Ui ’s may have complex geometry and explicit formulas are hard to derive (Mikkelsen & Hansen 2016). Moreover, it is difficult to even use Monte-Carlo samplers to approximate the integrals in (4.3) since the sets Ui ’s might be hard to describe and there are |M| integrals to evaluate, making it computationally expensive. dα provides an unbiased estimator for Errα at a much lower In contrast, Err computational cost. That is we only need to sample ω’s from N (0, ασ 2 I) and dα at each time and average over them. The major computation compute Err PREDICTION ERROR AFTER MODEL SEARCH 17 involved is re-selecting the model with y ∗ = y + ω. In practice, we choose the number of samples for ω’s to be less than the number of data points, so the computation involved will be even less than Leave-One-Out cross validation. 4.3. Prediction error after model selection. One key message of this work is that we can estimate the prediction error of the estimation rule µ̂ even if we have used some model selection procedure to construct the hat matrix HM̂ in µ̂. In practice, however, we need a priori information on σ 2 to compute dα . There are several methods for consistent estimation of σ 2 . In the low Err dimensional setting, we can simply use the residual sum of squares divided by the degrees of freedom to estimate σ 2 . In the high dimensional setting, the problem is more challenging, but various methods are derived including Reid et al. (2013), Sun & Zhang (2012), Tian et al. (2015). We also want to stress that the prediction error defined in this work is the in-sample prediction error that assumes fixed X. This is the same setup as in Cp (Mallows 1973), SURE (Stein 1981) and the prediction errors discussed in Efron (2012). A good estimator for the in-sample prediction error will allow us to evaluate and compare the predictive power of different estimation rules. However, in other cases, we might be interested in out-of-sample prediction errors. That is, the prediction errors are measured on a new dataset (Xnew , ynew ), Xnew ∈ Rn×p , ynew ∈ Rn where Xnew 6= X. In this case, assuming we observe some new feature matrix Xnew , and we are interested in the out-of-sample prediction error, h i 2 (4.4) Errout = E µ(Xnew ) − Xnew β̄(y) + nσ 2 , where T T β̄(y) = (XM XM )−1 XM y, M̂ (y) = M, where M̂ is the model selection procedure that depends on the data. Analogous to Errα , we define h i 2 (4.5) Errout,α = E µ(Xnew ) − Xnew β̄ ∗ (y) + nσ 2 , where T T β̄ ∗ (y) = (XM XM )−1 XM y, M̂ (y ∗ ) = M. We want to point out that we do not place any assumption on how the feature matrix is sampled. Specifically, we do not need to assume Xnew is sampled from the same distribution as X. Rather, we condition on the newly observed matrix Xnew . This is a distinction from cross validation which assumes the 18 XIAOYING TIAN HARRIS rows of the feature matrix X are i.i.d samples from some distribution. Such assumption may not be satisfied in practice. Then in the low dimensional setting where p < n, we are able to construct an unbiased estimator for Errout,α . Lemma 4.3. Suppose X ∈ Rn×p and rank(X) = p. Then if we further assume a linear model where µ(X) = Xβ 0 , where β 0 is the underlying coefficients. Assuming the homoscedastic model in (1.1), we have dout,α = H0 y − − H (4.6) Err M̂ (y ∗ ) y 2 + 2 tr(H0T HM̂ (y∗ ) )σ 2 + nσ 2   1 T − 2 tr(H0 H0 ) 1 + σ2 α is unbiased for Errout,α , where H0 = Xnew (X T X)−1 X T , T T . X )−1 XM̂ HM̂ (y∗ ) = Xnew,M̂ (y∗ ) (XM̂ (y ∗ ) (y ∗ ) M̂ (y ∗ ) The proof of the lemma is analogous to that of Theorem 2.1 noticing that H0 µ(X) = Xnew (X T X)−1 X T Xβ 0 = Xnew β 0 = µ(Xnew ). Lemma 4.3 provides an unbiased estimator for Errout,α for p < n and µ(X) being a linear function of X. To bound the difference Errout,α − Errout , we might need to assume conditions similar to those introduced at the beginning of Section 3. In the case where p < n, we might still hope that the matrices H0 , HM̂ will be close to projection matrices, and almost satisfy Assumptions 3.1. Thus, intuitively, Errout,α and Errout will be close and the estimator dout,α will be a good estimator of Errout when p < n. In simulations, Err we see that in the low-dimensional setting, the performance of Errout,α is comparable to that of cross validation. However in the high-dimensional setting where n < p, the estimation of out-of-sample errors remains a very challenging problem that we do not seek to address in the scope of this work. 4.4. Search degrees of freedom. There is a close relationship between (insample) prediction error and the degrees of freedom of an estimator. In fact, with a consistent estimator for the prediction error Err, we get a consistent estimator for the degrees of freedom. PREDICTION ERROR AFTER MODEL SEARCH 19 Under the framework of Stein’s Unbiased Risk Estimator, for any estimation rule µ̂, we have n h i X Err = E ky − µ̂(y)k2 + 2 Cov [µ̂i (y), yi ] , (4.7) i=1 where µ̂i is the i-th coordinate of µ̂. For almost differentiable µ̂’s, Stein (1981) showed the covariance term is equal to   ∂ µ̂i 2 . (4.8) Cov [µ̂i (y), yi ] = σ E ∂yi The sum of the covariance terms, properly scaled is also called the degrees of freedom.   n n X X ∂ µ̂i −2 (4.9) df = σ Cov [µ̂i (y), yi ] = E . ∂yi i=1 i=1 However, in many cases, the analytical forms of µ̂ are very hard to compute or there is none. In such cases, the computation of its divergence is only feasible for very special µ̂’s (Zou et al. 2007). Moreover, for discontinuous µ̂’s which are under consideration in this work, Mikkelsen & Hansen (2016) showed that there are further correction terms for (4.8) to account for the discontinuities. In general, these correction terms do not have analytical forms and are hard to compute. Intuitively, due to the search involved in constructing µ̂ = HM̂ y, it will have larger degrees of freedom than tr(HM̂ ) which treats the hat matrix as fixed. We adopt the name used in Tibshirani (2014) to call it “search degrees of freedom”. We circumvent the difficulty in computing ∂ µ̂i /∂yi by providing an asymptotically unbiased estimator for Err. Formally,   1 d(I) 2 b (4.10) df = 2 Errα − ky − µ̂k2 , σ (I) dα is defined as in (4.1). Using the discussion in Section 3.3, we where Err choose α = n−1/4 . Notice that such approach as above is not specific to any particular model search procedures involved in constructing µ̂. Thus it offers a unified approach to compute degrees of freedom for any µ̂ = HM̂ (y) y satisfying the appropriate assumptions in Section 3. We illustrate this flexibility by computing the search degrees of freedom for the best subset selection where there has been no explicitly computable formula. 20 XIAOYING TIAN HARRIS Prediction error estimates may also be used for tuning parameters. For example, if the model selection procedure M̂ is associated with some regularization parameter λ, we find the optimal λ that minimizes the prediction error of µ̂λ   2 (4.11) λoptimal = min E ynew − HM̂λ (y) · y , λ 2 where the expectation is taken over both ynew and y. Shen & Ye (2002) shows that this model tuning criterion will yield an adaptively optimal model which achieves the optimal prediction error as if the tuning parameter were given in advance. Using the relationship in (4.7) and (4.9), we easily see that the Cp type criterion (4.11) is equivalent to the AIC criterion using the definition of degrees of freedom (4.9). Analogously, we can also propose the BIC criterion as ky − µ̂k22 log n b BIC = + df nσ 2 n Yang (2005) points out that compared with the Cp or AIC criterion, BIC tends to recover the true underlying sparse model and recommends it if sparsity is the major concern. 5. Simulations. In this work, we propose a method for risk estimation for a class of “select and estimate” estimators. One remarkable feature of our method is that it provides a consistent estimator of the prediction error for a large class of selection procedures under general, mild conditions. To demonstrate this strength, we provide simulations for two selection procedure under various setups and datasets. The two estimators are the OLS estimator after best subset selection and relaxed Lasso, which we denote as µ̂best and µ̂relaxed . In particular, T T µ̂(y) = XM̂ (XM̂ XM̂ )−1 XM̂ y, where M̂ is selected by the best subset selection and Lasso at a fixed λ respectively using the original data y. In their Lagrangian forms, best subset selection and Lasso at fixed λ can be written as min β 1 ky − Xβk22 + λ kβkk 2 where k = 0 for best subset selection and k = 1 for Lasso. Thus, by showing the good performances (in simulation) of our estimator at both k = 0 PREDICTION ERROR AFTER MODEL SEARCH 21 and k = 1, we believe the good performance would persist for all the nonconvex optimization problems with 0 ≤ k < 1. In the simulation, we always marginalize over different randomizations to reduce variance. Specifically, we d(I) which we use in all of the comparisons use Algorithm 1 to compute Err α below. In the following simulations, we compare both the bias and variances of d(I) with the Cp estimator, cross validation as well as the our estimator Err α parametric bootstrap method proposed in Efron (2012). In particular, to ensure fairness of comparison, we use Leave-One-Out cross validation in all of our simulations. Most of our simulations are for in-sample prediction errors d(I) with some exceptions of comparing the out-of-sample estimator Err out,α in Section 4.3 to cross validation for estimating out-of-sample prediction errors. To establish a “known” truth to compare to, we use mostly synthetic data, with some of the synthetic datasets generated from a diabetes dataset. In the following simulations, we call our estimator “additive” due to the additive randomization used in the estimation. Cross validation is abbreviated as “CV”. The true prediction error is evaluated through Monte-Carlo sampling since we have access to the “true” underlying distribution. We assume the variance σ 2 is unknown and estimate it with the OLS residuals when p < n. In the high-dimensional setting, we use the methods in Reid et al. (2013) to estimate σ 2 . 5.1. Relaxed Lasso estimator. We perform simulation studies for the prediction error and degrees of freedom estimation for the relaxed Lasso estimator. Unless stated otherwise, the target of prediction error estimation is the in-sample prediction error: h i Err = E kynew − µ̂relaxed (y)k22 , ynew ∼ N (µ(X), σ 2 I) ⊥ y. According to the framework of SURE Stein (1981), the degrees of freedom of the estimator µ̂relaxed can be defined as df = n X Cov [µ̂relaxed,i , yi ] i=1 σ2 which is the target of our estimation. We first study the performance of the prediction error estimation. 5.1.1. Prediction error estimation. In the following, we describe our data generating distribution as well as the parameters used in the simulation. 22 XIAOYING TIAN HARRIS • The feature matrix X ∈ Rn×p is simulated from an equi-correlated covariance matrix with normal entries. The correlation is ρ = 0.3. • y is generated from a sparse linear model, y = Xβ 0 + ,  ∼ N (0, σ 2 I), where β 0 = (snr, . . . , snr, 0, . . . , 0) | {z } s and snr is the signal-to-noise ratio and s is the sparsity of β 0 . • We fit a Lasso problem with λ = κλ0 , where i h 0 0 ,  ∼ N (0, σ 2 I), λmin = E X T  ∞ is the level where noise below which noise starts to enter the Lasso path Negahban et al. (2009) and we choose κ > 1. 1 • The parameter α as defined in (2.1) is taken to be approximately n− 4 . We compare the performances of the estimators for different settings. We take n = 100, and p = 50, 200, 400 and sparsity to be s = 10, 20. Since 1 n− 2 = 10, s = 20 is the more dense signal situation. We take κ to be 1.1 for the low-dimensional setting and 1.5 for the high-dimensional setting. The randomization parameter α = 0.25 ≈ n−1/4 . We see in Figure 1 that in d(I) all settings Err α provides an unbiased estimator that has small variance. Remarkably, notice that the variance of our estimator is comparable to the dotted the black lines are the standard error of the true prediction error estimated from Monte-Carlo sampling, which is probably the best one can hope d(I) clearly outperforms both Cp and cross validation. Its performance for. Err α is comparable to the parametric bootstrap estimator in the sparse scenario although parametric bootstrap seems to have more extreme values. Our estimator also performs slightly better in the more dense scenario s = 20 in panel 3 of Figure 1. In the dense signal situation, the model selected by Lasso is often misspecified. We suspect that in this situation, that parametric bootstrap overfits the data in this situation, causing a slight bias downwards. The Cp estimator is always biased down because it does not take into account the “degrees of freedom” used for model search. On the other hand, cross validation has an upward bias for in-sample prediction error. However, this bias is two fold. First, the extra randomness in the new feature matrix will cause the out-of-sample prediction error to be higher. However, comparing panel 3 and 4 of Figure 1, we see that when the signal is more dense s = 20 in panel 3, cross validation has a much larger bias than when the dimension PREDICTION ERROR AFTER MODEL SEARCH 3.0 n=100, p=50, s=10, snr=7 3.5 n=100, p=200, s=10, snr=7 2.5 2.0 1.5 1.0 0.5 23 additive Cp CV bootstrap 9n=100, p=200, s=20, snr=10 8 7 6 5 4 3 2 1 additive Cp CV bootstrap 3.0 2.5 2.0 1.5 1.0 0.5 additive Cp CV bootstrap 6n=100, p=400, s=10, snr=10 5 4 3 2 1 additive Cp CV bootstrap Fig 1: Comparison of different estimators for different n, p, s, snr. The red horizontal line is the true prediction error estimated by Monte-Carlo simulation with the dashed black lines denoting its standard deviation. is higher p = 400 in panel 4. This suggests that cross validation might be susceptible to model misspecifications as well. With less sparse signals, the model selected by Lasso is not stable or consistent, causing cross validation to behave wildly even when we only leave out one observation at a time. In d(I) provides an unbiased contrast, in all of the four settings, our estimator Err α estimator with small variance. This phenomenon persists when we vary the penalty parameter λ. For a grid of λ’s with varying κ’s from [0.2, 1.6], we see from Figure 2 that cross validation error is always overestimates the in-sample prediction error. Moreover, the amount of over estimation highly depends on the data generating distribution. In both panels of Figure 2, n = 100, p = 200, snr = 7., and the only difference is the sparsity is s = 10 for Figure 2a and s = 20 for Figure 2b. Using the same dimensions for X, we seek to control the extra randomness by using a different X for the validation set. However, the change in the sparsity level alone has huge impact for the cross validation estimates of the prediction error. The curve by cross validation is also more 24 XIAOYING TIAN HARRIS 2.8 2.6 6.0 truth addtive CV 5.5 5.0 4.5 2.2 prediction error prediction error 2.4 2.0 1.8 1.6 4.0 3.5 3.0 2.5 1.4 1.2 0.2 truth addtive CV 2.0 0.4 0.6 0.8 λ/λmin 1.0 1.2 1.4 1.6 (a) n = 100, p = 200, s = 10 1.5 0.2 0.4 0.6 0.8 λ/λmin 1.0 1.2 1.4 1.6 (b) n = 100, p = 200, s = 20 Fig 2: Estimation of prediction errors for different λ’s. Cross validation is always biased upwards. However, the bias depends on the data generating distribution. (I) dα hugs the kinky due to its bigger variance. However, in both scenarios, Err true prediction error. 5.1.2. Degrees of freedom. In this section, we carry out a simulation study for our estimate of the degrees of freedom of the relaxed Lasso estimator µ̂relaxed . We take the 64 predictors in the diabetes dataset (Efron et al. 2004) to be our feature matrix X, which include the interaction terms of the original ten predictors. The positive cone condition is violated on the 64 predictors (Efron et al. 2004, Zou et al. 2007). We use the response vectors y to compute the OLS estimator β̂ols and σ̂ols and then synthetic data is generated through y = X β̂ols + , 2 I).  ∼ N (0, σols We choose λ’s to have different ratios κ ∈ {0.05, 0.1, 0.15, 0.2, 0.25}. Figure 3 shows the estimates of degrees of freedoms by our method as in (4.10) ˆ naive = |M̂ | compared with the truth computed and the naive estimate df by Monte-Carlo sampling. The naive Cp estimator always underestimate the degrees of freedom, not taking into account the inflation in degrees of freedom after model search. However, our estimator as defined in (4.9) provides an unbiased estimation for the true degrees of freedom for the relaxed Lasso estimator µ̂relaxed . 5.1.3. Out-of-sample prediction errors. Finally, we test the unbiasedness of the proposed estimator in Section 4.3 for out-of-sample prediction error. We compare with cross validation in the low-dimensional setting where p = 25 PREDICTION ERROR AFTER MODEL SEARCH 45 truth addtive Cp 40 degrees of freedom 35 30 25 20 15 10 5 0.05 0.10 0.15 λ/λmin 0.20 0.25 ˆ Fig 3: Comparison of estimates of degrees of freedom by cross validation, df −1/4 ˆ naive = M̂ at different λ’s. α = 0.25 ≈ n in (4.10) and df . 26 XIAOYING TIAN HARRIS 2.5 n=100, p=20, s=10, snr=7 3.0 n=100, p=50, s=10, snr=7 2.5 2.0 2.0 1.5 1.5 1.0 0.5 1.0 additive CV 0.5 additive CV d(I) and cross validation reFig 4: Out of sample prediction error by Err out,α spectively. 20 and p = 50 respectively. In this section only, our target is the out-ofsample prediction error   2 0 Errout = E Xnew β − Xnew,M̂ (y) β̄(y) + nσ 2 , where β̄ is the relaxed Lasso estimator and M̂ (y) is the nonzero set of the Lasso solution at λ. We still abbreviate our estimator as “additive” and compare with the out-of-sample prediction error by cross validation. We see in Figure 4 that the estimator proposed in Section 4.3 is roughly unbiased for out-of-sample prediction error. Its performance is comparable with cross validation in both settings, with a slightly larger variance. However, as pointed in Section 4.3, our estimator does not assume any assumptions on the underlying distribution of the feature matrix X. 5.2. Best subset selection. The Cp estimator was originally proposed for picking the model size in best subset selection. One aspect that often gets neglected is that for any k < p, where p is the number of features to choose from, there are more than one models of size k to choose from. And the best subset of size k already includes a selection procedure that needs to 27 PREDICTION ERROR AFTER MODEL SEARCH 1.06 prediction error 1.04 1.02 1.00 0.98 truth addtive Cp 1 2 3 4 subset size 5 6 Fig 5: Comparison of different estimates of prediction errors be adjusted for. To illustrate this problem, we generate a feature matrix X of dimension 100 × 6 with i.i.d standard normal entries. And y is generated from a linear model of X y = Xβ + N (0, 1), β = (1, 2, 3, 4, 5, 6). For each subset of size k = 1, . . . , 6, we estimate the prediction error of d(I) the best subset of size k using both Cp and Err α . The true prediction error is evaluated using Monte-Carlo sampling. From Figure 5, we see that Cp is indeed an under estimate for the prediction error for best subset selection. The bias is bigger when k = 2, 3, 4 when there are more potential submodels d(I) to select from. In contrast, Err α hugs the true prediction error at every subset size k. 6. Discussion. In this work, we propose a method for estimating the prediction error after some data snooping in selecting a model. Remarkably, our estimation is not specific to any particular model selection procedures so long as it does not select too many variables to include in the model and it picks up some signals in the data. Different examples are considered. 28 XIAOYING TIAN HARRIS In the following, we propose two more aspects of the problem that deserve attention but we do not seek to address in this work. • We mainly focus on “in-sample” prediction errors, with the exception of Section 4.3. But as pointed in Section 4.3, although we can provide a consistent estimator of the (in-sample) prediction error in high dimensions, the same is not true for out-of-sample errors. Klement et al. (2008) points out that the same difficulty exists for cross validation as well. Under what assumptions can we provide a good estimator for out-of-sample prediction error in high dimensions remains a very interesting question. • Throughout the work, we assume that the data comes from a homoscedastic normal model (1.1). Some simulations show that the performance of our estimator persists when the noise in the data is subGaussian. The authors of Tian & Taylor (2015) pointed out that it is important that the tail of the randomization noise is heavier than that of the data. Since we add Gaussian noise for randomization, we suspect that the normal assumption on the data can be replaced by a subGaussian assumption. Alternatively, we may investigate what other randomization noise we may add to the data when we have heaviertailed data. 7. Proof of the lemmas. The following lemmas are essential in proving the main theorems and lemmas which we introduce first. Lemma 7.1. Suppose Zi ∼ N (0, 1) but not necessarily independently distributed. Let Wn = max1≤i≤n Zi4 , then E [Wn ] ≤ 18(log n)2 . (7.1) Proof. For any integer k > 0, we have k [E [Wn ]] ≤ E ≤ h Wnk i  =E max i=1,...,n Zi4k  n h i h h ii X E Zi4k = n E Zi4k . i=1 PREDICTION ERROR AFTER MODEL SEARCH 29 Thus we have for any positive integer k > 0, (7.2) h h ii 1 k E [Wn ] ≤ nE Zi4k   1  √ k 4k 4k + 1 = n2 2 Γ / π 2 s # k1 "   4k+1 2 √ 4k 2.1π (4k + 1)/2 (Sterling’s formula) ≤ n2 2 / π (4k + 1)/2 e "    # k1 √ 1 4k + 1 2k 4k + 1 = 4n k 2.1 exp − 2 2  2   1 1 4k + 1 1 = 4n k (2.1) 2k exp −2 − 2 2k 2  1 4 1 4k + 1 . ≤ 2 n k (2.1) 2k e 2 We choose k to minimize the bound on the right hand side. Let "   #   1 1 4k + 1 2 1 1 4k + 1 f (k) = log n k (2.1) 2k = log n+ log(2.1)+2 log 2 k 2k 2 Take derivatives with respect to k, we have 0 f (k) = − 1 1 8 log n − 2 log(2.1) + , 2 k 2k 4k + 1 √ and it is easy to see the maximum of f (k) is attained at k = 21 log( 2.1n). And the minimum of f (k) is ! √   √ log( 2.1n) log(2.1) √ f =2+ + 2 log 2 log( 2.1n) + 1/2 . 2 log( 2.1n) Since (7.2) holds for any integer k > 0, it is easy to see E [Wn ] ≤ 4 × 4.5(log n)2 ≤ 18(log n)2 . Based on Lemma 7.1, we prove Lemma 3.7, 30 XIAOYING TIAN HARRIS Proof. First since HM̂ can only take values in {HMi , i = 1, . . . , |M|} , we have (7.3) HM̂ Z 2 2 ≤ max i=1,...,|M| kHMi Zk22 . We use the short hand Hi to denote HMi . Since for any i = 1, . . . , |M|, Hi has eigen decomposition, Hi = Vi DViT , D = diag(d1 , . . . , dK ), ViT Vi = IK×K where 0 ≤ di ≤ 1 and rank(Hi ) ≤ K. Thus it is easy to see kHi Zk22 = DViT Z 2 2 ≤ K X 2 ξij , ξij ∼ N (0, σ 2 ). j=1 Note here that we do not assume any independence structure between ξij ’s. In fact, they are most likely not independent. Combining with (7.3), we have   2 2 HM̂ Z 2 ≤ K max ξij i=1,...,|M|, j=1,...,K Therefore,  HM̂ Z  2 2 2 ≤ K2 max i=1,...,|M|, j=1,...,K 4 ξij . Thus, using Lemma 7.1, it is easy to get the conclusion of this lemma. Finally, we prove Lemma 3.6 as follows. Proof. First notice that for hat matrix of form in (3.2), we have h E kµ̂(y) − µk 2 i = |M| Z X i=1 kHi y − µk2 φ(y; µ, σ 2 )dy, Ui where Hi is short for HMi and φ(·; µ, σ 2 ) is the density for N (µ, σ 2 I). Let σ 2 ≤ τ ≤ (1 + δ)σ 2 , where δ is defined in Assumption 3.3, and we define, h i g(τ ) = E kµ̂(u) − µk2 , u ∼ N (µ, τ I). 31 PREDICTION ERROR AFTER MODEL SEARCH We note that g is differentiable with respect to τ and (7.4) i h i 1 h 1 g((1 + α)σ 2 ) − g(σ 2 ) , E kµ̂(y + ω) − µk2 − E kµ̂(y) − µk2 = n n 0 < α ≤ δ. Moreover, we have ∂τ g(τ ) = = (7.5) = |∂τ g(τ )| ≤ |M| Z X " !# ku − µk22 1 kHi u − µk ∂τ √ exp − du n 2τ ( 2πτ ) U i i=1 " # |M| Z 2 1 X 2 ku − µk kHi u − µk − n φ(u; µ, τ )du 2τ τ U i i=1 " " ## 2 ku − µk2 1 E HM̂ (u) u − µ −n 2τ τ " " #2  21 2 # 12 2 2 1 ku − µk E HM̂ (u) u − µ E −n  2τ τ 2 Using Assumptions 3.2 and 3.3, we assume there is a universal constant c1 > 0 such that √ K 2 log p ≤ c1 n, "  # 2 2 E ≤ c21 n, u ∼ N (µ, τ I), ∀ τ ∈ [σ 2 , (1 + δ)σ 2 ]. (I − HM̂ (u) )µ Thus assuming (I − HM̂ )HM̂ = 0, we have " "  # 2 2 =E E HM̂ (u) u − µ (I − HM̂ (u) )µ − HM̂ (u) 0  2 2 # , "    2 2 ≤ E 2 (I − HM̂ (u) )µ + 2 HM̂ 0 0 ∼ N (0, τ I)  2 2 # ≤ 2c21 n + 36c21 nτ 2 ≤ 38c21 n max(1, τ 2 ) where the last inequality uses Lemma 3.7 and the fact |M| ≤ pK . Moreover, note that ku − µk2 /τ is a χ2n distribution with mean n, thus " #2  2   ku − µk − n  = Var χ2n = 2n. E τ 32 XIAOYING TIAN HARRIS Combining the above inequalities with (7.5) and we have   9 1 |∂τ g(τ )| ≤ c1 max , 1 n, τ ∈ [σ 2 , (1 + δ)σ 2 ]. 2 τ Therefore, for any 0 < α ≤ δ, we have i h i 1 h E kµ̂(y + ω) − µk2 − E kµ̂(y) − µk2 n 1 = g((1 + α)σ 2 ) − g(σ 2 ) n C ≤ 2 · ασ 2 ≤ C · α, σ where we take C = σ2. 9 2 2 c1 max(σ , 1) is a universal constant assuming fixed Now we prove Lemma 3.9. Proof. Using the singular value decomposition of A, it is easy to see we can reduce the problem to the case where A is a diagonal matrix. Thus, without loss of generality, we assume A = diag(a1 , a2 , . . . , an ). Then we see that h i " E kAZk22 = E n X n X !# a2i Zi2 = i=1 h i2 E kAZk22 = E " n X ! a2i , i=1  !#2 a2i Zi2 = 3 i=1 n X i=1  a4i + X a2i a2j  . i6=j Thus we deduce h i Var kAZk22 = 2 n X ! a4i  = 2 tr A4 . i=1 Acknowledgement The author wants to thank Professor Jonathan Taylor, Professor Robert Tibshirani, Frederik Mikkelsen and Professor Ryan Tibshirani for useful discussions during this project. PREDICTION ERROR AFTER MODEL SEARCH 33 References. Dwork, C. (2008), Differential privacy: A survey of results, in ‘International Conference on Theory and Applications of Models of Computation’, Springer, pp. 1–19. Efron, B. (2012), ‘The estimation of prediction error’, Journal of the American Statistical Association . Efron, B., Hastie, T., Johnstone, I. & Tibshirani, R. (2004), ‘Least angle regression’, Annals of Statistics 32(2), 407–499. Klement, S., Mamlouk, A. M. & Martinetz, T. (2008), Reliability of cross-validation for svms in high-dimensional, low sample size scenarios, in ‘International Conference on Artificial Neural Networks’, Springer, pp. 41–50. Mallows, C. L. (1973), ‘Some comments on c p’, Technometrics 15(4), 661–675. Meinshausen, N. (2007), ‘Relaxed lasso’, Computational Statistics & Data Analysis 52(1), 374–393. Mikkelsen, F. R. & Hansen, N. R. (2016), ‘Degrees of freedom for piecewise lipschitz estimators’, arXiv preprint arXiv:1601.03524 . Negahban, S., Yu, B., Wainwright, M. J. & Ravikumar, P. K. (2009), A unified framework for high-dimensional analysis of m-estimators with decomposable regularizers, in ‘Advances in Neural Information Processing Systems’, pp. 1348–1356. Reid, S., Tibshirani, R. & Friedman, J. (2013), ‘A study of error variance estimation in lasso regression’, arXiv preprint arXiv:1311.5274 . Shen, X. & Ye, J. (2002), ‘Adaptive model selection’, Journal of the American Statistical Association 97(457), 210–221. Stein, C. M. (1981), ‘Estimation of the mean of a multivariate normal distribution’, The annals of Statistics pp. 1135–1151. Sun, T. & Zhang, C.-H. (2012), ‘Scaled sparse linear regression’, Biometrika p. ass043. Tian, X., Loftus, J. R. & Taylor, J. E. (2015), ‘Selective inference with unknown variance via the square-root lasso’, arXiv preprint arXiv:1504.08031 . Tian, X. & Taylor, J. E. (2015), ‘Selective inference with a randomized response’, arXiv preprint arXiv:1507.06739 . Tibshirani, R. (1996), ‘Regression shrinkage and selection via the lasso’, Journal of the Royal Statistical Society: Series B 58(1), 267–288. Tibshirani, R. J. (2014), ‘Degrees of freedom and model search’, arXiv preprint arXiv:1402.1920 . Tibshirani, R. J. & Taylor, J. (2012), ‘Degrees of freedom in lasso problems’, The Annals of Statistics pp. 1198–1232. Yang, Y. (2005), ‘Can the strengths of aic and bic be shared? a conflict between model indentification and regression estimation’, Biometrika 92(4), 937–950. Ye, J. (1998), ‘On measuring and correcting the effects of data mining and model selection’, Journal of the American Statistical Association 93(441), 120–131. Zou, H., Hastie, T., Tibshirani, R. et al. (2007), ‘On the degrees of freedom of the lasso’, The Annals of Statistics 35(5), 2173–2192. 390 Serra Mall, Stanford, California, USA E-mail: [email protected]
10math.ST
arXiv:1606.05608v2 [cs.DS] 8 Nov 2016 EXPLICIT CORRELATION AMPLIFIERS FOR FINDING OUTLIER CORRELATIONS IN DETERMINISTIC SUBQUADRATIC TIME MATTI KARPPA, PETTERI KASKI, JUKKA KOHONEN, AND PADRAIG Ó CATHÁIN Abstract. We derandomize G. Valiant’s [J. ACM 62 (2015) Art. 13] subquadratic-time algorithm for finding outlier correlations in binary data. Our derandomized algorithm gives deterministic subquadratic scaling essentially for the same parameter range as Valiant’s randomized algorithm, but the precise constants we save over quadratic scaling are more modest. Our main technical tool for derandomization is an explicit family of correlation amplifiers built via a family of zigzag-product expanders by Reingold, Vadhan, and Wigderson [Ann. of Math. 155 (2002) 157–187]. We say that a function f : {−1, 1}d → {−1, 1}D is a correlation amplifier with threshold 0 ≤ τ ≤ 1, error γ ≥ 1, and strength p an even positive integer if for all pairs of vectors x, y ∈ {−1, 1}d it holds that (i) |hx, yi| < τ d implies |hf (x), f (y)i| ≤ (τ γ)p D; hx,yi p γhx,yi p and (ii) |hx, yi| ≥ τ d implies D ≤ hf (x), f (y)i ≤ D. γd d Keywords. correlation, derandomization, outlier, similarity search, expander graph AMS classification. 68W01, 05C85 1. Introduction We consider the task of identifying outlier-correlated pairs from large collections of weakly correlated binary vectors in {−1, 1}d. In more precise terms, we are interested in the following computational problem. Problem 1 (Outlier correlations). We are given as input two sets X, Y ⊆ {−1, 1}d with |X| = |Y | = n, and two thresholds, the outlier threshold ρ > 0 and the background threshold τ < ρ. Our task is to output all outlier pairs (x, y) ∈ X × Y with |hx, yi| ≥ ρd, subject to the assumption that at most q of the pairs (x, y) ∈ X × Y satisfy |hx, yi| > τ d. Remark. This setting of binary vectors and (Pearson) correlation is directly motivated, among others, by the connection to Hamming distance. Indeed, for two vectors x, y ∈ {−1, 1}d we have hx, yi = d − 2DH (x, y), where DH (x, y) = |{u = 1, 2, . . . , d : x(u) 6= y(u)}| is the Hamming distance between x and y. A naı̈ve way to solve Problem 1 is to compute the n2 inner products hx, yi for (x, y) ∈ X × Y and filter out everything but the outliers. Our interest is in algorithms that scale subquadratically in n, when both d and q are bounded from above by slowly growing functions of n. That is, we seek running times of the form Helsinki Institute for Information Technology HIIT and Department of Computer Science, Aalto University, Helsinki, Finland ([email protected]). The last author’s current address is [email protected]. This research was funded by the European Research Council, under the European Union’s Seventh Framework Programme (FP/2007-2013) / ERC Grant Agreement 338077 “Theory and Practice of Advanced Search and Enumeration” (M.K., P.K., J.K.); and Academy of Finland, Grants 276031, 282938, 283262 and 283437 (P.Ó C.). Work done in part while the second author was visiting the Simons Institute for the Theory of Computing. 1 2 MATTI KARPPA, PETTERI KASKI, JUKKA KOHONEN, AND PADRAIG Ó CATHÁIN O(n2−ǫ ) for a constant ǫ > 0. Furthermore, we seek to do this without a priori knowledge of q. Running times of the form O(n2−cρ ) for a constant c > 0 are immediately obtainable using techniques such as the seminal locality-sensitive hashing of Indyk and Motwani [20] and its variants (see §1.5). However, such algorithms converge to quadratic running time in n unless ρ is bounded from below by a positive constant. Our interest is in algorithms that avoid such a “curse of weak outliers” and run in subquadratic time essentially independently of the magnitude of ρ, provided that ρ is sufficiently separated from τ . Such ability to identify weak outliers from large amounts of data is useful, among others, in machine learning from noisy data. One strategy to circumvent the curse of weak outliers is to pursue the following intuition: (i) partition the input vectors into buckets of at most s vectors each, (ii) aggregate each bucket into a single vector by taking the vector sum, and (iii) compute the inner products between the ⌈n/s⌉× ⌈n/s⌉ pairs of aggregate vectors. With sufficient separation between τ and ρ, at most q of these inner products between aggregates will be large, and every outlier pair is discoverable among the at most s × s input pairs that correspond to each large inner product of aggregates. Furthermore, a strategy of this form is oblivious to q until we actually start searching inside the buckets, which enables adjusting ρ and τ based on the number of large aggregate inner products. 1.1. Randomized amplification. Such bucketing strategies have been studied before with the help of randomization. In 2012, G. Valiant [36] presented a breakthrough algorithm that, before bucketing, replaces each input vector with a randomly subsampled1 version of its pth Kronecker power. Because of the tensor-power identity (1) hx⊗p , y ⊗p i = hx, yip , the ratio between outlier and background correlations gets amplified to essentially its pth power, assuming that the sample is large enough so that sufficient concentration bounds hold with high probability. This amplification makes the outliers stand out from the background even after bucketing, which enables detection in subquadratic time using fast matrix multiplication. A subset of the present authors [23] further improved on Valiant’s algorithm by a modified sampling scheme that simultaneously amplifies and aggregates the input by further use of fast matrix multiplication. With this improvement, Problem 1 can be solved in subquadratic time if the logarithmic ratio logτ ρ = (log ρ)/(log τ ) is bounded from above by a constant less than 1. Also this improved algorithm relies on randomization. 1.2. Explicit amplification. In this paper we seek deterministic subquadratic algorithms. As with the earlier randomized algorithms, we seek to map the ddimensional input vectors to a higher dimension D so that inner products are sufficiently amplified in the process. Towards this end, we are interested in explicit functions f : {−1, 1}d → {−1, 1}D that approximate the tensor-power identity eq. (1). Definition 2 (Correlation amplifier). Let d, D and p be positive integers, with p even, and let 0 ≤ τ ≤ 1 and γ ≥ 1. A function f : {−1, 1}d → {−1, 1}D is a correlation amplifier with parameters (d, D, p, τ, γ) if for all pairs of vectors 1The dimension is reduced by subsampling because the full dp -dimensional Kronecker power is too large to be manipulated explicitly to yield subquadratic running times. EXPLICIT AMPLIFIERS FOR OUTLIER CORRELATIONS 3 x, y ∈ {−1, 1}d we have (2) (3) if hx, yi < τ d, then hf (x), f (y)i ≤ (τ γ)p D ; and  p p  if hx, yi ≥ τ d, then hx,yi D ≤ hf (x), f (y)i ≤ γhx,yi D. γd d Remark. A correlation amplifier f guarantees by eq. (2) that correlations below τ in absolute value stay bounded; and by eq. (3) that correlations at least τ in absolute value become positive and are governed by the two-sided approximation with multiplicative error γ ≥ 1. In particular, eq. (3) implies that correlations at least τ cannot mask outliers under bucketing because all such correlations get positive sign under amplification. It is immediate that correlation amplifiers exist. For example, take f (x) = x⊗p , with p even, to obtain a correlation amplifier with D = dp , τ = 0, and γ = 1 by eq. (1). For our present purposes, however, we seek correlation amplifiers with D substantially smaller than dp . Furthermore, we seek constructions that are explicit in the strong2 form that there exists a deterministic algorithm that computes any individual coordinate of f (x) in time poly(log D, p) by accessing poly(p) coordinates of a given x ∈ {−1, 1}d. In what follows explicitness always refers to this strong form. 1.3. Our results. The main result of this paper is that sufficiently powerful explicit amplifiers exist to find outlier correlations in deterministic subquadratic time. Theorem 3 (Explicit amplifier family). There exists an explicit correlation ampliK fier f : {−1, 1}d → {−1, 1}2 with parameters (d, 2K , 2ℓ , τ, γ) whenever 0 < τ < 1, γ > 1, and d, K, ℓ are positive integers with 20ℓ+1  60 · 2ℓ   γ −1/2 −1 K 10 . (4) 2 ≥d 2 1−γ τ As a corollary we obtain a deterministic algorithm for finding outlier correlations in subquadratic time using bucketing and fast matrix multiplication. Let us write α for the limiting exponent of rectangular integer matrix multiplication. That is, for all constants η > 0 there exists an algorithm that multiplies an m × ⌊mα ⌋ integer matrix with an ⌊mα ⌋ × m integer matrix in O(m2+η ) arithmetic operations. In particular, it is known that 0.3 < α ≤ 1 [25]. Theorem 4 (Deterministic subquadratic algorithm for outlier correlations). For any constants 0 < ǫ < 1, 0 < τmax < 1, 0 < δ < α, and C > 60, there exists a deterministic algorithm that solves a given instance of Problem 1 in time   0.99ǫ(α−δ) 1.99ǫ(α−δ) (5) O n2− 4C+1 + qnδ+ 4C+1 assuming that the parameters n, d, ρ, τ satisfy the following three constraints (1) d ≤ nδ ,   −ǫ/100000 α−δ 0.99ǫ , c2 = 1 − 4C+1 (2) c1 n−c2 ≤ τ ≤ τmax , where c1 = τmax C , and (3) logτ ρ ≤ 1 − ǫ. Remarks. Observe in particular that eq. (5) is subquadratic regardless of the magnitude of ρ provided that the separation between ρ and τ via logτ ρ ≤ 1 − ǫ holds.3 The constants in eq. (4) and eq. (5) have not been optimized beyond our 2In comparison, a weaker form of explicitness could require, for example, that there exists a deterministic algorithm that computes the entire vector f (x) from a given x in time D·poly(log D, p). 3The technical constraint c n−c2 ≤ τ only affects inputs where the dimension d grows essen1 tially as a root function of n since τ ≥ 1/d. 4 MATTI KARPPA, PETTERI KASKI, JUKKA KOHONEN, AND PADRAIG Ó CATHÁIN desired goal of obtaining deterministic subquadratic running time when d and q are bounded by slowly growing functions of n. In particular, eq. (5) gives substantially worse subquadratic running times compared with the existing randomized strategies [23, 36]. The algorithm in theorem 4 needs no a priori knowledge of q and is oblivious to q until it starts searching inside the buckets. 1.4. Overview and discussion of techniques. A straightforward application of the probabilistic method establishes (lemma 21) that low-dimensional correlation amplifiers can be obtained by subsampling uniformly at random the dimensions of the tensor power x⊗p as long as the sample size D is large enough. Thus, in essence our theorem 3 amounts to derandomizing such a subsampling strategy by presenting an explicit sample that is, up to the error bounds eq. (2) and eq. (3), indistinguishable from the “perfect” amplifier x 7→ x⊗p under taking of inner products. The construction underlying theorem 3 amounts to an ℓ-fold composition of explicit squaring amplifiers (p = 2) with increasingly strong control on the error (γ) and the interval of amplification ([τ, 1]) at each successive composition. Towards this end, we require a flexible explicit construction of squaring amplifiers with strong control on the error and the interval. We obtain such a construction from an explicit family of expander graphs (lemma 11) obtainable from the explicit zigzag-product constructions of Reingold, Vadhan, and Wigderson [34]. In particular, the key to controlling the error and the interval is that the expander family gives Ramanujan-like 4 concentration λ/∆ ≤ 16∆−1/4 of the normalized second eigenvalue λ/∆ by increasing the degree ∆. In essence, since we are working with {−1, 1}-valued vectors, by increasing the degree we can use the Expander Mixing Lemma (lemma 10) and the Ramanujan-like concentration to control (lemma 13) how well the restriction xG to the edges of an expander graph G approximates the full tensor square x⊗2 under taking of inner products. Our construction has been motivated by the paradigm of gradually increasing independence [9, 14, 15, 21] in the design of pseudorandom generators. Indeed, we obtain the final amplifier gradually by successive squarings, taking care that the degree ∆i of the expander that we apply in each squaring i = 0, 1, . . . , ℓ−1 increases with a similar squaring schedule given by eq. (11) and eq. (15) to simultaneously control the error and the interval, and to bound the output dimension roughly by the square of the degree of the last expander in the sequence. Here the term “gradual” is not particularly descriptive since growth under successive squaring amounts to doubly exponential growth in the number of squarings. Yet such growth can be seen as gradual and controlled in the following sense: we obtain strong amplification compared with the final output dimension precisely because the first ℓ − 1 squarings “come for free” as ∆0 ∆1 · · · ∆ℓ−2 is (up to low-order multiplicative terms) no more than ∆2ℓ−1 , essentially because we are taking the sum of powers of 2 in the exponent. The analogy with pseudorandom generators can in fact be pushed somewhat further. Namely, a correlation amplifier can be roughly seen as a pseudorandom generator that by eq. (3) seeks to fool a “truncated family of uniform combinatorial rectangles” with further control requested by eq. (2) below the truncation threshold τ . To see the rough analogy, let z ∈ {−1, 1}d be the Hadamard product of the vectors x, y ∈ {−1, 1}d and observe that eq. (3) seeks to approximate (with multiplicative error) the expectation of a uniform random entry in the dp length Kronecker power z ⊗p by instead taking the expectation over an explicit D-dimensional sample given by f . The Kronecker power z ⊗p is a uniform special 4Actual Ramanujan graphs (see [18, 26]) would give somewhat stronger concentration λ/∆ = O(∆−1/2 ) and hence improved constants in eq. (4). However, we are not aware of a sufficiently fine-grained family of explicit Ramanujan graphs to comfortably support successive squaring. EXPLICIT AMPLIFIERS FOR OUTLIER CORRELATIONS 5 case (with z = z1 = z2 = · · · = zp ) of a “combinatorial rectangle” formed by a Kronecker product z1 ⊗ z2 ⊗ · · · ⊗ zp , and truncation means that we only seek apPd proximation in cases where | u=1 z(u)| ≥ τ d, and accordingly want constructions that take this truncation into account—that is, we do not seek to fool all combinatorial rectangles and accordingly want stronger control on the dimension D (that is, the “seed length” log D). For a review of the state of the art in pseudorandom generators we refer to Gopalan, Kane, and Meka [14] and Kothari and Meka [24]. Our goal to obtain a small output dimension D roughly corresponds to optimizing the seed length of a pseudorandom generator. While our explicit construction eq. (4) does not reach the exact output dimension obtainable by lemma 21, it should be observed that in our parameter range of interest (with γ > 1 a constant and 0 < τ ≤ τmax for a constant 0 < τmax < 1), both eq. (4) and eq. (32) are of the form D ≥ dτ −Θ(p) ; only the constants hidden by the asymptotic notation differ between the explicit and nonconstructive bounds. Moreover, using results of Alon [4] we show a lower bound (lemma 26) on the 1 p ) , when output dimension D of any correlation amplifier: namely, that D ≥ 51 ( γτ p is in the range governed by (γτ )p ≤ 1/100 and p ≤ (log e)τ 2 d . 1 8 log( γτ ) Thus, viewed as a pseudorandom generator with “seed length” log D, theorem 3 essentially does not admit improvement except possibly at the multiplicative constants. 1.5. Related work and applications. Problem 1 is a basic problem in data analysis and machine learning admitting many extensions, restrictions, and variants. A large body of work exists studying approximate near neighbour search via techniques such as locality-sensitive hashing (e.g. [5, 7, 6, 20, 13, 29, 30]), with recent work aimed at derandomization (see Pagh [31] and Pham and Pagh [33]) and resource tradeoffs (see Kapralov [22]) in particular. However, these techniques enable subquadratic scaling in n only when ρ is bounded from below by a positive constant, whereas the algorithm in theorem 4 remains subquadratic even in the case of weak outliers when ρ tends to zero with increasing n, as long as ρ and τ are separated. Ahle, Pagh, Razenshteyn, and Silvestri [1] show that subquadratic √ scaling in n is not possible for logτ ρ = 1 − o(1/ log n) unless both the Orthogonal Vectors Conjecture and the Strong Exponential Time Hypothesis [19] fail. In small dimensions, Alman and Williams [3] present a randomized algorithm that finds exact Hamming-near neighbours in a batch-query setting analogous to Problem 1 in subquadratic time in n when the dimension is constrained to d = O(log n). Recently, Chan and Williams [10] show how to derandomize related algorithm designs; also, Alman, Chan and Williams [2] derandomize the probabilistic polynomials for symmetric Boolean functions used in [3], achieving deterministic subquadratic batch queries in small dimensions. One special case of Problem 1 is the problem of learning a weight 2 parity function in the presence of noise, or the light bulb problem. Problem 5 (Light bulb problem, L. Valiant [37]). Suppose we are given as input a parameter 0 < ρ < 1 and a set of n vectors in {−1, 1}d such that one planted pair of vectors has inner product at least ρd in absolute value, and all other n − 2 vectors are chosen independently and uniformly at random. Our task is to find the planted pair among the n vectors. Remark. From e.g. the Hoeffding bound eq. (7) it follows that there exists a constant c such that when d ≥ cρ−2 log n the planted pair is with high probability (as n increases) the unique pair in the input with the maximum absolute correlation. 6 MATTI KARPPA, PETTERI KASKI, JUKKA KOHONEN, AND PADRAIG Ó CATHÁIN For a problem whose instances are drawn from a random ensemble, we say that an algorithm solves almost all instances of the problem if the probability of drawing an instance where the algorithm fails tends to zero as n increases. Paturi, Rajasekaran, and Reif [32], Dubiner [11], and May and Ozerov [27] present randomized algorithms that can be used to solve almost all instances of the light bulb problem in subquadratic time if we assume that ρ is bounded from below by a positive constant; if ρ tends to zero these algorithms converge to quadratic running time in n. G. Valiant [36] showed that a randomized algorithm can identify the planted correlation in subquadratic time on almost all inputs even when ρ tends to zero as n increases. As a corollary of theorem 4, we can derandomize Valiant’s design and still retain subquadratic running time (but with a worse constant) for almost all inputs, except for extremely weak planted correlations with ρ ≤ n−Ω(1) that our amplifier is not in general able to amplify with sufficiently low output dimension to enable an overall subquadratic running time. Corollary 6 (Deterministic subquadratic algorithm for the light bulb problem). For any constants 0 < δ < α, C > 60, 0 < ρmax < 1, and κ > 1, there exists a deterministic algorithm that solves almost all instances of Problem 5 in time   0.99(1−1/κ)(α−δ) 4C+1 O n2− assuming the parameters n, d, ρ satisfy the two constraints (1) 5ρ−2κ log n ≤ d ≤ nδ and (2) c1 n−c2 /κ ≤ ρ ≤ ρmax ,   −κǫ/100000 α−δ where c1 = ρmax and c2 = 1 − 0.99(1−1/κ) 4C+1 C . Corollary 6 extends to parity functions of larger (constant) weight in the presence of noise (cf. [16, 23, 36]). This generalized version of the problem is as follows. Problem 7 (Learning parity with noise). Let S ⊆ [v] with |S| = k be the support of a parity function and 0 < η < 1 the noise level. Our task is to determine the set S by drawing independent random examples (x, y) such thatQx ∈ {−1, 1}v is chosen uniformly at random, and the label y ∈ {−1, 1} is y = z ℓ∈S x(ℓ) where z ∈ {−1, 1} is an independent random variable with Pr(z = −1) = η. With no information on k, the trivial solution is to enumerate all 2v subsets of [v] to locate the support S. Blum, Kalai, and Wasserman [8] provide a non-trivial solua tion which runs in time and sample complexity poly |1 − 2η|2 , 2b for any positive integers a, b with ab ≥ v; this is 2O(v/ log v) when η 6= 1/2 is a constant independent of v. If we assert that k is a constant independent of v, the trivial complexity drops from exponential to v k , and non-trivial speed-ups seek to lower the coefficient 1 of k in the exponent. Randomized solutions for constant k include Valiant’s breakthrough algorithm [36] and our subsequent randomized improvement [23] which ω+ǫ 8ω 4 runs in time Õ(v 3 k |1 − 2η|− 9ǫ − 3 ) for any constant 0 < ǫ < ω/3. Our present contribution is a deterministic algorithm for learning constantweight parity functions with noise. Our interest is in the case where the noise level η approaches 1/2, and accordingly we assume that |1 − 2η| is bounded from above by a constant less than 1. We say that a deterministic algorithm solves almost all instances of Problem 7 if the probability of drawing an instance on which the algorithm fails tends to zero as v increases.5 5Observe that from an information-theoretic perspective it is a positive-but-negligibleprobability event that the drawn examples do not uniquely identify S. EXPLICIT AMPLIFIERS FOR OUTLIER CORRELATIONS 7 Corollary 8 (Deterministic algorithm for learning parity with noise). For all constants 0 < δ < α, C > 60, ξ > 1, 0 < θ < 1, there exists a constant k0 and a deterministic algorithm that for all constants k ≥ k0 draws d examples and finds the support of almost all instances of Problem 7 in time   2 2 −2 , (6) O v k 1−0.245025(α−δ) (1−1/ξ) (1+4C) assuming the parameters v, d, η satisfy the constraints 6k |1−2η|2(ξ2 +1) (1−θ ξ−1 )2 −2 −c2 ξ k/2 (1) d ≥ (2) c1 v log v, and ≤ |1 − 2η| ≤ θ,  where c1 = θ−(1−1/ξ)/100000 and c2 = 1 − 0.99(1−1/ξ) 4C+1  α−δ C . Algorithms for learning parity functions enable extensions to further classes of Boolean functions such as sparse juntas and DNFs (cf. [12, 28, 36]). 2. Preliminaries All vectors in this paper are integer-valued. For a vector x ∈ Zd we denote the entry u = 1, 2, . . . , d of x by x(u). For two vectors x, y ∈ Zd we write hx, yi = Pd u=1 x(u)y(u) for the inner product of x and y. We write log for the logarithm with base 2 and ln for the logarithm with base exp(1). In our proofs, we need the following bound due to Hoeffding [17, Theorem 2] which provides an exponentially small upper bound on the deviation of a sum of bounded independent random variables from its expectation. Theorem 9 (Hoeffding [17, Theorem 2]). Let Z1 , Z2 , . . . , ZD be independent ranPD dom variables satisfying ℓi ≤ Zi ≤ ui for all 1 ≤ i ≤ D, and let Z = i=1 Zi . Then, for all c > 0, the following holds: ! 2c2 (7) Pr (Z − E[Z] ≥ c) ≤ exp − PD . 2 i=1 (ui − ℓi ) 3. Explicit amplifiers by approximate squaring This section proves theorem 3. We start with preliminaries on expanders, show an approximate squaring identity using expander mixing, and then rely on repeated approximate squaring for our main construction. The proof is completed by some routine preprocessing. 3.1. Preliminaries on expansion and mixing. We work with undirected graphs, possibly with self-loops and multiple edges. A graph G is ∆-regular if every vertex is incident to exactly ∆ edges, with each self-loop (if present) counting as one edge. Suppose that G is ∆-regular with vertex set V , and let L be a set of ∆ labels such that the ∆ edge-ends incident to each vertex have been labeled with unique labels from L. The rotation map RotG : V × L → V × L is the bijection such that for all u ∈ V and i ∈ L we have RotG (u, i) = (v, j) if the edge incident to vertex u and labeled with i at u leads to the vertex v and has the label j at v. For S, T ⊆ V (G), let us write E(S, T ) for the set of edges of G with one end in S and the other end in T . Suppose that G has D vertices and let λ1 , λ2 , . . . , λD be the eigenvalues of the adjacency matrix of G with |λ1 | ≥ |λ2 | ≥ · · · ≥ |λD |. Let us say that a graph G is a (D, ∆, λ)-graph if G has D vertices, G is ∆-regular, and |λ2 | ≤ λ. For an excellent survey on expansion and expander graphs, we refer to Hoory, Linial, and Wigderson [18]. 8 MATTI KARPPA, PETTERI KASKI, JUKKA KOHONEN, AND PADRAIG Ó CATHÁIN Lemma 10 (Expander mixing lemma, [18, Lemma 2.5]). For all S, T ⊆ V (G) we have p ∆|S||T | ≤ λ |S||T | . |E(S, T )| − D We work with the following family of graphs obtained from the zig-zag product of Reingold, Vadhan, and Wigderson [34]. In particular lemma 11 gives us λ/∆ ≤ 16∆−1/4 , which will enable us to control relative inner products by increasing ∆. Lemma 11. For all integers t ≥ 1 and b ≥ 10 there exists a (216bt , 24b , 16 · 23b )graph whose rotation map can be evaluated in time poly(b, t).6 Proof. See appendix A.  3.2. Main construction. The main objective of this section is to prove the following lemma, which we will then augment to theorem 3 by routine preprocessing of the input dimension. Lemma 12 (Repeated approximate squaring). There exists an explicit correlation k K amplifier fˆ : {−1, 1}2 → {−1, 1}2 with parameters (2k , 2K , 2ℓ , τ0 , γ0 ) whenever 0 < τ0 < 1, γ0 > 1, and k, K, ℓ are positive integers with ℓ     −1 20ℓ γ0 40 · 2 −20 (8) 2K ≥ 2k 210 1 − γ0−1 . τ0 Approximate squaring via expanders. For a vector x ∈ {−1, 1}D , let us write 2 x⊗2 ∈ {−1, 1}D for the Kronecker product of x with itself. Our construction for correlation amplifiers will rely on approximating the squaring identity hx⊗2 , y ⊗2 i = hx, yi2 , for vectors in {−1, 1}D . In more precise terms, let G be a (D, ∆, λ)-graph and let xG ∈ {−1, 1}∆D be a vector that contains each coordinate x(u)x(v) of x⊗2 with (u, v) ∈ V (G) × V (G) exactly once for each edge of G that joins the vertex u to the vertex v. Equivalently, let RotG : V × L → V × L be a rotation map for G, and define xG for all u ∈ V and all i ∈ L by xG (u, i) = x(u)x(v) where v ∈ V is given by RotG (u, i) = (v, j). In particular, xG has exactly ∆D coordinates. Lemma 13 (Approximate squaring). For all x, y ∈ {−1, 1}D we have hxG , y G i − ∆ ⊗2 ⊗2 hx , y i ≤ 2λD . D Proof. Let S = {u ∈ V (G) : x(u) = y(u)} and let us write S̄ = V (G) \ S. Since x, y are {−1, 1}-valued, we have hxG , y G i = |E(S, S)| + |E(S̄, S̄)| − |E(S, S̄)| − |E(S̄, S)| . Observing that |S|2 + |S̄|2 − |S||S̄| − |S̄||S| = 2|S| − D and applying lemma 10 four times, we have hxG , y G i − 2 = hx, yi2 = hx⊗2 , y ⊗2 i p  ∆ ⊗2 ⊗2 hx , y i ≤ λ D + 2 |S|(D − |S|) ≤ 2λD . D  6Caveat. Reingold, Vadhan, and Wigderson [34] work with eigenvalues of the normalized adjacency matrix (with |λ1 | = 1) whereas we follow Hoory, Linial, and Wigderson [18] and work with unnormalized adjacency matrices (with |λ1 | = ∆) in the manuscript proper. appendix A works with normalized adjacency matrices for compatibility with Reingold, Vadhan, and Wigderson [34]. EXPLICIT AMPLIFIERS FOR OUTLIER CORRELATIONS 9 The amplifier function. We now construct an amplifier function fˆ that uses ℓ approximate squarings, ℓ ≥ 1, with the graphs drawn from the graph family in lemma 11. Accordingly, we assume that all vectors have lengths that are positive integer powers of 2. The input x = x̃0 ∈ {−1, 1}d0 to the amplifier has dimension d0 = 2k for a positive integer k. For i = 0, 1, . . . , ℓ − 1, suppose we have the vector x̃i ∈ {−1, 1}di . Let bi be a positive integer whose value will be fixed later. Let ti be the unique positive integer with di ≤ Di = 216bi ti < 216bi di . Note in particular that di divides Di since di is a power of 2. Let Gi be a (216bi ti , 24bi , 16 · 23bi )-graph from lemma 11. Take Di /di copies of x̃i to obtain i ∈ {−1, 1}di+1 with di+1 = ∆i Di and the vector xi ∈ {−1, 1}Di . Let x̃i+1 = xG i 4bi ˆ ∆i = 2 . The amplifier outputs f (x) = x̃ℓ with x̃ℓ ∈ {−1, 1}dℓ . Since the graph family in lemma 11 admits rotation maps that can be computed in time poly(b, t), we observe that fˆ is explicit. Indeed, from the construction it is immediate that to compute any single coordinate of fˆ(x) it suffices to (i) perform in total 2ℓ−1−i evaluations of the rotation map of the graph Gi for each i = 0, 1, . . . , ℓ − 1, and (ii) access at most 2ℓ coordinates of x. Since bi ti = O(log dℓ ) for all i = 0, 1, . . . , ℓ − 1, we have that we can compute any coordinate of fˆ(x) in time poly(log dℓ , 2ℓ ) and accessing at most 2ℓ coordinates of x. Parameterization and analysis. Fix τ0 > 0 and γ0 > 1. To parameterize the amplifier (that is, it remains to fix the values bi ), let us track a pair of vectors as it proceeds through the ℓ approximate squarings for i = 0, 1, . . . , ℓ − 1. We start by observing that copying preserves relative inner products. That is, for any pair of vectors x̃i , ỹi ∈ {−1, 1}di we have hx̃i , ỹi i = νi di if and only if hxi , yi i = νi Di for 0 ≤ νi ≤ 1. An easy manipulation of lemma 13 using the parameters in lemma 11 gives us additive control over an approximate squaring via (9) −1/4 νi2 − 32∆i −1/4 ≤ νi+1 ≤ νi2 + 32∆i . For all inner products that are in absolute value above a threshold, we want to turn this additive control into multiplicative control via (10) νi2 γ0−1 ≤ νi+1 ≤ νi2 γ0 . Let us insist this multiplicative control holds whenever |νi | ≥ τi for the threshold parameter τi defined for all i = 0, 1, . . . , ℓ − 1 by τi+1 = γ0−1 τi2 . (11) Enforcing eq. (10) via eq. (9) at the threshold, let us assume that (12) −1/4 τi2 γ0−1 ≤ τi2 − 32∆i . The next lemma confirms that assuming eq. (12) gives two-sided control of inner products which is retained to the next approximate squaring. The following lemma shows that small inner products remain small. Lemma 14. If τi ≤ |νi |, then νi2 γ0−1 ≤ νi+1 ≤ νi2 γ0 and τi+1 ≤ νi+1 . Proof. From eq. (9) and eq. (12), we have (13) −1/4 νi+1 − νi2 ≤ 32∆i ≤ (1 − γ0−1 )τi2 ≤ (1 − γ0−1 )νi2 . Observe that 1 − γ0−1 ≤ γ0 − 1. Thus, from eq. (13) we conclude that νi+1 ≤ νi2 + (1 − γ0−1 )νi2 ≤ νi2 + (γ0 − 1)νi2 = γ0 νi2 . 10 MATTI KARPPA, PETTERI KASKI, JUKKA KOHONEN, AND PADRAIG Ó CATHÁIN In the converse direction, from eq. (13) and eq. (11) we conclude that νi+1 ≥ νi2 − (1 − γ0−1 )νi2 ≥ γ0−1 νi2 ≥ γ0−1 τi2 = τi+1 .  Lemma 15. If |νi | < τi , then |νi+1 | ≤ τi2 γ0 . Proof. From eq. (9) and eq. (12), we have −1/4 νi+1 − νi2 ≤ 32∆i (14) ≤ (1 − γ0−1 )τi2 . Since 1 − γ0−1 ≤ γ0 − 1, from eq. (14) we conclude that |νi+1 | ≤ νi2 + (1 − γ0−1 )τi2 ≤ τi2 + (γ0 − 1)τi2 = γ0 τi2 .  Let us now make sure that eq. (12) holds. Solving for ∆i in eq. (12), we have 4 (15) ∆i ≥ 32(1 − γ0−1 )−1 τi−2 . In particular, we can make sure that eq. (15) and hence eq. (12) holds by simply choosing a large enough ∆i (that is, a large enough bi ). Before proceeding with the precise choice of bi for i = 0, 1, . . . , ℓ−1, let us analyze the input–output relationship of the amplifier fˆ using lemma 14 and lemma 15. Let x, y ∈ {−1, 1}d0 be two vectors given as input with hx, yi = ν0 d0 . The outputs fˆ(x), fˆ(y) ∈ {−1, 1}dℓ then satisfy hfˆ(x), fˆ(y)i = νℓ dℓ , where the following two lemmas control νℓ via ν0 . ℓ ℓ Lemma 16. If |ν0 | ≥ τ0 , then ν02 γ0−2 +1 ℓ ℓ ≤ νℓ ≤ ν02 γ02 −1 . Proof. Use induction on i, where lemma 14 gives the inductive step. ℓ ℓ Lemma 17. If |ν0 | < τ0 , then |νℓ | ≤ τ02 γ02 i −1  . i Proof. From eq. (11) we have τi = τ02 γ0−2 +1 . Let us show by induction on i i i that |νi | ≤ τ02 γ02 −1 . The base case i = 0 is immediate. For i ≥ 1, there are two cases to consider. First suppose that |νi | < τi . Then, by lemma 15 we have i+1 i+1 i+1 i+1 |νi+1 | ≤ τi2 γ0 ≤ τ02 γ0−2 +3 ≤ τ02 γ02 −1 since γ0 > 1. Next suppose that i+1 i+1 |νi | ≥ τi . Then, by lemma 14 we have |νi+1 | ≤ νi2 γ0 ≤ τ02 γ02 −1 .  Since γ0 > 1, from lemma 16 and lemma 17 it now follows that fˆ meets the required amplification constraints eq. (2) and eq. (3) with p = 2ℓ , τ = τ0 , and γ = γ0 . Let us now complete the parameterization and derive an upper bound for dℓ . For each i = 0, 1, . . . , ℓ − 1, take bi to be the smallest nonnegative integer so that bi ≥ 10 and ∆i = 24bi satisfies eq. (15). Since Di ≤ 216bi di = ∆4i di , we have di+1 = ∆i Di ≤ ∆5i di , and hence 5 dℓ ≤ (∆ℓ−1 ∆ℓ−2 · · · ∆0 ) d0 . Recall that d0 = 2k . From eq. (15) we have that 4  4 ≤ 210 (1 − γ0−1 )−1 τi−2 . ∆i = 24bi ≤ max 240 , 24 32(1 − γ0−1 )−1 τi−2 i i Since τi = τ02 γ0−2 +1 by eq. (11), it follows that ℓ+1     −1 20ℓ γ0 20(2 −1) . dℓ ≤ 2k 210 1 − γ0−1 τ0 Repeatedly taking two copies of the output as necessary, for all 2K with 2K ≥ dℓ we obtain a correlation amplifier with parameters (2k , 2K , 2ℓ , τ0 , γ0 ). This completes the proof of lemma 12.  EXPLICIT AMPLIFIERS FOR OUTLIER CORRELATIONS 11 3.3. Copy-and-truncate preprocessing of the input dimension. We still want to remove the assumption from lemma 12 that the input dimension is a positive integer power of 2. The following copy-and-truncate preprocessing will be sufficient towards this end. k Let x ∈ {−1, 1}d and let k be a positive integer. Define the vector x̂ ∈ {−1, 1}2 k by concatenating ⌈2 /d⌉ copies of x one after another, and truncating the result to the 2k first coordinates to obtain x̂. Let us study how the map x 7→ x̂ operates on a pair of vectors x, y ∈ {−1, 1}d. For notational compactness, let us work with relative inner products ν, ν̂ with hx, yi = νd and hx̂, ŷi = ν̂2k . Lemma 18. For any 0 < τ0 < 1, γ0 > 1, and 2k ≥ 2dτ0−1 (1 − γ0−1 )−1 we have that (1) |ν| < τ0 implies |ν̂| ≤ γ0 τ0 , (2) |ν| ≥ τ0 implies γ0−1 ν ≤ |ν̂| ≤ γ0 ν. Proof. Let ℓ and t be the unique integers such that 2k + ℓ = td with 0 ≤ ℓ < d. Since we are leaving out ℓ coordinates, we have 2−k (νtd − ℓ) ≤ ν̂ ≤ 2−k (νtd + ℓ) . Suppose that |ν| < τ0 . We have   |ν̂| ≤ 2−k |ν|td + ℓ ≤ 2−k |ν|2k + 2ℓ ≤ τ0 + 21−k d . Observe that 1 − γ0−1 ≤ γ0 − 1. Since by hypothesis 2k ≥ 2dτ0−1 (1 − γ0−1 )−1 ≥ 2dτ0−1 (γ0 − 1)−1 , we thus have |ν̂| ≤ γ0 τ0 . For ν ≥ τ0 we have ν − 2−k d ≤ 2−k (ν2k − d) ≤ ν̂ ≤ 2−k (ν2k + 2d) ≤ ν + 21−k d . Similarly, for ν ≤ −τ0 we have ν − 21−k d ≤ 2−k (ν2k − 2d) ≤ ν̂ ≤ 2−k (ν2k + d) ≤ ν + 2−k d . By hypothesis we have  2k ≥ 2dτ0−1 max (1 − γ0−1 )−1 , (γ0 − 1)−1 . Thus we have both γ0−1 ν ≤ ν̂ ≤ γ0 ν if ν ≥ τ0 , and γ0 ν ≤ ν̂ ≤ γ0−1 ν if ν ≤ −τ0 .  3.4. Completing the proof of theorem 3. Let d, K, ℓ, τ, γ be parameters meeting the constraints in theorem 3, in particular the constraint eq. (4). To construct a required amplifier f , we preprocess each input vector x with copy-and-truncate, k obtaining a vector x̂ of length 2k . We then then apply an amplifier fˆ : {−1, 1}2 → K K {−1, 1}2 given by lemma 12. In symbols, we define f : {−1, 1}d → {−1, 1}2 for all x ∈ {−1, 1}d by f (x) = fˆ(x̂). It is immediate from lemma 12 and lemma 18 that the resulting composition is explicit. We begin by relating the given parameters of theorem 3 to those of lemma 12. Take γ0 = γ 1/2 , τ0 = τ γ −1 , and select the minimal value of k so that the constraint in lemma 18 is satisfied; that is 2k is constrained as follows, 2d(1 − γ −1/2 )−1 γτ −1 ≤ 2k < 4d(1 − γ −1/2 )−1 γτ −1 . Substituting this upper bound into the bound of lemma 12, we get a lower bound for 2K , 20ℓ+1 γ  γ 60 2ℓ τ 20  K −8 10 −1/2 −1 (16) 2 ≥ 2 d 2 (1 − γ ) . τ τ 40 γ 30 12 MATTI KARPPA, PETTERI KASKI, JUKKA KOHONEN, AND PADRAIG Ó CATHÁIN Observe that an integer 2K satisfying eq. (4) also satisfies eq. (16). We have not attempted to optimise our construction, and prefer the the statement of theorem 3 as it is reasonably clean and is sufficient to prove theorem 4. Let us study how the map x 7→ f (x) operates on a pair of vectors x, y ∈ {−1, 1}d. For notational compactness, again we work with relative inner products ν, ν̂, φ with hx, yi = νd, hx̂, ŷi = ν̂2k , and hf (x), f (y)i = φ2K . Observe that in the notation of the proof of lemma 12, we have ν̂ = ν0 and φ = νℓ . ℓ Lemma 19. If |ν| < τ then |φ| ≤ (γτ )2 . Proof. First we show that |ν̂| ≤ γ0 τ , dividing into cases as in Lemma 18. If |ν| < τ0 then |ν̂| < γ0 τ0 = γ0−1 τ ≤ γ0 τ . If τ0 ≤ ν < τ then ν̂ ≤ γ0 ν ≤ γ0 τ . If −τ < ν ≤ τ0 then ν̂ ≥ γ0 ν ≥ −γ0 τ . To complete the proof, we condition on |ν̂|. If |ν̂| ≤ τ0 then Lemma 17 applies, and we have ℓ ℓ |φ| = |νℓ | ≤ τ02 γ02 −1 ℓ < (τ γ)2 . Otherwise, τ0 ≤ |ν̂| < τ and by Lemma 16 we have ℓ ℓ 0 < φ = νℓ ≤ ν02 γ02 −1 ℓ ℓ ≤ τ 2 γ02 −1 ℓ ≤ (τ γ)2 .  ℓ ℓ Lemma 20. If |ν| ≥ τ then (νγ −1 )2 ≤ φ ≤ (νγ)2 . Proof. It will be convenient to split the analysis according to whether ν is positive or negative. Suppose first that ν ≥ τ . Then by Lemma 18 we have that γ0−1 ν ≤ ν̂ ≤ γ0 ν . (17) Since ν̂ ≥ νγ0−1 ≥ τ γ0−1 = τ0 γ0 ≥ τ0 , Lemma 16 applies, yielding ℓ ν̂γ0−2 +1 ℓ ℓ ≤ νℓ ≤ ν̂ 2 γ02 −1 . Now, we substitute φ = νℓ and bound ν̂ as in eq. (17), νγ0−1 1/2 Substituting γ = γ0 2ℓ ℓ γ0−2 +1 ≤ φ ≤ νγ0−1 2ℓ ℓ γ02 −1 . and observing that γ ≥ 1 provides the required bound νγ −1 2ℓ 2ℓ ≤ φ ≤ (νγ) . The case that ν ≤ −τ essentially follows from multiplying all inequalities in the positive case by −1.  Now, f satisfies eq. (2) and eq. (3) with p = 2ℓ by lemmas 19 and 20 respectively. This completes the proof of theorem 3.  4. A deterministic algorithm for outlier correlations This section proves theorem 4. We start by describing the algorithm, then parameterize it and establish its correctness, and finally proceed to analyze the running time. EXPLICIT AMPLIFIERS FOR OUTLIER CORRELATIONS 13 4.1. The algorithm. Fix the constants ǫ, τmax , δ, C as in theorem 4. Based on these constants, fix the constants 0 < σ < 1 and γ > 1. (We fix the precise values of σ and γ later during the analysis of the algorithm, and stress that σ, γ do not depend on the given input.) Suppose we are given as input the parameters 0 < τ < ρ < 1 and X, Y ⊆ {−1, 1}d with |X| = |Y | = n so that the requirements in theorem 4 hold. We work with a correlation amplifier f : {−1, 1}d → {−1, 1}D with parameters (d, D, p, τ, γ). (We fix the precise values of the parameters p and D later during the analysis of the algorithm so that f originates from theorem 3.) The algorithm proceeds as follows. First, apply f to each vector in X and Y to obtain the sets Xf and Yf . Let s = ⌊nσ ⌋. Second, partition the n vectors in both Xf and Yf into ⌈n/s⌉ buckets of size at most s each, and take the vector sum of the vectors in each bucket to obtain the sets X̃f , Ỹf ⊆ {−s, −s + 1, . . . , s − 1, s}D with |X̃f |, |Ỹf | ≤ ⌈n/s⌉. Third, using fast rectangular matrix multiplication on X̃f and Ỹf , compute the matrix Z whose entries are the inner products hx̃, ỹi for all x̃ ∈ X̃f and all ỹ ∈ Ỹf . Fourth, iterate over the entries of Z, and whenever the detection inequality (18) hx̃, ỹi > n2σ (τ γ)p D holds, brute-force search for outliers among the at most s2 inner products in the corresponding pair of buckets. Output any outliers found. 4.2. Parameterization and correctness. Let us now parameterize the algorithm and establish its correctness. Since γ > 1 is a constant and assuming that p is large enough, by theorem 3 we can select D to be the integer power of 2 with  Cp  Cp γ 1 γ <D≤d . d 2 τ τ Recall that we write α for the exponent of rectangular matrix multiplication. To apply fast rectangular matrix multiplication in the third step of the algorithm, we want  α n (19) D≤2 , s so recalling that d ≤ nδ and nσ − 1 < s, it suffices to require that  Cp γ ≤ n(1−σ)α−δ . τ Let us assume for the time being that (1 − σ)α − δ > 0. (We will justify this assumption later when we choose a value for σ.) Let p be the unique positiveinteger power of 2 such that ((1 − σ)α − δ) log n ((1 − σ)α − δ) log n (20) <p≤ . 2C log γτ C log γτ We will later, when fixing σ and γ, make sure that the right-hand side in eq. (20) is at least 1, so that p exists and is positive. Let us now consider a single entry hx̃, ỹi in Z, and analyze how the corresponding (at most s2 ) inner products hx, yi between the two buckets of input vectors relate to the detection inequality eq. (18). We make two claims: Claim 1 (background case). If all of the inner products have |hx, yi| ≤ τ d, then eq. (18) does not hold, so the algorithm will not search inside this pair of buckets. This claim will be used to control the running time. The claim follows directly from eq. (2) and eq. (3), since there are at most s2 ≤ n2σ inner products, each having |hf (x), f (y)i| ≤ (τ γ)p D. 14 MATTI KARPPA, PETTERI KASKI, JUKKA KOHONEN, AND PADRAIG Ó CATHÁIN Claim 2 (outlier case). If at least one of the inner products has |hx, yi| ≥ ρd, then eq. (18) holds, so the algorithm searches inside this pair of buckets. This guarantees that the outliers are detected. Note that in the third case, namely, if some inner products have |hx, yi| > τ d but none has |hx, yi| ≥ ρd, we make no claim on whether (18) holds or not. The algorithm is not required to search inside such pairs of buckets (since there are no outliers there), but may so do without hindering our overall running time bound. We proceed to parameterize the algorithm so that Claim 2 holds. In the outlier case, by eq. (2) and eq. (3), there is at least one inner product with hf (x), f (y)i ≥ (ργ −1 )p D, and the remaining at most n2σ inner products have hf (x), f (y)i ≥ −(τ γ)p D. Thus in the outlier case we have hx̃, ỹi ≥ (ργ −1 )p D − n2σ (τ γ)p D. (21) For Claim 2 we need the detection inequality eq. (18) to hold whenever eq. (21) holds. Towards this end, it suffices to require that p p p ργ −1 − n2σ τ γ > n2σ τ γ . Rearranging and solving for p, we require that (22) p> 1 + 2σ log n . log τ ργ 2 From eq. (20) and eq. (22) we thus see that it suffices to have p> ((1 − σ)α − δ) log n 1 + 2σ log n , ≥ γ 2C log τ log τ ργ 2 or equivalently, log τ ργ 2 (23) log γ τ ≥ 2C log n + 4Cσ (1 − σ)α − δ . Let us derive a lower bound for the left-hand side of eq. (23). Fix the constant γ > 1 log τmax so that log γ = − ǫ 100000 . By our assumptions we have τ ≤ τmax and 1−logτ ρ ≥ ǫ, so we have the lower bound log τ ργ 2 log γτ = log γ 2 log γ 1 − logτ ρ + 2log ǫ + log log ρ − log τ − 2 log γ τ τmax > 0.99ǫ . = ≥ γ log γ log γ − log τ 1 − log 1 − log τ log τmax Thus, eq. (23) holds for all large enough n when we require 0.99ǫ ≥ 4Cσ . (1 − σ)α − δ Since αǫ < 1, we have that eq. (23) holds when we set σ= 0.99ǫ(α − δ) 0.99ǫ(α − δ) ≤ . 4C + 1 4C + 0.99αǫ We also observe that (1 − σ)α − δ > 0, or equivalently, σ < (α − δ)/α holds for our choice of σ. Having now fixed σ and γ, we observe that in terms of assumption 2 of the = c2 . Thus the assumption statement of theorem 4, we have γ = c1 and (1−σ)α−δ C τ ≥ c1 n−c2 guarantees that the right-hand side of eq. (20) is at least 1, which was required for the existence of p. This completes the parameterization of the algorithm. EXPLICIT AMPLIFIERS FOR OUTLIER CORRELATIONS 15 4.3. Running time. Let us now analyze the running time of the algorithm. The first and second steps run in time Õ(nD) since p = O(log n) by eq. (20) and f originates from theorem 3 and hence is explicit. From eq. (19) and nσ − 1 < s, we have nD ≤ 4n1+(1−σ)α ≤ 4n2−σ. Since eq. (19) holds, the third step of the algorithm runs in time O (n/s)2+η for any constant η > 0 that we are free to choose. Since n/s ≤ 2n1−σ for all large enough n, we can choose η > 0 so that (2 + η)(1 − σ) ≤ 2 − σ. Thus, the first, second, and third steps together run in time O(n2−σ ). The fourth step runs in time O(n2−σ + qs2 d). Indeed, observe from Claim 1 in §4.2 that the detection inequality eq. (18) holds for at most q entries in Z. We have qs2 d ≤ qn2σ+δ , which completes the running time analysis and the proof of theorem 4.  5. Applications This section proves corollaries 6 and 8. 5.1. The light bulb problem. A useful variant of the Problem 1 asks for all outlier pairs of distinct vectors drawn from a single set S ⊆ {−1, 1}d rather than two sets X, Y . We observe that the single-set variant reduces to ⌈log |S|⌉ instances of the two-set variant by numbering the vectors in S with binary numbers from 0 to |S| − 1 and splitting S into two sets Xi , Yi based on the value of the ith bit for each i = 0, 1, . . . , ⌈log |S|⌉ − 1. Proof of corollary 6. We reduce to (the single-set version of) Problem 1 and apply theorem 4. Towards this end, in theorem 4 set ǫ = 1 − 1/κ and τmax = ρκmax . Suppose we are given an instance of Problem 5 whose parameters n, d, ρ satisfy the constraints. Set τ = ρκ . We observe that the constraints in theorem 4 are satisfied since (i) d ≤ nδ holds by assumption, (ii) τ ≤ τmax holds since τ = ρκ ≤ ρκmax , (iii) the constants c1 and c2 here match those in theorem 4, and the constraint log ρ log ρ c1 n−c2 /κ ≤ ρ implies c1 n−c2 ≤ τ , and (iv) logτ ρ = log τ = log ρκ = 1/κ ≤ 1 − ǫ. We claim that q = 1 in almost all instances of Problem 5 whose parameters satisfy the constraints in Corollary 6. Indeed, by the Hoeffding bound eq. (7) and the union bound, the probability that some other pair than the planted pair in an instance has inner product that exceeds τ d in absolute value is at most   2n2 exp −τ 2 d/2 ≤ 2n2 exp −ρ2κ · 5ρ−2κ log n = 2n−1/2 , so q = 1 with high probability as n increases. The claimed running time follows by substituting the chosen constants and q = 1 to eq. (5).  5.2. Learning parities with noise. We now generalize the result for parity functions of larger constant weight, and prove corollary 8. Proof of corollary 8. Fix the constants 0 < δ < α, C > 60, ξ > 1, 0 < θ < 1. We will fix the value of the constant k0 later. Let k ≥ k0 be a constant. The algorithm first draws d examples from a given instance of Problem 7 and then transforms these to two collections of vectors that we feed to the algorithm of theorem 4 and then proceed to mimic the proof of corollary 6. Let us first set up some notation. For A, B ⊆ [v], let A△B = (A\B)∪(B \A) dev note the symmetric difference of A and Q B. Let x = (x(1), x(2), . . . , x(v)) ∈ {−1, 1} A be a Boolean n-vector. Let x = ℓ∈A x(ℓ) be the product of elements indexed Q Q Q by A, with x∅ = 1. Observe that xA xB = i∈A x(i) j∈B x(j) = ℓ∈A△B x(ℓ) =  xA△B . Let us write [n] for the set of all k-subsets of [v]. v Suppose we are now given as input an instance of Problem 7 with noise level η that satisfies |1 − 2η| ≤ θ < 1. Furthermore, we assume that η is part of the input. (If this is not the case, at the cost of increasing time complexity, we can search for 16 MATTI KARPPA, PETTERI KASKI, JUKKA KOHONEN, AND PADRAIG Ó CATHÁIN η using a geometric progression with limit 1/2.) With the objective of eventually applying theorem 4, set ρ = |1 − 2η|ξ (24) and 2 τ = ρξ = |1 − 2η|ξ . (25) In particular, we have τ < ρ since 0 < |1 − 2η| < 1 and ξ > 1. Let d be the least positive integer that satisfies d ≥ (2k + 1 + 4kζ)τ −2 (|1 − 2η| − ρ)−2 log v , (26) where 0 < ζ < 1/2 is constant whose value we will fix later. Draw from the given instance d example–label pairs (xi , yi ) ∈ {−1, 1}v × {−1, 1} with i = 1, 2, . . . , d. We use to define two collections X, Y ⊆ {−1, 1}d of vectors of sizes  these examples  v v ⌊k/2⌋ and ⌈k/2⌉ , respectively. For all k ≥ ⌈1/(2ζ)⌉ and all v ≥ 2k it is immediate that we have     v v ≤ ≤ v k(1/2+ζ) . ⌊k/2⌋ ⌈k/2⌉ In particular, we can assume that |X|, |Y | ≤ n for n = ⌊v k(1/2+ζ) ⌋. The set X consists of all the vectors with aJi 1 = xJi 1 vectors aJ1 = (aJ1 1 , aJ2 1 , . . . , aJd1 ) ∈ {−1, 1}d [v]  for all i = 1, 2, . . . , d and J1 ∈ ⌊k/2⌋ . The set Y consists of all the bJ2 = (bJ1 2 , bJ2 2 , . . . , bJd2 ) [v]  with bJi 1 = xJi 2 yi for all i = 1, 2, . . . , d and J2 ∈ ⌈k/2⌉ . Let us now study the distribution of inner products between vectors in X and Y . We write Bin±1 (d, β) for a random variable that is the sum of d independent random variables, each of which takes the value −1 with probability β, and the value 1 otherwise. Observe that the expectation of Bin±1 (d, β) is (1 − 2β)d. Let S ⊆ [v] with |S| = k be the support of the parity function that is unknown to us. Recall that yi = zi xSi with zi ∈ {−1, 1} getting value −1 with probability η. [v]  [v]  we have and J2 ∈ ⌈k/2⌉ For all J1 ∈ ⌊k/2⌋ haJ1 , bJ2 i = d X xJi 1 xJi 2 yi = d X xJi 1 △J2 xSi zi = xJi 1 △J2 △S zi . i=1 i=1 i=1 d X Now observe that there are two distinct cases: If J1 △ J2 6= S, then haJ1 , bJ2 i ∼ Bin±1 (d, 1/2) . (27) If J1 △ J2 = S, then (28) haJ1 , bJ2 i = d X i=1 xJi 1 △J2 △S zi = d X i=1 zi ∼ Bin±1 (d, η) . Hence, our task of finding the support S reduces to that of locating the inner products with distribution Bin±1 (d, η) from among those with Bin±1 (d, 1/2). We now argue that our choices eq. (24), eq. (25), and eq. (26) suffice for the algorithm in theorem 4 to distinguish between the two cases eq. (27) and eq. (28) for almost all draws of the d examples. Here we stress that the algorithm is deterministic, the randomness is over the draw of the examples. From the perspective of the algorithm in theorem 4, it suffices that (a) no pair with eq. (27) exceeds τ d in absolute-value inner product, and (b) at least one of EXPLICIT AMPLIFIERS FOR OUTLIER CORRELATIONS 17 the at most k k = O(1) pairs with eq. (28) has absolute-value inner product at least ρd. To control (a), from eq. (7) we observe that  2   τ d Pr | Bin±1 (d, 1/2)| ≥ τ d ≤ 2 exp − 2   (2k + 1 + 4kζ) log v ≤ 2 exp − 2(|1 − 2η| − ρ)2 = 2v −(2k+1+4kζ)(|1−2η|−ρ) −2 /2 . Since there are at most n2 ≤ (v k(1/2+ζ) )2 = v k+2ζ such pairs, we observe by the union bound that (a) holds with high probability as v increases since (29) n2 · 2v −(2k+1+4kζ)(|1−2η|−ρ) −2 /2 ≤ 2v −(1/2)(|1−2η|−ρ) −2 . To control (b), select any fixed pair with eq. (28). From eq. (7) we have  Pr | Bin±1 (d, η) − (1 − 2η)d| ≥ (|1 − 2η| − ρ)d   (|1 − 2η| − ρ)2 d ≤ 2 exp − 2   (30) (2k + 1 + 4kζ) log v ≤ 2 exp − 2τ 2 = 2v − 2k+1+4kζ 2τ 2 . Thus, (b) holds with high probability as v increases. It remains to verify the constraints for the parameters n, d, ρ, τ in theorem 4. Suppressing the constants, our choice of d in eq. (26) is Θ(k) · |1 − 2η|−Θ(1) · log v. For theorem 4 to apply, this must be bounded from above by nδ = v Θ(k) , which holds if |1 − 2η| ≥ v −Θ(k) . This holds by assumption for sufficiently large k. Select k0 so that this constraint holds and k0 ≥ ⌈1/(2ζ)⌉. We can choose τmax = θ and 2 ǫ = 1 − 1/ξ. We then have τ = |1 − 2η|ξ < τmax < 1 by assumption, as required. Since n ≥ v k/2 , we also have by assumption 2 2 τ = |1 − 2η|ξ ≥ c1 ξ v −c2 k/2 ≥ c1 n−c2 as required. The constants c1 and c2 here match those in theorem 4. Furthermore by the choice of ǫ we have log ρ log ρ = 1/ξ = 1 − ǫ , = log τ log ρξ as required. So the constraints of theorem 4 are satisfied. For brevity, let E = 0.99ǫ(α−δ) and take ζ = E/4. Thus, we have 4C+1 2−E  2 2 −2 ≤ v k(1−0.245025(α−δ) (1−1/ξ) (1+4C) ) . (31) n2−E ≤ v k(1/2+ζ) The claimed running time eq. (6) follows by observing that eq. (31) subsumes the time it takes to construct the collections X and Y together with the time it takes to search the q pairs of buckets with q ≤ k k = O(1) inside the algorithm of theorem 4. Inserting our choices eq. (24) and eq. (25) into eq. (26) and approximating up2 wards with ζ ≤ 1 and |1 − 2η|2ξ +2 (1 − θξ−1 )2 ≤ τ 2 (|1 − 2η| − ρ)2 yields d≥ 6k log v . |1 − 2η|2(ξ2 +1) (1 − θξ−1 )2  18 MATTI KARPPA, PETTERI KASKI, JUKKA KOHONEN, AND PADRAIG Ó CATHÁIN 6. Nonconstructive existence and a lower bound This section shows that nontrivial correlation amplifiers exist and establishes a lower bound on the output dimension D of any correlation amplifier. The former is done by a routine application of the Hoeffding bound and the latter by applying results of Alon [4]. 6.1. Low-dimensional amplifiers exist. By combining the Hoeffding bound with the union bound, we observe that low-dimensional amplifiers exist. Lemma 21 (Existence). There exists a correlation amplifier f : {−1, 1}d → {−1, 1}D with parameters (d, D, p, τ, γ) whenever 0 < τ < 1, γ > 1, and d, p, D are positive integers satisfying  γ 2p . (32) D ≥ 3d (γ p − 1)−2 τ Proof. Let f : {−1, 1}d → {−1, 1}D be the function which maps x onto D entries of x⊗p chosen independently at random. That is, each entry of the vector f (x) is the product of p entries of x, chosen independently and uniformly at random. Let x, y ∈ {−1, 1}d be a fixed pair of vectors, set c = D(1 − γ −p )τ p , and suppose that the following inequality holds,  p hx, yi (33) hf (x), f (y)i − D ≤ c. d Observe that if |hx, yi| < τ d then eq. (33) implies p  hx, yi |hf (x), f (y)i| ≤ D + D(1 − γ −p )τ p d ≤ Dτ p + D(1 − γ −p )τ p ≤ (τ γ)p D . The final inequality holds because 2−γ −p ≤ γ p is logically equivalent to (γ p −1)2 ≥ 0. Similarly, if |hx, yi| ≥ τ d then eq. (33) implies the following upper bound, p  hx, yi hf (x), f (y)i ≤ D + D(1 − γ −p )τ p d p p   hx, yi hx, yi −p + D(1 − γ ) ≤D d d  p γhx, yi ≤ D. d We also obtain a lower bound from eq. (33) when |hx, yi| ≥ τ d, p  hx, yi hf (x), f (y)i ≥ D − D(1 − γ −p )τ p d  p p  hx, yi hx, yi ≥D − D(1 − γ −p ) d d p  hx, yi D. ≥ γd In fact, eq. (33) implies conditions eq. (2) and eq. (3) in definition 2. So if the function f satisfies eq. (33) for all x, y ∈ {−1, 1}d, then f is a correlation amplifier. We use theorem 9 to bound the probability that eq. (33) fails, and take a union bound over the range of f to establish a non-constructive existence result for sufficiently large D. EXPLICIT AMPLIFIERS FOR OUTLIER CORRELATIONS 19 Define the random variable Zf = hf (x), f (y)i. Since f (x) is a restriction onto D entries of x⊗p chosen uniformly at random, we have  p hx, yi E[Zf ] = D . d PD Observe that Zf = i=1 Zf,i where Zf,i is the product of the ith entries of f (x) and f (y). In particular, −1 ≤ Zf,i ≤ 1 holds for i = 1, 2, . . . , D. Summing over the Zf,i in eq. (7), the probability that eq. (33) fails to hold is bounded above by c2 Pr (Zf − E[Zf ] ≥ c) ≤ e− 2D . Taking a union bound over all x, y ∈ {−1, 1}d, there exists a correlation amplifier with parameters (d, D, p, τ, γ) whenever c2 22d e− 2D < 1 . Solving for D, we get d ln 16 2 . τ 2p (1 − γ −p ) Simplifying this expression and approximating ln 16 by 3 completes the proof. D≥  6.2. Lower bound on output dimension. We next show a lower bound on the output dimension D of any correlation amplifier, when the other parameters d, p, τ and γ are given. The proof is based on taking a collection of N vectors xi ∈ {−1, 1}d, with all pairs below the background threshold τ , and then bounding the number of their images f (xi ) ∈ {−1, 1}D , whose absolute pairwise correlations are required to be below ǫ = (τ γ)p by definition 2. Lemma 22. There is a collection of N = exp(τ 2 d/4) vectors x1 , x2 , . . . , xN ∈ {−1, 1}d such that |hxi , xj i| < τ d for all i 6= j. Proof. We show this by the probabilistic argument. We call a pair of vectors bad if |hxi , xj i| ≥ τ d. Let a collection of vectors X1 , X2 , . . . , XN be chosen uniformly at random from {−1, 1}d. Consider a pair Xi , Xj with i 6= j, and let Zij = hXi , Xj i. Now Zij is a sum of d independent random variables in [−1, 1], with E[Zij ] = 0. Applying the two-sided Hoeffding bound with c = τ d, we observe that the pair Xi , Xj is bad with probability Pr(|hXi , Xj i| ≥ τ d) = Pr(|Zij − E[Zij ]| ≥ τ d) ≤ 2 exp(−τ 2 d/2). Since there are less than N 2 /2 = (1/2) exp(τ 2 d/2) pairs of vectors, the expected number of bad pairs is less than 1. Thus in at least one collection there are no bad pairs.  To bound the number of the image vectors, we use a combinatorial result from Alon [4] to bound the rank of their correlation matrix. Lemma 23. √ Let A = (aij ) be an N × N real, symmetric matrix with aii = 1 and |aij | ≤ 1/ N for all i 6= j. Then rank(A) ≥ N/2. √ Proof. Apply Alon’s Lemma 9.1 [4] with ǫ = 1/ N .  Lemma 24. Let B = (bij ) be an N × N matrix with rank(B) = D′ , and let  ′ A = (bkij ), where k is a positive integer. Then rank(A) ≤ D +k−1 . k Proof. Apply Alon’s Lemma 9.2 [4] with the polynomial P (x) = xk .  The next lemma is in essence Alon’s Theorem 9.3 [4], modified to avoid any asymptotic notation. All logarithms here are in base 2. 20 MATTI KARPPA, PETTERI KASKI, JUKKA KOHONEN, AND PADRAIG Ó CATHÁIN Lemma 25. Let B = (bij ) be √ an N × N real, symmetric matrix with bii = 1 and |bij | ≤ ǫ for all i 6= j, where 1/ N ≤ ǫ ≤ 1/100, and rank(B) = D′ . Then   2r/(r+1) 1 r ′ (34) D ≥ 5 ǫ where r = (log N )/(2 log(1/ǫ)). Proof. Choose r as stated. Note that by the assumed range of ǫ, we have r ≥ 1. Let further k = ⌈r⌉, so in particular 1 ≤ r ≤ k < r + 1. Let A = (aij ) = (bkij ). Since the off-diagonal elements of B satisfy |bij | < ǫ, it follows √ from the choice of k that the off-diagonal elements of A satisfy |aij | ≤ ǫk ≤ ǫr = 1/ N . Combining lemma 23 and lemma 24, we have k  r+1  ′   e(D′ + r) D +k−1 e(D′ + k − 1) ≤ . N/2 ≤ rank(A) ≤ ≤ k r k Taking logarithms and rearranging the inequality we obtain   log(N/2) D′ log N ≥ log 1 + − log e ≥ − 2, r r+1 r+1 implying D′ 2(log N )/(r+1) ≥ . r 4 Observing that log N = r log(1/ǫ2 ), we get  2r/(r+1) 1 1 D′ ≥ 1+ r 4 ǫ 1+ and, since ǫ ≤ 1/100 and r ≥ 1, this implies   2r/(r+1) r 1 D′ ≥ 5 ǫ as stated.  Remark. The parameter r measures, in a sense, √ the distance from the case of an extremely low correlation requirement ǫ = 1/ N . If r tends to infinity, the exponent 2r/(r + 1) approaches 2, matching the asymptotic form given by Alon [4]. However, with small r the exponent diminishes, reaching 1 in the limiting case √ r = 1, that is, when ǫ = 1/ N . In the limiting case a direct application of lemma 23 would give the better linear bound D′ ≥ N/2. We can now combine lemma 22 and 25 to get a lower bound on output dimension. Lemma 26 (Lower bound on output dimension). The output dimension of a correlation amplifier with parameters (d, D, p, τ, γ) is bounded by  p 1 1 D≥ 5 γτ when (γτ )p ≤ 1/100 and p ≤ (log e)τ 2 d . 1 8 log( γτ ) Proof. By lemma 22 there is a collection of N = exp(τ 2 d/4) vectors x1 , x2 , . . . , xN ∈ {−1, 1}d with correlations below τ in absolute value. By Definition 2 their images ui = f (xi ) ∈ {−1, 1}D have correlations below ǫ = (γτ )p in absolute value. EXPLICIT AMPLIFIERS FOR OUTLIER CORRELATIONS 21 Consider the N × N correlation matrix B = (bij ) = (hui , uj i/D). It is real and symmetric, with diagonal elements bii = 1 and off-diagonals satisfying |bij | ≤ ǫ. We observe that D′ = rank(B) ≤ D. Applying lemma 25 we have r= (log e)τ 2 d log N = ≥ 1, 1 2 log(1/ǫ) ) 8p log( γτ and (35) as claimed.   2r/(r+1)  p r 1 1 1 D≥D ≥ ≥ 5 ǫ 5 γτ ′  (log e)τ 2 d , 1 8 log( γτ ) Remark. At the limiting case where p = we have r = 1 and ǫ = √ 1/ N = exp(−t2 d/8), and the bound eq. (35) becomes D ≥ exp(τ 2 d/8). For p greater than the limit, one can essentially map all of the N = exp(τ 2 d/4) input vectors to orthogonal output vectors of dimension D ≤ 2N using a Hadamard matrix, in which case eq. (2) holds for arbitrary p > 1. Appendix A. An expander family This section proves lemma 11 following Reingold, Vadhan and Wigderson [34]; we present the proof for completeness of exposition only with no claim of originality. Following Reingold, Vadhan and Wigderson [34] we will work with normalized eigenvalues. To avoid confusion with the unnormalized treatment in the manuscript proper, we say that a graph is a [D, ∆, λ]-graph if the graph has D vertices, is ∆-regular, and |λ2 |/∆ ≤ λ. (Here |λ2 | is the unnormalized second eigenvalue as defined in the manuscript proper.) We refer to Sections 2.3 and 3.1 of Reingold, Vadhan, and Wigderson [34] for the definition of the square G2 of a graph G, the tensor product G1 ⊗ G2 of graphs G1 , G2 , and the zigzag product G Z H of graphs G, H. The following omnibus result collects elements of Propositions 2.3, Proposition 2.4, Theorem 3.2 and Theorem 4.3 of [34] which will be sufficient to control the second normalized eigenvalue for our present purposes. (We choose to omit the details of the rotation maps with the understanding that they can be found in [34].) Lemma 27 (Reingold, Vadhan, and Wigderson [34]). The following bounds hold. (1) If G is a [D, ∆, λ]-graph, then G2 is a [D, ∆2 , λ2 ]-graph. (2) If G1 is a [D1 , ∆1 , λ1 ]-graph and G2 is a [D2 , ∆2 , λ2 ]-graph, then G1 ⊗ G2 is a [D1 D2 , ∆1 ∆2 , max(λ1 , λ2 )]-graph. (3) If G is a [D1 , ∆1 , λ1 ]-graph and H a [∆1 , ∆2 , λ2 ]-graph, then G Z H is a [D1 ∆1 , ∆22 , f (λ1 , λ2 )]-graph with q  1 1 2 1 − λ22 λ1 + f (λ1 , λ2 ) = (1 − λ22 ) λ21 + 4λ22 ≤ λ1 + λ2 . 2 2 Let us study the following sequence of graphs. Let H be a [D, ∆, λ]-graph. Let G1 = H 2 , G2 = H ⊗ H, and for t = 3, 4, . . . let 2  Z H. (36) Gt = G⌈ t−1 ⌉ ⊗ G⌊ t−1 ⌋ 2 2 From lemma 27 it is easily seen that Gt is a [Dt , ∆2 , λt ]-graph with λt defined by λ1 = λ2 , λ2 = λ , λ2t−1 = λ + λ2t−1 , λ2t = max(λ + λ2t , λ for t = 2, 3 . . . , and + λ2t−1 ) , for t = 2, 3, . . . . 22 MATTI KARPPA, PETTERI KASKI, JUKKA KOHONEN, AND PADRAIG Ó CATHÁIN Lemma 28 (Reingold, Vadhan, and Wigderson [34, Theorem 3.3]). The rotation map RotGt can be computed in time poly(t, log D) and by making poly(t) evaluations of RotH . Lemma 29. If 0 ≤ λ ≤ 1/4 then λt ≤ λ + 4λ2 for all t ≥ 1. Proof. The conclusion is immediate for t ≤ 2. So suppose that the conclusion holds up to 2t − 2. We need to show that the conclusion holds for λ2t−1 and λ2t . By induction, it suffices to show that λ2t−1 ≤ λ + (λ + 4λ2 )2 ≤ λ + 4λ2 . Observing that λ2 + 8λ3 + 16λ4 ≤ 4λ2 holds for 0 ≤ λ ≤ 1/4 yields the desired conclusion. The proof for λ2t is identical.  Finally, we construct the expanders that we require in the manuscript proper. Lemma 30 (lemma 11 stated with normalized eigenvalue notation). For all integers t ≥ 1 and b ≥ 10 there exists a [216bt , 24b , 16 · 2−b ]-graph whose rotation map can be evaluated in time poly(b, t). Proof. Take q = 2b and d = 15 in Proposition 5.3 of Reingold, Vadhan, and Wigderson [34] to obtain a [216b , 22b , 15 · 2−b ]-graph H whose rotation map can be computed in time poly(b). (Indeed, observe that an irreducible polynomial to perform the required arithmetic in the finite field of order 2b can be constructed in deterministic time poly(b) by an algorithm of Shoup [35].) Let us study the sequence Gt given by eq. (36). The time complexity of the rotation map follows immediately from lemma 28. Since b ≥ 10, lemma 29 gives that λt ≤ λ + 4λ2 for all t ≥ 1. Take λ = 15 · 2−b and observe that since b ≥ 10 we have 2−b < 1/900. Thus, λt ≤ 15 · 2−b + 4(15 · 2−b )2 = 15 · 2−b + 900 · 2−2b ≤ 16 · 2−b .  References [1] Thomas D. Ahle, Rasmus Pagh, Ilya Razenshteyn, and Francesco Silvestri. On the complexity of inner product similarity join. arXiv, abs/1510.02824, 2015. [2] Josh Alman, Timothy M. Chan, and Ryan Williams. Polynomial representations of threshold functions and algorithmic applications. arXiv, abs/1608.04355, 2016. [3] Josh Alman and Ryan Williams. Probabilistic polynomials and Hamming nearest neighbors. In Proc. 56th Annual IEEE Symposium on Foundations of Computer Science (FOCS), pages 136–150, Los Alamitos, CA, USA, 2015. IEEE Computer Society. [4] Noga Alon. Problems and results in extremal combinatorics – I. Discrete Math., 273(1-3):31– 53, 2003. [5] Alexandr Andoni, Piotr Indyk, Huy L. Nguyen, and Ilya Razenshteyn. Beyond localitysensitive hashing. In Proc. 25th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1018–1028, Philadelphia, PA, USA, 2014. Society for Industrial and Applied Mathematics. [6] Alexandr Andoni, Thijs Laarhoven, Ilya P. Razenshteyn, and Erik Waingarten. Optimal hashing-based time-space trade-offs for approximate near neighbors. arXiv, abs/1608.03580, 2016. [7] Alexandr Andoni and Ilya Razenshteyn. Optimal data-dependent hashing for approximate near neighbors. In Proc. 47th ACM Annual Symposium on the Theory of Computing (STOC), pages 793–801, New York, NY, USA, 2015. Association for Computing Machinery. [8] Avrim Blum, Adam Kalai, and Hal Wasserman. Noise-tolerant learning, the parity problem, and the statistical query model. J. ACM, 50(4):506–519, 2003. [9] L. Elisa Celis, Omer Reingold, Gil Segev, and Udi Wieder. Balls and bins: Smaller hash families and faster evaluation. SIAM J. Comput., 42(3):1030–1050, 2013. [10] Timothy M. Chan and Ryan Williams. Deterministic APSP, orthogonal vectors, and more: Quickly derandomizing Razborov-Smolensky. In Robert Krauthgamer, editor, Proc. 27th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1246–1255, Arlington, VA, USA, 2016. Society for Industrial and Applied Mathematics. [11] Moshe Dubiner. Bucketing coding and information theory for the statistical high-dimensional nearest-neighbor problem. IEEE Trans. Inf. Theory, 56(8):4166–4179, 2010. EXPLICIT AMPLIFIERS FOR OUTLIER CORRELATIONS 23 [12] Vitaly Feldman, Parikshit Gopalan, Subhash Khot, and Ashok Kumar Ponnuswami. On agnostic learning of parities, monomials, and halfspaces. SIAM J. Comput., 39(2):606–645, 2009. [13] Aristides Gionis, Piotr Indyk, and Rajeev Motwani. Similarity search in high dimensions via hashing. In Malcolm P. Atkinson, Maria E. Orlowska, Patrick Valduriez, Stanley B. Zdonik, and Michael L. Brodie, editors, Proc. 25th International Conference on Very Large Data Bases (VLDB’99), pages 518–529, Edinburgh, Scotland, UK, 1999. Morgan Kaufmann. [14] Parikshit Gopalan, Daniek Kane, and Raghu Meka. Pseudorandomness via the Discrete Fourier Transform. In Proc. IEEE 56th Annual Symposium on Foundations of Computer Science (FOCS), pages 903–922, Berkeley, CA, USA, 2015. IEEE Computer Society. [15] Parikshit Gopalan, Raghu Meka, Omer Reingold, and David Zuckerman. Pseudorandom generators for combinatorial shapes. SIAM J. Comput., 42(3):1051–1076, 2013. [16] Elena Grigorescu, Lev Reyzin, and Santosh Vempala. On noise-tolerant learning of sparse parities and related problems. In Proc. 22nd International Conference on Algorithmic Learning Theory (ALT), pages 413–424, Berlin, Germany, 2011. Springer. [17] Wassily Hoeffding. Probability inequalities for sums of bounded random variables. J. Amer. Statist. Assoc., 58:13–30, 1963. [18] Shlomo Hoory, Nathan Linial, and Avi Wigderson. Expander graphs and their applications. Bull. Amer. Math. Soc., 43(4):439–561, 2006. [19] Russell Impagliazzo and Ramamohan Paturi. On the complexity of k-SAT. J. Comput. Syst. Sci., 62(2):367–375, 2001. [20] Piotr Indyk and Rajeev Motwani. Approximate nearest neighbors: Towards removing the curse of dimensionality. In Proc. 30th Annual ACM Symposium on the Theory of Computing (STOC), pages 604–613, New York, NY, USA, 1998. Association for Computing Machinery. [21] Daniel M. Kane, Raghu Meka, and Jelani Nelson. Almost optimal explicit JohnsonLindenstrauss families. In Proc. 14th International Workshop on Approximation, Randomization, and Combinatorial Optimization, RANDOM and 15th International Workshop on Algorithms and Techniques, APPROX, pages 628–639, Princeton, NJ, USA, 2011. [22] Michael Kapralov. Smooth tradeoffs between insert and query complexity in nearest neighbor search. In Proc. 34th ACM Symposium on Principles of Database Systems (PODS), pages 329–342, New York, NY, USA, 2015. Association for Computing Machinery. [23] Matti Karppa, Petteri Kaski, and Jukka Kohonen. A faster subquadratic algorithm for finding outlier correlations. In Proc. 27th Annual ACM-SIAM Symposium on Discrete Algorithms, (SODA), pages 1288–1305, Arlington, VA, USA, 2016. Society for Industrial and Applied Mathematics. [24] Pravesh K. Kothari and Raghu Meka. Almost optimal pseudorandom generators for spherical caps. In Proc. 47th Annual ACM Symposium on Theory of Computing (STOC), pages 247– 256, Portland, OR, USA, 2015. [25] François Le Gall. Faster algorithms for rectangular matrix multiplication. In Proc. 53rd Annual IEEE Symposium on Foundations of Computer Science (FOCS), pages 514–523, Los Alamitos, CA, USA, 2012. IEEE Computer Society. [26] A. Lubotzky, R. Phillips, and P. Sarnak. Ramanujan graphs. Combinatorica, 8:261–277, 1988. [27] Alexander May and Ilya Ozerov. On computing nearest neighbors with applications to decoding of binary linear codes. In Proc. EUROCRYPT 2015 - 34th Annual International Conference on the Theory and Applications of Cryptographic Techniques, pages 203–228, Berlin, Germany, 2015. Springer. [28] Elchanan Mossel, Ryan O’Donnell, and Rocco A. Servedio. Learning functions of k relevant variables. J. Comput. Syst. Sci., 69(3):421–434, 2004. [29] Rajeev Motwani, Assaf Naor, and Rina Panigrahy. Lower bounds on locality sensitive hashing. SIAM J. Discrete Math., 21(4):930–935, 2007. [30] Ryan O’Donnell, Yi Wu, and Yuan Zhou. Optimal lower bounds for locality-sensitive hashing (except when q is tiny). ACM Trans. Comput. Theory, 6(1):Article 5, 2014. [31] Rasmus Pagh. Locality-sensitive hashing without false negatives. In Proc. 27th Annual ACMSIAM Symposium on Discrete Algorithms (SODA), pages 1–9, Philadelphia, PA, USA, 2016. Society for Industrial and Applied Mathematics. [32] Ramamohan Paturi, Sanguthevar Rajasekaran, and John H. Reif. The light bulb problem. In Proc. 2nd Annual Workshop on Computational Learning Theory (COLT), pages 261–268, New York, NY, USA, 1989. Association for Computing Machinery. [33] Ninh Pham and Rasmus Pagh. Scalability and total recall with fast CoveringLSH. arXiv, abs/1602.02620, 2016. [34] Omer Reingold, Salil Vadhan, and Avi Wigderson. Entropy waves, the zig-zag graph product, and new constant-degree expanders. Ann. of Math., 155(1):157–187, 2002. 24 MATTI KARPPA, PETTERI KASKI, JUKKA KOHONEN, AND PADRAIG Ó CATHÁIN [35] Victor Shoup. New algorithms for finding irreducible polynomials over finite fields. Math. Comp., 54:435–447, 1990. [36] Gregory Valiant. Finding correlations in subquadratic time, with applications to learning parities and the closest pair problem. J. ACM, 62(2):Article 13, 2015. [37] Leslie G. Valiant. Functionality in neural nets. In Proc. 1st Annual Workshop on Computational Learning Theory (COLT), pages 28–39, New York, NY, USA, 1988. Association for Computing Machinery.
8cs.DS
Thermophysical Phenomena in Metal Additive Manufacturing by Selective Laser Melting: Fundamentals, Modeling, Simulation and Experimentation Christoph Meiera,∗, Ryan W. Pennya , Yu Zoua , Jonathan S. Gibbsa , A. John Harta,∗ arXiv:1709.09510v1 [physics.app-ph] 4 Sep 2017 a Mechanosynthesis Group, Department of Mechanical Engineering, Massachusetts Institute of Technology, 77 Massachusetts Avenue, Cambridge, 02139, MA, USA Abstract Among the many additive manufacturing (AM) processes for metallic materials, selective laser melting (SLM) is arguably the most versatile in terms of its potential to realize complex geometries along with tailored microstructure. However, the complexity of the SLM process, and the need for predictive relation of powder and process parameters to the part properties, demands further development of computational and experimental methods. This review addresses the fundamental physical phenomena of SLM, with a special emphasis on the associated thermal behavior. Simulation and experimental methods are discussed according to three primary categories. First, macroscopic approaches aim to answer questions at the component level and consider for example the determination of residual stresses or dimensional distortion effects prevalent in SLM. Second, mesoscopic approaches focus on the detection of defects such as excessive surface roughness, residual porosity or inclusions that occur at the mesoscopic length scale of individual powder particles. Third, microscopic approaches investigate the metallurgical microstructure evolution resulting from the high temperature gradients and extreme heating and cooling rates induced by the SLM process. Consideration of physical phenomena on all of these three length scales is mandatory to establish the understanding needed to realize high part quality in many applications, and to fully exploit the potential of SLM and related metal AM processes. 1. Introduction Additive manufacturing (AM) offers the opportunity to produce parts with high geometric complexity, without the requirement for dedicated tooling. AM processes for polymer parts are quite well established, yet melt-based processes for AM of metals still exhibit severe practical challenges, many of them resulting from the high melting temperatures of metals and their relatively low viscosities [104]. In selective laser melting (SLM) of metals, the arguably most prominent representative of powder bed AM methods, a 3D manufacturing task is digitally segmented into thin 2D layers [48]. A solid part is simply formed by selectively melting pre-defined contours in successive layers of powder using a focused laser beam. After one layer of powder has been scanned, the regions melted by the laser form the cross-section of the final part. Subsequently, the underlying build platform is lowered down and a further layer of powder is deposited by means of a powder coater mechanism. This procedure is successively repeated until the final 3D geometry is completed and the remaining unfused powder is then removed (see Figure 1). In this context, the so-called build direction denotes the direction normal to the powder bed. Further, the orientation of the part geometry with respect to this (vertical) build direction has crucial influence on the resulting part properties [65]. SLM of metals offers significant advantages including near-net-shape production without the need for expensive molds or time consuming post-processing, high material utilization rate and highest production flexibility [145, 143]. Most essentially, the layer-wise production leads to nearly unlimited freedom of design, which enables the generation of highly complex geometries that cannot be obtained by conventional manufacturing processes. This paradigm shift in mechanical design allows to integrate complex substructures such as lattice-based geometries enabling lightweight ∗ Corresponding authors Email addresses: [email protected] (Christoph Meier), [email protected] (A. John Hart) Preprint submitted to Annual Review of Heat Transfer September 28, 2017 mesoscale roller atmospheric gas pressure laser beam gas flow macroscale powder reservoir melt pool cyclic heating surface tension heat flux concentration spatter ejected particle vapor recoil melting wetting solidification microscale platform segregation powder layers columnar grains SLM part + support structure build platform heterogeneous nucleation Figure 1: Experimental setup of SLM process as well as schematic visualization of macroscale, mesoscale and microscale view. yet sufficiently stiff components. Possible fields of application include aerospace or medical engineering and basically all industries requiring highly complex and individualized parts (see Figure 2). Figure 2: Exemplary metal parts fabricated by SLM/EBM: (a) aircraft bracket (approx. 150 mm width), by Airbus; (b) acetibular cup for human hip implant (approx. 75 mm diameter), by Arcam; (c) injection mold tool with internal cooling passages (approx. 200 mm width), by Sodick. The parts are post-processed, including support removal, machining, and polishing, in addition to the SLM/EBM process as the construction method. The overall SLM process is highly complex and governed by a variety of (competing) physical mechanisms. The most important effects and physical phenomena occurring in the powder bed, the melt pool and the solidified phase of typical SLM systems are summarized in the following three subsections, as well as in Figure 1. A typical schematic parametrization of the powder layer and laser beam as well as a typical local scan pattern are shown in Figure 3. 1.1. Physical phenomena within the powder bed The incident laser beam is a collimated, polarized, monochromatic electro-magnetic wave, with a wavelength in the range of w ≈ 1µm. The spatial power density distribution of incident radiation on the powder bed is commonly assumed to follow a Gaussian distribution, with the associated 2σ value typically being taken as the laser beam spot size. Typical values for the nominal laser power and the employed laser beam velocities are in the range of P ≈ 50−1000W and V ≈ 0.1 − 3m/s [146]. The effective laser beam absorption within the powder bed is governed by multiple reflections of incident laser rays within the open-pore system of the powder bed, each with partial absorption of the incident radiation. The laser beam can penetrate to considerable depths, which can even reach the range of the powder layer thickness [58, 59]. Thus the net absorptivity of powder beds is considerably higher than the value known for 2 contour scan current melt track l V h subsequent melt track 2rB infill hatching powder layer substrate Figure 3: Typical geometrical dimensions and process parameters (left) as well as typical scan pattern (right) characterizing the SLM process. flat surfaces and, moreover, the laser beam energy source must to be thought of a volumetric heat source distributed over the powder bed thickness, as opposed to a surface heat source. The factors influencing overall absorption and local energy distribution are numerous, including the laser beam power, wavelength, polarization, angle of incidence, powder temperature, surface roughness, surface chemistry (e.g. oxidation) and contamination [13, 81]. Issues of powder bed morphology, determined by particle shape, size distribution and packing density, are also central to radiative transfer. A further important factor is given by the intra- and inter-particle heat transfer within the powder bed. The inter-particle heat transfer is typically governed by the gas in the powder bed pores, with commonly negligible overall conductivity contributions from particle-to-particle contact points as long as loose, i.e. not mechanically compressed, powder layers are considered. Consequently, the thermal conductivity of loose powder is comparable to the conductivity of gas and by orders of magnitude smaller than the conductivity in the solidified phase [132]. Also when considering the intra-particle heat transfer, it can be observed that the time scales governing this process are typically larger than the time scales governing particle melting. In other words, under typical SLM process conditions, there is not enough time for conductive homogenization of non-uniform energy and temperature distributions across the powder bed but also across individual particles. As consequence, partially molten particles may cause defects such as pores or inclusions [13]. A further characteristic of the SLM process is that the powder bed, when considering it as a homogenized continuum, shows mesoscopic heterogeneities (in form of individual particles) that are in the same order of magnitude as relevant macroscopic process length scales such as powder layer thickness and laser beam spot size. Typically, powder particle sizes in the range of 2R ≈ 10µm−50µm, layer thicknesses in the range of l ≈ 20µm−100µm and laser beam spot sizes in the range of 2rB ≈ 20µm−200µm are employed. Also, in standard linear scan patterns the distance between two successive laser tracks, denoted as hatch spacing h, is an important process parameter. This is typically chosen in the range of rB such that a sufficient overlap and remelting between two subsequent tracks is guaranteed (Figure 3). A good overview of the typically applied range of these process parameters is also given in [146]. The large size of individual powder grains as compared to powder layer thickness and laser beam spot size typically leads to non-uniform energy distributions, across the entire powder bed but also across individual particles, which may have considerable influence on the resulting melting behavior and melt pool hydrodynamics. Furthermore, these comparatively large heterogeneities cause differences in the resulting temperature fields and melt track shapes when considering different samples of stochastically equivalent powder layers (i.e. identical particle shape and size distribution, powder layer density and thickness etc.). Consequently, the variance of process results due to the stochastic nature of the powder layer is considerably greater than for comparable processes such as laser beam welding (LBW, see below). Besides the energy input by the laser beam, also possible energy losses in form of thermal radiation emission, thermal convection or heat conduction from the solidified material to the underlying built platform play an important role in the overall SLM process. For further information on powder bed radiation and heat transfer in the context of SLM, the interested reader is referred to [132, 59, 56, 13]. 1.2. Physical phenomena within the melt pool As soon as the melting temperature is reached at local positions on the powder grain surface, the phase transition from solid to liquid as well as the formation of a melt pool and, ideally, a continuous melt track is induced. Driven by surface tension and capillary forces tending to minimize surface energy, a coalescence of individual melt drops and a reshaping of the resulting melt pool is initiated [91, 92]. In addition, the wetting behavior of the low-viscosity melt 3 on the underlying substrate, i.e. the solidified material formed by previous layers, and surrounding powder grains influences the resulting melt pool shape, continuity and adhesion to the previous layer. The wetting behavior crucially depends on the material, temperature, surface roughness, and surface chemistry. Oxidation on the powder grain or substrate surfaces - either due to contaminated primary powder material or due to thermally induced oxidation during the process - is known to considerably decrease the wetting behavior of the melt which might result in instable, balled melt pools and rough surfaces, pores or delamination due to insufficient layer-to-layer adhesion [36]. The prevalent length and time scales essentially determine which physical effects govern the process and which are negligible. Typically, viscous and gravity forces can be considered as secondary effects while surface tension and capillary forces, wetting behavior but also inertia effects are the primary driving forces that influence the melt pool dynamics and shape as well as the surrounding powder morphology by attracting or rejecting individual grains [92]. The heat transfer within the melt pool is governed by convection rather than by heat conduction, with Marangoni convection, i.e. melt flow from hot to cool regions induced by temperature-dependent surface tension, playing a prominent role in this process [83]. Depending on the amount of absorbed energy density and the surrounding atmospheric pressure, the peak temperature within the melt pool might exceed the boiling temperature and considerable material evaporation may take place [63]. The evaporation itself as well as the gas flow induced by evaporation may influence the melt pool thermo-hydrodynamics and the overall process as consequence of an evaporative mass loss and additional cooling, of a recoil pressure considerably distorting the melt pool surface and representing a means of transport for potential pollutants, of melt drops spattered out of the pool and even of powder particles ejected away from the direct vicinity of the laser beam [83]. Driven by the so-called keyholing mechanism, cavitation resulting from evaporated material might considerably contribute to the overall material porosity or even burst the surrounding solidified material due to thermal expansion of trapped gas [105]. As soon as the melt pool has solidified, the evolution of possible defects on the mesoscopic scale is virtually established. A discussion of the physical phenomena governing the melt pool thermo-hydrodynamics can e.g. be found in [36, 97, 83, 92, 101]. 1.3. Physical phenomena within the solidified phase With the solidification of the melt pool, the development of the metallurgical microstructure, crucially determining the macroscopic properties of the final part, begins. The evolution of the solid-phase microstructure characterized by grain size, grain shape (morphology) and grain orientation (texture) are governed by the prevalent spatial temperature gradients, the cooling rates, as well as the velocity of the solidification front [49]. In SLM processes, two regimes can be distinguished: The first regime is given by the temperature field in the direct vicinity of the laser beam, in the so-called heat affected zone (HAZ) [161], which is controlled by highly complex mechanisms such as radiation absorption and heat conduction in the powder bed as well as convective heat transfer within the melt pool with all the individual physical phenomena and process parameters of influence as discussed in the two previous paragraphs. The material in this region is subject to a rapid heating above melting temperature due to the absorption of laser energy by powder grains, a high velocity of the melt pool front induced by the laser beam velocity as well as a rapid solidification of the molten material after the heat source has moved on, which is a direct consequence of the large ratio of solid material to hot molten material. These pronounced non-equilibrium conditions lead to meta-stable microstructures and compositions of the resulting phases, as well as smaller grain sizes - which typically result in higher material strengths - as compared to traditional melt-based manufacturing processes such as casting [68, 65]. A second regime of thermal evolution is prevalent in previously deposited material layers located below the current layer and further away from the heat source. Each solidified location experiences repeated heating and cooling cycles with decreasing amplitudes as the laser processes adjacent scan tracks, and as it processes consecutive new layers. The heat transfer in this regime is rather determined by global part properties, e.g. the global laser beam scanning strategy, the build direction, the fixation of the part on the built platform, the temperature of the built platform but also by the part porosity and the metallurgic microstructure distribution itself, which both influence the (effective) thermal conductivity. Of course, microstructure also depends on the specific part geometry, e.g. due to heat flux concentration at the transition region from bulk material to slender columns or thin walls (see e.g. Figure 1), which are surrounded by the low conductivity unfused powder [116, 106]. Also the evolution of columnar grain structures oriented in direction of the main temperature gradients, usually in build direction, is typical for SLM processes and often yields a strongly 4 anisotropic macroscopic material behavior with higher material strength in the build direction [68, 51]. With increasing distance from the top powder layer, the maximal temperature values and gradients experienced by a material layer during the repeated thermal cycles decrease and, similar to a heat treatment, these cycles might lead to a coarsening of the microstructure, a reduction of brittle non-equilibrium phases and, consequently, to more ductile material characteristics. As consequence of these effects, the final part will typically exhibit a change of microstructure in built direction: On the one hand, the initial creation of fine grain structures and non-equilibrium phases will be more pronounced in the first material layers deposited in the direct vicinity of the build platform where higher thermal conductivity and faster cooling rates are prevalent. On the other hand, these initially deposited powder layers are exposed to the heat treatment of repeated heating and cooling cycles for longer times, which might lead to longer evolution times for solid phase transformations and grain coarsening. Apart from the microstructure evolution considered so far, the high temperature gradients in the direct vicinity of the melt pool, but also at locations of heat flux concentration, are giving rise to considerable thermal strains induced by a successive thermal expansion and shrinkage of material. These thermal strains result in thermal stresses within the kinematically constrained SLM part. The magnitude of these stresses is essentially determined by the underlying solid microstructure and the resulting macroscopic material behavior. A ductile material can compensate these thermal strain variations by means of local plastic flow. On the contrary, brittle material behavior fostered by small grain sizes, the existence of certain non-equilibrium phases or a local segregation of alloying elements [145] might result in cracks at locations of stress concentration such as residual pores or inclusions. The complexity of this thermo-mechanical coupling is further increased by the fact that the microstructure does not only influence the amount of residual stresses, but that also the prevalence of residual stresses, which are for example known to mechanically stabilize the metastable austenitic phase in steels [68], will influence the evolution of the microstructure. In order to reduce residual stresses on large surface areas, these surfaces are typically subdivided in smaller islands that are completed successively. Figure 3 (right) illustrates the processing of a single island consisting of contour scan and infill hatching. During the SLM process, the amplitude of residual stresses might also decrease since stress relaxation is likely to occur during the repeated heating and cooling cycles at lower temperature levels. Furthermore, annealing is typically applied to the final part before removing support structures in order to relief residual stresses. While a neglect of support structures would result in reduced residual stresses, this advantage has to be paid for by dimensional warping as often observed in parts made by SLM. For further information on material aspects and microstructure evolution in the context of SLM, the interested reader is referred to [87, 15, 67, 68, 133]. Exemplary references focusing on the investigation of the residual stresses resulting during the SLM process are [70, 69, 38, 130, 161, 106]. All in all, it can be concluded that the entire thermal history between solidification and cooling down to the ambient temperature, governed by many heating and cooling cycles at different temperature levels and time scales, considerably determines the resulting metallurgic microstructure as well as the macroscopically observable material properties such as ductility, micro hardness, yield strength or tensile strength as well as their spatial distribution in a possibly inhomogeneous and anisotropic manner. On the one hand, a restriction of the overall process parameters is required in order to avoid undesirable material characteristics and possible part defects such as excessive residual stresses, dimensional warping, crack propagation or delamination of layers, effects which might destroy the SLM part during the build process or at least reduce the mechanical resilience of the final part considerably. In order to fully exploit the efficiency potential of SLM, this result has to be achieved without the need for time- and cost-intensive post-processing as required by alternative processes such as selective laser sintering (SLS, see below). On the other hand, the flexibility of the SLM process offers the unique opportunity to optimize process parameters in order to manufacture parts with prescribed inhomogeneous and anisotropic microstructures and macroscopic material properties in a controlled manner, contributing to the paradigm shift in design enabled by SLM. 1.4. Differentiation of related additive manufacturing processes for metals References and comparisons to other powder-bed metal AM processes such as electron beam melting (EBM) and selective laser sintering (SLS) will be useful from time to time in the foregoing discussion [97, 98, 50, 52, 68]. Similar to SLM, the EBM process represents a powder bed-based additive manufacturing process where pre-defined contours are selectively melted in successively deposited powder layers. While SLM applies a laser beam as energy source, EBM is based on an electron beam. EBM is only applicable to electrically conductive materials and has to be 5 performed in a near-vacuum environment in order to avoid an interaction of the electron beam with surrounding gas molecules. On the other hand, SLM is also suitable for dielectric materials and has to be performed in an inert gas atmosphere in order to prevent surface oxidation of (metallic) powder and substrates. Also the way of energy transfer into the powder bed is different: When considering one individual powder grain, the laser beam radiation is (in good approximation) absorbed at the powder grain surface while in EBM, electrons penetrate the powder grain surface, a process in which kinetic energy of the electrons is dissipated leading to melting of the grain. However, the energy of the electron beam is typically completely deposited to the powder grain of first incidence. On the contrary, only a part of the laser beam energy is directly absorbed at the powder grain of first incidence while the remaining part is reflected, leading to considerably higher powder bed penetration depths due to multiple reflections in the open pore system provided by the powder bed. Eventually, the laser beam in SLM is suitable for smaller, more focused spot sizes as well as smaller powder grains sizes, which enables finer geometrical resolutions and an improved surface quality. The higher concentration of incident energy in SLM might yield higher local cooling rates and, thus, smaller grain sizes of the resulting metallurgical microstructure. Similar to SLM, SLS is based on a laser beam energy source. However, in contrast to SLM and EBM, the particles within the powder bed are not fully molten. In this context, solid and liquid state sintering can be distinguished. Solid state sintering is governed by thermally activated diffusion of atoms at temperatures below the melting point leading to slowly growing necks between adjacent powder particles. However, the slow time scales governing this process and the resulting low output rates make it often infeasible from an economic point of view. On the contrary, liquid state sintering aims at a partial melting of powder. The molten liquid will typically spread almost instantaneously between the unmolten particles and acts as binder. In order to achieve defined ratios of molten and unmolten material, either a second material species with lower melt point is employed as binder, bi-modal powder mixtures of one and the same material are used such that the smaller grains melt earlier than the larger ones or the incident energy density has to be adjusted such that only the top surface of powder grains melts while the core remains in solid state. These sintering processes often result in so-called green parts with high porosity, which requires subsequent heat treatment processing. Related processes such as directed energy deposition (DED) and wire-feed laser and electron beam additive manufacturing methods share some similarities with SLM and EBM, yet they generally operate on larger length scales (mm − cm scale molten trajectory). Also the processes of Laser Beam Welding (LBW) and Electron Beam Welding (EBW) will be considered at some points in this work, since some of the underlying physical mechanisms are similar to SLM and EBM. While SLM and EBM aim at additively generating solid material structures out of layers of loose powder, LBW and EBW intend the connection of individual solid parts by partially melting their contact surface. 1.5. Organization of this article The remainder of this article is structured as follows: In Section 2, the governing physical mechanisms are further detailed, modeling as well as numerical (and analytical) solution procedures in the context of SLM processes are reviewed and findings derived by these approaches are discussed. Section 2.1 focuses on the modeling of radiation and heat transfer in powder beds. In Sections (2.2)-(2.4), a comprehensive overview of existing approaches classified by means of the three main categories of SLM modeling approaches found in the literature, namely macroscopic, mesoscopic and microscopic models, will be given. Section 3 focuses on experimental studies that are especially relevant to understand the thermophysical mechanisms. Following a similar structure, Section 3.1 deals with experimental approaches of powder bed characterization while Sections (3.2)-(3.4) refer to experimental investigations on effects visible on macroscopic, mesoscopic and microscopic level, thus representing the counterpart to the Sections (2.2)-(2.4) considering modeling and simulation. Finally, Section 4 provides recommendations concerning practical implementation of the SLM process and gives open questions and potentials for future process improvement. 2. Modeling and simulation approaches Approaches to modeling of SLM can be classified in macroscopic, mesoscopic and microscopic models. Macroscopic simulation models typically treat the powder phase as a homogenized continuum resulting in efficient numerical tools capable of simulating the manufacturing of entire parts by SLM. Macroscopic models commonly aim at determining 6 spatial distributions of temperature, residual stresses as well as dimensional warping within SLM parts. Mesoscopic models typically resolve individual powder grains and melt pool thermo-hydrodynamics in order to determine part properties such as adhesion between subsequent layers, surface quality as well as creation mechanisms of defects such as pores and inclusions. Microscopic models consider the evolution of the metallurgical microstructure involving resulting grain sizes, shapes and orientations as well as the creation of thermodynamically stable or unstable phases. The computational effort required for mesoscopic models currently limits the application of these models to single track simulations. However, the insights gained by mesoscopic models might serve as basis to further improve continuum models for powder and melt phase in macroscopic models. Similarly, microscopic models can readily address small areas of the part in the range of the powder layer thickness, and existing approaches are commonly limited to 2D. Nevertheless, they may serve as basis for the development of improved inhomogeneous and anisotropic continuum constitutive models whose quality is essential for the quality of simulation results derived by macroscopic simulation models. Macroscopic as well as mesoscopic models commonly require submodels for the radiation transfer into and the heat transfer within the powder phase. For this purpose, modeling approaches considering powder bed radiation and heat transfer are discussed in Section 2.1. A challenge more severe than the actual derivation of physical models, is the application and development of powerful discretization techniques and numerical solution schemes in order to enable a robust and efficient computational solution, two key factors in the simulation-based characterization of SLM processes. While the finite element method (FEM), the finite difference method (FDM) as well as the finite volume method (FVM) represent spatial discretization schemes typically employed in the considered models, temporal discretization is almost exclusively based on explicit or implicit finite difference time integration schemes. For implicit schemes, the fully discretized problem is typically represented by a system of equations that is nonlinear in the unknown discrete primary variables, which requires the application of a nonlinear solver, e.g. a Newton-Raphson scheme. Depending on the characteristics of the nonlinear system of equations to be solved, convergence of the nonlinear solution scheme might not always be guaranteed. On the contrary, explicit schemes allow for a direct extrapolation of the known configuration at time tn to the unknown configuration at time tn+1 . The resulting system of equations is linear in the discrete unknowns such that no iterative, nonlinear solution process is required. However, at least in the geometrically linear regime, implicit time integrators can be proven to be unconditionally stable, thus typically allowing for considerably larger time step sizes as compared to explicit schemes in order to preserve system stability, i.e. to keep the total system energy bounded during the simulation. Consequently, implicit schemes are favorable for problems that are dominated by a low frequency response, where the large time step sizes possible for these schemes are sufficient in order to resolve the low-frequent system answer. On the contrary, explicit schemes are rather suited to model high frequency responses and wave-like phenomena such as high velocity impacts. There, small time step sizes are required for both explicit and implicit schemes in order to accurately resolve the high-frequent system dynamics. Since the computational effort per time step is lower for explicit schemes and no nonlinear convergence has to be accounted for, these schemes are preferable in such scenarios. The stability-relevant time step sizes dictated for explicit schemes in the context of mesoscopic SLM models are typically considerably smaller than the time step sizes required to capture the relevant physical phenomena. Consequently, implicit schemes would have the potential to achieve substantial computational savings. However, the complexities arising from the multiple field and domain couplings as well as the geometrical characteristics prevalent in SLM make the implicit treatment of fully resolved mesoscopic models a challenging task. A further important question to be asked from a numerical point of view is the way different physical fields (e.g. thermal and mechanical fields) and domains (e.g. powder, melt and solid phase) are coupled. Here, three concepts of treating phase boundaries shall be distinguished: The first category models a sharp boundary between the phases resulting in a jump of material properties and physical fields under consideration of displacement / velocity continuity conditions and mechanical equilibrium at the interface. Numerical realizations of such models are typically based on explicit interface tracking, e.g. via level set schemes, and discrete solution spaces allowing for discontinuity in the primary variables, e.g., the extended finite element method (XFEM). The second category of schemes still identifies the interface in an explicit manner, e.g., by means of an additional phase variable φ taking on the value φ = 1 for the one phase and φ = 0 for the other. However, the interface is not represented by a sharp boundary, but rather by a transition range of finite thickness characterized by phase variable values φ ∈]0; 1[. An example for such a scheme is the volume of fluid method (VOF). A third category is given by methods that do not introduce additional variables in order to 7 distinguish the phases. These schemes typically treat the phase boundary implicitly by means of specific values of existing primary fields, e.g., by defining the phase boundary between liquid and solid as the isothermal contour T = T m representing the melt temperature T m . The rapid change of physical properties at the phase boundary is considered by these schemes in terms of high gradients in the applied material parameters. Thus, again, the interface is extended to a finite thickness. A too small transition region might lead to excessive gradients in the material parameters and, as consequence, to ill-conditioned discrete problems that are challenging to be solved numerically. The SLM modeling approaches reviewed in this work almost exclusively rely on the second and third category of approaches. Numerical schemes can also be distinguished concerning the succession of solving different physical fields and domains. Namely, monolithic schemes solve the entire multi-physics problem at once, and partitioned schemes solve the different physical fields/domains subsequently. The partitioned schemes can be further subdivided into so-called iterative or strongly coupled partitioned schemes that iterate between the different fields several times within a time step until achieving the solution of the monolithic problem statement and so-called staggered or weakly coupled partitioned schemes that solve for each field only once per time step without additional iterations, eventually leading to a solution that differs from the monolithic one. Monolithic and strongly coupled partitioned schemes are typically combined with implicit time integrators because these coupling schemes can preserve the desirable stability properties of implicit integrators at large time step sizes. On the other hand, weakly coupled schemes are often combined with explicit time integrators since these schemes also lead to low computational costs per time step. However, for weakly coupled partitioned schemes, even when combined with implicit time integrators on the individual fields, the stability requirement often becomes very restrictive typically dictating very small time step sizes and in some cases even prohibiting stability at all (see e.g. [44, 19] for a discussion in the context of fluid structure interaction (FSI) simulations). Since time step size correlates with computational costs, these considerations are of high practical interest. 2.1. Modeling of optical and thermal properties of powder phase The optical and thermal properties of the powder bed, in combination with the laser characteristics, crucially determine the heat distribution in the powder bed and the subsequent melt pool dynamics. This section addresses approaches to modeling the optical and thermal properties of the powder bed as well as the radiation-dominated energy transfer from the laser beam source into the powder bed and the conduction-dominated heat transfer within the powder bed. The considered approaches are typically employed as powder bed submodels within the three main categories of macroscopic, mesoscopic and microscopic models. Often, the laser beam can be described in good approximation by means of a Gaussian power distribution [40, 149, 131]. In a first step, the question of how the laser energy transfers into the SLM powder bed, a process that can be described via the principles of thermal radiation, geometrical optics and electro-magnetic wave propagation, will be considered. When it is not mechanically compressed after spreading, the powder bed has as high porosity as freely poured powder, which is in the range of 40% − 60% for typical SLM powders [132]. As shown in [58, 59], laser radiation penetrates into powder through pores to a depth of several particle diameters because of multiple reflections [154]. This depth is comparable with the powder layer thickness. Thus, laser energy is deposited not on the surface but in the bulk of the powder layer. The incident laser energy consists of portions that are reflected, absorbed and transmitted when impinging upon the surface of powder particles. The emission of radiation from the powder particles themselves can often be neglected. In terms of radiation reflection, it can be distinguished between specular (mirror-like) and diffuse reflection, with the latter exhibiting a reflection intensity that is equally distributed over all possible directions. Concerning radiation transmission, typically the classification of opaque, semi-transparent or transparent particles is made. The multiple reflection/absorption processes of the light in the powder bed is additionally influenced by powder characteristics such as mixture ratio, mean particle size and shape, size distribution, packing density, powder bed depth but also by the laser beam spot size. Also the polarization of the laser beam relative to the powder particle surfaces can lead to non-uniform energy absorption even across individual powder particles. Since the thermal conduction in the powder bed is typically weak due to the prevalent porosity and melting time scales are typically smaller than heat conduction time scales for individual grains, non-uniform energy absorptions across the powder bed but also across individual powder particles will in general have considerable influence on the melt pool shape and the properties of the solidified track. A further factor of influence is given by possible surface oxidation and contamination effects. 8 An extensive review on the topic of radiation transfer within heterogeneous/porous/dispersed media can be found in [150] as well as [10]. The following section will mainly focus on approaches in the context of SLM. According to [10], the modeling approaches for radiation transfer in heterogeneous media can be classified as models based on a continuum formulation of the radiation transfer equation (RTE) well-known for homogeneous (participating) optical media, and models based on a discrete formulation of the RTE, which typically leads to ray tracing schemes. Both approaches are based on the simplifying assumptions of geometrical optics which considers light rays that: propagate in rectilinear paths as they travel in a homogeneous medium; bend, and in particular circumstances may split in two at the interface between two dissimilar media; follow curved paths in a medium in which the refractive index changes; and may be absorbed or reflected. 2.1.1. Continuum model for powder bed radiation transfer The general radiation transfer equation (RTE) known for homogeneous media and underlying the homogenized, continuum models for powder bed radiation transfer is (see also [10, 59, 22]): σ Ω 5I(x, Ω) = −(σ + κ)I(x, Ω) + κIeb + 4π Z4π I(x, Ω0 )S c (Ω0 → Ω)dΩ0 . (1) 0 The RTE describes the rate of the direction- and position-dependent radiation energy flux density I(x, Ω) based on an energy balance considering radiation absorption (first term on the right-hand side), radiation emission (second term on the right-hand side) and scattering (first and third term on the right-hand side). In this context, Ω represents a directional unit vector, dΩ is an infinitesimal solid-angle increment whose magnitude is per definition identical to an infinitesimal surface element dA on a unit sphere, and the product ΩI(x, Ω)dΩ represents the vector of energy flux density transferred by photons in direction of the unit vector Ω within a solid angle increment dΩ at position x. The constants σ and κ are the scattering and absorption coefficients, which are, however, often replaced by the alternative constants extinction coefficient β = σ + κ and albedo (portion of reflected light) ω = σ/β. The mapping S c (Ω0 → Ω) is the (normalized) scattering phase function stating the probability that radiation in Ω0 -direction is scattered to the Ω-direction and Ieb is Planck’s blackbody function representing radiation emission at a certain position: 4 Ieb = kS B (T 4 − T re f ). (2) Here, T is the powder temperature, T re f is the temperature of the ambient gas atmosphere and kS B is the StefanBoltzmann constant. While the radiation transfer equation has originally been derived for homogeneous media, it is well-established to apply (1) also to heterogeneous systems, where the continuous quantities and parameters introduced so far have to be replaced by their effective counterparts, determined via spatial averaging over a reference volume that has to be much greater than the length scales of prevalent heterogeneities (e.g., of individual particles in powder beds). When considering radiation heat transfer in powder beds, the extinction coefficient β is typically determined by the structure of powder, i.e., by size and shape of particles and by their arrangement, but is independent of optical properties of the material of particles like reflectance. On the contrary, the scattering characteristics such as the albedo ω and phase function P(Ω0 → Ω) commonly depend on reflective properties of the material of particles. In the following, a coordinate system is chosen such that the positive z-axis points into powder layer thickness direction with z = 0 representing the upper surface of the powder bed and z = L the boundary between powder bed (with thickness L) and the underlying solidified phase (substrate). The net radiation heat flux density qr , following the typical definition as heat flux per unit area measured in [W/m2 ], results from the energy flux density per unit angle increment I(x, Ω) via integration over all directions of incident radiation. Expressed by means of solid-angle increments dΩ this is equivalent to an integration over the surface 4π of a unit sphere. Assuming that the x− and y−components of the heat flux density qr resulting from a solution of (1) are negligible, also the incident power density per unit volume u s can be determined based on the following relations: qr = Z4π IΩdΩ ≈ qrz ez → 0 9 u s := − ∂qrz . ∂z (3) Gusarov et al. [59] considered the general problem of normal incidence of collimated radiation on a thin powder layer consisting of opaque (i.e. optically non-transparent) particles of arbitrary shapes and dimensions placed on a reflecting substrate in order to derive resulting net absorptivities and deposited energy profiles. It is assumed that the particle size is sufficiently small as compared to the laser spot size and powder bed thickness as to allow for the employed homogenization procedure. Further, in SLM the particle size is typically much greater than the wavelength of radiation, thereby justifying the applicability of the geometrical optics theory. The contribution κIeb representing the emission from individual powder grains has been neglected due to the comparatively high energy density of the incident laser radiation. Gusarov et al. [59] propose a statistical model based on the powder porosity and specific powder surface areas in order to determine the model parameters required in (1). This statistical model accounts for dense powder beds where particles touch each other and radiation transfer cannot be treated as scattering by independent particles, an effect that is referred to as dependent scattering (see e.g. [42, 24, 79, 76, 78, 144]). Moreover, an analytical solution for the radiation transfer equation has been derived via the so-called two-flux method with the boundary conditions of normally collimated incident flux qr0 := qr (z = 0) and specular reflection at z = L. Among others, the additional assumption of the laser spot size being much larger than the absorption depth has been made. Thus, the radiation transfer problem could be considered as 1D-problem in z−direction with an incident flux qr0 = qr0 (x, y)ez . Figure 4: Comparison of deposited energy profiles resulting from the RTE (full lines) and ray tracing simulation (broken lines) for WC-Co or Fe-Cu powders with particles sizes 2R = 50µm (Fe and WC), 2R = 30µm (Cu) and 2R = 20µm (Co) and laser beam wave lengths of 1.06µm or 10.6µm, [59]. In Figure 4, the resulting energy deposition profiles derived by the RTE solution as proposed in [59] and by means of ray tracing simulation are compared for three examples considering either WC-Co or Fe-Cu mixed powders with particles sizes 2R = 50µm (Fe and WC), 2R = 30µm (Cu) and 2R = 20µm (Co) as well as laser beam wave lengths of 1.06µm or 10.6µm. Accordingly, these two modeling approaches are in good agreement since a sufficiently deep powder bed and a sufficiently large laser beam size has been chosen. The total/integrated laser energy absorbed in a thin powder layer on a reflective substrate increases with its thickness while the deposited energy density per unit volume decreases. The absorption depths in Figure 4 seem to be comparatively high as compared to typical values known from SLM powder layers, which might be an indication for rather low packing densities considered in this study. Subsequently, the model of [59] has e.g. been applied in the independent works [149], [81] and [70]. In [56], the model of [59] has been further extended and refined. There, it has been shown that the application of the model derived in [59] is only reasonable if either one of the two phases (powder particles and pores) are opaque, one phase has a much larger volume fraction, or the radiation properties of the two phases are at least similar. On the contrary, an application of this model, which simply averages the radiation intensities across both phases, to problems where the two phases capture comparable volume fractions but show considerably different optical properties at the same time (e.g., mixed powders), seems unjustified. For mixed materials, a model based on partial homogenized radiation intensities and denoted as Vector Radiation Transfer Equation (VRTE) has been proposed in [56]. In this model, the partial values are obtained by averaging over each individual phase. It consists of two transport equations for the partial homogenized radiation intensities. The 10 equations are similar to the conventional RTE, but contain additional terms that take into account the exchange of radiation between the individual phases. The structure of the medium is specified by volume fractions of phases and by specific surfaces of phase boundaries. The optical properties are described by refractive indices and absorption coefficients of phases. It has been verified that the vector RTE model reduces to the conventional RTE model in case one of the two phases is opaque or one phase prevails in volume. Compared to the RTE in [59], an analytical solution was no longer achievable for the VRTE [56]. Instead, a discrete-ordinate method [18] was exploited. 2.1.2. Ray tracing model for powder bed radiation transfer As pointed out in Boley et al. [13], the assumption of a RTE continuum model [59] is questionable for very thin, low-porosity metal powder layers with a layer thickness in the range of a few powder particles and/or a laser spot size comparable to the size of the powder particles. In such scenarios, which are often prevalent in SLM processes, the averaging process underlying the RTE continuum models lead to considerable model errors as compared to the exact solutions, which typically yield an energy absorption that is highly non-uniform with respect to the spatial position of the laser beam. Ray tracing modeling is one possible approach to resolve such heterogeneities. However, it shall be emphasized that depending on the spatial resolution required by the physical question to be answered (e.g. when global residual stress distributions but not the specific locations of individual pores are of interest), also the application of RTE continuum models might be reasonable, which typically require a considerably lower computational effort as compared to ray tracing simulations. In ray tracing simulations, the total energy emitted by the laser beam in a certain time interval is represented by a discrete ensemble of rays with defined spatial position, orientation and energy. The position and energy associated with the individual rays is typically chosen such that the overall energy emission but also the spatial energy distribution resulting from the entire ensemble equals the corresponding characteristics of the laser beam (e.g. a Gaussian energy distribution). After defining the individual rays, the path of each ray is traced until striking an obstacle (powder particle). Based on the optical properties of the obstacle surface, part of the ray energy is absorbed whereas the remaining part of the ray energy is represented by a reflected ray with defined energy and orientation, which will further be traced trough the powder bed (Figure 5(a)). Commonly, several reflections are considered for each ray until the remaining energy drops below some predefined threshold. The individual ray energy contributions absorbed by each particle are accumulated during the simulation. Also the ray-tracing model is based on the principles of geometrical optics, thus, the corresponding requirements mentioned above have to be fulfilled. In particular, the particle radius R has to be considerably larger than the laser wavelength. (a) Global view on ray tracing scheme. (b) Absorptivity across spherical particle, [13]. Figure 5: Modeling of laser energy absorption on powder grain scale based on ray tracing approach, [13]. Ray tracing is a well-established approach for studying radiation problems in countless scientific disciplines. Several contributions have applied this approach in order to study the radiation transfer trough porous media and packed 11 beds [21, 164, 139, 140, 141, 4]. In one of the first ray-tracing approaches in the context of SLM/SLS, Wang et al. [153, 154] studied a powder mixture consisting of two species of spherical particles differing in particle size and material. In [154], the dimensions of the sintering/melting zone are estimated by determining the accumulated energy absorbed by each particle. However, no inter- or intra-particle mechanisms of heat transfer (such as heat conduction or convection) besides radiation has been considered. Based on a powder bed consisting of 50µm Tungsten Carbide (WC) particles and 20µm Cobalt (Co) particles and typical SLS system parameters, the ray tracing simulations as well as accompanying experiments came to the consistent result that > 96% of the absorbed energy is concentrated within a depth of 400µm from the powder bed surface. The resulting spatial distribution of the absorbed energy E, normalized by the total incident energy Etotal , is plotted over the powder bed depth coordinate in Figure 6. While the highest amount of primary laser beam radiation obviously arrives directly at the top surface of the powder bed, the maximum of the absorbed energy distribution occurs slightly beneath the top surface since at this location also secondary radiation contributions, stemming from multiple reflections in the powder bed, are prevalent. Figure 6: Absorped to incident energy for a mixture of 50µm Tungsten Carbide (WC) and 20µm Cobalt (Co) particles, [154]. In [13], the model of [153, 154] has been further refined by additionally considering the angular- and polarizationdependence of the laser absorption. Thereto, the absorptivity of the laser beam components in perpendicular (index S ) and parallel (index P) polarization states have been distinguished according to the Fresnel formulas (see [100]): cos θ − (n2 − sin2 θ)1/2 αS (θ) = 1 − cos θ + (n2 − sin2 θ)1/2 " #2 n2 cos θ − (n2 − sin2 θ)1/2 αP (θ) = 1 − n2 cos θ + (n2 − sin2 θ)1/2 " , #2 . (4) This description takes into account the model of laser beam radiation as an electromagnetic wave with speed c and direction of propagation k. This wave is assumed to be (linearly) polarized with the constant unit vector e representing the direction of the electric field normal to k. In (4), the complex refraction index n = nc + inα , describing the speed of electromagnetic wave propagation (real part nc ), absorption of the electromagnetic wave (imaginary part nα ) and the angle of incidence θ between k and the normal vector n onto the powder particle surface have been employed. The perpendicular (index S ) and parallel (index P) components of the polarized electromagnetic wave are typically determined by means of a projection of the polarization vector e into a reference plane spanned by the vectors k and n. In Figure 5(b), the resulting spatial absorptivity distribution across one isolated sphere of stainless steel irradiated by a horizontally polarized laser beam is depicted. Correspondingly, the distribution of energy absorption can vary 12 considerably across one individual particle. The high importance and practical relevance of these fluctuations shall be illustrated by the following time scale estimate. The time scales governing the homogenization of energy on a sphere with radius R due to thermal conduction can be estimated by τc = R2 /a, where a is the thermal diffusivity of the metal. The time scales governing the melting of a spherical particle can be approximated by τm = RHm /(α0 I), where Hm is the melting enthalpy per volume, I is the laser irradiance, and α0 is the flat-surface absorptivity. For parameter values typical for SLM processes, the melting time is often considerably shorter as compared to the diffusion time. Consequently, non-uniformity of energy absorption results in only partial melting of the particle, which might in turn considerably influence melt pool dynamics or creation mechanisms of pores [13]. However, in cases where melting takes place on larger time scales than heat conduction, the initial energy and temperature distributions might homogenize before melting occurs, and, consequently, approaches considering a constant energy distribution across a particle (see e.g. [154]) might yield reasonable results. The fact that thermal conductivity of loose powder is typically governed by the gas filling the powder bed pores [132] yields even larger time scales for the inter-particle heat conduction. (a) Laser beam path across powders with Gaussian (average size 27µm, top) and bimodal distribution (particle ratio 1:7,bottom.) (b) Variation of absorptivity along laser path for two different laser beam sizes with a radius of a = 8µm and a = 24µm. Figure 7: Dependence of absorptivity on laser beam size and particle size distribution investigated for a stainless steel alloy, [13]. According to [13], there are two general factors of influence leading to absorption non-uniformity, one related to the non-uniformity of absorption within a single particle as considered above, and the other related to the non-uniformity of the powder bed. In order to investigate this second factor of influence, ray tracing simulations on laser tracks across powder beds (thickness ≈ 50µm) resting on a solid substrate and consisting of two different powder mixtures, a Gaussian distribution (average size 27µm, Figure 7(a), top) and a bimodal distribution with maximal packing density (particle ratio 1:7, Figure 7(a), bottom), have been conducted in [13]. For both powder mixtures, typical distributions of the total absorbed energy along the scan track length as illustrated in Figure 7(b) can be observed. While large laser beam sizes (see curve a = 24µm in Figure 7(b)) lead to a comparatively homogeneous spatial energy absorption, laser beam sizes in the range of the particle diameter (see curve a = 8µm in Figure 7(b)), lead to strongly fluctuating spatial absorption values. Furthermore, the overall absorptivity turned out to be higher for the bimodal distribution, although the degree of fluctuations was comparable for both mixtures. Again, these fluctuations are not considered in homogenized continuum models such as [59]. Thus, their applicability requires sufficiently small particle sizes as compared to laser spot size and powder layer thickness. In [14], the results derived in [13] have been extended to a considerable range of different materials and verified experimentally. Concretely, the simulated power bed absorptivity and the flat surface absorptivity of the considered materials have been compared. Plotting the powder bed absorptivity over the flat surface absorptivity has revealed a smooth functional relation between these two quantities as already observed in [57, 62]. This interesting observation 13 allows to extract the assumed powder bed absorptivity for a material with given flat surface absorptivity, a procedure that can be considered to be of highest practical relevance (Figure 8). In [172], a comparable ray tracing simulation model has been proposed in the context of SLS. Figure 8: Dependence of powder layer absorptivity on flat surface absorptivity for different materials, [14]. 2.1.3. Continuum models for powder bed heat conduction The macroscopic simulation models discussed later in Section 2.2 regard the powder bed as a homogenized continuum. Consequently, these approaches typically rely on a model for an effective, homogenized powder bed conductivity. On the contrary, mesoscopic simulation models as considered in Section 2.3 resolve individual powder grains by spatial discretization. Consequently, each powder particle can be considered as a separate solid body, which simplifies the powder bed heat conduction problem to the two subproblems of intra-solid heat conduction as well as inter-solid heat conduction across the contact interfaces. Study of thermal transport in two phase systems has long been a topic of scientific inquiry, with Maxwell and Rayleigh each publishing models of the problem in 1873 (see [112]) and 1892 (see [128]), respectively. These early contributions have been based on comparatively strong model assumptions. For example, Rayleigh’s treatment idealizes the geometry to regularly spaced spherical particles in non-contact within the dispersed phase. An extension of the spherical particles to a larger spectrum of geometric primitives has been proposed in the contribution by Bruggeman [16]. In general, the aim of more recent work has been to further relax such assumptions and match the model to expected behavior in limit conditions, often in the context of interpreting experimental results [148]. For example, an often cited model has been published by Meredith and Toblas to explain the conductivity of water-propylenecarbonate emulsions. This approach is assumed to yield more realistic predictions than Maxwell’s or Bruggeman’s models at high volume fractions of the dispersed phase [117]. In the review by Tostsas and Martin [148], models are further divided by how they treat primary parameters of bulk conductivities and porosity, and secondary parameters including radiative transfer, pressure dependence, contact between particles and deformation as well as convective effects. The effective thermal conductivity of loose metallic powders, i.e. of material with negligible particle-to-particle contact surfaces, is typically controlled by the gas in the pores [132, 63]. On the contrary, in [60], a model has been derived for the heat conduction in power beds in case of small but finite contact areas between the particles. Such a scenario typically arises during the sintering process in SLS, at the beginning of the powder particle melting process in SLM or already in the non-melted solid state of the powder bed in case it is not loose but pre-compressed. Now, the effective thermal conductivity λe within a powder bed in direction of a given unit vector nλ can be defined 14 on the basis of the heat flux density q and the temperature gradient 5T according to: ||nλ ·q|| = λe ||nλ ·5T ||. (5) For illustration, the effective thermal conductivity shall briefly be presented for two well-known, but conceptionally different, heat transfer models for heterogeneous / particle-based systems, the Maxwell model mentioned above (index M) and the Reimann-Weber model (index RW). According to these models, the effective thermal conductivities are: λe,M = λg 1 − 2Φ 1 − λ s /λg 2 + λ s /λg !, 1+Φ ! 1 − λ s /λg , 2 + λ s /λg λe,RW = λ 1 1 2 + ln x π x !−1 with x = rc . R (6) The well-known Maxwell model describes the effective thermal conductivity of systems consisting of randomly packed spheres with conductivity λ s embedded in a continuous medium with conductivity λg and is applicable for sufficiently small volume fractions Φ of spheres. It can easily be verified that this model yields the expected results for limit conditions such as λg = 0, Φ = 0 or Φ = 1. On the contrary, the Reimann-Weber model considers the effective thermal conductivity between two isolated spheres (bulk material conductivity λ, radius R) exhibiting a circular contact area of radius rc . The packing densities prevalent in powder bed AM make the application of the Maxwell model to these processes virtually impossible. For that reason, Gusarov et al. [60] derived formulations for the thermal conductivity of regular and random packings of spherical particles based on models of the ”isolated-spheres” type comparable to the Reimann-Weber model. As such, the effective conductivity of ordered powder beds has been determined by summing up the conductivity contributions of all individual sphere-to-sphere pairs. In the case of random packings, the models of ”isolated-spheres” type have been homogenized by spatial integration and the geometrical quantities rc and R are replaced by effective quantities such as the particle volume fraction Φ, the mean coordination number n (describing the average number of closest neighbors) and the average dimensionless contact size ratio x leading to the following effective thermal conductivity (index G for Gusarov): λe,G = λ nΦ x. π (7) While a strong dependence on the contact size x, at least qualitatively, has already been known from experiments, experimental characterization approaches typically consider only the powder density Φ, but not the coordination number n. However, the derivations in [60] recommended for future experiments to treat these two parameters independently since for randomly packed powder beds different coordination numbers, and according to (7) considerably differing effective conductivities, are possible for the same powder bed density Φ. Besides modeling approaches for the determination of effective powder bed conductivities, often also experimentally determined effective conductivities are employed in macroscopic SLM models as discussed in Section 2.2. An extensive comparison of experimental and modeling approaches for effective powder bed conductivities can also be found in Yagi et al. [163]. 2.1.4. Conclusion In summary, ray tracing models are based on a comparatively simple theory, yield a high degree of detailedness in the obtained solutions, but also require considerable computational resources to be applied to scenarios of practically relevant size. On the other hand, heat transfer continuum models are based on a more complex theory, are in general not capable of accurately resolving details on the length scales of prevalent heterogeneities (i.e., on the length scale of individual powder particles in the case of SLM), but are considerably less expensive for computation. Some simple cases even allow for the derivation of analytical solutions, for example by Gusarov [59]. Importantly, ray tracing models require specification of the powder bed structure by particle shape, dimension, and their coordinates, which is a non-trivial problem itself and often requires additional assumptions. In principle, the distinction of these two model classes is comparable to the distinction of atomistic and continuum models in other physical disciplines such as electricity or mechanics. Unfortunately, in the characterization of SLM processes, one is often interested in length scales comparable to the powder layer thickness. Because powder bed heterogeneities in form of individual particles are on the same length scale, the application of continuum models can often only be considered as an estimate of the underlying radiation transfer processes. Moreover, the assumption of a laser beam penetrating into a powder bed by means of multiple reflection actually only applies for sintering or for SLM processes based on point-wise modulated 15 scanning strategies as long as no melt pool is prevalent. When SLM is performed with linear/continuous scanning, however, the resulting melt pool front typically proceeds the laser beam position and, thus, the laser beam impacts directly on the melt pool and not on an undisturbed powder bed. In order to accurately resolve the impact of the laser beam on the melt pool, whose actual surface contour is typically considerably distorted and part of the solution itself, ray tracing techniques seem to be indispensable. Similar conclusions can be drawn by comparing mesoscopic and homogenized macroscopic models for the powder bed heat conduction. Because the local energy and temperature distribution within the powder bed but also within individual powder grains crucially influence the melting behavior on the mesoscopic scale and the creation of defects such as pores or inclusions, also sufficiently resolved mesoscopic models for the laser beam radiation transfer (e.g. via ray tracing) and the inter- and intra-particle heat conduction (e.g. by resolving these particles in the spatial discretization of the thermal problem) are mandatory if questions on this scale are relevant. However, when applying macroscopic SLM models, i.e. continuum approaches suffering from a powder homogenization error anyways, heat transfer continuum models appear to be a reasonable choice. 2.2. Macroscopic simulation models Macroscopic simulation models in the context of SLM processes typically treat the powder phase as a homogenized continuum described by means of effective, i.e., spatially averaged, thermal and mechanical properties, without resolving individual powder grains. This homogenization procedure yields efficient numerical tools capable of simulating entire SLM builds of practically relevant size across practically relevant time scales. The thermal problem lies typically in the focus of interest when these models are applied. In some works, an additional solid-mechanics problem statement is considered aiming at the assessment of global residual stress distributions or dimensional warping effects based on the full thermo-mechanical interactions. The thermal problem is commonly given by the following set of balance equations, boundary and initial conditions: =:−qk z}|{ ∂(ρc p T ) + 5·(ρc p T v) = 5·(k·5T ) + u s ∂t T = T̄ ∀(x, t) ∈ Ω×]0, tend ], (8a) ∀(x, t) ∈ ΓT ×]0, tend ], (8b) q = q̄ ∀(x, t) ∈ Γq ×]0, tend ], (8c) ∂T s ∂T l − kl = Hm ρ n sl · ẋ ∂x ∂x ∀(x, t) ∈ Γm ×]0, tend ], (8d) ∀(x, t) ∈ Ω × 0. (8e) ! T = Tm and n sl · k s T = T0 Here, (8a) represents a variant of the energy equation as typically employed in the macroscopic and mesoscopic SLM models on the problem domain Ω within the considered time interval ]0, tend ], where ρ is the density, c p is the specific heat at constant pressure, T is the temperature and v represents the velocity field. In (8a), k is the thermal conductivity tensor and u s represents a heat source term. The two terms on the left-hand side of (8a) represent the material time derivative of the thermal energy density constituted of the local and the convective time derivative. It has to be mentioned that the thermodynamically consistent statement of thermo-mechanical problems involving compressible materials would require additional thermo-mechanical coupling terms in (8a). However, most of the models discussed in the following typically assume either exact or approximate incompressibility and resign these additional terms. Typically, isotropic conductivity is assumed, which yields k = λI. Often, the source term u s is modeled based on a solution I(x, Ω) of the RTE according to (3), e.g., the one proposed in [59]. Equation (8b) accounts for essential boundary conditions with prescribed temperature T̄ on the boundary ΓT whereas (8c) represents natural boundary conditions with prescribed heat fluxes q̄ on the boundary Γq typically accounting for thermal convection q̄c and radiation emission q̄e on the powder surface with normal n p : q̄c = c(T − T re f )n p , 4 q̄e = kS B (T 4 − T re f )n p . (9) Here, T re f denotes the temperature of the ambient gas atmosphere, c is the convection coefficient and  the emissivity. The Stefan-Neumann equation (8d) describes the phase change due to melting at the interface between the powder phase and the melt pool. In this equation, the indices s and l refer to the temperature gradients and thermal 16 conductivities in the solid and liquid phase, respectively. Moreover, T m is the melting temperature, Hm is the latent heat of melting and n sl is the normal vector of the solid-liquid interface defining the melt process as a pure interface phenomenon. Eventually, (8e) prescribes the initial temperature T 0 . If the solid mechanics problem is also solved as well, the following system has to be considered in addition to (8): ρv̇ = 5·σ + ρb  u+I − u−I xI =0 and ∀(x, t) ∈ Ω×]0, tend ], (10a) u(xu , t) = ū ∀(x, t) ∈ Γu ×]0, tend ], (10b) σ · n (x ,t) = t̄ ∀(x, t) ∈ Γt ×]0, tend ], t  σ+ − σ− · nI (x ,t) = 0 ∀(x, t) ∈ ΓI ×]0, tend ], (10c) I (10d) u(x, 0) = u0 ∀(x, t) ∈ Ω × 0, (10e) v(x, 0) = v0 ∀(x, t) ∈ Ω × 0. (10f) Equation (10a) represents the mechanical equilibrium of linear momentum. In this equation, the Cauchy stress tensor σ(u, T ) = σ(e (u),  p (u), T (T )) is related to the primary displacement field u and temperature field T by constitutive parameters with contributions from elastic e , plastic  p and thermal T strains. The mechanical equilibrium of angular momentum is fulfilled by definition due to the symmetry of the Chauchy stress tensor. The total time derivative v̇ represents the material acceleration vector and b is the vector of volume forces acting on the physical domain Ω. Similar to the thermal problem, essential and natural boundary conditions are given by equations (10b) and (10c) prescribing displacements ū and tractions t̄. Equation (10d) represents the requirement of displacement continuity and mechanical equilibrium at the relevant interfaces, e.g. the interface powder-melt or the interface melt-solid, characterized by a normal vector field nI . The superscripts (.)+ and (.)− denote quantities on the two different sides of the interface. However, most of the considered macroscopic models do typically not resolve these interfaces in the sense of a sharp 2D interface with discontinuous material parameters, but are rather based on a smooth, homogenized transition between the different phases, with the latter variant being easier to realize numerically. Finally, the initial position and velocity field is given by equations (10e) and (10f). The spatial discretization of the heat equation (8a) and the momentum equation (10a) is typically based on the finite element method (FEM), which requires transfer of these equations into the (equivalent) weak form. For time integration, explicit as well as implicit approaches can be found. The coupling between the thermal and the solid-mechanical problem is often realized in a staggered partitioned manner. In the following, some of the most important representatives of macroscopic models available in the literature will briefly be discussed. As long as macroscopic models are considered, melt pool dynamics are not explicitly resolved and porosity-dependent, effective material properties are employed. The principal strategies of deriving macroscopic models are comparable for SLS, SLM and EBM processes in many aspects. Thus, scientific contributions related to all of these processes will be considered. In recent years, considerable research effort in the development of macroscopic, mesoscopic and microscopic models has been conducted at the Lawrence Livermore National Laboratory (LLNL), including a macroscopic homogenized continuum model for SLM processes described by Hodge et al. [70]. In this model, the laser beam heat source term has been taken from [59]. Heat losses occurring for example due to thermal emission, evaporation or mass ejection are taken into account by reducing the nominal total laser power P to an effective total laser power, Pe = 2/3P. In order to describe the phase transition, two spatial phase function fields φ p (x) ∈ [0; 1] and φc (x) ∈ [0; 1] with φ p (x)+φc (x) = 1 ∀ x are introduced representing the powder and the consolidated phase., i.e. φ p (x) = 1, φc (x) = 0 for pure powder and φ p (x) = 0, φc (x) = 1 for pure consolidated material. Importantly, the consolidated phase considered in this model is assumed to have a vanishing porosity and captures both, the melting as well as the solidified phase. In other words, only the phase transition from powder to liquid during melting is explicitly considered by means of a phase function, while the phase boundary between melt pool and solidified material is accounted for by means of high gradients in the material properties at these spatial locations. All of the relevant thermal material parameters (e.g. conductivity, heat capacity etc.) are considered to be a function of both temperature and phase (thus, of porosity). In the spatial discretization process based on the finite element method (FEM) and the subsequent numerical implementation, the phase boundary between powder and melt pool is not resolved in a sharp manner. Instead, there 17 is typically a small band of finite elements where both phases are prevalent, a fact that is reflected by non-vanishing values 0 < φ p (x), φc (x) < 1. Correspondingly, the Stefan-Neumann equation is not evaluated in its original form (8d) at a 2D interface, but rather in an equivalent form within 3D volume elements. While the phase boundary between powder and melt pool is tracked by means of phase variables, the Stefan-Neumann equation (8d) (or a spatially average version of it) is not considered for the phase transition from melting to solid. Instead, a high value of the heat capacity is chosen in the range of the melting temperature in order to account for the released latent heat during the solidification process. Thus, the typical temperature plateau observable in enthalpytemperature diagrams of elementally pure materials at the melting point is replaced by an interval of very slowly increasing temperature. From a physical point of view, this might rather be comparable to the melting behavior of alloys between liquidus and solidus temperatures T l and T s . From a numerical point of view, it can be interpreted as a (commonly employed) regularization of the constraint equation T = T m in (8d) of elementally pure materials in order to simplify the numerical solution process. Hodge et al. [70] considered the thermo-mechanical problem by coupling the thermal and mechanical simulations in an iteratively partitioned manner. There, an elasto-plastic material law, additionally accounting for thermal expansion and consolidation shrinkage, has been employed while the boundaries between powder, melt and solidified phase have been considered implicitly via strong gradients of the associated mechanical material parameters with respect to temperature and porosity. This means, the interfaces given in equation (10d) are not explicitly resolved in the mechanical simulation. Furthermore, this elasto-plastic material model has been applied to all three phases, i.e. to the powder phase, to the solidified phase, and also to the liquid phase. Thus, a detailed modeling of the melt pool fluid dynamics, as prevalent in the mesoscopic models discussed in Section 2.3, is missing. This statement applies in a similar manner to most of the macroscopic SLM models considered in this section. (a) [63]: low velocity. (b) [70]: low velocity. (c) [63]: high velocity. (d) [70]: high velocity. Figure 9: Comparison between melt pool geometries derived via macroscopic SLM models in [63] and [70] for different scan velocities. The thermal simulations in [70] confirmed the results of Gusarov et al. [63] for 316L stainless steel concerning maximal temperature, top surface shape as well as cross-section shape of the melt pool. It was observed that the melt pool shape gets narrower and longer with increasing scan speed and that at some point a concave region at the tail of the melt pool boundary occurs, an effect that can be explained by the higher thermal conductivity of the solidified material behind the melt pool as compared to the powder material to the sides (Figure 9). Nevertheless, when considering the similarities between the two models proposed in [63] and in [70], it has to be emphasized that both approaches are based on the same radiation transfer model for the incident laser energy, and, that the chosen submodel for the heat source might considerably influence the overall results since the thermal conductivity within the powder phase is strongly limited by the low conductivity of the atmospheric gas filling the powder pores. Because the model in [70] explicitly considered the powder-melting phase change via φ p (x) and φc (x), the inertia of 18 (a) Step 1, driving to the left. (b) Step 2, driving to the left. (c) Step 3, driving to the right. (d) Step 4, driving to the right. Figure 10: Laser beam scanning across an overhang. First, the beam moves to the left (steps 1+2), than back to the right (steps 3+4). Blue color represents loose powder, red color represents consolidated material (melt and solid). The melt pool shape is indicated by black contour lines, [70]. the melting process could be visualized by means of a melt pool boundary lagging behind the isothermal of the actual melting temperature in the range of high scan velocities. Furthermore, [70] investigated the melt pool behavior when scanning across overhangs supported by loose (poorly conductive) powder (Figure 10). Experimental results [29, 151] could be confirmed that typically larger melt pool sizes, excessive overheating and evaporation occur when applying the same scan parameters in domains of low thermal conductivity such as overhangs or thin-walled features. Commonly, this problem is addressed by adapting the laser parameters and/or increasing the net thermal conductivity in this regions by means of additional support columns. (a) Experimental measurements. (b) Simulation: no plastic hardening. (c) Simulation: with plastic hardening. Figure 11: Comparison of predicted and actual displacements in x−direction arising from dimensional warping of SLM triangular prisms, [69]. Further validation of the Hodge simulation via experiments is given in the follow-up work [69], which also investigates the effects of build orientation, infill scan strategy, and preheating on residual stresses and dimensional warping. Digital Image Correlation (DIC) was used to measure the displacements of 316L triangular prisms fabricated by SLM as test artifacts. Moreover, neutron diffraction experiments were used to measure the stress distribution. Figure 11 illustrates typical results for a triangular prism printed vertically, with the shortest edge of the triangle forming one edge of the rectangular contact area with the build platform. The prisms are 5mm along the bottom edge, 6mm tall, and 1mm thick, in part to enable comparison to a previous study in [158]. In order to measure dimensional warping, the prism is removed from the build platform leading to a considerable distortion of the part geometry due to the previously present residual stresses. In Figure 11(a), the experimentally measured displacements in x−direction, taking on maximal values in the range of 0.01mm, are illustrated. For comparison, in Figures 11(b) and 11(c) the 19 corresponding simulation results either based on a material law with or without plastic hardening, are illustrated. Besides a qualitative agreement of numerical and experimental results, the relative error of the simulation without plastic hardening lies in the range of 20% in the maximal displacement while the variant with plastic hardening overestimates the maximal displacement approximately by a factor of two. For the displacements in y−direction (not illustrated), the behavior turned out to be (almost) the opposite, i.e. rough agreement of the simulation based on a material law with plastic hardening, large deviations for the variant without plastic hardening. This behavior might indicate that the comparatively simple isotropic constitutive laws applied in [69], and in many other state-of-the-art macroscopic SLM models, are still not accurate enough. Additionally, the simulation results in Figure 12 show the x−component of the Cauchy stresses resulting from the two chosen preheating temperatures of 303K and 1000K. Again, qualitative results can be captured quite well confirming that preheating reduces the peak values of residual stresses. Figure 12: Simulated Cauchy stress in x−direction for the cases (a) non-heated and (b) preheated built platform, [69]. A slightly different macroscopic SLM model has been proposed in [63], where the radiation transfer model derived in the authors’ earlier work [59] has been supplemented by a thermal simulation framework based on equations (8). In contrast to [70], no additional phase function field is introduced in order to explicitly solve (8d) for the interface between powder and melt pool. Thus, this interface is only given implicitly by the isothermal line T = T m . Because the effective thermal conductivity k of loose metallic powders at low temperatures is controlled by gas in the pores [132], the effective thermal conductivity of the 10 - 50µm powders considered in [63] has been chosen to k = k p I employing the ”gas conductivity” k p = 0.3W/(mK) at temperatures below the melting point. Surface contacts between particles are formed when powder melts, so that the thermal conductivity of the powder rapidly approaches the value of the solid material, chosen as k = kd I with kd = 20W/(mK) for the investigated stainless steel type 316 L. In the model of [63], a sharp change from k p to kd at the melting point is assumed. Based on the simulated temperature fields and the derived isothermals T = T m , the melt pool geometry has been analyzed for different laser beam scan velocities (see Figure 9(a) and 9(c) as well as the corresponding discussion in the paragraph above). In [61], the pure thermal simulation framework of [63] has been extended to thermo-mechanics incorporating a simple linear-elastic, plain strain version of (10) with elastic constitutive parameters jumping at the phase boundaries. In [129], the pure thermal problem (8) has been solved for the EBM process based on the finite element method for spatial discretization and an implicit Runge-Kutta method for temporal discretization. Additionally, adaptive mesh refinement strategies have been applied in the direct vicinity of the electron beam. In this model, the penetration of the electron beam into the material is described by means of the model [88] and thermal losses due to emission have been considered. The different phases of the investigated Ti-6Al-4V alloy are considered in the thermal model via phase- and temperature-dependence of the thermal material parameters. The heat capacity c p is prescribed as a non20 linear function of temperature with high gradients around the melting point in order to account for the latent heat of melting (Figure 13(a)). The temperature-dependence of the heat conductivity k has been described by a linear law that differs for the three different phases powder, melt and solidified material (Figure 13(b)). In [130], the same authors considered the coupled thermo-mechanical problem (8) and (10), wherein the energy equation (8a) has been derived in a thermodynamically consistent manner from a free energy density functional leading to the well-known thermomechanical coupling term of compressible materials not only in the momentum equation (10) but also in the energy equation (8a). The focus of this work lay in comparing the overall computational efficiency of two different numerical coupling schemes, a monolithic coupling scheme solving the thermal and structural mechanics problem simultaneously as well as a staggered partitioned scheme proposed in [6]. The staggered partitioned scheme allowed for lower computational costs per time step due to the smaller system sizes resulting from a separate treatment of the thermal and structural mechanics problem. However, for stability reasons, the partitioned scheme required a considerable smaller time step size resulting in a higher overall computational effort as compared to the monolithic scheme. Consequently, the authors recommend the latter category of coupling schemes for future application to problem classes such as SLM. (a) Temperature-dependence of specific heat c p . (b) Temperature-dependence of thermal conductivity k. Figure 13: Dependence of thermal material parameters c p and k of Ti-6Al-4V on current temperature and considered phase, [129]. Much work by the group of Stucker et al. [121] has focused on the development of FEM-based simulation tools for SLM processes, and has been commercialized as the software package 3DSIM. In [122, 120, 169], the thermal problem (8) is solved, modeling the laser beam as surface heat flux with Gaussian distribution and accounting for heat losses on the powder bed surface due to thermal convection. In this work, an implicit Crank Nicolson scheme for time integration and a spatial finite element discretization combined with a dynamic adaptive mesh refinement and derefinement algorithm have been employed. While the physical model is comparatively simple and only treats the pure thermal problem, the dynamic mesh adaption algorithm might allow for considerable efficiency gains as compared to uniform meshing strategies. In [170], further steps in improving the aforementioned simulation models are conducted by developing homogenization strategies based on representative volume elements (RVE) for specific part geometries with lattice-like support structures. Among others, these tools have been applied for predicting dimensional warping and suggesting strategies of geometrical compensation. For the homogenized, macroscopic continuum models as considered in this section, it can generally be expected that high spatial gradients in the thermo-mechanical solution of SLM problems will predominantly occur in the direct vicinity of the laser beam and at spatial locations where also the individual part geometry is characterized by very fine length scales (e.g. transition from high to very low wall thicknesses etc.). Similarly, high temporal gradients are primarily expected in the direct neighborhood of the laser beam position. Consequently, thermo-mechanical simulation approaches based on time- and space-dependent adaption schemes for the spatial (see e.g. [169] as discussed above) and for the temporal discretization are very promising to save computational costs. 21 A further commercial FEM software for SLM simulation is offered by Pan Computing, and was acquired by Autodesk in 2016. For example, the work of [38] extended the contributions discussed in this section so far by allowing for residual stress relaxation. Specifically, the stress and plastic strain of the underlying thermo-elasto-plastic constitutive model were reset to zero, when the temperature exceeded a prescribed annealing temperature. Based on experimental verifications and thermo-mechanical simulations of Ti-6Al-4V parts, it was claimed that temperatureinduced stress-relaxation plays a crucial role in EBM processes and that the error in maximal residual stresses and geometrical distortion could reach a value up to 500% as compared to their model without annealing effects. In [39], the computational efficiency of the model [38] has been increased by employing an inactive element activation strategy as well as a dynamic mesh coarsening algorithm. In the context of FEM-based SLM process simulations, typically two different strategies of material deposition are distinguished: The notion of quiet elements refers to a scheme were all elements representing the final part are already present in the beginning of the SLM process, whereat elements in power layers that have not been deposited yet are characterized by artificial material properties (e.g. very low thermal conductivity k, specific heat c p , Young’s modulus E etc.) that are intended to rarely influence the results within the already deposited layers. On the other hand, the notion of inactive elements refers to a scheme that adds new finite elements with every newly deposited layer while elements associated with subsequent layers are not yet present. Of course, quiet element schemes result in a higher computational effort since a large system of equations (containing all degrees of freedom of the final part) has to be solved in every time step and the resulting system matrices might be ill-conditioned as consequence of the artificial material parameters. On the other hand, inactive element schemes are often not supported by commercial FEM codes. There is an abundant number of further scientific contributions solving the thermal or thermo-mechanical problem prevalent in SLM processes based on macroscopic models and FEM discretization schemes in a manner similar to the examples already given above, with [70] being an exception that explicitly tracks the interface between powder and melting phase. While the SLM models discussed so far typically assume temperature- and porosity-dependent thermal properties of the powder based on the initial powder bed porosity, in [20] the energy equation (8) as well as an evolution equation for the time- and temperature dependent powder bed porosity field during the sintering process based on a model according to [134, 107] have been solved simultaneously. These two equations have been coupled by means of a model for porosity-dependent powder bed conductivity proposed in [163]. In contrast to this SLS model, an explicit solution of the powder sintering state and porosity in order to accurately determine the current powder conductivity is rather untypical for approaches considering SLM. This may be explained by the fact that the time interval separating the states of loose powder and fully molten liquid is rather short and the modeling error by assuming a simpler ad-hoc relation between powder temperature, porosity and conductivity seems not to be essential compared to the basic assumptions underlying typical macroscopic SLM models. In the works of Dai et al. [32, 33, 34, 35], a 3D model for SLS / SLM has been proposed to solve the thermo-mechanical problem (8) and (10) defined by a small number of subsequent powder layers with a special emphasis on multi-material parts. The thermal problem accounts for heat losses due to radiation and convection and employs a Gaussian laser energy distribution. The structural mechanics problem is defined via a constitutive law accounting for elastic, plastic and thermal strains. The effect of volume shrinkage during melting of loose powder is considered by means of a rather heuristic model that moves finite element nodes in direction of gravity by a distance defined via the initial powder porosity and mass conservation. In the work [110], a very simple 2D thermo-mechanical model considering one powder layer has been proposed. The mechanical model assumes a linear-elastic state of plane stress and accounts for elastic and thermal strains. The thermal and structural-mechanics problems are solved in a partitioned manner based on the FEM. The contribution [27] proposes a finite element model for the thermal problem of SLS / SLM processes. There, two different empiric laws have been proposed for the temperature-dependence of powder porosity during the sintering process. Moreover, a porosity-dependent law for the powder conductivity as observed experimentally in [138] has been considered. The laser beam has been modeled as volumetric heat source with an intensity decreasing linearly with the penetration depth. In [131], a computational model for determining the temperature history in a practical SLM process involving the scanning of (up to five) multiple layers has been proposed. The model builds on the previously developed concepts in [20] and [90] and extends these approaches to the FEM-based simulation of multiple layers. A comparatively rough yet efficient model is proposed in [167] to estimate residual stresses and dimensional warping effects occurring during the cooling phase of SLM processes. The principal idea is to combine several layers of the SLM process and to apply an equivalent thermal load to this integrated, virtual layer within the heating cycle and to determine residual stresses 22 and dimensional warping effects in a subsequent cooling cycle of the staggered thermo-mechanical scheme. More recent thermal and thermo-mechanical FEM-based SLM models similar to the ones discussed above are e.g. given by [136, 137, 71, 26, 157]. In [136, 137], an increased thermal conductivity of the melt pool has been considered to account for the convective melt pool heat transfer, however, without modeling hydrodynamics. In contrast to the finite element models considered so far, in [149] the finite volume method (FVM) has been employed for spatial discretization in combination with an explicit Euler time integration scheme. The underlying model considers the pure thermal problem (8), including the phases of powder, melt, solid and evaporated gas, in order to determine melt pool shapes. Furthermore, the radiation transfer model originally proposed by [59] has been employed. Based on the resulting temperature field, the volume fractions of the phases powder, melt, solid and gas have been determined for each computational cell defined by the FVM. Based on comparisons with experimental data, it was found that the consideration of evaporative heat loss is essential for realistic temperature field results. An interesting approach that crucially differs from the ones discussed so far has been proposed in [74, 75]. This model somehow combines the procedures typically applied to macroscopic and mesoscopic SLM models. The powder phase is modeled as homogenized continuum with effective, porosity-dependent properties as typical for macroscopic models, whereas the fluid dynamics within the melt pool are explicitly modeled as it is common for mesoscopic SLM models. The melt pool fluid flow is modeled as a compressible, laminar Newtonian liquid flow governed by the NavierStokes equations (11) accounting e.g. for temperature-dependent surface tension, temperature-dependent buoyancy forces as well as frictional dissipation in the mushy zone typical for alloys such as the considered Ti-6Al-4V material at the powder-fluid interface. The thermal problem is described on the basis of (8) with a laser beam energy source as proposed in [168] as well as heat losses due to thermal radiation. After solving the thermo-hydrodynamics problem based on a commercial, FVM-based computational fluid dynamics (CFD) solver, the resulting temperature fields are used as input data for a subsequent structural mechanics simulation in the sense of a partitioned field coupling. By employing a commercial FEM code, the structural mechanics problem (10) in combination with a thermo-elastoplastic material law is solved in order to determine residual stress distributions in the solidified phase. All thermal and mechanical properties relevant for the model have been assumed to be temperature- and porosity-dependent. While the employed laser beam model [168] is only valid for EBM processes, the remaining model constituents seem to be directly transferable to related processes such as SLM. Although the applied field and domain coupling schemes appear to be comparatively rudimentary, the general idea of employing efficient, homogenized macroscale powder models while simultaneously considering the convection-driven heat transfer in the melt pool on the basis of CFD simulations can be regarded as very appealing. Some basic physical phenomena typically resolved by the mesoscopic models discussed in the next section but not accessible by the models considered above, e.g. increased melt pool widths and decreased melt pool peak temperatures as consequence of an outward-directed Marangoni flow, could already be captured by this model, although at a considerably lower resolution - and numerical cost - as compared to the mesoscopic models. A considerable number of similar approaches that might also be relevant for SLM can e.g. be found in the fields of laser and electron beam welding (see e.g. [9, 23, 45, 47, 86, 85, 103, 127, 135]). Future SLM simulation models could strongly benefit from these works. 2.3. Mesoscopic simulation models While macroscopic models do not consider physical phenomena on the length scale of individual powder particles, mesoscopic models explicitly resolve these scales in order to study melt pool dynamics, melt pool heat transport as well as the wetting of melt on substrate and powder particles. Typically, these models aim at the prediction of part properties such as layer-to-layer adhesion, surface quality and defects on the mesoscale (pores, inclusions etc.). In these models, the initial powder grain distribution is either determined on the basis of contact mechanics simulations, e.g. by employing the discrete element method (DEM, see e.g. [101]), or by means of more generic packing algorithms such as rain models [91, 114]. Subsequently, thermo-mechanical simulations are performed considering the heat transfer within the powder bed, the melting process as well as the heat transfer and the hydrodynamics in the melt pool. Often, the unmelted powder grains are assumed to be spatially fixed and the powder phase is only solved for the pure thermal problem based on a proper laser beam model (see Section 2.1.1 and 2.1.2). The melt pool fluid dynamics are commonly modeled by means of energy equation (8) considering the latent heat of melting as well as the following 23 set of balance equations (11), accounting for conservation of mass and momentum of an incompressible flow: 5·v = 0 ∀(x, t) ∈ Ω×]0, tend ], (11a) ∀(x, t) ∈ Ω×]0, tend ], (11b) ∀(x, t) ∈ Γv ×]0, tend ], (11c) σ·n = t̄ ∀(x, t) ∈ Γt ×]0, tend ], (11d) ∂(ρv) + 5·(ρv ⊗ v) = 5p + 5·(2µ˜ ) + ρb ∂t v = v̄  u+I − u−I = 0 and  dγ σ+ − σ− ·nI = [γκI nI + (I − n ⊗ n) 5T ] ∀(x, t) ∈ ΓI ×]0, tend ], dT v = v0 ∀(x, t) ∈ Ω × 0. (11e) (11f) In the above system of equations, (11a) is the continuity equation requiring conservation of mass, (11b) is the NavierStokes momentum equation, (11c) and (11d) represent essential and natural boundary conditions on the boundaries Γv and Γt , while (11f) prescribes the initial velocity field at t = 0. The flow is typically assumed to be laminar and incompressible and the fluid to be Newtonian with a stress tensor σ(v, p) = −pI + 2µ˜ (v). Here, ˜ (v) represents the rate-of-deformation tensor, p is the pressure and µ is the kinematic viscosity. The body force term b typically accounts for gravity. This term might contain additional contributions, e.g. related to the phase change from metal powder to liquid or from temperature-induced buoyancy forces, as discussed below. The equation (11e) accounts for surface tension effects on the free-surface ΓI between melt pool and ambient gas. While the velocity field is assumed to be continuous (left side of (11e)) the surface stress exhibits a jump at the interface (right side of (11e)). The jump in interface-normal direction (first term in square brackets) is already present for the case of spatially constant surface tension γ leading to fluid surfaces with curvature κI and capillary effects already for hydrostatic problems. The jump in the interface-tangential direction (second term in square brackets) only occurs for spatially varying surface tension values, here solely considered due to spatial temperature gradients, and will for Newtonian fluids (with shear stresses being proportional to velocity gradients) always induce flow. The Marangoni effect resulting from such surface tension variations crucially influences the convective heat transfer within the melt pool. All in all, surface tension significantly determines the melt pool shape and the resulting solidified surfaces (see [83, 124, 101]). For the numerical solution of the mathematical problem defined by (8) and (11), often a staggered solution scheme relying on a sequential solution of the thermal problem in the powder bed and the fluid dynamics problem in the melt pool is applied. In such a partitioned solution procedure, the powder particle zones with T > T m are typically identified as mass increments per time step that are transferred from the solid to the fluid phase due to melting. Consequently, the contour lines T = T m define the boundary conditions for the subsequent fluid dynamics solution step. Many of the existing approaches consider a spatial discretization via the finite volume method (FVM) in combination with the volume of fluid (VOF) method in order to track the fluid-gas interface. Temporal discretization of the fluid dynamics problem is predominantly based on explicit time integration schemes. The required resolution of the spatial discretization and the small time step sizes admissible for explicit time stepping schemes lead typically to high computational costs, allowing for an application of these 3D mesoscopic models only to single-/dual-track simulations so far. Arguably one of the best-known mesoscopic models of this kind has been presented by Khairallah et al. [81]. There, the continuum radiation transfer model [59] as discussed in Section 2.1.1 has been employed as a laser beam source term, while individual particles are resolved on the mesoscale in order to explicitly consider the conductive heat transfer within the powder bed limited by the atmospheric gas in the pores and the point contacts between the powder grains (see [132]). For spatial discretization, the finite element method based on a uniform Cartesian mesh with element size 3µm has been applied. The numerical solution process is based on a staggered solution of the thermal and hydrodynamics problem. While the applied explicit time stepping scheme allowed for a robust treatment of this complex simulation problem in case proper stability requirements are fulfilled, the latter restriction limited the achievable time step sizes and consequently the observable spatial and temporal problem dimensions considerably leading to a computational effort in the range of 100.000 cpu h in order to simulate a single laser track with length dimensions in the range of 1mm and time spans in the range of several 100µs [81]. The model of [81] includes surface tension effects prevalent in the melt pool flow as well as fluid viscosity and gravity effects, although the latter effect turned out 24 to be negligible given the short time scales and the dominating surface tension effects at the considered length scales. However, the Marangoni effect, i.e., temperature gradient-induced melt flow due to temperature-dependent surface tension characteristics was not considered. (a) Simulation with surface tension effects. (b) Simulation without surface tension effects. (c) Balling effect attributed to Plateau-Rayleigh instabilities of a liquid cylinder in the range of high scanning velocities. Figure 14: Single-track mesoscopic simulation of thermo-hydrodynamics prevalent in SLM process of stainless steel 316 powder, [81]. In order to validate the thermal building block of the simulation framework, effective powder conductivities have been calculated with the developed model and compared to analytical solutions and experimental observations. In agreement to [132], the thermal conductivity of stainless steel 316 powder with a typical packing density of 45% turned out to be only by a factor of three higher than the conductivity of air filling the powder pores. Coupled thermohydrodynamics simulations revealed the importance of considering surface tension, which results in a smoother, less granular melt pool geometry and eventually leads to higher effective thermal conductivities within the pool but also from the pool to the underlying substrate. This effect, in turn, led to increased heat transfer, faster cooling rates and consequently smaller melt pool sizes (Figure 14). The rapid thermal expansion directly below the laser beam turned out to induce high melt flow velocities, especially in backward direction. Also the well-known balling effect [102, 98, 63], attributed to the Plateau-Rayleigh instability of a long cylindrical fluid jet breaking up into individual droplets, has been observed for high scanning velocities (see Figure 14(c) and [62] for a theoretical analysis). In [83], the model proposed in [81] has been extended to also (implicitly) account for the effects of recoil pressure in the evaporation zone below the laser beam on the subjacent melt pool flow, Marangoni convection, as well as evaporative and radiative surface cooling (see also [82] for more details of the model). Moreover, a more detailed resolution of the laser energy absorption has been achieved based on a ray-tracing model as discussed in Section 2.1.2. The surface tension has been assumed to decrease linearly with temperature, which accounts for Marangoni effects driving the melt flow from the hot laser spot toward the cold rear. As already observed in [63], the surface temperatures below the laser spot can easily reach boiling values. The resulting vapor recoil pressure adds extra forces to the melt pool surface that create a depression below the laser (Figure 15, bottom right). These observations motivated the differentiation of three different melt pool regions, namely the depression region, which is governed by recoil forces, the tail end region, which is governed by surface tension forces as well as the transition region (see Figure 16). The approach of [83] does not explicitly resolve vapor flow but employs a traction boundary condition (11d) based on 25 a recoil pressure model proposed in [2]. Accordingly, the recoil pressure p(T ) depends exponentially on temperature: p(T ) = 0.54pre f exp (λev /kS B (1/T − 1/T b )), (12) In this model, pre f is the ambient pressure, λev is the evaporation energy per particle, kB is the Stefan-Boltzmann constant, T is the surface temperature of the melting and T b is the boiling temperature. In order to more accurately model heat losses at the melt pool surface, the evaporation heat flux q̄v (again based on [2]) as well as the heat flux due to thermal radiation emission (see q̄e in (9)) has been considered according to q (13) q̄v = (0.82A s p(T )/ 2πMRg T )n. Here, A s represents a so-called sticking coefficient, p(T ) is the recoil pressure according to (12), M is the molar mass, Rg is the gas constant and T again the surface temperature. However, the mass loss due to evaporation has not been considered in the continuity equation (11a) underlying this model. The effect of considering / neglecting these different model components is illustrated in Figure 15. As compared to a continuum radiation transfer model, the employed ray-tracing model can resolve non-uniformities in the radiation absorption across powder particles leading to more narrow, neck-shaped connections to the underlying substrate and in turn to higher heat accumulations within particles as consequence of the reduced thermal conductivity. The ray-tracing model also captures scenarios such as particles that are only partially melted and might contribute to surface and pore defects more accurately. The principal surface tension effect fostering strongly curved melt pool surfaces with minimized surface energy is observable in Figure 15b), where additional melt flow is generated by temperature-induced buoyancy forces. The consideration of temperaturedependent surface tension allows for Marangoni convection, which induces a melt flow from the hot laser spot toward the cold rear (see Figure 15c)). This effect, in turn, increases the melt pool depth, and contributes to melt flow recirculation and melt spattering when colder liquid metal with lower viscosity is ejected from the pool. The additional consideration of recoil pressure due to evaporation leads to a considerably deeper melt pool, increased convective heat transfer within the melt pool and, in combination with the additional heat sinks due to evaporative and radiative surface cooling, to an increased overall cooling rate. Consequently, this case exhibits the lowest amount of stored heat. Figure 15: Influence of considering the individual modeling components ”laser ray tracing”, ”constant surface tension”, ”temperature-dependent surface tension” and ”recoil pressure” (from top left to bottom right) on resulting melt pool dynamics. Visualized is the temperature field in the range from 293K (blue) to 4000K (red) as well as the fluid velocity vector field. The red contour line represents the melt pool front, [83]. In [83], it has been stated that the overall melt pool dynamics in the direct vicinity of the laser beam is strongly influenced by the recoil pressure, whose magnitude increases exponentially with temperature, leading to the aforementioned depression directly below the laser. The depth of depression due to the recoil pressure is considered to 26 increase with increasing laser power, an effect that is closely related to the so-called keyhole mechanism observed in laser and electron beam welding (see e.g. [127]). On the other hand, the fluid dynamics of the cooler back end of the melt pool are governed by surface tension fostering e.g. balling of the tail during cooling [81]. Via this analysis, [83] enabled also theoretical considerations concerning the creation mechanism of so-called denudation zones. These ”low powder ditches” alongside the melt pool result from lateral particles that are dragged into the melt pool by surface tension and are responsible for the so-called elevated ”edge effect” [30] typically observed for the first laser track of a powder layer. Accordingly, the effects of denudation are stronger in the first track of a layer since powder particles can be attracted from two lateral sides of the melt pool. In [111], this simulation model has been combined with experimental approaches in order to further investigate the physical phenomena and mechanisms responsible for the occurrence of the denudation zone. Accordingly, the ambient gas flow induced by vapor rising above the melt pool may eject surrounding particles from the powder layer and contribute to denudation. Figure 16: Fluid velocity magnitudes and vectors in three different melt pool regions: depression region, tail end region and transition region. The laser beam has been scanning from the left to the right and is currently located in top of the depression region, [83]. In an alternative model proposed in [101], the discrete element method (DEM) is applied in order to study the spreading (i.e., recoating) of Inconel 718 powder accounting for frictional particle-to-particle, particle-to-wall and particleto-coater contact interaction as well as gravity as driving forces. Subsequently, the thermo-hydrodynamics problem of melt pool (free surface) flow is solved by means of a commercial computational fluid dynamics (CFD) solver based on the finite difference method (FDM), which resolves the fluid-gas phase boundary between melt pool and surrounding atmosphere on the basis of the volume of fluid (VOF) method. Also in this model, temperature-dependent surface tension effects, but no evaporation-related phenomena, have been considered. The laser beam energy absorption within the powder bed is modeled in a simplified manner based on prescribed heat flux boundary conditions on the top surface of the powder layer following the Gaussian distribution of the laser beam. Similar to [83], a depression of the melt pool directly underneath the laser beam could be observed. Since no evaporation-induced recoil pressure has been considered in [101], this observation has essentially been attributed to backward flow of molten metal due to Marangoni convection (comparable to Figure 15c) of [83]). Also the balling effect has been observed in these numerical simulations and attributed to Plateau-Rayleigh instabilities occurring in dependence of the melt pool dimensions and the powder particle arrangement. Accordingly, a higher packing density was found to reduce the likeliness of the balling effect and to increase the overall surface smoothness while a faster travel speed and lower laser power have been identified as possible sources fostering this effect. On the one hand, [101] claimed to obtain these simulation results on the basis of a comparable spatial resolution at considerably lower computational effort as compared to alternative approaches such as [81, 92]. On the other hand, this reference does not discuss the employed time integrator and the chosen time step size, which have crucial influence on computational performance. In the review article [115], Megahed et al. discuss several mesoscopic simulation results taken from previous contributions of the authors. The model of Megahed et al. is comparable to the mesoscopic models discussed in this section 27 so far. Among others, it has been shown that lower energy densities yielded smoother melt pool surfaces within the considered range of processing parameters. This effect has been observed in a similar manner in reference [83], where it has been argued that the melt pool associated with the high energy density reaches higher peak temperature values and, in turn, the surface tension can act for longer time spans leading to a more pronounced balling effect. On the contrary, [101] has reported about an increasing balling effect with decreasing energy density. The authors of [101] have argued that the total absorbed energy density decreases when decreasing the laser power at constant velocity, similar, as the total absorbed energy density decreases when increasing the laser velocity at constant laser power, with the latter scenario being known to foster the balling effect and resulting surface roughness. However, the difference between these two cases is that the latter scenario results in a longer, more slender melt pool shape (since less time is available for the melt pool to cool down while the laser beam moves by certain distance along the track), which fosters hydrodynamic instabilities according to the Plateau-Rayleigh theory. All in all, the observations made in [101] and the observations made in [115] and [83] might not necessarily be contradictory. Instead, these observations could be associated with an upper and a lower bound of an energy density interval that allows for stable processing [162, 62], which is consistent with experimental findings. In contrast to the continuous laser scan mode as considered so far, some SLM machines use point-wise modulation melting strategies. Megahed et al. have also investigated the melt pool characteristics resulting from such a modulated laser (see Figure 17 for the resulting melt pool shapes at different points in time). It can be seen that the laser modulation and the corresponding exposure time leads to singular deep melt pools. As the melt pools grow in diameter during the exposure time, they join into one continuous melt track on the upper surface of the powder bed down to a depth as typical for a continuous laser beam. However, as argued in [115], there are deeper melt pool peaks that might imply additional anchorage for the new layer to the previous ones. In addition to the discussion in [115], these modulated strategies might also allow for higher energy absorption as the laser beam can penetrate deeply into the pores of the powder layer via multiple reflections instead of directly interacting with the melt pool surface as it is typically the case in continuous scan mode. Potential benefits could arise in terms of lower melt pool surface temperatures and a decreased evaporative mass loss and recoil pressure. Figure 17: Melt pool shapes for a modulated laser: Temperature at three different points in time ranging from 500K (blue) to 2000K (yellow), [115]. In [124], the melt flow fluid dynamics have been solved by means of the finite volume method (FVM), employing the volume of fluid method (VOF) in order to resolve the interface of the free-surface flow. Besides recoil pressure, temperature-dependent surface tension and temperature-induced buoyancy forces as already considered in [83], [124] additionally supplemented the Navier-Stokes momentum equations (11b) by drag force contributions due to solid/liquid transition in the mushy zone of the melting Ti-6Al-4V alloy based on Darcys term for porous media. In the energy equation (8), the additional heat losses due to evaporation, convection as well as radiation emission have been considered. As laser beam heat source, the model proposed in [159] in the context of laser beam welding has been applied. Since the radiation transfer mechanisms prevalent in SLM might strongly depend on the mode of operation, the application of this, but also of other ad-hoc continuum radiation transfer laser models (see Section 2.1.1), has to be questioned critically. Also in [124], a strong influence of the recoil pressure on the melt pool dynamic in the direct vicinity of the laser beam as well as thereby induced material spattering could be observed, even though significantly decreased laser track sizes as well as structured instead of random powder packings have been considered in this reference. Similar mesoscopic modeling approaches are for example given in [55, 165, 166]. 28 Figure 18: Simulation results based on a 2D Lattice-Boltzmann model with discretization cells of size 5µm. Illustrated are initial powder grains and melt pool shapes resulting from two different prescribed wetting angles of θ0 = 10◦ (=good wetting, left) and θ0 = 160◦ (worse wetting, left), [91]. While the mesoscopic models considered so far have been discretized by means of either the FEM, FDM or the FVM, an entirely different approach has been proposed in [91]. Based on the contributions [7] and [8], the Lattice-Boltzmann method in combination with an explicit time integration scheme has been applied in order to discretize a 2D mesoscopic free-surface, incompressible melt flow model incorporating gravity, surface tension and wetting effects as well as the melting and solidification process prevalent in SLM. The initial powder bed generation is based on a random packing algorithm employing a rain model scheme [114]. While the general simulation framework might be applicable to SLM as well as EBM processes, the actual simulations and experiments conducted in this reference focused on the EBM process. Consequently, the energy absorption has been considered to predominantly take place at the powder particle surfaces of first incidence. This fact is reflected by the chosen exponential Lambert-Beer law of light attenuation in dense matter, which is not applicable to the SLM process. The free surface between melt pool and ambient gas is numerically described on the basis of averaged, fluid volume fraction-dependent material properties prevalent in discretization cells at the boundary fluid-gas in a manner similar to the VOM. In [91], temperature-dependent surface tension, and thus Marangoni convection, has been neglected. Figure 18 shows the 2D melt pool shape resulting from a spatially fixed laser beam position and different wetting angles θ0 , determined by the considered material combination at the triple point solid-fluid-gas, for otherwise identical system parameters. According to Figure 18, the wetting behavior expressed via the wetting angle θ0 considerably influences the resulting melt pool shape. More detailed investigations revealed, however, that the resulting melt pool shape is considerably influenced by the powder packing density and, at least in the range of lower packing densities, also by the individual stochastic realization of different powder beds with identical packing density. Consequently, the local powder topology as well as the wetting behavior resulting from the prevalent material combination solid-liquid-gas might have considerable influence on balling of the solidifying tail. In addition to these single point simulations, also single-track simulations have been performed in [91]. The melt pool shapes resulting from different scan velocities, once at fixed laser beam power and once at fixed spatial energy density are illustrated in Figure 19. It can be observed that the melt pool stability at constant laser power increases with decreasing laser beam velocity, which, in turn, yields a higher spatial energy density. Small melt pool sizes in the range of the powder grain diameter, which results from low energy input, limits the wetting behavior and fosters round melt pool shapes with minimal surface energy. With increasing size, the relative contribution of surface tension, wetting and gravity effects changes and the melt pool is able to interconnect the powder particles visible in form of a pronounced wetting behavior. On the other hand, comparable melt track topologies are found at constant spatial energy density, advocating this quantity as most relevant factor of influence for the given process conditions. All in all, it has been concluded that the packing density of the powder bed has the most significant effect on the melt pool characteristics. 29 (a) Constant laser beam power. (b) Constant energy density. Figure 19: 2D Single-track mesoscopic simulation results: Influence of laser beam velocity on wetting behavior and melt track shape, [91]. In [92], the model of [91] has been applied to study a layer-wise buildup process, again for Ti-Al6-V4 powder material. The model simplifies the actual SLM process to 2D by investigating only the plane spanned by the build and scan direction. Furthermore, dimensionless factors characterizing the different physical phenomena and time scales prevalent in SLM have been analyzed in this reference. Based on a comparison of laser beam interaction time and thermal diffusion time, a recommendation for the optimal laser beam velocity has been made in order to avoid overheating and excessive evaporation on the one hand as well as unnecessarily high energy losses on the other hand. It has been argued that surface tension dominates gravity in terms of driving forces whereas inertia effects dominate viscous damping in terms of resistance forces. Furthermore, the Rayleigh time scale, characterizing the relaxation of an interface perturbation under the action of inertia and surface tension forces, turned out to be more than one order of magnitude smaller than thermal diffusion time scales. That is, surface tension-driven coalescence of powder particles follows the melting process nearly instantaneously. Further, in [11] this model has been employed in order to illustrate the formation of pores. In [109, 88], the simulation framework proposed in [91] has been supplemented by means of alternative models for electron energy dissipation in EBM processes. In reference [1] the method has been extended to 3D problems with a special emphasize on code parallelization strategies to increase computational efficiency. 2.4. Microscopic simulation models The mechanical and physical properties of metals, and therefore of all parts made by SLM, are influenced by their local composition and microstructure characterized by grain size, grain morphology (shape) and grain orientation (texture). In general, the microstructure evolution during solidification processes is governed by the spatial temperature gradients (G), cooling rates Ṫ and solidification front velocities v [99, 49]. As visualized in Figure 20, a qualitative solidification map can approximately be divided into areas of high and low cooling rates and into regimes of columnar and equiaxed structures. Typically, high cooling rates lead to rapid nucleation of grains and non-equilibrium microstructures, characterized by small grain sizes or small dendrite arm spacings. The competition of thermal gradients and solidification front velocities lead to columnar / elongated, dendritic grain morphologies if high thermal gradients are dominating since in this scenario the liquid-solid interface becomes unstable and preferred orientations grow faster than others to finally form dendritic morphologies. On the contrary, equiaxed, rather spherical grain morphologies are expected if high solidification front velocities are dominating as consequence of a more stable liquid-solid interface. In contrast to conventional metallurgy such as casting, SLM is based on a highly localized energy supply yielding extremely high heating rates. As consequence of the localized energy supply, the solidified material of previous melt tracks and layers remains at comparatively low temperatures, which in turn yields extreme cooling rates after the laser 30 Figure 20: Microstructure versus thermal gradients and melt front velocity, [49]. has passed by. The magnitude of heating and cooling rates typically lies in the range of 103 −108 K/s. The succession of extreme heating and cooling rates eventually also results in high spatial temperature gradients (103 −108 K/m) and solidification front velocities (1−30m/s) [108, 53]. Such a large range of cooling rates and thermal gradients induces a large variety of microstructures of as-deposited materials differing by grain size, morphology, and orientation. The high cooling rates often lead to meta-stable and non-equilibrium phases and comparatively small grain sizes. As a trend, the grain sizes in SLM are typically smaller and the material strength is higher as compared to alternative manufacturing processes such as casting. With the build direction commonly representing the direction of highest temperature gradients, depending on the employed material often elongated, dendritic grain morphologies accompanied by a higher material strength in build direction can be observed. Additionally, solid state transformations and grain growth contribute to crucial microstructure changes during the repeated thermal cycles (while remaining below the melting point) experienced by previously deposited material after the initial solidification process (see e.g. the transformation from martensite to alpha and beta phases in Ti-6Al-4V as discussed in Section 3.4). The knowledge of the resulting micro structure characteristics might provide important details for the formulation of process-specific continuum constitutive laws, which are crucial for accurate residual stress predictions as intended by macroscopic simulation models. On the other hand, global temperature distributions provided by macroscopic continuum models represent essential input variables for studying the solidification process by means of microstructure models. So-called phase field models based on various spatial and temporal discretization strategies can be considered as one of the most established modeling approaches to study solidification processes [155, 41, 96, 25, 142, 3, 54, 89, 123, 152] but also diffusionless solid-solid phase transformations occurring for example during the formation of martensitic non-equilibrium phases [118, 5]. Phase field models are based on the definition of a free-energy functional, which typically has the following form: Π(φ, c, T, Ψ) = Z Ω Π̃(φ, c, T, Ψ)dΩ with Π̃(φ, c, T, Ψ) = f (φ, c, T ) + c2φ 2 || 5 φ||2 + Π̃o (φ, T, || 5 Ψ||). (14) Here, φ represents the phase field variable or order parameter taking on the value φ = 1 in the solid phase, φ = 0 in the liquid phase, and a value φ ∈]0; 1[ in the finitely thick interface region of smooth phase transition. Furthermore, c represents the (volume / mass) fraction of a certain chemical species prevalent in the problem of interest. For 31 binary systems, the fraction of the second species is given by 1 − c while for general systems consisting of n species, n − 1 variables ci are required. As before, T is the absolute temperature. Also grain orientation is considered in the phase field model (14). In this context, Ψ represents a vector-valued parametrization of grain orientation, e.g. given by rotation vectors, quaternions or Euler parameters. The first term of Π̃(φ, c, T, Ψ) in (14) represents the temperature-dependent Gibbs free energy either associated with the liquid or solid phase at a certain location in the liquid or solid domain, or a proper interpolation of these values for locations on the phase boundary region. The second term represents energy contributions stemming from liquid-solid phase boundaries characterized by an existing gradient 5φ of the phase field. The choice of the parameter cφ determines the resulting phase boundary thickness and requires a compromise between an accurate resolution of the small boundary layer thickness typically prevalent in physical systems (high value of cφ ) and a certain degree of artificially increased thickness for reasons of computational efficiency and robustness (low value of cφ ). Eventually, the last term Π̃o (φ, T, || 5 Ψ||) yields energy contributions due to crystallographic orientation gradients 5Ψ. This contribution fosters uniform growth within individual grains and penalizes the misorientations at grain boundaries, making larger grain sizes with reduced overall grain boundary surface favorable in configurations of thermodynamical equilibrium. Variation of the functional (14) yields the EulerLagrange equations of the variational problem, determining the equilibrium solution φ∗ , c∗ , Ψ∗ of the variables φ, c, Ψ: Π(φ∗ , c∗ , T, Ψ∗ ) = extremum → ∂Π ∗ ∗ (φ , c , T, Ψ∗ ) = 0, ∂φ ∂Π ∗ ∗ (φ , c , T, Ψ∗ ) = 0, ∂c ∂Π ∗ ∗ (φ , c , T, Ψ∗ ) = 0. ∂Ψ (15) Starting with a non-equilibrium system, e.g. an undercooled melt, the equilibrium solution defined by the functional (14) is found in practical simulations by transforming (15) into a transient problem and searching for the associated steady state solution. Thereto, the so-called time-evolution phase-field approach applies additional rate terms proportional to φ̇, ċ and Ψ̇ on the right-hand sides of equations (15). In SLM, the temperature field is required as input variable for the phase field model and might e.g. be provided by a macroscopic model based on (8). Figure 21: 3D phase field simulation of solidification and grain growth process assuming isotropic boundary energy and mobility, [96]. The spatial discretization resolution required for a sufficiently accurate phase boundary representation typically results in a considerable numerical effort. For that reason, many of the available phase field models are applied to 2D scenarios. For illustration of the general capability of phase field models, two existing 3D approaches - even though not applied in the context of SLM - shall briefly be presented. One has been proposed in [96], based on a finite difference discretization scheme in space and in time. In this work, only one chemical species (variable c not required) and only a discrete set of possible grain orientations is considered. Consequently, the variable Ψ representing a continuous spectrum of possible orientations is replaced by additional phase field variables φi , one for each possible orientation. Figure 21 shows different time steps of a solidification problem allowing for 20 discrete grain orientations. Considering the employed 180x180x180 grid and 8000 time steps, the computational effort seems to be large. As visualized in Figure 21, the initial number of 6000 grains reduces to approximately 200 at the end of the simulation, yielding a decreased overall grain boundary energy in thermodynamical equilibrium. 32 A further 3D PFM based on a spatial FVM discretization and an implicit backward difference time stepping scheme is given in [41]. There, two different chemical species (reflected by one concentration variable c) as well as a continuous grain orientation field Ψ as given in (14) has been employed. Figure 22 shows the results of a simulation considering grain growth within a two-phase region (solid phases  and δ) of a binary alloy (chemical species A and B) phase diagram under non-equilibrium conditions. Starting with randomly distributed δ-phase nuclei (see Figure 22, left), this phase begins growing within the -phase matrix (see Figure 22, middle, for the phase distribution at the end of the simulation). Due to the lower diffusion constant of the δ-phase and the high cooling rates chosen in this example, the so-called coring effect, i.e. a thermodynamical non-equilibrium state characterized by an inhomogeneous concentration distribution in the δ-phase (see Figure 22, right), has been observed. The representability of such non-equilibrium configurations is very relevant for the modeling of SLM, given the high cooling rates present. The numerical simulation was based on a 256x256x256 finite volume grid and 1400 time steps in order to simulate 2s of physical time, leading to 14h computation time on 64 processors. Figure 22: 3D phase field simulation considering the grain growth process within the -δ solid phase region of a binary alloy phase diagram: Initial distribution of δ-phase nuclei (left), final distribution of - and δ-phase (middle) as well as concentration of the second species B (right), [41]. Only a few approaches to microstructure modeling can be found for SLM/EBM, all of them restricted to 2D. In Gong et al. [51], the solidification and growth of primary β-phase grains during the EBM processing of Ti-6Al-4V has been studied. Thereto, a phase field model similar to (14) has been employed with c representing the concentration of Ti and 1 − c the concentration of the solute, treated as the combination of Al and V. Thus, the ternary alloy Ti-6Al-4V has been simplified as a binary system, and the influence of grain orientation by means of a parameter Ψ has not been considered. The required temperature field has been provided via a finite element solution of the energy equation (8). Figure 23 shows the simulated growth of columnar structures at different time steps during the solidification process. Initially, a pre-defined number of nuclei is placed on the substrate below the molten powder layer. The pre-defined number of nuclei has been determined via a regression equation accounting for an increasing nucleation density with increasing undercooling. Initial dendrite growth occurred primarily in scan direction (horizontal) and build direction (vertical) until neighboring dendrites started to contact each other. Subsequently, vertical growth was observed, resulting in a grain orientation parallel to the build direction as typically observed in experiments. In Figure 24, the final configurations resulting from different scan speeds are depicted. As expected and in agreement with experiments considered in [51], the higher cooling rates induced by higher scan velocities result in a higher nucleation density and, eventually, in a finer grain structure, which is still oriented in the build direction. A second category of approaches often applied to study solidification processes is given by cellular automaton (CA) schemes. Like PFMs, CA schemes require the solution of the thermal field as input. However, the CA schemes do not rely on an additional phase field variable φ that explicitly describes the location of the individual phases and phase boundaries. Instead, the evolution of the grain growth is typically constructed in a rather geometrical manner by tracking the solid-liquid interface based on so-called cell capturing schemes under consideration of a given initial nucleus orientation and the current interface geometry. A comprehensive overview of methodologies for the modeling of microstructure evolution, among others considering PFM and CA schemes, is also given by [12]. 33 Figure 23: 2D phase field simulation of β-structure growth during EBM of Ti-6Al-4V with nucleation on substrate at different time steps, [51]. Figure 24: 2D phase field modeling of dendrite growth for different scan velocities: (a) low, (b) medium, (c) high, [51]. In [126], the columnar grain growth during solidification of Inconel 718 alloy is studied based on a 2D realization of the CA approach. As visualized in Figure 25, in every time step, the temperature field as well as the phase distribution, are exchanged in a weakly coupled manner. The required temperature field is derived by means of a 2D mesoscopic Lattice Boltzmann simulation model [91] already discussed in Section 2.3. This study predicts columnar grain growth in the build direction. In contrast to [51], different grain orientations are distinguished by this model and the effect of stray grain formation due to partially molten powder particles on the left and right boundary of the computational domain could be taken into account thanks to the powder bed resolution on the mesoscopic scale. Also an epitaxial grain growth in build direction, i.e. grain growth on a newly deposited layer continuing the crystallographic structures of grains prevalent in the previous layer [68], is predicted, leading to grain sizes that considerably exceed the powder layer height. The effect of grain coarsening in the solid phase, i.e. competition between solid grains of different orientation leading to a growth of certain grains, has not been addressed by this CA scheme. Figure 25: Exchange of temperature field and phase information between mesoscopic and microscopic simulation model, [126]. 34 Also in the recent contribution [171], first steps are made towards a 2D simulation framework predicting the resulting microstructure evolution during solidification of 316L in SLM processing. The thermal field has been simulated based on a FEM discretization of (8) while the microstructure model relies on a CA approach. A problem during the SLM processing of Inconel 718 is the segregation of Nb and the formation of so-called Laves phase grains in the interdendritic region, which may considerably degrade mechanical properties. Nie et al. [119] employed a numerical scheme based on a stochastic approach for studying the nucleation and growth of dendrites, the segregation of niobium (Nb) and the formation of Laves phase particles during the solidification. Additionally, a FEM-based discretization of the heat conduction equation has been employed for solving the thermal problem, which is required as input for the stochastic microstrucutre evolution model. Based on numerical simulations, Nie et al. demonstrated that low cooling rates as well as a high ratio G/v of temperature gradients (G) to solidification front velocities (v) yield microstructures tending to large columnar dendrite arm spacing and, consequently, continuously distributed coarse Laves phase particles. Thus, it is suggested to increase cooling rates and to decrease temperature gradients in order to obtain small equiaxed dendrite arm spacings and discrete Laves phase particles (Figure 26). Figure 26: Formation of Laves phase particles under different solidification conditions, [119]. 2.5. Summary of existing macro-, meso- and microscale models and potential for future developments Computational efficiency represents one of the key requirements for SLM simulation tools to be capable of capturing practically relevant time and length scales. Potential improvements of existing simulation frameworks comprise temporally and spatially adaptive time and space discretization schemes, the use of implicit time integrators, efficient code design suitable for high-performance parallel computing as well as the development of efficient and consistent strategies for material deposition and dynamic adaption of the computational domain. Besides computational performance, existing simulation tools require further improvements in terms of model and discretization accuracy. In the following, such improvements but also a potential coupling of the three model classes will be discussed (Figure 27). 35 Macroscopic model Output: Temperature field; residual stresses; dimensional warping Needs: Improved material laws; models for heat absorption and transport in powder and melt Effective powder/melt properties Mesoscopic model Thermal boundary conditions Temperature field Reduction of computational costs: Mesh adaptivity Time-step adaptivity Implicit schemes Material law Microscopic model Output: Grain size, morphology and texture; phase distribution Needs: 3D models; solid state transformations; multiple chemical species Output: Surface quality; layer adhesion; porosity Needs: Powder dynamics and collisions; vapor and gas flow Figure 27: Potential for future improvements and coupling of submodels on macro-, meso- and microscale. Macroscopic models: Macroscopic SLM models typically consider entire parts in order to predict global temperature fields, residual stresses or dimensional warping. The melt pool shape and the resulting temperature distribution crucially depend on an accurate model of laser radiation transfer into and heat conduction transfer within the powder phase. In this context, macroscopic models considering the powder phase in a spatially homogenized sense could be improved by transferring information from mesoscopic models. Also model information related to the melt pool hydrodynamics, which determine the convection-dominated heat transfer within the melting phase, could either be gained from mesoscopic models or explicitly considered in macroscopic models following existing approaches in the fields of laser and electron beam welding. These measures might considerably improve the accuracy of the predicted temperature field in the direct vicinity of the melt pool. While macroscopic models tend to strongly abstract the complex physics prevalent in this regime, they are expected to provide good predictions of the temperature field in regions further away from the melt pool, where heat flow is predominantly governed by the global part geometry, thermal boundary conditions as well as solid material characteristics. However, one of the ultimate goals of macroscopic models is the prediction of residual stress distributions. While temperature-induced thermal strains can be identified as the kinematic origin of these residual stresses, the actual magnitude of these stresses as well as the magnitude of maximally admissible stresses (given e.g. by the yield strength) is essentially determined by the solid material properties, and, thus, by the underlying metallurgic microstructure, which is, however, part of the unknown solution variables of the SLM process. On the contrary, even if derived on the basis of exact temperature solutions, the significance of residual stress predictions is strongly limited as long as comparatively simple material laws are employed. Supplying macroscopic constitutive models with further details concerning material inhomogeneity, anisotropy and temperature-history-dependence, e.g. by means of microscale SLM models, could drastically improve the quality of predictive macroscopic SLM models. Mesoscopic models: Mesoscopic SLM models resolve length scales of individual powder grains and below in order to accurately describe radiation and heat transfer within the powder bed as well as heat and mass transfer within the melt pool governed by surface tension, wetting effects and evaporation-induced recoil pressure. The ultimate goal pursued by these highly resolved models is to gain an essential understanding of underlying physical phenomena responsible for melt track stability, resulting adhesion between subsequent layers, surface quality as well as creation mechanisms of pores and other types of volume and surface defects. This high spatial resolution naturally goes along with a considerable computational effort. So far, the computational challenge of solving the resulting numerical problems involving complex field and domain couplings has been addressed by explicit time integration limiting the 36 currently observable length scales to single tracks of a few mm length and the observable time scales to several 100µs despite massive usage of high performance computing resources. Concerning model accuracy, recent experimental investigations suggest that the evaporation-induced gas flow above the melt pool might considerably affect melt pool and powder bed dynamics [111]. While some first contributions have already considered evaporation in an implicit manner via recoil pressure and heat transfer models, an explicit modeling of the ambient gas flow, at least in the direct vicinity of the melt pool, could allow for considerable insights into the governing physical mechanisms and suggest strategies of influencing these (typically undesirable) gas dynamics. Also effects such as powder particle ejection or denudation are not sufficiently understood and require further investigation. While existing mesoscopic models typically consider spatially fixed powder particles, models that account for powder particle dynamics and collision could allow to study and understand these effects in detail. Eventually, mesoscopic models could benefit from improved thermal and mechanical boundary conditions at the interface between the representative mesoscopic volume and the global SLM part. In this context, macroscopic SLM models might provide useful data. Microscopic models: Microscopic SLM models investigate the metallurgical microstructure evolution resulting from the high temperature gradients and extreme heating and cooling rates during the SLM process and aim at the prediction of resulting grain sizes, morphologies and textures as well as phase distributions. While mesoscopic models intend rather universal statements concerning optimal adjustment of parameters such as laser beam velocity and power, powder layer thickness and packing density as well as grain shape and size distribution for a given powder material, global temperature distributions, residual stress fields and microstructure evolutions as derived by macroscopic and microscopic SLM models strongly depend on the specific part geometry. Consequently, efficient numerical tools are required in order to enable the simulation of real-size parts in acceptable simulation time and a possible exchange of information between the macroscopic and microscopic scale. Macroscopic models typically provide the temperature field solution as input for microscopic SLM models. Compared to the variety of existing macroscopic and mesoscopic SLM models, microscopic modeling approaches that take into account the specific thermal conditions prevalent in SLM processes can be regarded as being less elaborated. The few existing approaches rely on a simplified 2D representation of the problem although many of the physical mechanisms governing metallurgical microstructure evolution are three-dimensional in nature. Current approaches do not consider more than two individual chemical species. However, an explicit modeling of all relevant alloying elements might be desirable in order to describe experimentally observed effects such as the segregation of individual alloying components. Also mechanisms of solid state phase transformation and grain growth, which have been observed in experiments and are likely to considerably change the material properties during repeated thermal cycles after the initial solidification process, might be desirable supplementations for future models. 3. Experimental studies of thermophysical mechanisms in SLM In the following, the modeling approaches discussed in Section 2 shall be supplemented by representative methods of experimental characterization. The next Section 3.1 focuses on powder bed radiation and heat transfer as discussed from a rather theoretical point of view in Section 2.1. In the subsequent Sections 3.2, 3.3 and 3.4, the experimental characterization of melt track stability, surface quality and defects, of residual stresses and dimensional warping effects as well as of metallurgical microstructures and grain morphology will be considered. These sections represent the counterpart to the Sections 2.2, 2.3 and 2.4, where exactly these aspects, prevalent at different length scales of the SLM-manufactured part, have been described by means of macroscopic, mesoscopic and microscopic models. 3.1. Optical and thermal characterization of powders Initial research on powder-laser interaction began not with AM in mind, but rather with surface modification techniques such as cladding and hard-facing processes [64]. Haag et al. [64] describe the archetypal experiment to measure absorption coefficients of metal powder, wherein a laser of known power is used to irradiate a powder bed, with the absorption coefficient calculated from the time-derivative of the powder bed temperature. In this study, a CO2 laser (w = 10.6µm) was used to heat aluminum, iron, titanium aluminide, and copper powders. Selected particle sizes range from 42µm to 200µm, validated by SEM. The key results are summarized in Figure 28: In the range of low laser power, material-dependent absorption coefficients as well as the consideration of different powder sizes cause only 37 Figure 28: Effects of powder grain size, material as well as laser power on relative powder layer absorption (w = 10.6µm), [64]. slight variations in the overall powder bed absorption leading to values of 30% − 45% absorption as compared to the incident energy radiation. The increase in absorption of iron and copper at higher laser powers, and correspondingly high temperatures, is attributed to surface oxidation of the material. A practically identical experiment was performed by Tolochko [147] to catalog nominal absorption coefficients for materials relevant to SLS and SLM processing at wavelengths of 1.06µm and 10.6µm. Results indicate that the wavelength 1.06µm represents a superior choice for processing metals, with absorption coefficients generally lying between 0.6 and 0.7, as opposed to values around 0.25 to 0.45 for the longer wavelength. Conversely, absorption of 10.6µm radiation is greater for the cataloged polymer materials, and ceramic materials are shown to vary widely. Figure 29: Optical depth versus powder layer thickness for iron powder of different grain size (w = 1.064µm), [113]. 38 McVey presented an approach to measure the optical power delivered, reflected, and transmitted through a layer of powder using integrating spheres above and below the powder bed [113]. In Figure 29, the effective optical depth (the logarithm of the ratio of measured / non-absorbed to incident light), is plotted versus the powder bed thickness for an applied laser beam of wavelength w = 1.064µm. Accordingly, the absorption of 1.064µm light has a strong dependence on particle size, with finer particles yielding stronger absorption. More recently, Rombouts et al. [132] have studied the thermal conductivity of powdered 316 stainless steel, iron, and copper. Their apparatus relies on a modulated laser beam to heat an optically opaque sample container, which subsequently conducts heat into a powder bed within the container. Under these circumstances, heat flows through the powder bed to pyro-electric detectors, which are read out using lock-in amplifiers indexed to the laser’s modulation. Experimental results are shown in Figure 30. Accordingly, the powder conductivities are two to three orders of magnitude lower than in the fused material or build platform with bulk conductivities of 15, 74 and 395 W/m K for stainless steel, iron, and copper, respectively. It is argued that the thermal transport in powder beds of such fine particles occurs primarily through the gaseous phase (Knudsen diffusion) while direct particle-to-particle conduction is the governing effect for larger particle sizes (millimeter scale). Figure 30: Thermal conductivities of metallic powder beds for different materials and grain sizes distributions, [132]. 3.2. Measurement of residual stresses and dimensional warping As already stated earlier, basically, two thermal regimes responsible for the creation of residual stresses can be distinguished in SLM. First, stresses might be induced in the solid substrate just underneath the melt pool due to high thermal gradients and cooling rates in this region. These residual stresses are influenced by the powder structure and melt pool thermo-hydrodynamics and might result in a delamination of the current material layer. Secondly, residual stresses might also result as consequence of the repeated thermal cycles prevalent in material layers in larger distance from the top layer. The main factors of influence in this region are given by the specific geometry of the part (e.g. slender columns, thin walls and overhanging structures) as well as thermal boundary conditions e.g. in form of fixations on the build platform. Even for primitive geometries such as cubes this second category of thermal stresses might occur due to the temperature gradient in build direction, which typically results in compressive stresses in lower layers induced when the layers of higher temperatures above cool down. In all cases, these thermal stresses need to be reduced, or better avoided, in order to fully exploit the mechanical strength of a part when exposed to external loads, in order to avoid dimensional warping after removing the parts from the build platform and eventually in order to avoid crack propagation typically initiated at locations with residual stress peaks. On the one hand, residual stresses can be relieved by annealing after the SLM process (before build plate removal). On the other hand, such a heat treatment can compromise dimensional accuracy and impose grain growth demanding a tradeoff between these diverse requirements. Lu et al. [106] studied the effect of laser scan pattern on the properties of Inconel 718. The laser beam scanning strategy is depicted in Figure 31, where part infill is fused by sequentially scanning square regions known as islands. 39 Specifically, dimensions of 2mm x 2mm, 3mm x 3mm, 5mm x 5mm, and 7mm x 7mm have been studied, using a 180W laser, 600mm/s scan speed, 30µm layer thickness, and a 150µm spot size. The study begins with a discussion of part density, in which the 5mm and 7mm specimens are shown to have 99.1% of their theoretical density. Both the 2mm and 3mm specimens featured lower densities at 98.67% and 98.86%, respectively (see also the next section for a discussion of mesoscale defects such as pores). These findings are congruent with surface micrographs, which show increased porosity within the smaller island sizes, and even cracking within the 2mm specimen. Stress-strain data is also included, which shows comparatively little change in offset yield strength and ultimate tensile strength, yet elongation at break increases from 16.85% to 25.25% with increasing island size. Again, this is in agreement with the surface micrographs, as the increased density of defects serve as sites to propagate cracks and initiate fracture. Finally, micro-hardness and residual stresses were evaluated. Counterintuitively, the 2mm specimen featured the lowest residual stress. This observation is believed to result from stress relaxation due to cracking. As expected, the 3mm specimens showed greater residual stresses than the 5mm and 7mm specimens. The greater thermal gradients of the smaller 3mm specimen, arising from a greater degree of heat remaining from the previous pass, can explain the greater residual stresses as compared to the 5 mm specimens. Figure 31: Determination of optimal island sizes for an island-based SLM scanning strategy, [106]. In Hodge et al. [69], a combined simulation and experimental approach has been applied in order to access residual stresses as well as the amount of dimensional warping observed in AM components, and to find how those stresses are dependent upon the build orientation and infill scan strategy. As described earlier, Digital Image Correlation (DIC) has been used to measure stress-induced dimensional changes in triangular test components. This experimental technique is based upon time-correlating successive images to extract strain measurements as an object undergoes a physical change, such as in temperature or external loading [84]. Hodge uses this technique to monitor the surface of triangular test components as they are removed from the build platform. This process changes the residual stress distribution within the component, thereby inducing a corresponding change in strain. These measurements and previously discussed simulation results are augmented using neutron diffraction experiments to directly measure the residual stress distribution in the test pieces. Like DIC, this experimental method is well established in solid mechanics for mapping stress distributions in crystal-like materials [156]. The full strain tensor may be determined by approaching the same volume from a plurality of directions. Of course, this measurement represents the average strain tensor over the volume probed, commonly between 0.2 and 1000mm3 . Hodge et al. mainly employed the described experimental techniques for verification of their simulation model (see Section 2.2 for a discussion of the results). Mercelis and Kruth [116] evaluated residual stresses as dependent on scan strategy, build height, and degree of base plate pre-heating. They relied on sectioning their parts followed by X-Ray diffraction studies to determine residual stresses. X-Ray diffraction determines residual stresses on identical principles to neutron diffraction. However, the comparatively low penetration of the X-Ray technique makes it only useful as a surface technique, hence the need for sectioning. The authors clearly showed that residual stresses increase with component height. As long as the parts stayed connected to the baseplate, very high stress levels, typically in the range of the materials yield strength, could be observed. Also the exposure strategy that is being used to fuse the powder layers has a large influence on the residual stress levels being developed. It was found that stresses are larger perpendicular to the scan direction than 40 in scan direction. A scanning strategy subdividing the surface in small islands resulted in a lower maximum stress value. Preheating the baseplate to 200◦C was shown to reduce induced residual stresses by nearly 10%. Finally, the authors showed that removing the part from the baseplate relaxes approximately 50% of the residual stresses in the part, which negatively impacts the dimensional accuracy of the component once removed. Havermann et al. [66] presented a unique method in which fiber optics were embedded within SLM parts. Thereto, the authors fabricated rectangular test pieces. At the location where a strain measurement was desired, a fiber optic strain sensor (Bragg grating) was placed into the part, thereby enabling the strains induced by residual stresses in the component to also deform the optical fiber. This changes the optical transmission of the embedded Bragg grating, enabling the quantification of strain. The employed small fiber diameter enabled measurements within the first several layers of SLM parts, where interaction with the baseplate is central to the development of stresses. The authors observed increasing compressive residual stresses as the first 6 layers above the fiber were fused. After building 14 layers above the fiber and allowing the component to cool, residual stresses in the range of 70MPa were measured. Kempen et al. [80] studied the effect of build platform preheating temperature in order to lower thermal gradients and to finally yield crack-free parts with high density. Figure 32 shows three samples, which were built using three different preheating temperatures (90◦C (left), 150◦C (middle), and 200◦C (right)). It can be seen that the higher preheating temperature clearly results in lower residual stresses and a lower degree of crack formation. In practical applications, too high preheating temperatures and too long preheating times might however lead to an undesirable sintering of loose powder, which complicates its removal from the final part. Figure 32: Blocks of M2 high speed steel (HSS), manufactured by SLM at build platform temperatures of 90◦ C (left), 150◦ C (middle), and 200◦ C (right), showing the critical influence of preheating temperature on cracking, [80]. 3.3. Characterization of melt track quality and defect detection In this section, some exemplary experimental approaches for characterizing melt track stability and quality as well as possible defects will be discussed. In recent years, much effort has been invested in monitoring SLM and EBM processes for process and quality control [43]. Defect detection centers on identifying pores, balling, unfused powder, and cracking as a result of sub-optimal process parameters. In parallel, high speed monitoring has been researched with the aim of real time process control. The results of these efforts have manifested in several commercially available systems for monitoring metal AM processes as shown in Figure 33. In the following, we will discuss several of the key works behind these products, and refer to the review by Everton et al. [43] for complete treatment. Figure 33: Summary of different commercially available monitoring systems for SLM, EBM and DED processing, [43]. 41 Craeghs et al. [30] studied melt pool heat transfer using a radiometer and high speed camera, arranged coaxially with the laser. In this work, they describe the so-called edge effect that occurs when a track is scanned without an adjacent contour as already discussed in Section 2.3. Heat from scan segments adjacent to neighboring contours are shown to conduct both into the underlying layers and into the previously solidified, adjacent track in the same layer. However, heat from the initial contour does not have this second path, meaning the bulk of the heat flux must be directed into the underlying layer. As a result, the melt pool is larger, reaches higher peak temperatures, stays hot longer, and incorporates more material than when subsequent contours are melted. Overhanging features were also studied, where it was found that scan parameters suitable for standard processing conditions delivered too much power to the initial layers of overhanging features fabricated above areas of unfused powder. The extra heat enhances Rayleigh instabilities, leading to balling, poor surface finish, and undesirable mechanical properties. Finally, acute corners were shown to present many of the same challenges with temperature control. Specifically, the case of a scan pattern incorporating a 180 degree U-turn was investigated. Again, due to the reduced area for heat to flow out of the melt pool and the partly scan path re-heating, this extra heat results in blob-like formations at the apex of the U-turn. This same experimental hardware was also used to investigate the influence of support strategies on process temperatures for fabrication of overhanging features [93]. Here, the authors printed T-shaped structures, in which the overhung regions were supported with uniformly spaced and sized column like additions. The spatial density of the columns was shown to greatly influence the maximum temperatures achieved in the overhanging layer, and therefore porosity and surface finish, as the support structure serves as a heat sink. Krauss et al. [94] performed long-wave infrared (LWIR) imaging to determine component porosity in Inconel-718 specimens. The study begins by deriving an approximation for the thermal diffusivity a, which is given by !2 !2 1 AP 1 . (16) a≈ √ t T (t) − T re f Vhc p ρ π In this expression, P and V represent laser power and scan velocity, h is the hatch spacing and A represents the efficiency of energy absorption in the powder layer (assumed 60%). Moreover, c p and ρ are the specific heat capacity and density of the material, as usual. Finally, T re f is the ambient temperature of the build environment, and T (t) is the radiometrically determined temperature of the pixel as a function of time t. The authors assign an imperfection level based upon the observed porosity and degree of binding error. The imperfection level IP was determined to be inversely proportional to the diffusivity a of cubic specimens at a layer height of 4mm with a correlation coefficient of 0.7. Based on a proper calibration of the IP-a relation and (16), the proposed procedure enables the in-situ determination of part imperfection based on measured temperature profiles. The method is shown to be sensitive to both delamination and porosity arising from sub-optimal process parameters, with sensitivity demonstrated over a range from nearly fully dense material to a void fraction of 40%. Ultimate sensitivity of this method is not calculated. However, the authors note that the spatial and temporal resolution of imaging sensors represents the greatest challenge. In the follow-up study [95], more detailed information on laser power and scan rate is provided, and the resulting effects on thermal diffusivity, maximal temperature, and observation of melt sputter have been presented. Figure 34: Detection of porosity from melt pool size and temperature: Comparison of a) predicted and b) actual defect locations (black areas), [28]. Simultaneously, Clijsters et al. [28] have studied in-situ monitoring for quality control and are working towards feedback for process control. Using an InGaAs NIR camera and a pair of photo diodes with differing band pass filters, 42 the authors extract signals representing melt pool radiance (intensity of emitted light), area, length, and width. They report success in detection of overheating of the melt pool, specifically at the extremes of fill scan vectors, along with increased porosity in these regions. Using data-fusion techniques, the authors are able to predict the locations of porosity within AM parts, as verified by CT cross sectional images (see Figure 34 for illustration). Experimentally, process maps can be constituted by gathering data on part density and other properties as related to the scan rate, hatch spacing, laser power, etc. While ultimate control of SLM will arguably require determination of local scan parameters according to the local part geometry and thermal boundary conditions, the present approach is to determine a generalized ”process window” which can be applied to approximately achieve full density. The literature in this field is abundant, considering a wide variety of materials and processes. Thomas et al. [146], propose a methodology of normalizing process parameters for SLM and EBM. The methodology builds on the previous work [72], where a dimensionless laser beam power P∗ and velocity V ∗ are defined: P∗ := AP , rB λ(T m − T 0 ) (17a) VrB . (17b) a Originally, these dimensionless quantities have been identified from analytical solutions derived for the temperature profiles resulting from a moving heat source on a semi-infinite body, a model that has been applied in order to study heat transport in laser beam welding. In (17a), the effective laser power, defined as laser power P times absorption coefficient A, is related to the temperature difference between initial powder bed temperature T 0 and melting temperature T m , the thermal conductivity λ and the laser beam radius rB . Equation (17b) relates the velocity V of the laser beam heat source to the thermal diffusivity a. According to [72], the choice of the parameters P∗ and V ∗ controls the peak temperature and heating rate. This representation matches Figure 3. V ∗ := Thomas et al. supplemented the parameters P∗ and V ∗ proposed in [72] by the following dimensionless quantities: E ∗ := ∗ Emin := l∗ := 2l rB (18a) h∗ := h rB (18b) AP 1 · 2VlrB 0.67ρc p (T m − T 0 ) (18c) 3 ∗ AP 1 E = · 2 2VlrB ρc p (T m − T 0 ) (18d) Here, the dimensionless thickness of the powder layer l∗ is defined via the physical powder bed depth l and the laser beam radius rB . Also the hatch spacing h is normalized by the laser beam radius. Assuming a powder bed porosity of 0.67, the term 0.67ρc p (T m −T 0 ) represents the energy density required to increase the temperature of powder with effective density 0.67ρ and heat capacity c p from the initial temperature T 0 to the melt temperature T m . Again, AP represents the effectively absorbed laser power. The normalization of the laser power with the product 2VlrB yields an energy per unit volume. By assuming a latent heat of melting in the range of Hm ≈ 0.5ρc p (T m −T 0 ) for typically employed metals and preheating temperatures, Thomas postulated the energy density required to increase the powder temperature from T 0 to T m and to melt the powder subsequently as the value 0.67ρc p (T m −T 0 ) + Hm = ρc p (T m −T 0 ). ∗ Consequently, the last dimensionless parameter Emin represents the ratio of incident energy density to the energy density required for heating and melting the powder. The parameter combinations found in different experimental contri∗ butions from the literature have been plotted in a double-logarithmic (Emin )−(1/h∗ ) diagram as illustrated in Figure 35. In order to gain further understanding concerning this specific choice of diagram axes, a slightly different version of normalized hatch spacing h̃∗ := 0.5h∗ = h/(2rB ) shall be considered in the following. In this context, h̃∗ = 0 represents 43 Figure 35: Normalized processing diagram showing different SLM/EBM realizations from the literature. The inverse ∗ , [146]. dimensionless hatch spacing h∗ is plotted over the dimensionless energy ratio Emin ∗ full overlap and h̃∗ = 1 represents vanishing overlap of two successive scan tracks. In a (Emin )−(1/h̃∗ ) diagram fol∗ ∗ ∗ lowing this modified definition, constant ratios Ẽ0 := Emin /h̃ = const. are represented by straight lines with slope −1. For a powder bed surface of dimensions Lv in scan direction and Lh in hatch direction, Lh /h scan tracks are required in order to scan the entire surface and Lv /V represents the time of scanning one track. The total amount of energy absorbed in one powder layer is given by Lh /h · Lv /V · AP. On the other hand, based on the assumptions in [146], the amount of energy required to heat and melt the entire powder bed is given by (Lh Lv l) · ρc p (T m −T 0 ). It can easily be ∗ verified that the ratio Ẽ0∗ = Emin /h̃∗ exactly represents the ratio of these two energies, i.e. of the energy absorbed in the entire powder bed and the energy required to heat and melt the entire powder bed. The statements made in [146] on the basis of the alternative dimensionless hatch spacing h∗ = 1 are equivalent, when considering the relation h̃∗ = 0.5h∗ or Ẽ0∗ = 2E0∗ . Consequently, the constants E0∗ , denoted as isopleths and plotted in Figure 35, have to be multiplied by a factor of 2 in order to yield the practically relevant ratio of absorbed energy to required energy for heating and melting the entire powder bed. For completeness, the experimental raw data from the literature underlying the data points in Figure 35 is summarized in Table 1 focusing mostly on Ti-6Al-4V and 316L SS. Based on the failure mechanisms reported in the literature and included into Figure 35, excessively high energy deposition may result in overheating, cracking and part swelling whereas insufficient energy deposition may yield an incomplete melting and incorporation of the powder bed and consequently undesired porosity and void formation. From these graphical results, Thomas et al. deduce that across different processes and alloys, SLM/EBM is most effective in the realm of 2 < E0∗ < 8. Thus, the amount of effectively absorbed energy per powder layer was typically chosen to be at least by a factor of four higher than the minimally required energy for heating and melting the powder of one layer (since E0∗ > 2 is equivalent to Ẽ0∗ > 4; see also the discussion in the paragraph above). The factor of four might become reasonable when considering that substantial parts of neighboring tracks and underlying layers have to be remelted in order to provide good adhesion as well as additional effects such as thermal losses due to emission, 44 Heat source (type) Processing parameters AM platform Alloy system Bed Temp., T 0 Power, P [K] [W] Thomas et. al. Electron Beam Arcam A2 Juechter et al. Electron Beam Arcam S12 Vranken et al. Laser (SMYb:YAG) In-house LM-Q Qui et al. Laser Concept Laser Xu et al. Laser SLM 250 HL Kamath et al. Laser Concept Laser M2 Ziolkowski Laser et al. SLM Realizer II Unpublished Laser Data Cooper et al. Laser In-house Boswell et al. Laser SLM 280HL Carter et al. Laser Concept Laser M2 Boswell et al. Laser EOS M270 Brif et al. Laser AM 125 Velocity, V Layer height, l [m/s] [m] Hatch spacing, h [m] Beam radius, rB [m] Ti–6Al–4V 923 600 – – – 150x10−6 Ti–6Al–4V 923 0.2 − 6.4 50x10−6 100x10−6 150x10−6 Ti–6Al–4V 298 60− 1400 250 1.6 30x10−6 60x10−6 52x10−6 Ti–6Al–4V 298 0.8 − 1.5 20x10−6 75x10−6 75x10−6 Ti–6Al–4V 498 Ti–6Al–4V 298 150− 200 175− 375 150− 400 0.7− 1.029 0.5 − 2.5 30− 90x10−6 30x10−6 120− 180x10−6 112x10−6 70− 120x10−6 27x10−6 316L SS 298 97 0.2 50x10−6 125x10−6 100x10−6 316L SS 298 – – – – – Ti–6Al–4V 298 0.1 − 0.12 100x10−6 500x10−6 850x10−6 Ti–6Al–4V 298 800− 1000 – – – – – −6 Ti–6Al–4V 298 100− 200 0.4 − 2.0 20x10 Ti–6Al–4V 298 – – – Ti–6Al–4V 298 200 0.33 − 0.60 20 − 50x10 −6 −6 75x10 75x10−6 – – 90x10 −6 −6 25x10 Thermophysical Properties Al-Bermani et al. Al-Bermani et al. ASM International ASM International ASM International ASM International ASM International Al-Bermani et al. ASM International ASM International Mukai et al. Avala et al. Mukai et al. Avala et al. Brif et al. Table 1: Literature survey of SLM / EBM process parameters required for normalization and creation of process maps, [146]. Figure 36: Porosity in SLM of Ti-6Al-4V vs. energy density for different laser powers: (a) P = 160W, (b) P = 120W, (c) P = 80W, (d) P = 40W, [50]. 45 convection and evaporation on the powder bed surface and heat conduction across the build platform. Thomas et al. went on to conduct experiments using EBM of Ti-6Al-4V to develop a material specific processing map. They found that high quality (absence of defect voids) is obtained at highest E0∗ of the study trade space. This region sacrifices production speed with slow scanning speed and more scans per a given part planar area (low h∗ and v∗ ). At higher speeds and wider hatch spacings the Vickers hardness number (VHN) has been observed to increase (as consequence of higher cooling rates and the possible creation of non-equilibrium phases) and structural integrity to be compromised by the presence of undesirable microstructural features such as lack of fusion defects. While the isopleth lines of constant E ∗ might provide a rough process window as well as an evaluation of the process energy efficiency, this approach does not allow precise assessment of the stability and surface quality of individual melt tracks. For example, it is obvious that in practice an excessively too high energy deposition per melt track cannot be compensated by a larger hatch spacing, which, however, would be suggested by the definition of E0∗ . ∗ Similar to the discussion of Ẽ0∗ above, the dimensionless energy Emin , represented by vertical lines in Figure 35, can be shown to represent the ratio of absorbed energy per melt track to the energy required for heating and melting a track of ∗ cross-section 2rB · l. However, even a comparison of results related to constant values of Emin exhibit difficulties. Considering for example the mechanisms of powder bed radiation transfer discussed in Section 2.1, it becomes obvious that an excessively high beam power can typically not be compensated by a very high powder layer thickness. Due to the typical radiation transfer depth determined by the powder layer structure and the very low thermal conductivity of the powder, it is rather likely that the intended compensation results in excessive evaporation at the powder bed surface in combination with unmolten powder at the bottom of the too thick powder layer. For specific materials and process parameter sets, direct optimization of the scan parameters and their relation e.g. to porosity can readily be performed. Gong et al. [50] varied the energy density for SLM-processing Ti-6Al-4V via laser beam power and velocity, however, at constant powder bed depth and density. In Figure 36, each subplot representing a constant laser power shows the relation between incident energy density and part quality using porosity as the metric of quality. Accordingly, Gong et al. identified four broad regimes of process parameters through comparing porosity resulting from differing combinations of beam speed and beam power. These regimes are shown in Figure 37 and denoted as I, II, III and OH. Region I is described as the ”processing window” producing fully dense SLM parts. Process parameters in regions II and III produce porosity defects, due to underheating and incomplete melting of powder particles, or due to overheating, evaporation and gas cavitation, respectively. The region labeled as OH, referring to the ”overheating” regime, induces such severe accumulated thermal strains and part distortions so as to inhibit operation of the SLM recoater. A similar approach has been taken in [77] for 316L stainless steel. Figure 37: Processing regimes for SLM of Ti-6Al-4V depending on laser power and scanning velocity: (I) ”processing window”, (II) ”incomplete melting”, (III) ”overheating”, (OH) ”severe overheating”, [50]. 46 Figure 38: Visualization of physical mechanisms associated with melt pool dynamics and keyhole formation, [105]. Figure 39: Results of x-ray micro computed tomography of Ti-6Al-4V samples processed by SLM, ordered by increasing effective track penetration depth (red arrows). The scans show the transition from large and numerous pores generated by keyhole formation to highly dense parts to highly porous samples generated by lack of fusion. It has to be noted that the inhomogeneity observed in these scans stems from the use of a different set of machine processing parameters (”top skin”) for high part surface quality, [31]. 47 Commonly, porosities < 0.5% are desired for SLM processes. A high residual porosity may be regarded as a significant defect of SLM-manufactured parts for many applications because pores facilitate crack formation and propagation, which is particularly deleterious to strain-to-failure and cyclic fatigue life. In this context, the formation of gas-induced pores typically occurs at too high energy densities and can be explained by the so-called keyhole mechanism as illustrated in Figure 38. It is generally believed that the recoil pressure resulting from excessive heating and attended evaporation is able to form a deep depression of the melt pool directly below the laser beam (see also the discussion in Section 2.3). With increasing depth of this depression, the incident laser radiation is reflected on the sides of the depression, leading to a more concentrated energy input and higher temperatures at the bottom of the depression, further evaporation at this location and consequently to the creation of even deeper, narrow corridors. Typically, this keyhole mechanism can lead to considerable penetration depths of the laser beam into the underlying material. Due to the more focused energy source, SLM commonly results in deeper keyholes than EBM, whereas EBM yields larger melt pool sizes [105]. When the laser beam moves on, also the recoil pressure as driving force diminishes and, at some point, the keyhole will collapse under the action of surface tension and gravity (which, however, is typically negligible for the relevant range of length scales). The incomplete collapse of the vapor cavity at the bottom of the keyhole typically leaves voids with a characteristic spherical shape. Recent micro x-ray computed tomography (µSXCT) studies have clearly shown the relationship between SLM processing energy density and the resulting porosity in the as-built SLM part. Cunningham et al. [31] performed µSXCT using the Advanced Photon Source at Argonne National Laboratory to image Ti-6Al-4V samples fabricated by SLM as shown in Figure 39. The 12 samples studied were processed on an EOS M290 machine with varying laser powers (100 - 370 W), scan speeds (400 - 1200 mm/s), and hatch spacings (40 - 240 µm) to achieve widely varying melt-pool penetration overlap depths d (0 - 128 µm as calculated in the idealized geometry of two hemispherical melt tracks separated by the hatch spacing). As expected, lack-of-fusion defects were effectively eliminated for layer thicknesses t less than d. These results correlate well with the density and porosity results shown above by Gong et al. [50]. 3.4. Microstructure evolution during the SLM process In this section, we summarize representative studies of the process-microstructure relationships of the materials Ti6Al-4V, Inconel 718 as well as stainless steels. On the one hand, these materials are commonly studied in the context of SLM. On the other hand, they exhibit importantly different thermomechanical behavior. Figure 40: (a) Temperature distribution profile in the xz-plane for a single Ti-6Al-4V powder layer on top of a Ti-6Al-4V substrate derived by means of a macroscopic SLM simulation model. (b) SEM micrograph (side view) of a Ti-6Al-4V reference sample processed with the same scan pattern, i.e. from right to left, indicating grain growth along the direction of highest temperature gradients, [145]. 48 Thijs et al. [145] studied the microstructural evolution of Ti-6Al-4V during selective laser melting. Due to the high cooling rates and high-temperature gradients during the SLM process, an acicular martensitic phase, i.e. a hexagonal close-packed structure, was revealed in as-melted Ti-6Al-4V layers. They also found that the grains grew epitaxially, i.e. the material from a crystalline overlayer has a well-defined orientation with respect to the crystal structure of the underlying substrate, and observed elongated grains in the side and front views of the samples. It is interesting to note that the direction of the elongated grains depends on the local heat transfer condition, i.e. the grains are oriented towards the melt pool as shown in Figure 40. Based on a simple thermal model, they indicated that the grain direction is parallel to the local conductive heat transfer given by the highest thermal gradient. Since the thermal field can be controlled by the scan strategy, also the resulting grain orientation can be influenced. Another interesting phenomenon occurring due to fast solidification during the SLM process is the segregation of aluminum. It is distinguishable as dark bands, which are believed to represent an intermetallic Ti-3Al phase. Further, when a higher energy density (e.g. due to higher laser power) is applied to the material, the volume of precipitates, i.e. second phases, will increase as consequence of lower cooling rates and longer diffusion times for the formation of second phases. Although the microstructure of Ti-6Al-4V is dominated by martensite due to rapid cooling rates, post heat treatment can transform the alpha prime martensite into alpha and beta phases. Owing to the prevalent thermal profiles, in particular through cycled heating, the SLM process can allow the formation of alpha and beta phases, and may form a graded microstructure in build direction. Recently, Xu et al. [160] observed the decomposition of alpha prime martensite into ultra-fine lamellar (alpha and beta) phases from the bottom to the top of a SLM-fabricated Ti-6Al-4V sample, which, in turn, might result in a decreasing material strength from the bottom to the top. Figure 41: Grain morphology of Inconel 718 samples built by EBM with (a) 10 mA and (b) 20 mA beam settings, [125]. Raghavan et al. [125] analyzed the microstructures resulting from EBM-processing of Inconel 718. As shown in Figure 41, the sample processed by means of a beam current of 10mA exhibits directional grains (columnar grains) along the build direction, while the one with an employed beam current of 20mA shows equiaxed grains. The reason for this observation is that a higher beam current induces a higher sample temperature (acting as local preheat temperature in the unmolten material regions), and therefore, the thermal gradient is lower, consequently resulting in equiaxed grains. Dehoff et al. [37] demonstrate how to use computer-aided design to change the local crystal orientation. Three scanning strategies, relying on different beam powers and velocities, were applied in order to vary thermal gradients, cooling rates and solidification front velocities, which eventually allows to achieve tailored grain orientations. The working principle of this strategy has been proven by ”writing” prescribed letters in a block of Inconel 718 such that 49 these letters are represented by misoriented equiaxed grains while the surrounding bulk materials exhibits oriented [001] columnar grains. This methodology provides a fundamentally new design tool to tailor microstructures. Also several classes of stainless steel alloys are of practical interest for SLM processing. Most steel parts manufactured by means of SLM satisfy typical requirements of general-purpose applications, especially in terms of high-strength and high-hardness, for example for mould and tool applications. Steel grades available for AM are typically austenitic stainless steels such as 304L and 316L. They commonly exhibit a completely austenitic microstructure after the AM process characterized by elongated and textured grains. For example, in 304L stainless steel, a typical microstructure with elongated grains in build direction has been observed. The precipitation of chromium carbides at grain boundaries, which would be observable for lower cooling rates prevalent in traditional processes such as casting, can typically be avoided by means of SLM. Similar observations have also been made for SLM-processed 316L stainless steel. Moreover, in [46], a gradient microstructure has been observed for SLM-manufactured 316L parts. Accordingly, an increasing distance from the build platform resulted in coarser microstructures due to lower cooling rates. 4. Future directions and practical implementation Indisputably, modeling and simulation approaches will play a key role in enabling SLM and EBM powder bed processes to build highly accurate complex parts, and to achieve stringent quality specifications. Current implementations of macroscopic, mesoscopic, and microscopic models have shed further light on the underlying physical mechanisms of SLM and relations between the process parameters and the part characteristics, e.g. in terms of residual porosity, residual stresses or metallurgical microstructure, resulting on these different length scales. The long-term goal must be to combine these three model classes, based on a proper exchange of information, to yield an integrated modeling scheme that is capable of predicting final part characteristics on all of these three length scales (see also Figure 27). Clearly, computational efficiency and robustness, not only of the sub-models but also of the coupling schemes, will play a key role in this regard. With such an integrated process model at hand, not only the prediction of process output for given input data, but also the inverse problem of determining locally optimal process parameters in order to optimize the resulting part characteristics, measured by a properly defined objective function and realized via methods of numerical optimization, is of highest practical interest (see e.g. [17] for an overview on inverse analysis schemes). The combination of integrated simulation and optimization methodologies with the capability of SLM to locally control temperature gradients and cooling rates opens the door for improved dimensional quality as well as engineered microstructures. This includes, but is not limited to, interest in metallurgical microstructure designs that are optimal with respect to prescribed requirements such as: inhomogeneous and anisotropic distributions of material strength and ductility, part density and porosity, material dissipation resulting in mechanical or acoustic damping, thermal or electrical conductivity, or internal flow/cooling capacity. Of course, the predictive accuracy of the respective modeling approaches is strongly related to their degree of abstraction. In this context, correlation of in situ metrology data to models is essential to build both accurate and computationally efficient representations of the process physics with a proper degree of abstraction. Furthermore, by employing real-time in situ measurements as a means of fault indentification, close-loop process control can be achieved [43]. Eventually, a comparison of real-time in situ data for relevant physical fields (e.g. the temperature field) with the desired/optimal evolutions of these fields as predicted by numerical simulation opens the door for a proper manipulation of the process input parameters according to the principles of control theory. Both, simulations and experiments must be pursued to understand the fundamental limits to SLM processes. For instance, the process and machine parameters, as well as the heating and cooling rates resulting from the laser-material interaction and boundary conditions of the melt pool, impose a range of accessible heating and cooling rates which in turn limit the potential for microstructure and stress control. Also, the ability to perform SLM with powders having a broader candidate set of shape and size distributions is of practical interest, to increase the range of processible materials and geometrical resolutions and to access powder production techniques that have lower intrinsic cost. Currently, two main approaches of powder manufacturing are typically applied: Water-atomized powder is favorable in terms of lower production costs. However, water-atomized powder has irregular particle shape, and therefore typically leads to 50 powder with poor flow and packing characteristics. Gas-atomized powder is more costly, yet is advantageous in producing spherical particles with higher purity [73]. Consideration of powder purity and transient composition changes is critical, because oxidation on powder and substrate surfaces might considerably decrease the wetting behavior of the liquid melt, possibly leading to melt pool balling and reduced surface quality, defects such as pores and inclusions, and reduced layer-to-layer adhesion, which might induce delamination [36]. Furthermore, in practice, SLM powder must be kept dry since evaporation of water contents during the melt process induces undesirable recoil pressure and fosters surface oxidation. The water content is commonly reduced as far as possible by combined approaches of preheating and a controlled inert gas flow flushing out possible sources of contamination from the build chamber. A further paramount need is to assist quality control and certification of the resulting parts and materials. First, the complexity of the SLM process and the underlying physical phenomena leads to a higher sensitivity of part quality with respect to process parameter variations than is the case for rather traditional manufacturing technologies such as casting or forging. Second, in combination with the limited practical experience in this comparatively new technology, the frequency of defective parts is considerably higher. Third, the types of defects as well as the resulting mechanisms of failure are specific for the physical phenomena governing the SLM process, and consequently, require sophisticated surface and bulk metrology such as X-ray tomography. In this context, in-situ measurement and recording of certain physical fields, e.g. the temperature field, might be useful in locating different types of defects within SLM parts. The success of SLM as manufacturing technology is of course also strongly determined by economic aspects. Currently, the limited understanding of underlying physical phenomena, leads to strong restrictions on the SLM process in form of tight process windows. For example, a maximum scan velocity may not be exceeded in order to avoid melt pool instabilities resulting in high surface roughness and possible porosity of the final part. Further, excessive energy input as consequence of high power densities can lead to overheating and decreased surface quality, high residual stresses, and crack propagation. However, for high process throughput, increased scan velocities, beam spot sizes and power densities would be desirable. In this regard, the study of novel scanning strategies, laser configurations, and beam shaping approaches seem to be very promising in order to find regimes of stable processing yet higher production throughput. A rough estimate for the lower energy bound of SLM per layer is given by the amount of energy required to heat and melt the powder material of one layer due to its heat capacity. The amount of energy required for the actual manufacturing process is considerably higher because substantial parts of neighboring tracks and underlying layers have to be remelted in order to provide good adhesion, and because of various thermal losses, e.g. due to emission, convection and evaporation on the powder bed surface as well as heat conduction across the build platform. Energy consumption is insignificant to the overall cost of SLM at present, yet will become critical as the other cost drivers, namely material and equipment cost, improve. Besides the pure energy balance, the avoidance of evaporation is also essential in terms of material savings, melt track quality and avoidance of machine and instrument contamination due to condensation of metallic vapor, which is especially undesirable when remelted during the processing of different alloys in the same machine. Conductive and emissive losses have to be minimized via proper thermal insulation and shielding of build platform and machine casing, while research in the field of powder material production may yield improved grain surface properties in terms of heat conduction and radiation absorption. All in all, it is certain that SLM and related AM technologies present a rich set of multidisciplinary scientific and practical challenges, along with opportunity to transform the design of advanced products and manufacturing systems. Acknowledgements Financial support for preparation of this article was provided by the German Academic Exchange Service (DAAD) to C. Meier; Honeywell Federal Manufacturing & Technologies to R. Penny and A.J. Hart; the Swiss National Science Foundation Early Postdoctoral Mobility Fellowship to Y. Zou (Grant: P2EZP2 165278), and United States Navy and the Engineering Duty Officer program to J.S. Gibbs. This work has been partially funded by Honeywell Federal Manufacturing & Technologies, LLC which manages and operates the Department of Energy’s Kansas City National Security Campus under Contract No. DE-NA-0002839. The United States Government retains and the publisher, by accepting the article for publication, acknowledges that the United States Government retains a nonexclusive, paid up, irrevocable, world-wide license to publish or reproduce the published form of this manuscript, or allow others to do so, for the United States Government purposes. 51 Nomenclature Abbreviations AM CA CAD DEM DIC EBM EBW FDM FEM FS I FV M HAZ LBW LWIR NIR PF M RT E S LM S LS V HN VOF XFEM Additive Manufacturing Cellular Automation Comuter Aided Design Discrete Element Method Digital Image Correlation Electron Beam Melting Electron Beam Welding Finite Difference Method Finite Element Method Fluid Structure Interaction Finite Volume Method Heat Affected Zone Laser Beam Welding Long Wave Infrared Near Infrared Phase Field Method Radiation Transport Equation Selective Laser Melting Selective Laser Sintering Vickers Hardness Number Volume of Fluid Method Extended Finite Element Method SLM process L/l R P V h rB (.)∗ Lv Lh A Powder bed thickness of process material Radius of spherical particle Laser power Laser velocity Hatch spacing Laser beam radius Associated dimensionless quantity Powder bed dimension in scan direction Powder bed dimension in hatch direction Factor describing fraction of effectively absorbed laser energy Mechanical problem σ Cauchy stress tensor b Vector of volume forces u Displacement vector  Solid strain tensor ˜ Rate-of-deformation tensor γ Surface tension θ0 Wetting angle κI Interface curvature p Pressure in melt pool pre f Pressure of ambient gas λev Evaporation energy per particle As Sticking coefficient Rg Gas constant M Molar mass 52 Thermal problem T T re f T0 Tm Tb Tl Ts Ω x I σ κ β ω Sc kS B q qr qe qv qλ qc us αS αP n θ k w e n λ/k λe a α α0 Hm Φ n rc x ρ cp t v k Ω Γ c Temperature of processes material Temperature of ambient atmosphere Initial temperature Melting temperature Boiling temperature Liquidus temperature Solidus temperature Unit vector describing direction of radiation transfer Spatial position vector Radiation intensity Scattering coefficient Absorption coefficient Extinction coefficient Albedo Scattering phase function Stefan-Boltzmann constant General heat flux density Radiation heat flux density Emissive heat flux density Evaporation heat flux density Conductive heat flux density Convective heat flux density Incident energy density Perpendicular component of polarized absorptivity Parallel component of polarized absorptivity Complex refraction index Angle of radiation incidence Direction of electro-magnetic wave propagation wave length of electro-magnetic wave Direction of electric field Surface normal vector / general unit vector Thermal conductivity Effective thermal conductivity Thermal diffusivity General absorptivity Flat surface absorptivity Latent heat of melting Powder density Coordination number Radius of spherical contact area Dimensionless contact size Density Specific heat at constant pressure Time Velocity vector Thermal conductivity tensor Problem domain Problem boundary Convection coefficient 53  φ c Ψ Π Π̃ cφ Radiation emissivity Phase field Concentration Rotation vector describing grain orientation Free energy Volume-specific free energy Boundary layer thickness parameter in phase field method References [1] R. Ammer, M. Markl, U. Ljungblad, C. Körner, and U. Rüde. Simulating fast electron beam melting with a parallel thermal free surface lattice Boltzmann method. Computers & Mathematics with Applications, 67(2):318–330, 2014. [2] S. I. Anisimov and V. A. Khokhlov. Instabilities in laser-matter interaction. CRC Press, Boca Raton, 1995. [3] M. Apel, J. Eiken, and U. Hecht. Phase field models for heterogeneous nucleation: Application to inoculation in alpha-solidifying Ti-Al-B alloys. European Physical Journal: Special Topics, 223(3):545–558, 2014. [4] C. Argento and D. Bouvard. A Ray Tracing Method For Evaluating The Radiative Heat Transfer In Porous Media. International journal of heat and mass transfer, 39(15):3175–3180, 1996. [5] T. T. Arif and R. S. Qin. A Phase-Field Model for the Formation of Martensite and Bainite. Advanced Materials Research, 922(1):31–36, 2014. [6] F. Armero and J. C. Simo. A new unconditionally stable fractional step method for non-linear coupled thermomechanical problems. International Journal for Numerical Methods in Engineering, 35(4):737–766, 1992. [7] E. Attar and C. Körner. Lattice Boltzmann method for dynamic wetting problems. Journal of Colloid and Interface Science, 335(1):84–93, 2009. [8] E. Attar and C. Körner. Lattice Boltzmann model for thermal free surface flows with liquid-solid phase transition. International Journal of Heat and Fluid Flow, 32(1):156–163, 2011. [9] M. Bachmann, V. Avilov, A. Gumenyuk, and M. Rethmeier. Numerical assessment and experimental verification of the influence of the Hartmann effect in laser beam welding processes by steady magnetic fields. International Journal of Thermal Sciences, 101:24–34, 2016. [10] D. Baillis. Thermal radiation properties of dispersed media: theoretical prediction and experimental characterization. Journal of Quantitative Spectroscopy and Radiative Transfer, 67(5):327–363, 2000. [11] A. Bauereiß, T. Scharowsky, and C. Körner. Defect generation and propagation mechanism during additive manufacturing by selective beam melting. Journal of Materials Processing Technology, 214(11):2497–2504, 2014. [12] W. J. Boettinger, S. R. Coriell, a. L. Greer, a. Karma, W. Kurz, M. Rappaz, and R. Trivedi. Solidification microstructures: recent developments, future directions. Acta Materialia, 48(1):43–70, 2000. [13] C. D. Boley, S. A. Khairallah, and A. M. Rubenchik. Calculation of laser absorption by metal powders in additive manufacturing. Applied optics, 54(9):2477–2482, 2015. [14] C. D. Boley, S. C. Mitchell, A. M. Rubenchik, and S. S. Q. Wu. Metal powder absorptivity: modeling and experiment. Applied Optics, 55(23):6496–6500, 2016. [15] D. L. Bourell. Perspectives on Additive Manufacturing. Annual Review of Materials Research, 46:1–18, 2016. [16] D. A. G. Bruggeman. Berechnung verschiedener physikalischer Konstanten von heterogenen Substanzen. I. Dielektrizitätskonstanten und Leitfähigkeiten der Mischkörper aus isotropen Substanzen. Annalen der Physik, 416(7):636, 1935. [17] M. Calvello and R. J. Finno. Selecting parameters to optimize in model calibration by inverse analysis. Computers and Geotechnics, 31(5):411–425, 2004. [18] B. G. Carlson and K. D. Lathrop. Transport theory: the method of discrete ordinates. Los Alamos Scientific Laboratory of the University of California, 1965. [19] P. Causin, J. F. Gerbeau, and F. Nobile. Added-mass effect in the design of partitioned algorithms for fluid-structure problems. Computer Methods in Applied Mechanics and Engineering, 194(42-44):4506–4527, 2005. [20] G. B. M. Cervera and G. Lombera. Numerical prediction of temperature and density distributions in selective laser sintering processes. Rapid Prototyping, 5(1):21–26, 1999. [21] C. K. Chan and C. L. Tien. Radiative transfer in packed spheres. Journal of Heat Transfer, 96(1):52–58, 1974. [22] S. Chandrasekhar. Radiative transfer. Courier Corporation, 2013. [23] B. Chang, C. Allen, J. Blackburn, P. Hilton, and D. Du. Fluid Flow Characteristics and Porosity Behavior in Full Penetration Laser Welding of a Titanium Alloy. Metallurgical and Materials Transactions B: Process Metallurgy and Materials Processing Science, 46(2):906–918, 2015. [24] J. C. Chen and S. W. Churchill. Radiant heat transfer in packed beds. AIChE Journal, 9(1):35–41, 1963. [25] L.-Q. Chen. Phase -Field Models for Microstructure Evolution. Annual Review of Materials Research, 32(1):113–140, 2002. [26] Y. X. Chen, X. J. Wang, and S. B. Chen. The effect of electron beam energy density on temperature field for electron beam melting. Advanced Materials Research, 900:631–638, 2014. [27] T. Childs. Selective laser sintering (melting) of stainless and tool steel powders: experiments and modelling. Proceedings of the Institution of Mechanical Engineers, Part B: Journal of Engineering Manufacture, 219(4):339–357, 2005. [28] S. Clijsters, T. Craeghs, S. Buls, K. Kempen, and J.-P. Kruth. In situ quality control of the selective laser melting process using a high-speed, real-time melt pool monitoring system. International Journal of Advanced Manufacturing Technology, 75(5):1089–1101, 2014. [29] T. Craeghs, F. Bechmann, S. Berumen, and J. P. Kruth. Feedback control of Layerwise Laser Melting using optical sensors. Physics Procedia, 5(PART 2):505–514, 2010. 54 [30] T. Craeghs, S. Clijsters, E. Yasa, F. Bechmann, S. Berumen, and J. P. Kruth. Determination of geometrical factors in Layerwise Laser Melting using optical process monitoring. Optics and Lasers in Engineering, 49(12):1440–1446, 2011. [31] R. Cunningham, S. P. Narra, C. Montgomery, J. Beuth, and A. D. Rollett. Synchrotron-Based X-ray Microtomography Characterization of the Effect of Processing Variables on Porosity Formation in Laser Power-Bed Additive Manufacturing of Ti-6Al-4V. Jom, 69(3):479–484, 2017. [32] K. Dai and L. Shaw. Thermal and stress modeling of multi-material laser processing. Acta Materialia, 49(20):4171–4181, 2001. [33] K. Dai and L. Shaw. Thermal and mechanical finite element modeling of laser forming from metal and ceramic powders. Acta Materialia, 52(1):69–80, 2004. [34] K. Dai and L. Shaw. Finite element analysis of the effect of volume shrinkage during laser densification. Acta materialia, 53(18):4743–4754, 2005. [35] K. Dai and L. Shaw. Parametric studies of multi-material laser densification. Materials Science and Engineering A, 430(1-2):221–229, 2006. [36] S. Das. Physical Aspects of Process Control in Selective Laser Sintering of Metals. Advanced Engineering Materials, 5(10):701–711, 2003. [37] R. R. Dehoff, M. M. Kirka, W. J. Sames, H. Bilheux, A. S. Tremsin, L. E. Lowe, and S. S. Babu. Site specific control of crystallographic grain orientation through electron beam additive manufacturing. Materials Science and Technology, 31(8):931–938, 2015. [38] E. R. Denlinger, J. C. Heigel, and P. Michaleris. Residual stress and distortion modeling of electron beam direct manufacturing Ti-6Al-4V. Proceedings of the Institution of Mechanical Engineers, Part B: Journal of Engineering Manufacture, 229(10):1803–1813, 2015. [39] E. R. Denlinger, J. Irwin, and P. Michaleris. Thermomechanical Modeling of Additive Manufacturing Large Parts. Journal of Manufacturing Science and Engineering, 136(6):61007, 2014. [40] L. Dong, A. Makradi, S. Ahzi, and Y. Remond. Three-dimensional transient finite element analysis of the selective laser sintering process. Journal of Materials Processing Technology, 209(2):700–706, 2009. [41] M. R. Dorr, J.-L. Fattebert, M. E. Wickett, J. F. Belak, and P. E. A. Turchi. A numerical algorithm for the solution of a phase-field model of polycrystalline materials. Journal of Computational Physics, 229(3):626–641, 2010. [42] B. L. Drolen and C. L. Tienf. Independent and dependent scattering in packed-sphere systems. Journal of Thermophysics and Heat Transfer, 1(1):63–68, 1987. [43] S. K. Everton, M. Hirsch, P. Stravroulakis, R. K. Leach, and A. T. Clare. Review of in-situ process monitoring and in-situ metrology for metal additive manufacturing. Materials {&} Design, 95:431–445, 2016. [44] C. Förster, W. A. Wall, and E. Ramm. Artificial added mass instabilities in sequential staggered coupling of nonlinear structures and incompressible viscous flows. Computer Methods in Applied Mechanics and Engineering, 196(7):1278–1293, 2007. [45] A. Ganser, J. Pieper, S. Liebl, and M. Zaeh. Numerical Simulation of the Thermal Efficiency During Laser Deep Penetration Welding. Physics Procedia, 83:1377–1386, 2016. [46] M. Garibaldi, I. Ashcroft, M. Simonelli, and R. Hague. Metallurgy of high-silicon steel parts produced using Selective Laser Melting. Acta Materialia, 110:207–216, 2016. [47] M. Geiger, K. H. Leitz, H. Koch, and A. Otto. A 3D transient model of keyhole and melt pool dynamics in laser beam welding applied to the joining of zinc coated sheets. Production Engineering, 3(2):127–136, 2009. [48] I. Gibson, D. W. Rosen, and B. Stucker. Additive manufacturing technologies, volume 238. Springer, 2010. [49] M. E. Glicksman. Principles of solidification: an introduction to modern casting and crystal growth concepts. Springer Science {&} Business Media, 2010. [50] H. Gong, K. Rafi, H. Gu, T. Starr, and B. Stucker. Analysis of defect generation in Ti-6Al-4V parts made using powder bed fusion additive manufacturing processes. Additive Manufacturing, 1:87–98, 2014. [51] X. Gong and K. Chou. Phase-field modeling of microstructure evolution in electron beam additive manufacturing. JOM, 67(5):1176–1182, 2015. [52] X. Gong, J. Lydon, K. Cooper, and K. Chou. Beam speed effects on Ti-6Al-4V microstructures in electron beam additive manufacturing. Journal of Materials Research, 29(17):1951–1959, 2014. [53] D. Gu. Laser additive manufacturing of high-performance materials. Springer, 2015. [54] Z. Guo, J. Mi, S. Xiong, and P. S. Grant. Phase field simulation of binary alloy dendrite growth under thermal- and forced-flow fields: An implementation of the parallel-multigrid approach. Metallurgical and Materials Transactions B: Process Metallurgy and Materials Processing Science, 44(4):924–937, 2013. [55] F.-J. Gürtler, M. Karg, K.-H. Leitz, and M. Schmidt. Simulation of Laser Beam Melting of Steel Powders using the Three-Dimensional Volume of Fluid Method. Physics Procedia, 41:881–886, 2013. [56] A. V. Gusarov. Homogenization of radiation transfer in two-phase media with irregular phase boundaries. Physical Review B - Condensed Matter and Materials Physics, 77(14):1–14, 2008. [57] A. V. Gusarov, E. H. Bentefour, M. Rombouts, L. Froyen, C. Glorieux, and J.-P. Kruth. Normal-directional and normal-hemispherical reflectances of micron-and submicron-sized powder beds at 633 and 790nm. Journal of applied physics, 99(11):113528, 2006. [58] A. V. Gusarov and E. P. Kovalev. Model of thermal conductivity in powder beds. Physical Review B - Condensed Matter and Materials Physics, 80(2):16–20, 2009. [59] A. V. Gusarov and J. P. Kruth. Modelling of radiation transfer in metallic powders at laser treatment. International Journal of Heat and Mass Transfer, 48(16):3423–3434, 2005. [60] A. V. Gusarov, T. Laoui, L. Froyen, and V. I. Titov. Contact thermal conductivity of a powder bed in selective laser sintering. International Journal of Heat and Mass Transfer, 46(6):1103–1109, 2003. [61] A. V. Gusarov, M. Pavlov, and I. Smurov. Residual stresses at laser surface remelting and additive manufacturing. Physics Procedia, 12(PART 1):248–254, 2011. [62] A. V. Gusarov and I. Smurov. Modeling the interaction of laser radiation with powder bed at selective laser melting. Physics Procedia, 5(PART 2):381–394, 2010. [63] A. V. Gusarov, I. Yadroitsev, P. Bertrand, and I. Smurov. Heat transfer modelling and stability analysis of selective laser melting. Applied 55 Surface Science, 254(4):975–979, 2007. [64] M. Haag, H. Hügel, C. E. Albright, and S. Ramasamy. CO2 laser light absorption characteristics of metal powders. Journal of Applied Physics, 79(8):3835–3841, 1996. [65] P. Hanzl, M. Zetek, T. Bakša, and T. Kroupa. The influence of processing parameters in The mechanical properties of SLM parts. Energy Procedia, 100(C):1405–1413, 2015. [66] D. Havermann, J. Mathew, W. N. MacPherson, D. P. Hand, and R. R. J. Maier. Measuring residual stresses in metallic components manufactured with fibre bragg gratings embedded by selective laser melting. In International Conference on Optical Fibre Sensors (OFS24), pages 96340T–96340T. International Society for Optics and Photonics, 2015. [67] R. J. Hebert. Viewpoint: metallurgical aspects of powder bed metal additive manufacturing. Journal of Materials Science, 51(3):1165–1175, 2016. [68] D. Herzog, V. Seyda, E. Wycisk, and C. Emmelmann. Additive manufacturing of metals. Acta Materialia, 117:371–392, 2016. [69] N. Hodge, R. Ferencz, and R. Vignes. Experimental comparison of residual stresses for a thermomechanical model for the simulation of selective laser melting. Additive Manufacturing, 12:159–168, 2016. [70] N. E. Hodge, R. M. Ferencz, and J. M. Solberg. Implementation of a thermomechanical model for the simulation of selective laser melting. Computational Mechanics, 54(1):33–51, 2014. [71] A. Hussein, L. Hao, C. Yan, and R. Everson. Finite element simulation of the temperature and stress fields in single layers built withoutsupport in selective laser melting. Materials {&} Design, 52:638–647, 2013. [72] J. C. Ion, H. R. Shercliff, and M. F. Ashby. Diagrams for laser materials processing. Acta Metallurgica Et Materialia, 40(7):1539–1551, 1992. [73] H. Irrinki, M. Dexter, B. Barmore, R. Enneti, S. Pasebani, S. Badwe, J. Stitzel, R. Malhotra, and S. V. Atre. Effects of Powder Attributes and Laser Powder Bed Fusion (L-PBF) Process Conditions on the Densification and Mechanical Properties of 17-4 PH Stainless Steel. JOM, 68(3):860–868, 2016. [74] M. Jamshidinia. Temperature distribution and fluid flow modeling of electron beam melting R (EBM). In ASME 2012 International Mechanical Engineering Congress and Exposition, pages 3089–3101, 2012. [75] M. Jamshidinia, F. Kong, and R. Kovacevic. The Coupled CFD-FEM Model of Electron Beam Melting(EBM). In ASME District F-Early Career Technical Conference, pages 163–171, Birmingham, AL, 2013. [76] P. D. Jones, D. G. McLeod, and D. E. Dorai-Raj. Correlation of measured and computed radiation intensity exiting a packed bed. Journal of heat transfer, 118(1):94–102, 1996. [77] C. Kamath, B. El-Dasher, G. F. Gallegos, W. E. King, and A. Sisto. Density of additively-manufactured, 316L SS parts using laser powderbed fusion at powers up to 400 W. International Journal of Advanced Manufacturing Technology, 74(1-4):65–78, 2014. [78] K. Kamiuto. Correlated radiative transfer in packed-sphere systems. Journal of Quantitative Spectroscopy and Radiative Transfer, 43(1):39– 43, 1990. [79] K. Kamiuto, M. Iwamoto, M. Sato, and T. Nishimura. Radiation-extinction coefficients of packed-sphere systems. Journal of Quantitative Spectroscopy and Radiative Transfer, 45(2):93–96, 1991. [80] K. Kempen, B. Vrancken, S. Buls, L. Thijs, J. Van Humbeeck, and J.-P. Kruth. Selective Laser Melting of Crack-Free High Density M2 High Speed Steel Parts by Baseplate Preheating. Journal of Manufacturing Science and Engineering, 136(6):61026, 2014. [81] S. A. Khairallah and A. Anderson. Mesoscopic simulation model of selective laser melting of stainless steel powder. Journal of Materials Processing Technology, 214(11):2627–2636, 2014. [82] S. A. Khairallah, A. Anderson, A. M. Rubenchik, J. Florando, S. Wu, and H. Lowdermilk. Simulation of the main physical processes in remote laser penetration with large laser spot size. AIP Advances, 5(4):047120, 2015. [83] S. A. Khairallah, A. T. Anderson, A. Rubenchik, and W. E. King. Laser powder-bed fusion additive manufacturing: Physics of complex melt flow and formation mechanisms of pores, spatter, and denudation zones. Acta Materialia, 108:36–45, 2016. [84] S.-W. Khoo, S. Karuppanan, and C.-S. Tan. A Review of Surface Deformation and Strain Measurement Using Two-Dimensional Digital Image Correlation. Metrology {&} Measurement Systems, 23(3):461–480, 2016. [85] H. Ki, J. Mazumder, and P. S. Mohanty. Modeling of laser keyhole welding: Part I. mathematical modeling, numerical methodology, role of recoil pressure, multiple reflections, and free surface evolution. Metallurgical and Materials Transactions A, 33(June):1817–1830, 2002. [86] H. Ki, J. Mazumder, and P. S. Mohanty. Modeling of laser keyhole welding: Part II. simulation of keyhole evolution, velocity, temperature profile, and experimental verification. Metallurgical and Materials Transactions A, 33(6):1831–1842, 2002. [87] W. King, A. T. Anderson, R. M. Ferencz, N. E. Hodge, C. Kamath, and S. A. Khairallah. Overview of modelling and simulation of metal powder bed fusion process at Lawrence Livermore National Laboratory. Materials Science and Technology, 31(8):957–968, 2015. [88] A. Klassen, A. Bauereiß, and C. Körner. Modelling of electron beam absorption in complex geometries. Journal of Physics D: Applied Physics, 47(6):065307, 2014. [89] R. Kobayashi. Modeling and numerical simulations of dendritic crystal growth. Physica D, 63:410–423, 1993. [90] S. Kolossov, E. Boillat, R. Glardon, P. Fischer, and M. Locher. 3D FE simulation for temperature evolution in the selective laser sintering process. International Journal of Machine Tools and Manufacture, 44(2):117–123, 2004. [91] C. Körner, E. Attar, and P. Heinl. Mesoscopic simulation of selective beam melting processes. Journal of Materials Processing Technology, 211(6):978–987, 2011. [92] C. Körner, A. Bauereiß, and E. Attar. Fundamental consolidation mechanisms during selective beam melting of powders. Modelling and Simulation in Materials Science and Engineering, 21(8):85011, 2013. [93] H. Krauss, C. Eschey, and M. F. Zaeh. Thermography for monitoring the selective laser melting process. Proc. of the 23rd Annual International Solid Freeform Fabrication Symposium, pages 999–1014, 2012. [94] H. Krauss, T. Zeugner, and M. F. Zaeh. Layerwise Monitoring of the Selective Laser Melting Process by Thermography. Physics Procedia, 56:64–71, 2014. [95] H. Krauss, T. Zeugner, and M. F. Zaeh. Thermographic Process Monitoring in Powderbed Based Additive Manufacturing. AIP Conference Proceedings, 1650(1):177–183, 2015. 56 [96] C. E. Krill III and L.-Q. Chen. Computer simulation of 3-D grain growth using a phase-field model. Acta Materialia, 50(12):3059–3075, 2002. [97] J. Kruth, P. Mercelis, J. V. Vaerenbergh, L. Froyen, and M. Rombouts. Binding mechanisms in selective laser sintering and selective laser melting. Rapid Prototyping Journal, 11(1):26–36, 2005. [98] J. P. Kruth, G. Levy, F. Klocke, and T. H. C. Childs. Consolidation phenomena in laser and powder-bed based layered manufacturing. CIRP Annals - Manufacturing Technology, 56(2):730–759, 2007. [99] W. Kurz and D. J. Fisher. Fundamentals of solidification. Trans Tech Publications, Switzerland, 1986. [100] L. D. Landau. Electrodynamics of Continuous Media 2nd edition, Landau and Lifshitz Course of Theoretical Physics, 1984. [101] Y. Lee. Simulation of Laser Additive Manufacturing and its Applications. PhD thesis, The Ohio State University, 2015. [102] G. N. Levy, R. Schindel, and J.-P. Kruth. Rapid manufacturing and rapid tooling with layer manufacturing (LM) technologies, state of the art and future perspectives. CIRP Annals-Manufacturing Technology, 52(2):589–609, 2003. [103] J. F. Li, L. Li, and F. H. Stott. A three-dimensional numerical model for a convection-diffusion phase change process during laser melting of ceramic materials. International Journal of Heat and Mass Transfer, 47(25):5523–5539, 2004. [104] H. Lipson and M. Kurman. Fabricated: The new world of 3D printing. John Wiley {&} Sons, 2013. [105] Y. J. Liu, S. J. Li, H. L. Wang, W. T. Hou, Y. L. Hao, R. Yang, T. B. Sercombe, and L. C. Zhang. Microstructure, defects and mechanical behavior of beta-type titanium porous structures manufactured by electron beam melting and selective laser melting. Acta Materialia, 113:56–67, 2016. [106] Y. Lu, S. Wu, Y. Gan, T. Huang, C. Yang, L. Junjie, and J. Lin. Study on the microstructure, mechanical property and residual stress of SLM Inconel-718 alloy manufactured by differing island scanning strategy. Optics {&} Laser Technology, 75:197–206, 2015. [107] J. K. Mackenzie and R. Shuttleworth. A phenomenological theory of sintering. Proceedings of the Physical Society. Section B, 62(12):833– 852, 1949. [108] J. D. Majumdar and I. Manna. Introduction to laser assisted fabrication of materials, pages 1–67. Springer, 2013. [109] M. Markl, R. Ammer, U. Ljungblad, U. Rüde, and C. Körner. Electron beam absorption algorithms for electron beam melting processes simulated by a three-dimensional thermal free surface lattice boltzmann method in a distributed and parallel environment. Procedia Computer Science, 18:2127–2136, 2013. [110] M. Matsumoto, M. Shiomi, K. Osakada, and F. Abe. Finite element analysis of single layer forming on metallic powder bed in rapid prototyping by selective laser processing. International Journal of Machine Tools and Manufacture, 42(1):61–67, 2002. [111] M. J. Matthews, G. Guss, S. A. Khairallah, A. M. Rubenchik, P. J. Depond, and W. E. King. Denudation of metal powder layers in laser powder bed fusion processes. Acta Materialia, 114:33–42, 2016. [112] J. C. Maxwell. Electricity and Magnetism. Clarendon Press, Oxford, UK, 1873. [113] R. W. McVey, R. M. Melnychuk, J. A. Todd, and R. P. Martukanitz. Absorption of laser irradiation in a porous powder layer. Journal of Laser Applications, 19(4):214–224, 2007. [114] P. Meakin and R. Jullien. Restructuring effects in the rain model for random deposition. Journal de Physique, 48(10):1651–1662, 1987. [115] M. Megahed, H.-W. Mindt, N. N’Dri, H. Duan, and O. Desmaison. Metal additive-manufacturing process and residual stress modeling. Integrating Materials and Manufacturing Innovation, 5(1):1–33, 2016. [116] P. Mercelis and J. Kruth. Residual stresses in selective laser sintering and selective laser melting. Rapid Prototyping Journal, 12(5):254–265, 2006. [117] R. E. Meredith and C. W. Tobias. Conductivities in emulsions. Journal of the Electrochemical Society, 108(3):286–290, 1961. [118] M. Militzer. Phase field modeling of microstructure evolution in steels. Current Opinion in Solid State and Materials Science, 15(3):106– 115, 2011. [119] P. Nie, O. A. Ojo, and Z. Li. Numerical modeling of microstructure evolution during laser additive manufacturing of a nickel-based superalloy. Acta Materialia, 77:85–95, 2014. [120] D. Pal, N. Patil, K. H. Kutty, K. Zeng, A. Moreland, A. Hicks, D. Beeler, and B. Stucker. A Generalized Feed-Forward Dynamic Adaptive Mesh Refinement and Derefinement Finite-Element Framework for Metal Laser SinteringPart II: Nonlinear Thermal Simulations and Validations. Journal of Manufacturing Science and Engineering, 138(6):61003, 2016. [121] D. Pal, N. Patil, K. Zeng, and B. Stucker. An Integrated Approach to Additive Manufacturing Simulations Using Physics Based, Coupled Multiscale Process Modeling. Journal of Manufacturing Science and Engineering, 136(6):61022, 2014. [122] N. Patil, D. Pal, K. R. Haludeen Kutty, K. Zeng, A. Moreland, A. Hicks, D. Beeler, and B. Stucker. A generalized feed forward dynamic adaptive mesh refinement and de-refinement finite element framework for Metal Laser Sintering : Part I (Formulation and Algorithm Development). Journal of Manufacturing Science and Engineering, 137(August):1–15, 2015. [123] R. S. Qin and H. K. Bhadeshia. Phase field method. Materials Science and Technology, 26(7):9, 2010. [124] C. Qiu, C. Panwisawas, M. Ward, H. C. Basoalto, J. W. Brooks, and M. M. Attallah. On the role of melt flow into the surface structure and porosity development during selective laser melting. Acta Materialia, 96:72–79, 2015. [125] N. Raghavan, R. Dehoff, S. Pannala, S. Simunovic, M. Kirka, J. Turner, N. Carlson, and S. S. Babu. Numerical modeling of heat-transfer and the influence of process parameters on tailoring the grain morphology of IN718 in electron beam additive manufacturing. Acta Materialia, 112:303–314, 2016. [126] A. Rai, M. Markl, and C. Körner. A coupled Cellular Automaton-Lattice Boltzmann model for grain structure simulation during additive manufacturing. Computational Materials Science, 124:37–48, 2016. [127] R. Rai, P. Burgardt, J. O. Milewski, T. J. Lienert, and T. DebRoy. Heat transfer and fluid flow during electron beam welding of 21Cr6Ni9Mn steel and Ti6Al4V alloy. Journal of Physics D: Applied Physics, 42(2):25503, 2008. [128] L. Rayleigh. On the influence of obstacles arranged in rectangular order upon the properties of a medium. Philosophical Magazine Series 5, 34(211):481–502, 1892. [129] D. Riedlbauer, T. Scharowsky, R. F. Singer, P. Steinmann, C. Körner, and J. Mergheim. Macroscopic simulation and experimental measurement of melt pool characteristics in selective electron beam melting of Ti-6Al-4V. International Journal of Advanced Manufacturing Technology, pages 1–9, 2016. 57 [130] D. Riedlbauer, P. Steinmann, and J. Mergheim. Thermomechanical finite element simulations of selective electron beam melting processes : performance considerations. Computational Mechanics, 54(1):109–122, 2014. [131] I. A. Roberts, C. J. Wang, R. Esterlein, M. Stanford, and D. J. Mynors. A three-dimensional finite element analysis of the temperature field during laser melting of metal powders in additive layer manufacturing. International Journal of Machine Tools and Manufacture, 49(12-13):916–923, 2009. [132] M. Rombouts, L. Froyen, A. V. Gusarov, E. H. Bentefour, and C. Glorieux. Photopyroelectric measurement of thermal conductivity of metallic powders. Journal of Applied Physics, 97(2):0–9, 2005. [133] W. J. Sames, F. A. List, S. Pannala, R. R. Dehoff, and S. S. Babu. The metallurgy and processing science of metal additive manufacturing. International Materials Reviews, 6608(March):1–46, 2016. [134] G. W. Scherer. Sintering of Low-Density Glasses: I, Theory. Journal of The American Ceramic Society, 60(5-6):236–239, 1976. [135] V. V. Semak, W. D. Bragg, B. Damkroger, and S. Kempka. Transient model for the keyhole during laser welding. Journal of Physics D: Applied Physics, 32:L61–L64, 1999. [136] N. Shen and K. Chou. Numerical Thermal Analysis in Electron Beam Additive Manufacturing With Preheating Effects. In Proceedings of the 23rd Solid Freeform Fabrication Symposium, pages 774–784, Austin, Texas, USA, 2012. [137] N. Shen and K. Chou. Thermal modeling of electron beam additive manufacturing process - powder sintering effects. In Proceedings of the ASME 2012 International Manufacturing Science and Engineering Conference, pages 1–9, Notre Dame, Indiana, USA, 2012. [138] M. Shiomi, A. Yoshidome, F. Abe, and K. Osakada. Finite element analysis of melting and solidifying processes in laser rapid prototyping of metallic powders. International Journal of Machine Tools and Manufacture, 39(2):237–252, 1999. [139] B. P. Singh and M. Kaviany. Independent theory versus direct simulation of radiation heat transfer in packed beds. International Journal of Heat and Mass Transfer, 34(11):2869–2882, 1991. [140] B. P. Singh and M. Kaviany. Modelling radiative heat transfer in packed beds. International Journal of Heat and Mass Transfer, 35(6):1397– 1405, 1992. [141] B. P. Singh and M. Kaviany. Effect of solid conductivity on radiative heat transfer in packed beds. International Journal of Heat and Mass Transfer, 37:2579–2583, 1994. [142] I. Steinbach. Phase-Field Model for Microstructure Evolution at the Mesoscopic Scale. Annual Review of Materials Research, 43(1):89–107, 2013. [143] B. Stucker. Additive Manufacturing Technologies: Technology Introduction and Business Implications. Frontiers of Engineering 2011: Reports on Leading-Edge Engineering from the 2011 Symposium, pages 19–21, 2012. [144] M. Tancrez and J. Taine. Direct identification of absorption and scattering coefficients and phase function of a porous medium by a Monte Carlo technique. International Journal of Heat and Mass Transfer, 47(2):373–383, 2004. [145] L. Thijs, F. Verhaeghe, T. Craeghs, J. V. Humbeeck, and J. P. Kruth. A study of the microstructural evolution during selective laser melting of Ti-6Al-4V. Acta Materialia, 58(9):3303–3312, 2010. [146] M. Thomas, G. J. Baxter, and I. Todd. Normalised model-based processing diagrams for additive layer manufacture of engineering alloys. Acta Materialia, 108:26–35, 2016. [147] N. K. Tolochko, T. Laoui, Y. V. Khlopkov, S. E. Mozzharov, V. I. Titov, and M. B. Ignatiev. Absorptance of powder materials suitable for laser sintering. Rapid Prototyping Journal, 6(3):155–160, 2000. [148] E. Tsotsas and H. Martin. Thermal conductivity of packed beds: a review. Chemical Engineering and Processing: Process Intensification, 22(1):19–37, 1987. [149] F. Verhaeghe, T. Craeghs, J. Heulens, and L. Pandelaers. A pragmatic model for selective laser melting with evaporation. Acta Materialia, 57(20):6006–6012, 2009. [150] R. Viskanta. Radiative transfer in dispersed media. Applied Mechanics Reviews, 42(9):241–259, 1989. [151] D. Wang, Y. Yang, Z. Yi, and X. Su. Research on the fabricating quality optimization of the overhanging surface in SLM process. International Journal of Advanced Manufacturing Technology, 65(9-12):1471–1484, 2013. [152] S. L. Wang, R. F. Sekerka, A. A. Wheeler, B. T. Murray, S. R. Coriell, R. J. Braun, and G. B. McFadden. Thermodynamically-consistent phase-field models for solidification. Physica D: Nonlinear Phenomena, 69(1-2):189–200, 1993. [153] X. C. Wang and J. P. Kruth. A simulation model for direct selective laser sintering of metal powders. In B. H. V. Topping, editor, Computational techniques for materials, composites and composite structures, pages 57–71. Civil-Comp, Edingburg, 2000. [154] X. C. Wang, T. Laoui, J. Bonse, J. P. Kruth, B. Lauwers, and L. Froyen. Direct Selective Laser Sintering of Hard Metal Powders : Experimental Study and Simulation. The International Journal of Advanced Manufacturing Technology, 19(5):351–357, 2002. [155] A. A. Wheeler, W. J. Boettinger, and G. B. McFadden. Phase-field model for isothermal phase transitions in binary alloys. Physical Review A, 45(10):7424–7439, 1992. [156] P. J. Withers. Neutron scattering/Diffusion de neutrons: Mapping residual and internal stress in materials by neutron diffraction. Comptes rendus - Physique, 8:806–820, 2007. [157] W. W. Wits, R. Bruins, L. Terpstra, R. A. Huls, and H. J. M. Geijselaers. Single scan vector prediction in selective laser melting. Additive Manufacturing, 9:1–6, 2016. [158] A. S. Wu, D. W. Brown, M. Kumar, G. F. Gallegos, and W. E. King. An Experimental Investigation into Additive Manufacturing-Induced Residual Stresses in 316L Stainless Steel. Metallurgical and Materials Transactions A, 45(13):6260–6270, 2014. [159] G. X. Xu, C. S. Wu, G. L. Qin, X. Y. Wang, and S. Y. Lin. Adaptive volumetric heat source models for laser beam and laser + pulsed GMAW hybrid welding processes. International Journal of Advanced Manufacturing Technology, 57(1-4):245–255, 2011. [160] W. Xu, M. Brandt, S. Sun, J. Elambasseril, Q. Liu, K. Latham, K. Xia, and M. Qian. Additive manufacturing of strong and ductile Ti6Al4V by selective laser melting via in situ martensite decomposition. Acta Materialia, 85:74–84, 2015. [161] I. Yadroitsava, S. Grewar, D. Hattingh, and I. Yadroitsev. Residual Stress in SLM Ti6Al4V Alloy Specimens. Materials Science Forum, 828-829:305–310, 2015. [162] I. Yadroitsev, P. Bertrand, and I. Smurov. Parametric analysis of the selective laser melting process. Applied Surface Science, 253(19):8064– 8069, 2007. 58 [163] S. Yagi and D. Kunil. Studies on effective thermal conductivities in packed beds. A.I.Ch.E. Journal, 3(3):373–381, 1957. [164] Y. S. Yang, J. R. Howell, and D. E. Klein. Radiative Heat Transfer Through a Randomly Packed Bed of Spheres By the Monte Carlo Method. Journal of Heat Transfer, 105(2):325–332, 1983. [165] G. Yu, D. Gu, D. Dai, M. Xia, C. Ma, and Q. Shi. On the role of processing parameters in thermal behavior, surface morphology and accuracy during laser 3D printing of aluminum alloy. Journal of Physics D: Applied Physics, 49(13):135501, 2016. [166] P. Yuan and D. Gu. Molten pool behaviour and its physical mechanism during selective laser melting of TiC/AlSi10Mg nanocomposites: simulation and experiments. Journal of Physics D: Applied Physics, 48(3):35303, 2015. [167] M. F. Zaeh and G. Branner. Investigations on residual stresses and deformations in selective laser melting. Production Engineering, 4(1):35–45, 2010. [168] M. F. Zäh and S. Lutzmann. Modelling and simulation of electron beam melting. Production Engineering, 4(1):15–23, 2010. [169] K. Zeng, D. Pal, N. Patil, and B. E. Stucker. A New Dynamic Mesh Method Applied to the Simulation of Selective Laser Melting. In Proceedings of the Solid Freeform Fabrication Symposium, pages 12–14, Austin, Texas, USA, 2013. [170] K. Zeng, D. Pal, C. Teng, and B. E. Stucker. Evaluations of effective thermal conductivity of support structures in selective laser melting. Additive Manufacturing, 6:67–73, 2015. [171] J. Zhang, F. Liou, W. Seufzer, J. Newkirk, and Z. Fan. Probabilistic Simulation of Solidification Microstructure Evolution During LaserBased Metal Deposition. Proceedings of the Int. Solid Freeform Fabr. Symp., pages 739–748, 2013. [172] J. Zhou, Y. Zhang, and J. K. Chen. Numerical simulation of laser irradiation to a randomly packed bimodal powder bed. International Journal of Heat and Mass Transfer, 52(13-14):3137–3146, 2009. 59
5cs.CE
DEGENERATIONS OF GRADED COHEN-MACAULAY MODULES arXiv:1302.1643v1 [math.AC] 7 Feb 2013 NAOYA HIRAMATSU Abstract. We introduce a notion of degenerations of graded modules. In relation to it, we also introduce several partial orders as graded analogies of the hom order, the degeneration order and the extension order. We prove that these orders are identical on the graded Cohen-Macaulay modules if a graded ring is of graded finite representation type and representation directed. 1. Introduction The notion of degenerations of modules appears in geometric methods of representation theory of finite dimensional algebras. In [15] Yoshino gives a schemetheoretical definition of degenerations, so that it can be considered for modules over a Noetherian algebra which is not necessary finite dimensional. Now, a theory of degenerations is considered for derived module categories [10] or stable module categories [16]. The degeneration problem of modules has been studied by many authors [6, 12, 14, 15, 18, 19]. For the study, several order relations for modules, such as the hom order, the degeneration order and the extension order, were introduced, and the connection among them has been studied. In the previous paper [8], the author give the complete description of degenerations over a ring of even-dimensional simple hypersurface singularity of type (An ). In the present paper we consider degenerations of graded Cohen-Macaulay modules over a graded Gorenstein ring with a graded isolated singularity. First we consider an order relation on a category of graded Cohen-Macaulay modules which is called the hom order (Definition 2.2). We shall show that it is actually a partial order if a graded ring is Gorenstein with a graded isolated singularity. In section 3 we propose a definition of degenerations for graded modules and we state several properties on it. We show that if the graded ring is of graded finite representation type and representation directed, then the hom order, the degeneration order and the extension order are identical on the graded CohenMacaulay modules. We also consider stable analogue of degenerations of graded Cohen-Macaulay modules in section 4. Date: April 7, 2018. 2000 Mathematics Subject Classification. Primary 16W50 ; Secondary 13D10. Key words and phrases. degeneration, graded Cohen-Macaulay module, finite representation type. 1 2 NAOYA HIRAMATSU 2. Hom order on graded modules Throughout the paper R = ⊕∞ i=0 Ri be a commutative Noetherian N-graded Cohen-Macaulay ring with R0 = k a field of characteristic zero. A graded ring is said to be ∗ local if the set of graded proper ideals has a unique maximal element. Thus R is ∗ local since m = ⊕i>0 Ri is a unique maximal ideal of R. We denote by modZ (R) the category of finitely generated Z-graded modules whose morphisms are homogenous morphisms that preserve degrees. For i ∈ Z, M(i) ∈ modZ (R) is defined by M(i)n = Mn+i . Then HomR (M, N(i)) consisting of homogenous morphisms of degree i, and we set ∗ HomR (M, N) = ⊕i∈Z HomR (M, N(i)). For a graded prime ideal p of R, we denote by R(p) a homogenous localization of R by p. For M ∈ modZ (R), take a graded free resolution d d l 1 · · · → Fl − → Fl−1 → · · · → F1 − → F0 → M → 0. We define an lth syzygy module Ωl M of M by Imdl . We say that a graded R-module M is said to be a graded Cohen-Macaulay R-module if ∗ ExtiR (R/m, M) = 0 for any i < d = dim R. In particular, this condition is equivalent to ∗ ExtiR (M, ωR ) = 0 for any i > 0, where ωR is a ∗ canonical module of R. We denote by CMZ (R) the full subcategory of modZ (R) consisting of graded Cohen-Macaulay R-modules. By our assumption on R, modZ (R) and CMZ (R) are Krull-Schmidt, namely each object can be decomposed into indecomposable objects up to isomorphism uniquely. For M ∈ modZ (R) we denote by h(M) a sequence (dimk Mn )n∈Z of non-negative integers. By the definition, it is easy to see that h(M) = h(N) if and only if they have the same Hilbert series. Moreover, we also have that h(M) = h(N) if and only if h(M ∗ ) = h(N ∗ ) where (−)∗ = ∗ HomR (−, ωR ). See [7, Theorem 4.4.5.]. Remark 2.1. If M and N give the same class in the Grothendieck group, i.e. [M] = [N] as an element of K0 (modZ (R)) then h(M) = h(N). However the converse does not hold in general. Let R = k[x, y]/(x2 −y 2 ) with deg x = deg y = 1 and set M = R/(x + y) and N = R/(x − y). Then h(M) = h(N) and [M] 6= [N] in K0 (modZ (R)). In fact, let p be an ideal (x + y)R. Then rankR(p) M(p) = 1 and rankR(p) N(p) = 0. Note that [M] = [N] on K0 (modZ (R(p) )) yields that [M(p) ] = [N(p) ] on K0 (modZ (R(p) )). Thus [M] = [N] can never happen. Our motivation of the paper is to investigate the graded degenerations of graded Cohen-Macaulay modules in terms of some order relations. For the reason we consider the following relation on CMZ (R) that is known as the hom order. DEGENERATIONS OF GRADED COHEN-MACAULAY MODULES 3 Definition 2.2. For M, N ∈ CMZ (R) we define M ≤hom N if [M, X] ≤ [N, X] for each X ∈ CMZ (R). Here [M, X] is an abbreviation of dimk HomR (M, X). Remark 2.3. For M, N ∈ modZ (R), [M, N] is finite, and thus we can consider the above relation. As a consequence in [5], if R is of dimension 0, 1 or 2, ≤hom is a partial order on CMZ (R) since CMZ (R) is closed under kernels in such cases. Lemma 2.4. Let R be a graded Gorenstein ring an let M and N be indecomposable graded Cohen-Macaulay R-modules. Suppose that h(M) = h(N). For Y ∈ modZ (R) which is of finite projective dimension, we have [M, Y ] = [N, Y ]. Proof. For a graded R-module Y which is of finite projective dimension, Ωi Y is also of finite projective dimension. Since R is Gorenstein, we have Ext1R (M, Ωi Y ) = 0 for all graded Cohen-Macaulay R-modules M. Thus, take a graded free resolution of Y and apply HomR (M, −) to the resolution, we get an exact sequence 0 → HomR (M, Fl ) → HomR (M, Fl−1 ) → · · · → HomR (M, F0 ) → HomR (M, Y ) → 0. Hence [M, Y ] = l X (−1)i [M, Fi ]. i=0 We also have [N, Y ] = l X (−1)i [N, Fi ]. i=0 Since R is Gorenstein, ωR = R(l) for some l ∈ Z. Then we have an equality [M, F ] = [N, F ] for each free module. Therefore we have [M, Y ] = l X i=0 i (−1) [M, Fi ] = l X (−1)i [N, Fi ] = [N, Y ]. i=0  In the paper we use the theory of Auslander-Reiten (abbr. AR) sequences of graded Cohen-Macaulay modules. For the detail, we recommend the reader to look at [9, 3, 4] and [13, Chapter 15.]. We denote by CMZ0 (R) the full subcategory of CMZ (R) consisting of M ∈ CMZ (R) such that M(p) is R(p) -free for any graded prime ideal p 6= m. Theorem 2.5. [3, 4, 13, 9] Let (R, m) be a Noetherian Z-graded Gorenstein ∗ local ring. Then CMZ0 (R) admits AR sequences. Definition 2.6. We say that (R, m) is a graded isolated singularity if each graded localization R(p) is regular for each graded prime ideal p with p 6= m. It is easy to see that CMZ (R) = CMZ0 (R) if R is a graded isolated singularity, so that CMZ (R) admits AR sequences. We denote by µ(M, Z) the multiplicity of Z as a direct summand of M. 4 NAOYA HIRAMATSU Theorem 2.7. Let R be a graded Gorenstein ring with R0 is an algebraically closed field and let M and N graded Cohen-Macaulay R-modules. Assume that R is a graded isolated singularity. Then [M, X] = [N, X] for each X ∈ CMZ (R) if and only if M ∼ = N. Particularly, ≤hom is a partial order on CMZ (R). µ(M,M ) i Proof. We decompose M as M = ⊕Mi where Mi are indecomposable graded Cohen-Macaulay R-modules. If Mi is not free, we can take the AR sequence ending in Mi 0 → τ Mi → Ei → Mi → 0, where τ Mi is an AR translation of Mi . Apply HomR (M, −) and HomR (N, −) to the sequence, since k is an algebraically closed field, we have 0 → HomR (M, τ Mi ) → HomR (M, Ei ) → HomR (M, Mi ) → k µ(M,Mi ) → 0 and 0 → HomR (N, τ Mi ) → HomR (N, Ei ) → HomR (N, Mi ) → k µ(N,Mi ) → 0. Counting the dimensions of terms, we conclude that µ(M, Mi ) = µ(N, Mi ). If Mi is free, we may assume that Mi = R. Let m be a ∗ maximal ideal of R. We consider the Cohen-Macaulay approximation (see [2]) of m 0 → Y → X → m → 0. We note that X is a graded Cohen-Macaulay R-module and Y is of finite injective dimension. We also note that Y is of finite projective dimension since R is Gorenstein. Let f be a composition map of the approximation X → m and a natural inclusion m → R. Then we get the following commutative diagram. 0 −−−→ K −−−→ x g  f X −−−→ x =  R x ⊆  0 −−−→ Y −−−→ X −−−→ m −−−→ 0. By a diagram chasing we see that g is surjective, so that K is of finite projective dimension. Now we obtain an exact sequence 0 → HomR (M, K) → HomR (M, X) → HomR (M, R) → k µ(M,R) → 0. According to Lemma 2.4, [M, K] = [N, K]. Thus [M, R] + [M, K] − [M, X] = [N, R] + [N, K] − [N, X]. Hence µ(M, R) = µ(N, R). Consequently M ∼ = N.  DEGENERATIONS OF GRADED COHEN-MACAULAY MODULES 5 3. Graded degenerations of graded Cohen-Macaulay modules We define a notion of degenerations for graded modules. Definition 3.1. Let R be a Noetherian N-graded ring where R0 = k is a field and let V = k[[t]] with a trivial gradation and K be the localization by t, namely K = Vt = k(t). For finitely generated graded R-modules M and N, we say that M gradually degenerates to N or N is a graded degeneration of M if there is a finitely generated graded R ⊗k V -module Q which satisfies the following conditions: (1) Q is flat as a V -module. (2) Q ⊗V V /tV ∼ = N as a graded R-module. (3) Q ⊗V K ∼ = M ⊗k K as a graded R ⊗k K-module. In [15], Yoshino gives a necessary and sufficient condition for degenerations of (non-graded) modules. One can also show its graded version in a similar way. See also [12, 19]. Theorem 3.2. [15, Theorem 2.2] The following conditions are equivalent for finitely generated graded R-modules M and N. (1) M gradually degenerates to N. (2) There is a short exact sequence of finitely generated graded R-modules 0 −−−→ Z −−−→ M ⊕ Z −−−→ N −−−→ 0. Remark 3.3. (1) As Yoshino has shown in [15], the endomorphism of Z in the sequence of Theorem 3.2 is nilpotent. Note that we do not need the nilpotency assumption here. Actually, since EndR (Z) is Artinian, by using Fitting theorem, we can describe the endomorphism as a direct sum of an isomorphism and a nilpotent morphism. See also [15, Remark 2.3.]. (2) Assume that M and N are graded Cohen-Macaulay modules. Then we can show that Z is also a graded Cohen-Macaulay R-module. See [15, Remark 4.3.]. (3) Assume that there is an exact sequence of finitely generated graded Rmodules 0 −−−→ L −−−→ M −−−→ N −−−→ 0. Then M gradually degenerates to L⊕N. See [15, Remark 2.5] for instance. (4) Let M and N be finitely generated graded R-modules and suppose that M gradually degenerates to N. Then the modules M and N give the same class in the Grothendieck group. We can also prove in a similar way to the proof of [17, Theroem 2.1.] that, for L, M, N ∈ modZ (R), if L gradually degenerates to M and if M gradually degenerates to N then L gradually degenerates to N. And one can show that if L gradually degenerates to M then L ≤hom M. Consideration into this fact we define partial orders as follows. 6 NAOYA HIRAMATSU Definition 3.4. For finitely generated graded R-modules M, N, we define the relation M ≤deg N, which is called the degeneration order, if M gradually degenerates to N. We also define the relation M ≤ext N if there are modules Mi , N ′ i , N ′′ i and short exact sequences 0 → N ′ i → Mi → N ′′ i → 0 in CMZ (R) so that M = M1 , Mi+1 = N ′ i ⊕ N ′′ i , 1 ≤ i ≤ s and N = Ms+1 for some s. For M ∈ CMZ (R), we take a first syzygy module of M ∗ 0 → Ω1 M ∗ → F → M ∗ → 0. Applying ∗ HomR (−, ωR ) to the sequence, we have 0 → M ∗∗ ∼ = M → F ∗ → (Ω1 M ∗ )∗ → 0. Then we denote (Ω1 M ∗ )∗ by Ω−1 M. Lemma 3.5. Let R be a graded isolated singularity with ωR and let M and N be graded Cohen-Macaulay R-modules. Assume that h(M) = h(N) and M ≤hom N. Then, for each graded Cohen-Macaulay R-module X, there exists an integer lX ≫ 0 such that [M, X(±lX )] = [N, X(±lX )]. Proof. For each X ∈ CMZ (R), we can take an exact sequence as above 0 → X → E → Ω−1 X → 0. Note that E is a direct sum of ωR (n) for some integers n. Applying ∗ HomR (M, −) to the sequence, we have 0 → ∗ HomR (M, X) → ∗ HomR (M, E) → ∗ HomR (M, Ω−1 X) → ∗ ExtR (M, X) → 0. Since R is a graded isolated singularity, dim ∗ ExtR (M, X) is finite. Thus ∗ ExtR (M, X)±l1 = 0 for sufficiently large l1 ≫ 0. Similarly there also exists an integer l2 ≫ 0 so that ∗ ExtR (N, X)±l2 = 0. Set l = max{l1 , l2 }. Then we have 0 → ∗ HomR (M, X)±l → ∗ HomR (M, E)±l → ∗ HomR (M, Ω−1 X)±l → 0. Therefore we obtain the equation [M, X(±l)] = [M, E(±l)] − [M, Ω−1 X(±l)]. We also have [N, X(±l)] = [N, E(±l)] − [N, Ω−1 X(±l)]. Suppose that [M, X(±l)] < [N, X(±l)]. Then the following inequality holds. [M, E(±l)] − [M, Ω−1 (X)(±l)] < [N, E(±l)] − [N, Ω−1 X(±l)]. Since h(M) = h(N), [M, E(±l)] = [N, E(±l)]. Hence we see that [M, Ω−1 X(±l)] > [N, Ω−1 X(±l)]. This is a contradiction since M ≤hom N. Therefore we have some integer l such that [M, X(±l)] = [N, X(±l)].  DEGENERATIONS OF GRADED COHEN-MACAULAY MODULES 7 We say that the category CMZ (R) is of graded finite representation type if there are only a finite number of isomorphism classes of indecomposable graded CohenMacaulay modules up to shift. We note that if CMZ (R) is of finite representation type, then R is a graded isolated singularity. See [13, Chapter 15.] for the detail. As an immediate consequence of Lemma 3.5, we have the following. Corollary 3.6. Let R be of finite representation type and let M and N be graded Cohen-Macaulay R-modules. Assume that h(M) = h(N) and M ≤hom N. Then there are only finitely many indecomposable graded Cohen-Macaulay R-modules X such that [N, X] − [M, X] > 0. For graded Cohen-Macaulay R-modules M and N, we consider the following set of all the isomorphism classes of indecomposable graded Cohen-Macaulay modules FM,N = {X | [N, X] − [M, X] > 0 }/ ∼ =. Note from Corollary 3.6 that F is a finite set if h(M) = h(N) and M ≤hom N. We also note that ωR 6∈ F in the case. Proposition 3.7. [12] Let R be a graded Gorenstein ring which is of graded finite representation type and let M and N be graded Cohen-Macaulay R-modules. Assume that h(M) = h(N) and M ≤hom N. Then there exists some graded Cohen-Macaulay R-module L such that M ⊕ L degenerates to N ⊕ L. Proof. Although a proof of the proposition is given in [12], we refer the argument of the proof in the present paper. For this reason we briefly recall the proof of the proposition. Since R is a graded isolated singularity, CMZ (R) admits AR sequences. For each X ∈ FM,N , we can take an AR sequence starting from X. ΣX : 0 → X → EX → τ −1 X → 0. Now we consider a sequence which is a direct sum of [N.X] − [M, X] copies of ΣX where X runs through all modules in FM,N . Namely M [N.X]−[M,X] ΣX = 0 → U → V → W → 0. X∈FM,N For any indecomposable Z ∈ CMZ (R), we obtain 0 → HomR (W, Z) → HomR (V, Z) → HomR (U, Z) → k [N,Z]−[M,Z] → 0. This implies that the equality [U, Z] + [W, Z] − [V, Z] = [N, Z] − [M, Z], thus [U, Z] + [W, Z] + [M, Z] = [N, Z] + [V, Z] holds for all Z ∈ CMZ (R). Hence this yields that M ⊕U ⊕W ∼ = N ⊕ V. 8 NAOYA HIRAMATSU Since V degenerates to U ⊕ W , therefore M ⊕ V degenerates to M ⊕ U ⊕ W ∼ = N ⊕V.  Now we focus on the case that a graded Gorenstein ring is of graded finite representation type and representation directed. We say that a graded CohenMacaulay ring R is representation directed if the AR quiver of CMZ0 (R) has no oriented cyclic paths. Bongartz [6] has studied such a case over finite dimensional k-algebras. In our graded settings, the similar results hold. Actually we shall prove the following. Theorem 3.8. Let R be a graded Gorenstein ring which is of graded finite representation type and representation directed. Then the following conditions are equivalent for M and N ∈ CMZ (R). (1) h(M) = h(N) and M ≤hom N. (2) M ≤deg N. (3) M ≤ext N. To prove the theorem, we modify the arguments in [6]. Lemma 3.9 (Cancellation property). Let M, N and X be finitely generated graded R-modules. (1) Assume that [X, M] = [X, N]. If M ⊕ X gradually degenerates to N ⊕ X, M gradually degenerates to N. (2) Assume that R is Gorenstein and M and N graded Cohen-Macaulay Rmodules. If M gradually degenerates to N ⊕ F for some graded free Rmodule F then M/F gradually degenerates to N. Proof. (1) Since M ⊕ X gradually degenerates to N ⊕ X, there exist an exact sequence 0 → W → M ⊕ X ⊕ W → N ⊕ X → 0. We construct a pushout diagram. 0 0 x x     X x   X x   0 −−−→ W −−−→ M ⊕ X ⊕ W −−−→ N ⊕ X −−−→ 0 x x     0 −−−→ W −−−→ E x   0 −−−→ N x   0. −−−→ 0 DEGENERATIONS OF GRADED COHEN-MACAULAY MODULES 9 For the middle column sequence, [X, X] + [X, E] − [X, M ⊕ X ⊕ W ] = [X, E] − [X, M] − [X, W ] ≥ 0. On the other hand, for the bottom row sequence, since [X, M] = [X, N], [X, N] + [X, W ] − [X, E] = [X, M] + [X, W ] − [X, E] ≥ 0. Thus we have [X, X] + [X, E] − [X, M ⊕ X ⊕ W ] = 0. This implies that the middle column sequence splits, so that X ⊕E ∼ = M ⊕X ⊕W . ∼ Therefore E = M ⊕ W and we get 0 → W → M ⊕ W → N → 0. Namely M gradually degenerates to N. (2) Since M gradually degenerates to N ⊕ F , we have an exact sequence 0 → Z → M ⊕ Z → N ⊕ F → 0. Suppose that Z contains a graded free module G as a direct summand. We construct a pushout diagram 0 0 x x     0 −−−→ Z/G −−−→ x   0 −−−→ Z x   G x   E x   −−−→ N ⊕ F −−−→ 0 −−−→ M ⊕ Z −−−→ N ⊕ F −−−→ 0 x   G x   0 0. The left column sequence is a split sequence induced by the decomposition Z ∼ = Z/G ⊕ G. Note that E is also a graded Cohen-Macaulay module. Since R is Gorenstein, the middle column sequence is also split. Hence E⊕G∼ = M ⊕ Z/G ⊕ G. = M ⊕Z ∼ This yields that E ∼ = M ⊕ Z/G. Hence we may assume that Z has no graded free modules as direct summands. Consider a composition of the surjection M ⊕ Z → N ⊕ F and the projection N ⊕ F → F . Then the composition mapping is split, so that M contains F as a direct summand. Hence M ∼ = M/F ⊕ F gradually degenerates to N ⊕ F . Since [F, M/F ⊕ F ] = [F, M] = [F, N ⊕ F ], by (1), we conclude that M/F gradually degenerates to N.  10 NAOYA HIRAMATSU For indecomposable graded Cohen-Macaulay modules M and N, we write X  Y if X ∼ = Y or if there exists a finite path from X to Y in the AR quiver of CMZ0 (R). Lemma 3.10. Let R be a graded Gorenstein ring which is of graded finite representation type and representation directed and let M and N ∈ CMZ (R). Assume that h(M) = h(N), M ≤hom N and M and N have no common direct summands. Let X ∈ CMZ (R) be an indecomposable module such that -minimal with the property [N, X] − [M, X] > 0 and E be a middle term of an AR sequence starting from X. Then [E, N] = [E, M]. Proof. As in the proof of Proposition 3.7, we can construct the sequence 0→U →V →W →0 in CMZ (R) such that U ⊕W ⊕M ∼ = V ⊕N via taking a direct sum of AR sequences starting from modules in FM,N . This isomorphism implies that [Z, U] + [Z, W ] − [Z, V ] = [Z, N] − [Z, M] for each Z ∈ CMZ (R). Thus it is enough to show that the equality [E, U] + [E, W ] − [E, V ] = 0 holds. If there is a Y ∈ FM,N such that, for the AR sequence 0 → Y → G → τ −1 Y → 0, [E, Y ] + [E, τ −1 Y ] − [E, G] > 0. If Y ∼ = X, then one can show that τ −1 X is a direct summand of E, so that there is a cyclic path. This is a contradiction since R is represented directed. Thus Y is not isomorphic to X. The inequality also show that τ −1 Y is a direct summand of E. Thus there is an irreducible map from X to τ −1 Y . Hence X is a direct summand of G, so that Y  X. This is a contradiction. Consequently, for each Y ∈ FM,N , [E, Y ] + [E, τ −1 Y ] − [E, G] = 0, therefore [E, N] = [E, M].  Proof of Theorem 3.8. The implication (3) ⇒ (2) ⇒ (1) is trivial. Note that ExtR (X, X) = 0 for an indecomposable X ∈ CMZ (R) since R is representation directed. Thus the implication (2) ⇒ (3) can be shown by the same argument in [18, 3.5.]. Now we shall show (1) ⇒ (2). Set V = ⊕W ∈FM,N W . Since [N, X] − [M, X] = 0 for any X 6∈ FM,N , we can show the implication by induction on d = [N, V ] − [M, V ]. If d = 0, [N, Z] = [M, Z] for each X ∈ CMZ (R). Thus M∼ = N. Hence assume that d > 0 and M and N have no summand in common in the inductive step. We take X ∈ CMZ (R) in Lemma 3.10 and let E be a middle term of the AR sequence starting from X. By virtue of Lemma 3.10 and Lemma 3.9 (1) it is enough to show that E ⊕ M ≤deg E ⊕ N. Now we have E ⊕ M ≤ext X ⊕ τ −1 X ⊕ M. By the property of AR sequence, for each indecomposable Y , [X, Y ] + [τ −1 X, Y ] = [E, Y ] + δX,Y where δX,Y = 1 if X ∼ = Y and otherwise 0. Thus X ⊕ τ −1 X ⊕ M ≤hom E ⊕ N. We should remark that FX⊕τ −1 X⊕M,E⊕N is contained in FM,N since X ∈ FM,N . DEGENERATIONS OF GRADED COHEN-MACAULAY MODULES 11 Then [E ⊕ N, V ] − [X ⊕ τ −1 X ⊕ M, V ] = [N, V ] − [M, V ] + [E, V ] − [X ⊕ τ −1 X, V ] = d−1 By the induction hypothesis, E ⊕ M ≤ext X ⊕ τ −1 X ⊕ M ≤deg E ⊕ N so that E ⊕ M ≤deg E ⊕ N.  Remark 3.11. The implication M ≤ext N ⇒ M ≤deg N does not hold in general. Let R = k[x, y]/(x2 ) with deg x = deg y = 1. Then R(−1) gradually degenerates to (x, y 2 )R. In fact we have an exact sequence   x 0 (x y 2 ) 0 −−−→ R/(x)(−2) −−−→ R(−1) ⊕ R/(x)(−2) −−−−−→ (x, y 2 )R −−−→ 0. Since (x, y 2 )R is an indecomposable graded Cohen-Macaulay module which is not isomorphic to R(−1), R(−1) ≤ext (x, y 2 )R can never happen. See also [8, Remark 2.5.]. Proposition 3.12. Let R be a graded Gorenstein ring which is of graded finite representation type and let M and N be graded Cohen-Macaulay R-modules. Assume that h(M) = h(N) and M ≤hom N. Then for each indecomposable non-free graded Cohen-Macaulay R-module X we have the following equality. [N, X] − [M, X] = [τ −1 X, N] − [τ −1 X, M]. Proof. Under the assumption, as in the proof of Proposition 3.7, we can construct an exact sequence 0→U →V →W →0 Z in CM (R) such that [U, X] + [W, X] − [V, X] = [N, X] − [M, X]. It is enough to show that the equality [U ⊕ W, X] − [V, X] = [τ −1 X, U ⊕ W ] − [τ −1 X, V ] holds for each indecomposable X. Moreover the sequence is a direct sum of AR sequences, we may assume that 0 → U → V → W → 0 is an AR sequence starting from U. Let X ∈ CMZ (R) be indecomposable and non projective. By the property of an AR sequence, we have [U ⊕ W, X] − [V, X] = δU,X [τ −1 X, U ⊕ W ] − [τ −1 X, V ] = δW,τ −1 X . Since W ∼ = τ −1 U, we can get the equality.  4. Remarks on stable degenerations of graded Cohen-Macaulay modules In the rest of the paper we consider the stable analogue of degenerations of graded Cohen-Macaulay modules. Let R be a graded Gorenstein ring where R0 = k is a field and let V = k[[t]] with a trivial gradation and K = k(t). Note that R ⊗k V and R ⊗k K are 12 NAOYA HIRAMATSU graded Gorenstein rings as well. Then CMZ (R ⊗k V ) and CMZ (R ⊗k K) are triangulated categories. We denote by L : CMZ (R ⊗k V ) → CMZ (R ⊗k K) (resp. R : CMZ (R ⊗k V ) → CMZ (R)) the triangle functor defined by the localization by t (resp. taking − ⊗V V /tV ). See also [16, Definition 4.1]. Definition 4.1. Let M , N ∈ CMZ (R). We say that M stably degenerates to N if there exists a graded Cohen-Macaulay module Q ∈ CMZ (R ⊗k V ) such that L(Q) ∼ = M ⊗k K in CMZ (R ⊗k K) and R(Q) ∼ = N in CMZ (R). One can show the following characterization of stable degenerations similarly to the proof of [16, Theorem 5.1]. Theorem 4.2. Let R be a graded Gorenstein ring where R0 = k is a field. The following conditions are equivalent for graded Cohen-Macaulay R-modules M and N. (1) F ⊕ M degenerates to N for some graded free R-module F . (2) There is a triangle in CMZ (R) Z −−−→ M ⊕ Z −−−→ N −−−→ Z[1]. (3) M stably degenerates to N . Proof. We should note that the implication (3) ⇒ (1). In our setting, R ⊗k V and R ⊗k K are ∗ local. Then graded projective R ⊗k V (resp. R ⊗k K) -modules are graded free R ⊗k V (resp. R ⊗k K) -modules. Hence we can show the implication as in the Artinian case of the proof of [16, Theorem 5.1].  Remark 4.3. A theory of degenerations for derived categories has been studied in [10]. They have shown that, for complexes M, N in the bounded derived category of a finite dimensional algebra, M degenerates to N if and only if there exists a triangle of the form which appears in the above theorem. Let R be a graded Gorenstein ring with R0 = k is an algebraically closed field. As shown in [1, 11, 9], suppose that R has a simple singularity then there exists a Dynkin quiver Q such that we have a triangle equivalence CMZ (R) ∼ = Db (kQ) where Db (kQ) is a bounded derived category of the category of finitely generated left modules over a path algebra kQ. By virtue of Theorem 4.2, we can describe the degenerations for Db (kQ) in terms of the graded degenerations for CMZ (R). Since the graded ring R is of graded finite representation type and representation directed, we have already seen them in Theorem 3.8. Acknowledgments The author express his deepest gratitude to Tokuji Araya and Yuji Yoshino for valuable discussions and helpful comments. DEGENERATIONS OF GRADED COHEN-MACAULAY MODULES 13 References [1] T. Araya, Exceptional sequences over graded Cohen-Macaulay rings. Math. J. Okayama Univ. 41 (1999), 81–102 (2001). [2] M. Auslander and R.-O. Buchweitz, The homological theory of maximal CohenMacaulay approximations. Mem. Soc. Math. France (N.S.) No. 38 (1989), 5–37. [3] M. Auslander and I. Reiten, Almost split sequences for Z-graded rings. Singularities, representation of algebras, and vector bundles (Lambrecht, 1985), 232–243, Lecture Notes in Math., 1273, Springer, Berlin, 1987. [4] M. Auslander and I. Reiten, Almost split sequences for abelian group graded rings. J. Algebra 114 (1988), no. 1, 29–39. [5] K. Bongartz, A generalization of a theorem of M. Auslander. Bull. London Math. Soc. 21 (1989), no. 3, 255–256. [6] K. Bongartz, On degenerations and extensions of finite-dimensional modules. Adv. Math. 121 (1996), 245–287. [7] W. Bruns and J. Herzog, Cohen-Macaulay Rings, Cambridge Studies in Advanced Mathematics, 39. Cambridge University Press, Cambridge, 1993. xii+403 pp. Revised edition, 1998. [8] N. Hiramatsu and Y. Yoshino, Examples of degenerations of Cohen-Macaulay modules, to appear in Proc. Amer. Math. Soc., arXiv1012.5346. [9] O. Iyama and R. Takahashi, Tilting and cluster tilting for quotient singularities, to appear in Math. Ann., arXiv:1012.5954. [10] B. Jensen, X, Su and A. Zimmermann, Degenerations for derived categories. J. Pure Appl. Algebra 198 (2005), no. 1-3, 281–295. [11] H. Kajiura, K, Saito and A. Takahashi, Matrix factorization and representations of quivers. II. Type ADE case. Adv. Math. 211 (2007), no. 1, 327–362. [12] C. Riedtmann, Degenerations for representations of quivers with relations. Ann. Scient. École Norm. Sup. 4e sèrie 19 (1986), 275–301. [13] Y. Yoshino, Cohen-Macaulay Modules over Cohen-Macaulay Rings, London Mathematical Society Lecture Note Series 146. Cambridge University Press, Cambridge, 1990. viii+177 pp. [14] Y. Yoshino, On degenerations of Cohen-Macaulay modules. J. Algebra 248 (2002), 272– 290. [15] Y. Yoshino, On degenerations of modules. J. Algebra 278 (2004), 217–226. [16] Y. Yoshino, Stable degenerations of Cohen-Macaulay modules, J. Algebra 332 (2011), 500–521. [17] G. Zwara, A degeneration-like order for modules. Arch. Math. 71 (1998), 437–444. [18] G. Zwara, Degenerations for modules over representation-finite algebras. Proc. Amer. Math. Soc. 127 (1999), 1313–1322. [19] G. Zwara, Degenerations of finite-dimensional modules are given by extensions. Compositio Math. 121 (2000), 205–218. Department of general education, Kure National College of Technology, 2-2-11, Agaminami, Kure Hiroshima, 737-8506 Japan E-mail address: [email protected]
0math.AC
Build a Compact Binary Neural Network through Bit-level Sensitivity and Data Pruning Yixing Li and Fengbo Ren School of Computing, Informatics, and Decision Systems Engineering Arizona State University {yixingli, renfengbo}@asu.edu Abstract Convolutional neural network (CNN) has been widely used for vision-based tasks. Due to the high computational complexity and memory storage requirement, it is hard to directly deploy a full-precision CNN on embedded devices. The hardwarefriendly designs are needed for resource-limited and energy-constrained embedded devices. Emerging solutions are adopted for the neural network compression, e.g., binary/ternary weight network, pruned network and quantized network. Among them, Binarized Neural Network (BNN) is believed to be the most hardware-friendly framework due to its small network size and low computational complexity. No existing work has further shrunk the size of BNN. In this work, we explore the redundancy in BNN and build a compact BNN (CBNN) based on the bit-level sensitivity analysis and bit-level data pruning. The input data is converted to a high dimensional bit-sliced format. In post-training stage, we analyze the impact of different bit slices to the accuracy. By pruning the redundant input bit slices and shrinking the network size, we are able to build a more compact BNN. Our result shows that we can further scale down the network size of the BNN up to 3.9x with no more than 1% accuracy drop. The actual runtime can be reduced up to 2x and 9.9x compared with the baseline BNN and its full-precision counterpart, respectively. 1 Introduction Vision-based applications can be found in many embedded devices for classification, recognition, detection and tracking tasks [Tian et al., 2015; Hu et al., 2015]. Specifically, convolutional neural network (CNN) has become the core architecture for those vision-based tasks [LeCun et al., 2015]. Since it can outperform conventional feature selection-based algorithm in terms of accuracy, it becomes more and more popular. Advanced driver-assistance system (ADAS) can either used CNNs for guiding autonomous driving or alerting the driver of predicted risk [Tian et al., 2015]. It is obvious that ADAS depends on a low-latency system to get a timely reaction. Artificial intelligence (AI) applications also explode in smartphones, such as automatically tagging the photos, face detection and so on [Schroff et al., 2015; Hu et al., 2015]. Apple has been reported working on an “Apple Neural Engine” to partially move their AI processing module on device [Lumb et al., 2017]. If processing the users’ requests through sending them to the data center, there will be much overhead of the latency and power consumption caused by the commutation. As such, on-device AI processing is the future trend to balance power efficiency and latency. However, CNNs are known to have high computation complexity, which makes it hard to directly deploy on embedded devices. Therefore, compressed CNNs are in demand. In the early stage, research work of hardware-friendly CNNs have focused on reducing the network precision down to 8-16 bits in the post-training stage [Suda et al., 2016], which either has a limited reduction or suffers from severer accuracy drop. Lately, in-training techniques have been brought up, achieving much higher compression ratio. BinaryConnect, BNN, TernaryNet and XNOR-Net [Courbariaux et al., 2015; Hubara et al., 2016; Rastegari et al., 2016; Zhu et al., 2016] have pushed to reduce the weight to binary or ternary (-1, 0, +1) values. Network pruning [Han et al., 2015] reduces the network size (the memory size for all the parameters) by means of reducing the number of connections. Regarding the network size, pruned network and TernaryNet can achieve 13x and 16x reduction [Zhu et al., 2016; Han et al., 2015], respectively. While BinaryConnect, BNN and XNOR-Net can achieve up to 32x reduction. In terms of computational complexity, only BNN and XNOR-Net with both binarized weights and activations can simply replace convolution operation with bitwise XNOR and bit count operation. However, XNOR-Net has additional scaling factor filters in each layer, which brings overhead to both memory and computation cost. From the above, BNN is the optimal solution for hardware deployment when considering both network size and computational complexity. In addition, BNN has drawn a lot of attention in hardware community [Li et al., 2017]. However, no existing study has explored scaling down BNN for a more efficient inference stage. This work is the first one that explores and proves that there is still redundancy in BNN. The proposed flow to reduce the network size is triggered by conversion and analysis of input data rather than the network body, which is rarely seen in previous work. A novel flow is proposed to prune out the redundant input bit slices and rebuild a compact BNN through bit-level sensitivity analysis. Experiment results show that the compression ratio of the network size is achieving up to 3.9x with no more than 1% accuracy drop. The rest of the paper is organized as follows. Section 2 discusses the related work for network compression and explains why BNN is a more superior solution be deployed on the hardware. Section 3 validates the hypothesis that BNN has redundancy and proposes a novel flow to build a compact BNN. Experiment results and discussion are shown in Section 4. Section 5 concludes the paper. 2 Related work Regarded to hardware-friendly oriented designs, it is not fair only emphasize compressing the network size. Other than that, the computational complexity is also essential. In this section, we first discuss and evaluate the related work for network compression by emphasizing both factors. We also present a simple benchmark study to help the reader better understand the computational complexity in terms of hardware resource of existing work. It can reveal why BNN is a more superior solution to be deployed on the hardware. BinaryConnect [Courbariaux et al., 2015] is a study in the early stage of exploring the binarized weight neural network. In the BinaryConnect network, the weights are binary values while the activations are still non-binary. Arbitrary value multiplies +1/-1 is equivalent to a conditional bitwise NOR operation. Hence, the convolution operation can be decomposed into conditional bitwise NOR operations and accumulation. It is a big step moving from full-precision multiplication to much simpler bitwise operations. BNN [Hubara et al., 2016] is the first one that builds a network with both binary weights and activations. The convolution operation has been further simplified as bitwise XNOR (Exclusive-NOR) and bit count operations. The hardware resource cost is minimized for GPU, FPGA and ASIC implementation. For GPU implementation, 32-bit bitwise XNOR can be implemented in a single clock cycle with one CUDA core. For FPGA and ASIC implement, there is no need to use DSP (Digital Signal Processor) resources anymore, which is relatively costly. Simple logic elements (LUTs, Look Up Tables) can be used to map bitwise XNOR and bit count operations, which makes it easy to map highly parallel computing engines to achieve high throughput and low latency. XNOR-Net [Rastegari et al., 2016] also builds the network based on binary weights and activations. However, it introduces a filter of full-precision scaling factors in each convolutional layer to ensure a better accuracy rate. Additional nonbinary convolution operations are needed in each convolutional layer, which cost additional time and computing resources. Thus, the XNOR-Net is not strictly a fully binarized network. TernaryNet [Zhu et al., 2016] holds ternary (-1, 0, +1) weights for its network. By increasing the precision level of the weights, it enhances the accuracy rate. Since ternary weights have to be encoded in 2 bits, the computational complexity will at least double, compared with BinaryConnect. Network pruning [Han et al., 2015] is revealed as the most popular technique for compressing pre-trained full-precision or reduced-precision CNNs (weights of the reduced-precision CNN are usually in the range of 8 bit - 16 bit [Suda et al., 2016]). It compresses the network by pruning out the useless weights, which gains speedup mainly by reducing the network size. Unlike all the other technique mentioned above, neither the weights nor activations of a pruned network are binary or ternary. Still, the computation complexity of the full-precision or reduced-precision multiply-add operation is much higher than that of the BNN. We implement a W(10,10)×A(10,10) matrix multiplication on a Xilinx Virtex-7 FPGA board for analyzing the computational complexity of the different architecture that mentioned above. The precision of elements in W and A are the same as the precision of weights and activations in each architecture. The matrix multiplication is fully mapped onto the FPGA. In other words, we don’t reuse any hardware resource. So the resource utilization is a good reflection of computational complexity. Since 16 bits are enough to maintain the same accuracy rate as the full precision network [Suda et al., 2016], we set precision to 16 bits for any full precision weights or activations. For the pruned network, we set 84% of the W in the pruned network as zero for a fair comparison. (Since pruned network can get up to 13x reduction [Han et al., 2015] while BNN can get 32x, the size of the pruned network is 32/13=2.5x larger. With 16-bit weights, the total number of non-zero weights of the pruned network is 2.5/16=16% of that of the binarized weight cases.) As shown in Fig. 1, BNN apparently consumes the least amount of hardware resource among all these architecture. In summary, for all the methods mentioned above, pruning can be categorized as connection reduction, while the rest can be categorized as precision reduction. However, both kinds of methods cannot be applied to the BNN. For pruning, it prunes the weights that are close to zero value. However, the weights of BNN are already constrained to -1/+1. For precision reduction, BNN has already reached the lower bound. Since CNNs are believed to have huge redundancy, we hypothesize that BNN also has redundancy and it is able to get a more compact BNN. To our best knowledge, no existing work is inspired by analyzing and reducing the input precision. We are the 1st to explore the BNN redundancy by the bit-level analysis of the input data. We will validate our hypothesis step by step in the next section. Figure 1. Resource consumption of W(10,10)×A(10,10) multiplication on a Xilinx Virtex-7 FPGA for different architecture In the following paragraphs, BNN is referring to Hubara et al.’s work, a 9-layer binarized CNN, which is our baseline model. The reconstructed BNN and CBNN is referring to the reconstructed model we used for sensitivity analysis and the final compact BNN with shrunk network size, respectively. 194 11000010 118 72 H 01110110 01001000 H Build a compact BNN 3.1 BNN reconstruction In this section, we first illustrates reformatting the input and modifying the first layer for the BNN reconstruction. Then, the training method is presented. Bit-sliced binarized input All the existing binarized or ternarized neural networks take the non-binary format data as the network input, which means the computation of the first layer is the inner product of a nonbinary matrix and a binary matrix. In all the other layers, the convolution operation can be implemented as the XNOR dotproduct operation, which is simplified as bit count in GPU or XNOR logic in FPGA/ASIC. By contrast, the computation of the first layer is much more complicated. Intuitively, the bitsliced input can enable XNOR dot-product operation in the first layer. This inspires us to explore the feasibility of converting the dataset into the bit-sliced format. A single image in the dataset can be represented as 𝑑𝑎𝑡𝑎(𝑊,𝐻,𝐶) , where W is the width, H is the height, and C is the number of channels, as shown in Fig. 2. The raw data is usually stored in the format of a non-negative integer with the maximum value of 𝐴. Then a lossless conversion from integer (fixed-point) to N-bit binary format is defined as the int2b function. = 𝑖𝑛𝑡2𝑏(𝑑𝑎𝑡𝑎(𝑊,𝐻,𝐶) , 𝑁), 0 0 0 1 1 1 1 1 1 1 1 0 0 1 0 1 0 1 1 1 1 0 C’=CxN W W First, we need to reconstruct and train a new model for sensitivity analysis. Section 3.1 demonstrates the model reconstruction of BNN. Then, with the reconstructed BNN, Section 3.2 will prove the redundancy exists in BNN and decide the prunable bit slices through bit-level sensitivity analysis in post-training stage. Finally, Section 3.3 presents the guide to rebuilding a more compact BNN (CBNN) that triggered by input data pruning. 𝑏 𝑑𝑎𝑡𝑎(𝑊,𝐻,𝐶 ′) 1 1 0 0 N bits C 3 1 0 0 0 Figure 2. Conversion from fixed-point input to bit-sliced binary input Although the network size increases, the growth is somewhat limited. For example, in a 9-layer BNN [Hubara et al., 2016], the size of the first layer is only 0.02% of the entire network. It has been proved that 16-bit quantization for the weights is enough to preserve the accuracy [Suda et al., 2016]. With the bit-slice input, the network size will slightly increase by 3%, which can be negligible. With the bit-sliced input and non-binary first layer, we reconstruct the BNN model and refer it as the reconstructed BNN. Although the computational complexity is the same, the new structure helps to reduce the redundancy in BNN, which will be elaborated in the following sections. Binary constrained training We adopt the training method proposed by Hubara et al. [Hubara et al., 2016]. The objective function is shown in Eq. 2, where 𝑊1 represents the weights in the non-binary first layer and 𝑊𝑙 represents the weights in all the other binary layers. The loss function 𝐿 here is a hinge loss. In the training stage, the full-precision reference weights 𝑊𝑙 are used for the backward propagation, and. the binarized weights 𝑊𝑙𝑏 = 𝑐𝑙𝑖𝑝(𝑊𝑙 ) [Hubara et al., 2016] are used in the forward propagation. As Tang et al. propose in [Tang et al., 2017], the reference weights in the binary layers 𝑊𝑙 (𝑙 ≥ 2) should be punished if they are not close to +1/-1. Also, a 𝐿2 regularization term is applied for the non-binary first layer. 𝐿 𝐽(𝑊𝑙 , 𝑊1 , 𝑏) = 𝐿(𝑊𝑙𝑏 , 𝑊1 , 𝑏) + 𝜆(𝑎𝑣𝑔 ∑(1 − ‖𝑊𝑙 ‖22 ) + 𝑎𝑣𝑔(‖𝑊1 ‖22 )) 𝑙=2 (2) (1) where 𝐶’ = 𝐶 × 𝑁 and 𝑁 = 𝑐𝑒𝑖𝑙(𝑙𝑜𝑔2 (𝐴 + 1)). After int2b conversion, each channel of an image is expanded to N channels in binary format. Non-binary first layer Experimental observation shows that the bit-sliced input has a negative impact on the accuracy rate. There are two main reasons. Since the input data is in the bit-sliced format, the data-preprocessing methods, e.g., mean removal, normalization, ZCA whitening, cannot be applied here, which results in an accuracy drop. In addition, compared with a standard first layer in BNN, the computational complexity drops, which may hurt the accuracy rate. Therefore, we assign the first layer with full-precision float-point weights to keep the computational complexity of the first layer the same as a standard first layer in BNN. 3.2 Sensitivity analysis We use the training method in Section 3.1 to train a reconstructed BNN model with the bit-sliced input and non-binary first layer. In the post-training stage, we evaluate the sensitivity of the bit-sliced input to the accuracy performance. As shown in Fig. 3, the reconstructed BNN is pre-trained as initial. Then, the Nth bit (Nth least significant bit) slices in RGB channels are substituted with binary random bit slices. The reason why we use binary random bit slices other than pruning is that, pruning will reduce the size of the network. We want to eliminate any other factors that can influence the accuracy performance. If the difference between the actual inference error 𝐸𝑅𝑅𝑖𝑛𝑓 and the reference point 𝐸𝑅𝑅𝑟𝑒𝑓 (∆𝐸𝑅𝑅 = 𝐸𝑅𝑅𝑖𝑛𝑓 − 𝐸𝑅𝑅𝑟𝑒𝑓 ) is less than an error-tolerant threshold 𝐸𝑅𝑅𝑡ℎ , the Nth bit slices are classified as prunable. X = binary random bits th N bit slices in RGB channels X X X X X X X X X X X X X X X XX X X X X X X X X X X XX X X X X X X XX X X X X X X X X X X X X X X X X X XX X X X X X X X X X X X X X X X X X X X X X X X X X X X X ERRth ERRref ERR Sensitive or not X X ERRinf X X X X Pre-trained reconstructed BNN Figure 3. Sensitivity analysis of the reconstructed BNN with distorted input Without retraining the network, the error brought by random bit slices will propagate throughout the entire network. Even so, there can be merely no accuracy drop in the inference stage. Then, it can be inferred that these bit slices with less sensitivity to the accuracy performance are useless in the training stage. It also indicates that the existing redundancy in BNN allows us to further shrink the network size. After evaluating the sensitivity of each bit slice, we can also analyze the sensitivity of a stack of slices by using the same method. Then we can find a collection of insensitive bit slices which are prunable in the training stage. If P out of N slices are categorized as accuracy insensitive, the number of channels C’ can be reduced by N/P times. That is to say, the size of the input array is reduced by N/P times. 3.3 Rebuild a compact network In the most popular CNN architectures, such as AlexNet [Krizhevsky et al., 2012], VGG [Simonyan and Zisserman, 2014] and ResNet [He et al., 2016], the depth incremental ratio of feature map from one layer to the next layer is either keeping as 2x or remaining the same. Intuitively speaking, it is useful to keep the same depth incremental ratio across the entire network. Thus, a good starting point of rebuilding a compact BNN (CBNN) is shrinking the depth of all the layers by N/P times. Since there is a quadratic relation between depth and the network size, the reduction of the network size of the CBNN is expected to be (N/P)2 times. Although we haven’t explored how to build an accurate model to optimize the network compression ratio, we emphasize the entire flow (presented in Section 3) that reduces the redundancy of the BNN and enables speedup in the inference stage with the CBNN. In Section 4, we will present and discuss the performance results corresponding to each subsection in Section 3. 4 Result and discussion We will first walk through the flow presented in Section 3 with experimental results on the Cifar-10 classification task in Section 4.1. Section 4.2 will present additional results on SNVH, Chars74K and GTSRB datasets. For the experiment Table 1. Performance comparison with different input format and 1st layer configuration Arch. Input First layer Network size Error rate BNN full precision binary 1x 11.6% FBNN bit slices binary 1.01x 14.0% Reconstructed BNN bit slices non-binary 1.1x 10.1% setup, we build our model based upon Hubara et al.’s BNN in Theano. The description of each dataset is listed as follow. Cifar-10 [Krizhevsky and Hinton, 2009]. This is a dataset for 10-category classification task with 32×32 RGB images. The training dataset contains 50,000 images and the testing dataset contains 20,000. SVHN (The Street View House Numbers) [Netzer et al., 2011]. This dataset is a real-world house number dataset from Google Street View images. It has 73257 digits for training and 26032 digits for testing, with the image size of 32×32. Char74K [Campos et al., 2009]. This dataset contain 62 characters (0-9, A-Z and a-z) from both natural images and synthesized images. 80% of the Char74K images serve as the training set and the rest 20% serve as the testing set, with the image size of 56×56. GTSRB (The German Traffic Sign Benchmark) [Stallkamp et al., 2011]. This dataset includes 43-class traffic signs. We resize the traffic sign images to 32×32. It has 39209 training data and 12630 testing data. 4.1 Experiment on Cifar-10 Subsections of 4.1 show the experimental results corresponding to the methodology in Section 3.1-3.3, respectively. BNN reconstruction Following the input data conversion method in Section 3.1, the raw data of Cifar-10 dataset can be denoted as Cifar(32,32,3).. Each pixel value is represented by a non-negative integer with magnitude A=255. Thus, N=ceil(log2(255+1))=8 bits are enough for lossless binary representation. Then, the bit-sliced input can be denoted as Cifarb(32,32,24). As illustrated in Section 3.1, the proposed structure of the reconstructed BNN is different from the original BNN in both input format and the first layer. Table 1 compares the performance results of three network structure with different input and 1st layer. The baseline BNN design is the one in [Hubara et al., 2016], with full precision input and a binarized 1st layer. Here we define a CNN with bit-sliced input, binarized weights and activations as FBNN. FBNN has bit slices input but BNN does not. By training with the method in Section 3.1, FBNN shows 2.4% in the accuracy drop, compared with BNN. The accuracy here is affected by computational complexity degradation in the 1st layer and unnormalized input data. It also gives us some insights that The FBNN is hard to get a good accuracy rate, which is in accord with Tang et al.’s Table 2. Sensitivity analysis of single bit slice in each channel with random noise injected Arch. N th bit ERR/% ∆ERR/% Arch. N th bit ERR/% ∆ERR/% BNN 0 11.6 0.0 FNN 0 10.4 0.0 0 10.1 -1.5 0 10.4 0.0 1 9.8 -1.9 1 10.4 0.0 2 10.0 -1.6 2 10.4 0.1 Recon. 3 10.1 -1.6 3 10.4 0.1 FNN BNN 4 10.5 -1.2 4 10.9 0.5 5 12.5 0.8 5 13.0 2.6 6 20.9 9.2 6 21.4 11.1 7 40.3 28.6 7 43.8 33.4 Table 3. Sensitivity analysis of 1-Nth multiple bit slices in each channel with random noise injected th th Arch. 1-N ERR/% ∆ERR/% Arch. 1-N bits bits BNN 0 11.6 0.0 FNN 0 0 10.1 -1.5 0 1 9.8 -1.9 1 1-2 9.9 -1.7 1-2 Recon. 1-3 9.9 -1.8 1-3 FNN BNN 1-4 10.7 -0.9 1-4 1-5 13.6 1.9 1-5 1-6 24.3 12.6 1-6 1-7 46.1 34.5 1-7 ERR/% ∆ERR/% 10.4 10.4 10.4 10.5 10.5 11.3 14.4 23.3 54.1 0.0 0.0 0.1 0.2 0.2 1.0 4.1 13.0 43.7 opinion in [Tang et al., 2017]. By introducing bit slices input and non-binary 1st layer to reconstruct the BNN (as we proposed in Section 3.1), the accuracy drop can be compensated as shown in Table 1. We can even get a better error rate than the baseline BNN with a slightly increased network size. It also gives more margin in compressing the network. Sensitivity analysis of the reconstructed BNN With a pre-trained reconstructed BNN presented in the last section, now we can do bit-level sensitivity analysis as stated in Section 3.2. First, we analyze the sensitivity of a single bit slice. The results are shown in Table 2. The data shows in Table 2 is the Figure 4. Error rate of randomizing one or multiple bit slices in sensitivity analysis Table 4. Performance of CBNNs on Cifar-10 Network size GOPs th ERR ∆ERR Arch. 1-N % MB CP. ratio # CP. ratio bits % BNN CBNN 0 1 2 3 4 5 11.6 10.3 10.6 10.8 11.8 14.2 0.0 -1.3 -1.0 -0.8 0.2 2.6 1.75 1.38 1.02 0.71 0.45 0.25 1x 1.3x 1.7x 2.5x 3.9x 7.0x 1.23 0.98 0.72 0.50 0.32 0.18 1x 1.3x 1.7x 2.5x 3.8x 6.8x average over 10 trials. In addition to the reconstructed BNN, we also evaluate the bit-level sensitivity of the input with its full-precision counterpart, which is denoted as FNN. With FNN, we intend to show that the data itself has redundancy, which can be reflected in both binary domain or fixed-point domain with the same pattern. We take the architecture in the first row as the reference design. The 1st row of ERR column is the ERRref and the others are ERRinf. ∆ERR=ERRinf - ERRref. BNN is the reference design for the reconstructed BNN. FNN with non-distorted input is the reference design for the fullprecision ones. It is interesting that the 1st, 2nd and 3rd bit slices are at the same sensitivity level, concluded from the almost unchanged ∆ERR. We define the turning point of error in sensitivity analysis as the point where ∆ERR flips the sign or increases abruptly. The turning point here is the 5 th bit. Second, we analyze the sensitivity of bit slices stacks. Each stack contains 1st to Nth bit slices in each color channel. The results are shown in Table 3. For the 1st, 2nd and 3rd bit slices, it makes no difference if distortion is injected in one of them or all of them. The 4th makes a slight difference of around 0.5% accuracy drop and the 5th bit is also the turning point with around 3% accuracy drop. Even when we randomize 50% of the entire input values (1st to 4th bit slices) and the variation propagates through the entire network, the accuracy doesn’t change much. Therefore, these bits are useless in the training stage. This validates the hypothesis that the BNN still has redundancy. In Fig. 4, the error rate turning point is circled at the 5 th bit slice. The trend of error rate in the binary domain and full-precision domain (shown in Fig. 4) align well. In order to make the entire process be automatic, we can simple set an error-tolerant threshold ERRth to determine how many bits are prunable. Here, ERRth is set to 1%. We can conclude that 1st-4th bit slices here are redundant and prunable through bit-level sensitivity analysis. Accordingly, the reconstructed BNN can be shrunk to reduce the redundancy and get a more compact architecture. Rebuild a compact BNN (CBNN) Since 4 out of 8 bit slices are prunable, we can rebuild a compact BNN with the depth of each layer shrunk by half. The performance of CBNN is shown in Table 4. CP. Ratio represents compression ratio and GOPs stands for Giga operations (one operation is either an addition or a multiplication). Regarding the network size, we use 16 bits for measuring non-binary weights in the 1st layer, since it has been proved that 16 bit is enough to maintain the same accuracy [Suda et Table 5. Performance results of CBNNs on SVHN/Chars47k/GTSRB datasets Dataset SVHN Chars47k GTSRB th 1-N ERR ∆ERR % bits % 0 1 2 3 4 0 1 2 3 4 0 1 2 3 4 4.8 4.9 5.1 5.0 6.6 15.4 15.3 15.3 15.2 16.3 1.0 1.0 1.2 1.6 2.0 0.0 0.1 0.3 0.2 1.8 0.0 -0.1 -0.1 -0.2 1.0 0.0 0.0 0.2 0.6 1.0 Network size CP. MB ratio 0.44 1x 0.36 1.2x 0.26 1.7x 0.18 2.4x 0.12 3.7x 0.44 1x 0.36 1.2x 0.26 1.7x 0.18 2.4x 0.12 3.7x 1.81 1x 1.39 1.3x 1.02 1.8x 0.71 2.5x 0.46 3.9x GOPs CP. # ratio 0.31 1x 0.26 1.2x 0.19 1.6x 0.13 2.4x 0.08 3.7x 0.31 1x 0.26 1.2x 0.19 1.6x 0.13 2.4x 0.08 3.7x 3.89 1x 2.98 1.3x 2.19 1.8x 1.52 2.6x 0.97 4.0x al., 2016]. We also show the alternatives of pruning 1-Nth (𝑁 = 1, 2, … , 5) bit slices and shrink the layerwise depth by 1/8 to 5/8. The results align with the sensitivity analysis that 1-3rd bit slices have little impact on the classification performance. The choice of pruning 1-4th bit slices is the optimal one to maximize the compression ratio with <1% accuracy drop. Since the size of the 1st layer is larger than that of BNN, we cannot achieve the ideal network size compress ratio (4x) regarding the entire network. The actual compression ratio of the network size is 3.9x and the compression ratio of number of GOPs is 3.8x. 4.2 Experiment on SVHN/Chars74K/GTSRB datasets In this section, we will skip the sensitivity analysis and just show the result comparison between the baseline and the final CBNNs we get in the same procedure. For SVHN and Char74K datasets, we use a baseline architecture that has half of the depth in each layer as the one for Cifar-10. For GTSRB, we use a baseline architecture that has Unit: ms 53 30 31 1.7x 15 2x 15 9 15 11 1.7x 1.4x (a) FNN 1x Pruned+quantized 4x BNN 5x 9.9 x 8.7 x 8.4 x 7.0x (b) CBNN-Cifar10 CBNN-GTSRB CBNN-Char47k CBNN-SVHN Figure 5. Runtime comparison of different network compression technique the same filter configuration as the one for Cifar-10. Since the input size of GTSRB is larger than Cifar-10, so the network for GTSRB has the same depth but larger width and height in each layer. In Table 5, it shows the performance results of CBNNs evaluating on different datasets and network setting. The baseline for each dataset is shown in the first row of each dataset region. For Chars47k and GTSRB, the CBNNs are able to maintain no more than 1% accuracy drop, achieving 3.7x and 3.9x network size reduction, respectively. For SVHN dataset, the accuracy drop between pruning 1-3rd bits and pruning 1-4th bits is large. In order to preserve no more than 1% accuracy drop, the network size reduction is yield to 2.4x. 4.3 Runtime evaluation We evaluate the actual runtime performance of CBNNs by Nvidia GPU Titan X. The batch size is fixed as 128 in all the experiments. We use the same XNOR-based GPU kernel as [Hubara et al., 2016] for CBNN implementation. The computational time is calculated by averaging over 10 runs. Fig. 5 (a) illustrates the actual runtime and runtime speedup of 4 CBNNs compared with their baseline BNNs. The configurations are the same as the highlight ones in Table 4 and Table 5. For the CBNNs processing Cifar10, GTSRB and Char47k datasets, their network size and total GOPs shrink 3.7-4.0x, resulting in the speedup of 1.7-2.0x. For the CBNN processing the SVHN dataset, its network size and total GOPs shrinks 2.4x, resulting in a speedup of 1.4x. In Fig. 5 (b), we normalize the runtime performance to a full-precision CNN (FNN). It is proved in [Han et al., 2016], combining pruning, quantization and Huffman coding technique, an FNN can achieve up to 4x speedup, which is shown in the green bar. Hubara et al. demonstrate that a multilayer perceptron BNN can get 5x speedup compared with its full-precision counterpart. On top of the BNN, the proposed CBNN can give extra 1.4-2.0x speedup. Therefore, the CBNN can achieve 7.0-9.9x speedup compared with FNN. 5 Conclusion In this paper, we propose a novel flow to explore the redundancy of BNN and remove the redundancy by bit-level sensitivity analysis and data pruning. In order to build a compact BNN, one should follow these three steps. Specifically, first reconstruct a BNN with bit-sliced input and non-binary 1st layer. Then, inject randomly binarized bit slices to analyze the sensitivity level of each bit slice to the classification error rate. After that, prune P accuracy insensitive bit slices out of total N slices and rebuild a CBNN with depth shrunk by (N/P) times in each layer. The experiment results show that the error variation trend in sensitivity analysis of the reconstructed BNN is well aligned with that of CBNN. In addition, the CBNN is able to get 2.4-3.9x network compression ratio and 2.4-4.0x computational complexity reduction (in terms of GOPs) with no more than 1% accuracy loss compared with BNN. The actual runtime can be reduced by 1.4-2x and 7.09.9x compared with the baseline BNN and its full-precision counterpart, respectively. References [Courbariaux et al., 2015] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. "Binaryconnect: Training Deep Neural Networks with Binary Weights During Propagations." In Advances in Neural Information Processing Systems, pages 3123-3131, 2015. [Hubara et al., 2016] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized Neural Networks. In Advances in Neural Information Processing Systems, pages 4107-4115, 2016. [Rastegari et al., 2016] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. XNOR-net: Imagenet Classification Using Binary Convolutional Neural Networks. In European Conference on Computer Vision, pages 525-542. Springer, Cham, 2016. [Zhu et al., 2016] Chenzhuo Zhu, Song Han, Huizi Mao, and William J. Dally. Trained Ternary Quantization. arXiv preprint arXiv:1612.01064, 2016. [Han et al., 2015] Song Han, Jeff Pool, John Tran, and William Dally. Learning both Weights and Connections for Efficient Neural Network. In Advances in Neural Information Processing Systems, pages 1135-1143. 2015. [Krizhevsky et al., 2012] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet Classification with Deep Convolutional Neural Networks. In Advances in neural information processing systems, pages 1097-1105. 2012. [Simonyan and Zisserman, 2014] Karen Simonyan, and Andrew Zisserman. Very Deep Convolutional Networks for Large-scale Image Recognition. arXiv preprint arXiv:1409.1556, 2014. [He et al., 2016] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778. 2016. [Suda et al., 2016] Naveen Suda, Vikas Chandra, Ganesh Dasika, Abinash Mohanty, Yufei Ma, Sarma Vrudhula, Jae-sun Seo, and Yu Cao. Throughput-optimized OpenCL-based FPGA Accelerator for Large-scale Convolutional Neural Networks. In Proceedings of the 2016 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays, pages 16-25. ACM, 2016. [Tian et al., 2015] Yonglong Tian, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Pedestrian Detection ASided by Deep Learning Semantic Tasks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5079-5087. 2015. [Schroff et al., 2015] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A Unified Embedding for Face Recognition and Clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. [Hu et al., 2015] Guosheng Hu, Yongxin Yang, Dong Yi, Josef Kittler, William Christmas, Stan Z. Li, and Timothy Hospedales. When Face Recognition Meets with Deep Learning: an Evaluation of Convolutional Neural Networks for Face Recognition. In Proceedings of the IEEE international conference on computer vision workshops, pages 142-150, 2015. [Tang et al., 2017] Wei Tang, Gang Hua, and Liang Wang. How to Train a Compact Binary Neural Network with High Accuracy? In AAAI, pages 2625-2631, 2017. [Lumb et al., 2017] D. Lumb. Apple 'Neural Engine' Chip Could Power AI on iPhones. https://www.engadget.com/2017/05/26/apple-neural-engine-chip-couldpower-ai-on-iphones/. [Li et al., 2017] Yixing Li, Zichuan Liu, Kai Xu, Hao Yu, and Fengbo Ren. A 7.663-TOPS 8.2-W Energy-efficient FPGA Accelerator for Binary Convolutional Neural Networks. In Proceedings of the 2017 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays, pages 290-291. ACM, 2017. [Krizhevsky and Hinton, 2009] Alex Krizhevsky, and Geoffrey Hinton. Learning Multiple Layers of Features from Tiny Images. http://www.cs.utoronto.ca/~kriz/learningfeatures-2009-TR.pdf. 2009. [LeCun, 1998] Yann LeCun. The MNIST Database of HandWritten Digits. http://yann. lecun. com/exdb/mnist/, 1998. [Netzer et al., 2011] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading Digits in Natural Images with Unsupervised Feature Learning. In NIPS workshop on deep learning and unsupervised feature learning, vol. 2011, no. 2, p. 5. 2011. [LeCun et al., 2015] Yann LeCun, Yoshua Bengio, and G. Hinton. Deep learning. Nature 521, no. 7553 (2015): 436444. [Alpaydin et al., 2014] Ethem Alpaydin. Introduction to Machine Learning. MIT press, 2014. [Campos et al., 2009] Teo de Campos, Bodla Rakesh Babu, and Manik Varma. In Proceedings of the International Conference on Computer Vision Theory and Applications, 2009 [Stallkamp et al., 2011] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. The German Traffic Sign Recognition Benchmark: a Multi-class Classification Competition. In Neural Networks (IJCNN), The 2011 International Joint Conference on, pages 1453-1460. IEEE, 2011. [Han et al., 2016] Song Han, Huizi Mao, and William J. Dally. Deep compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding. International Conference on Learning Representations, 2016.
1cs.CV
Hierarchical Reinforcement Learning: Approximating Optimal Discounted TSP Using Local Policies Tom Zahavy 1 2 Avinatan Hasidim 1 3 Haim Kaplan 1 4 Yishay Mansour 1 4 arXiv:1803.04674v1 [cs.LG] 13 Mar 2018 Abstract In this work, we provide theoretical guarantees for reward decomposition in deterministic MDPs. Reward decomposition is a special case of Hierarchical Reinforcement Learning, that allows one to learn many policies in parallel and combine them into a composite solution. Our approach builds on mapping this problem into a Reward Discounted Traveling Salesman Problem, and then deriving approximate solutions for it. In particular, we focus on approximate solutions that are local, i.e., solutions that only observe information about the current state. Local policies are easy to implement and do not require substantial computational resources as they do not perform planning. While local deterministic policies, like Nearest Neighbor, are being used in practice for hierarchical reinforcement learning, we propose three stochastic policies that guarantee better performance than any deterministic policy. 1. Introduction One of the unique characteristics of human problem solving is the ability to represent the world on different granularities. When we plan a trip, we first choose the destinations we want to visit and only then decide what to do at each destination. Hierarchical reasoning enables us to map the complexities of the world around us into simple plans that are computationally tractable to reason. Nevertheless, the most successful Reinforcement Learning (RL) algorithms are still performing planning with only one abstraction level. RL provides a general framework for optimizing decisions in dynamic environments. However, scaling it to real-world problems suffers from the curses of dimensionality; that is, coping with exponentially large state spaces, action spaces, and long horizons. One approach deals with large state spaces by introducing a function approximation to the value 1 Google 2 The Technion 3 Bar Ilan Univeristy 4 Tel Aviv University. Correspondence to: Tom Zahavy <[email protected]>. function or policy, making it possible to generalize across different states. Two famous examples are TD-Gammon (Tesauro, 1995) and the Deep Q Network (DQN) (Mnih et al., 2015), both introduced a Deep Neural Network (DNN) to approximate the value function leading to a high performance in solving Backgammon and video games. A different approach deals with long horizons by using a policy network to search among game outcomes efficiently (Silver et al., 2016), leading to a super-human performance in playing Go, Chess, and Poker (Silver et al., 2016; 2017; Moravčı́k et al., 2017). However, utilizing this approach when it is not possible to simulate the environment by doing model-based RL is still an open problem (Oh et al., 2015). A long-standing approach for dealing with long horizons is to introduce hierarchy into the problem (see Barto & Mahadevan (2003) for a survey). We will focus on the options framework (Sutton et al., 1999), a two-level hierarchy formulation where options (local policies that map states to actions) are learned to achieve subgoals, while the policy over options selects among options to accomplish the final goal of a task. Recently, it was demonstrated that learning a selection rule among pre-defined options using a DNN delivers promising results in challenging environments like Minecraft and Atari (Tessler et al., 2017; Kulkarni et al., 2016; Oh et al., 2017); other studies have shown that it is possible to learn options jointly with a policy-over-options end-to-end (Vezhnevets et al., 2017; Bacon et al., 2017). In this work, we focus on a specific type of hierarchy - reward function decomposition - that dates back to the works of (Humphrys, 1996; Karlsson, 1997) and has been studied among different research groups recently (van Seijen et al., 2017). In this formulation, each option i learns to maximize a local reward function Ri , while the P final goal is to maximize the sum of rewards RM = Ri . Each option is trained separately and provides a value function and an option oi for the policy over options, which then uses the local values to select among options. That way, each option is responsible for solving a simple task, and the options are learned in parallel across different machines. While the higher level policy can be trained using SMDP algorithms (Sutton et al., 1999), different research groups suggested using pre-defined rules to select among options. Approximating Optimal Discounted TSP Using Local Policies For example, choosing the option with maximal value function (Humphrys, 1996; Barreto et al., 2017), or choosing the action that maximizes the sum of option value functions (Karlsson, 1997). By using pre-defined rules, we can derive policies for MDP M by learning options (and without learning in MDP M ), such that learning is fully decentralized. Although in many cases one can reconstruct from the options the original MDP, doing it would defeat the entire purpose of using options. In this work we concentrate on local rules that select among available options. Even more specifically, we consider a set of n MDPs {Mi }ni=1 = {S, A, P, γ, Ri }ni=1 with deterministic dynamics that share all components but the reward. Given a set of options, one per reward, with an optimal policy for collecting the reward, we are interested in deriving an optimal policy for collecting all the rewards, i.e., solving MDP Pk M = {S, A, P, RM = i=1 Ri }. In this setting, an optimal policy for M can be derived by solving the SMDP Pk Ms = {S, O, P, RM = i=1 Ri }, whose actions are the optimal policies for collecting single rewards. Specifically, we focus on collectible rewards, a special type of reward that is very common in 2D and 3D navigation domains like Minecraft (Tessler et al., 2017; Oh et al., 2017), DeepMindLab (Teh et al., 2017; Beattie et al., 2016) and VizDoom (Kempka et al., 2016). The challenge when dealing with collectible rewards is that the state space changes each time we collect a reward (One can think of the subset of available rewards is part of the state). Since all the combinations of remaining items have to be considered, the state space grows exponentially with the number of rewards. Here, we show that solving an SMDP under these considerations is equivalent to solving a Reward Discounted Traveling Salesman Problem (RD-TSP) (Blum et al., 2007). Similar to the classical Traveling Salesman Problem (TSP), computing an optimal solution to RD-TSP is NP-hard, and furthermore, it is NP-hard to get an approximate solution of value at least 99.5% the optimal discounted return (Blum et al., 2007) in polynomial time.1 A brute force approach for solving the RD-TSP requires evaluating all the n! possible tours connecting the n rewards. We can also adapt the Bellman–Held–Karp dynamic programming algorithm for TSP (Bellman, 1962; Held & Karp, 1962) to solve RD-TSP (see Algorithm 4 in the appendix). This scheme is identical to tabular Q-learning on SMDP Ms , and still requires exponential time.2 1 That is, any algorithm for approximating the optimal return for RD-TSP to within a factor larger than 0.995 must have a worstcase running time that grows faster than any polynomial (assuming the widely believed conjecture that P 6= NP). 2 The Hardness results for RD-TSP do not rule out efficient solutions for special MDPs. For example, we provide, in the appendix, exact polynomial-time solutions for the case in which Blum et al. (2007) proposed a polynomial time planning algorithm for RD-TSP that computes a policy which collects at least 0.15 fraction of the optimal discounted return, which was later improved to 0.19 (Farbstein & Levin, 2016). These planning algorithms need to know the entire SMDP in order to compute their approximately optimal policies. In contrast, in this work, we focus on deriving and analyzing policies that use only local information3 to make decisions; such local policies are simpler to implement, more efficient, and do not need to learn in M nor Ms . The reinforcement learning community is already using simple local approximation algorithms for RD-TSP. We hope that our research will provide important theoretical support for comparing local heuristics, and in addition introduce new reasonable local heuristics. Specifically, we prove worst-case guarantees on the reward collected by these algorithms relative to the reward of the optimal RD-TSP tour. We also prove bounds on the maximum relative reward that such local algorithms can collect. In our experiments, we compare the performance of these local algorithms. In particular, our main contributions are as follows. Our results: We establish impossibility results for local policies, showing that no deterministic local policy can guarantee a reward larger than 24OPT/n for every MDP, and no √ stochastic policy can guarantee a reward larger than 8OPT/ n for every MDP. These impossibility results imply that the Nearest Neighbor (NN) algorithm that iteratively collects the closest reward (and thereby a total of at least OP T /n reward) is optimal up to constant factor amongst all deterministic local policies. On the positive side, we propose three simple stochastic policies that outperform NN. The best of them combines NN with a Random Depth First Search √(RDFS) and guarantees performance of at least Ω (OPT/ n) when OPT achieves 2 Ω (n), and at least Ω(OPT/n 3 ) in the general case. Combining NN with jumping to a random reward and sorting the rewards by their distance from it, has a slightly worse guarantee. A simple modification of the NN to first jump to a random reward and continues NN from there, already improves the guarantee to O(OP T log(n)/n). 2. Problem formulation We now define our problem explicitly, starting from a general transfer framework in Definition 1, and then the more specific transfer learning setting of of collectible reward decomposition in Definition 2. Definition 1 (General Transfer Framework) Given a set of MDPs {Mi }ni=1 = {S, A, P, γ, Ri }ni=1 and an MDP M = {S, A, P, γ, RM = f (R1 , ..., Rn )} that the MDP is a line and when it is a star. 3 Observe only the value of each option from the current state. Approximating Optimal Discounted TSP Using Local Policies differ only by their reward signal, derive an optimal policy for M given the optimal policies for {Mi }ni=1 . Definition 1 describes a general transfer learning problem in RL. Similar to (Barreto et al., 2017), our transfer framework assumes a set of MDPs sharing all but the reward signal. We are interested in transfer learning, i.e., using quantities that were learned from the MDPs {Mi }ni=1 on the MDP M . More specifically for model-free RL, given a set of optimal options and their value functions {o∗i , Vi∗ }ni=1 , we are interested in zero-shot transfer to MDP M , i.e., deriving policies for solving M without learning in M . Definition 2 (Collectible Reward Decomposition) 1. Reward Decomposition: the reward Pn in M represents the sum of the local rewards: RM = i=1 Ri . 2. Collectible Rewards: each reward signal {Ri }ni=1 represents a collectible prize, i.e., Ri (s, a) = 1 iff s = si , a = ai for some particular state si and action ai and Ri (s, a) = 0 otherwise. In addition, each reward can only be collected once. 3. Deterministic Dynamics: P is a deterministic transition matrix, i.e., for each action a each row of P a has exactly one value that equals 1, and all its other values equal zero. Property 1 in Definition 2 requires RM to be a decomposition of the previous rewards, and Property 2 requires each local reward to be a collectible prize. While limiting the generality, models that satisfy these properties has been investigated in theory and simulation (Oh et al., 2017; Barreto et al., 2017; Tessler et al., 2017; Higgins et al., 2017; van Seijen et al., 2017; Humphrys, 1996). Now, given that the value functions of the local policies are optimal, the shortest path from a reward i to a reward j is given by following option oj from state i. In addition, the length of the shortest path from i to j, denoted by di,j , is given by the value function, since Vj (i) = γ di,j . Notice that in any state, an optimal policy on M will always follow the shortest path to one of the rewards. To see this, assume there exists a policy µ that is not following the shortest path from some state k to the next reward-state k 0 . Then, we can improve µ by taking the shortest path from k to k 0 , contradicting the optimality of µ. The last observation implies that an optimal policy on M is a composition of the local options {oi }ni=1 . Property 3 in Definition 2 requires deterministic dynamics. This property is perhaps the most limiting of the three, but again, it appears in numerous domains including many maze navigation problems, the Arcade Learning Environment (Bellemare et al., 2013), and games like Chess and Go. Given that P is a deterministic transition matrix, an optimal policy on M will make decisions only at states which contain rewards. In other words, once the policy arrived at a reward state i and decided to go to a reward-state j it will follow the optimal policies πj until it reaches j.4 For collectible reward decomposition (Definition 2), an optimal policy for M can be derived on an SMDP (Sutton et al., 1999) denoted by Ms . The state space of Ms contains only the initial state s0 and the reward states {si }ni=1 . The action space is replaced by the set of options {oi }ni=1 , where oi corresponds to following the optimal policy in Mi until reaching state si . In addition, in this action space, the transition matrix P is deterministic since ∀s, ai , ∃s0 such ai ai that Ps,s 0 = 1, and otherwise Ps,· = 0. Finally, the reward signal and the discount factor remain the same. In general, optimal policies for SMDPs are guaranteed to be optimal in the original MDP only if the SMDP includes both the options and the regular (primitive) actions (Sutton et al., 1999). In a related study, Mann et al. (2015) analyzed landmark options, a specific type of options that plan to reach a state on a deterministic MDP. While landmark options are not related to reward decomposition or collectible rewards, they also represent policies that plan to reach a specific state in an MDP. Given a set of landmark options, Mann et al. analyzed the errors from searching the optimal solution on the SMDP (planning with landmark options) instead of searching it in the original MDP (planning with primitive actions). Under Definition 2, and given that all of our policies and value functions are optimal, all these errors equal zero. Thus, the solution to the SMDP is guaranteed to be optimal to the MDP. In addition, the analysis of Mann et al. (2015) provides bounds for dealing with sub-optimal options and nondeterministic dynamics that may help to extend our analysis of these cases in future work. Finally, an optimal policy on Ms can be derived by solving an RD-TSP (Definition 3). To see this, look at a graph that includes only the initial state and the reward states. Define the length di,j of an edge ei,j in the graph to be Vj (i), i.e., the value of following option j from state i. A path in the graph is defined by a set of indices {it }nt=1 and the length of Pn−1 the path is given by t=0 dit ,it+1 . Definition 3 (RD-TSP) Given an undirected graph with n nodes and edges of length ei,j , find a path in the graph that maximizes the discounted cumulative return: {i∗t }nt=1 = arg max {it }n t=1 ∈perm{1,..,n} n−1 X γ Pj t=0 dit ,it+1 j=0 To summarize, our modeling approach allows us to deal with the curse of dimensionality in three different ways. First, each option can be learned with function approximation techniques, e.g., (Tessler et al., 2017; Bacon et al., 2017) to 4 Note that this is not true if P is stochastic. To see this, recall that stochastic shortest path is only shortest in expectation. Thus, stochasticity may lead to states that require changing the decision. Approximating Optimal Discounted TSP Using Local Policies deal with raw, high dimensional inputs like vision and text. Second, formulating the problem as an SMDP reduces the state space to include only the reward states and effectively reduces the planning horizon (Sutton et al., 1999; Mann et al., 2015). And third, under the RD-TSP formulation, we derive approximate solutions for dealing with the exponentially large state spaces that emerge from modeling one-time events like collectible rewards. The basic idea of the proof (in the supplementary material) is that the NN has the first reward larger than any reward that OPT collects. Next, we propose a simple, easy to implement, stochastic adjustment to the vanilla NN algorithm with a better upper bound which we call R-NN (for Random-NN). The algorithm starts by collecting one of the rewards, say s1 , at random, and continues by executing NN (Algorithm 1). 3. Local heuristics Algorithm 1 R-NN: NN with a First Random Pick We start by defining a local policy. All the policies which we analyze are local policies. A local policy is a mapping who inputs are: (1) the current state x, (2) the history h containing the previous steps taken by the policy; in particular h encodes the collectible states that we had already visited, and (3) the discounted return for each reward from the current state, i.e., {Vi (x)}ni=1 , and whose output is a distribution over the options. Notice that a local policy does not have full information on the graph (but only on local distances). Formally, Definition 4 (Local policy) A local policy πlocal is a mapping: πlocal (x, h, {Vi (x)}ni=1 ) → ∆({oi }ni=1 ), Input: Graph G, with n nodes, and s0 the first node Flip a coin if outcome = heads then # Perform NN Visit a node at random, denote it by s1 end if Follow by executing NN The following theorem shows that our stochastic modification to NN improves its guarantees by a factor of log(n). Theorem 2 (R-NN Performance) For a discounted reward  traveling salesman graph with n log(n) R-NN nodes: . OPT ≥ Ω n Due to space considerations, most of the proofs are found in the supplementary material. While the improvement over NN may seem small (log(n)) the observation that stochasticity improves the performance guarantees of local policies is essential to our work. In the following sections, we derive more sophisticated randomized algorithms with better performance guarantees. 3.1. NN performance 3.2. Impossibility Results We start with an analysis of one of the most natural heuristics for the TSP, the famous NN algorithm. In the context of our problem, NN is the policy selecting the option with the highest estimated value, exactly like GPI (Barreto et al., 2017). We shall abuse the notation slightly and use the same name (e.g., NN) for the algorithm itself and its value; no confusion will arise. For TSP (without discount) in general graphs, we know that (Rosenkrantz et al., 1977): 3.2.1. D ETERMINISTIC L OCAL P OLICIES where ∆(X) is the set of distributions over a finite set X. 4 NN 1 1 1 log2 (n + 1) + ≤ ≤ dlog2 (n)e + 3 9 OPT 2 2 However, for RD-TSP the NN algorithm only guarantees a value of OPn T , as Theorem 1 states. In the next subsection, we prove a lower bound for deterministic local policies (such as NN) of O OPn T , that implies that NN is optimal for deterministic policies. The observation that NN is at least OP T /n of OPT motivated us to use NN as a component in our stochastic algorithms. Theorem 1 (NN Performance) For any discounted reward traveling salesman graph with n NN 1 nodes, and any discount factor γ: OPT ≥ n . In the previous subsection, we saw that the NN heuristic guarantees performance of at least OPT n . Next, we show an impossibility result for all deterministic local policies, indicating that no such policy can guarantee more than OPT n , which makes NN optimal over such policies. Theorem 3 (Impossibility for Deterministic Local Policies) For any deterministic local policy D-Local, there exists a graph with n nodes and a discount factor γ = 1 − n1 such D-Local 24 that: OPT ≤ n . Proof sketch. Consider a family of graphs, G, each of which consists of a star with a central vertex and n leaves at a distance d. The starting vertex is the central vertex, and there is a reward at each leaf. Each graph of the family G corresponds to a different subset of n/2 of the leaves which we connect (pairwise) by edges of length 1  d. (The other n 2 leaves are only connected to the central vertex.) While at the central vertex, local policy cannot distinguish among the n rewards (they all at the same distance from the origin), and therefore its choice is the same for all graphs in G. It Approximating Optimal Discounted TSP Using Local Policies follows that, for any given policy, there exists a graph in G such that the adjacent n/2 rewards are visited last. It 24 follows from simple algebra that D-Local OPT ≤ n . 3.2.2. S TOCHASTIC L OCAL P OLICIES In the previous subsection, we saw that deterministic local policies could only guarantee OPT n . We then showed that NN is optimal over such policies and that a small stochastic adjustment can improve its guarantees. These observations motivated us to look for better local policies in the broader class of stochastic local policies. We begin by providing a better impossibility result for such policies in Theorem 4. Theorem 4 (Impossibility for Stochastic Local Policies) For each stochastic local policy S-Local, there exists a graph with n nodes and a discount factor γ = 1 − √1n such S-Local √8 that: OPT ≤ n . The proof (in the supplementary material) is similar to √ the previous one but considers a family of graphs where n leaves are connected to form a clique (instead of n/2). We do not have a policy that achieves this lower bound, but we now propose and analyze two stochastic policies (in addition to the R-NN) that substantially improve over the deterministic upper bound. As we will see, these policies satisfy the Occam’s razor principle, i.e., policies with better guarantees are also more complicated and require more computational resources. 3.3. NN with Randomized DFS (RDFS) We now describe the NN-RDFS policy (Algorithm 2), the best performing local policy we were able to derive. The policy performs NN with probability 0.5 and local policy which we call RDFS with probability 0.5. RDFS starts at a random node and continues by performing a DFS on edges shorter than θ, where θ is chosen at random as we specify later. When it runs out of edges shorter than θ then RDFS continues by performing NN. Algorithm 2 NN with RDFS Input: Graph G, with n nodes, and s0 the first node Let x = log 1 (2) γ Flip a coin if outcome = heads then # Perform RDFS Visit a node at random, denote it by s1 Choose at random i ∼ Uniform{1, 2, ..., log2 (n)} Fix n0 = 2ni and set θ = √xn0 Initiate a DFS from s1 on edges shorter than θ end if Follow by executing NN The performance guarantees for the NN-RDFS method are stated in Theorem 5. The analysis is conducted in three steps. In the first two steps, we assume that OPT achieved a value of Ω(n0 ) by collecting n0 rewards at a segment of length x ≤ log 1 (2). The first step considers the case γ where n0 = Ω(n), and in the second step we remove this requirement and analyze the performance of NN-RDFS for the worst value of n0 . The third step considers all the value collected by OPT (not necessarily in a segment of length x) and completes the proof. In the second and the third steps we loose two logarithmic factors. One since we use a segment of length x in which OPT collects value of at least OPT/ log(n), and the second for guessing a good enough approximation for n0 (for setting θ). Theorem 5 (NN-RDFS Performance) For any instance of RD-TSP with n rewards    1  n− 2  , if OPT = Ω(n). Ω log  2 (n) NN-RDFS   ≥ 2  OPT n− 3  , otherwise. Ω log 2 (n) Proof. Step 1. Assume that OPT collects a set SOPT of αn rewards for some 0 ≤ α ≤ 1, in a segment p of length x = log 1 (2) (i.e. x is the distance from the first reward γ to the last reward – it does not include the distance from the starting point to the first reward). Let dmin , dmax the shortest and longest distances from s0 to a reward in SOPT respectively. By the triangle inequality, dmax − dmin ≤ x. We further assume that OP T ≤ O(γ dmin αn) (i.e., That is the value that OPT collects from rewards which are √ not in SOPT is√ negligible). We now show that RDFS is Ω( n) for θ = x/ n. We start with the following Lemma. Lemma 1 For any path p of length x, and ∀θ ∈ [0, x], there are less than xθ edges in p that are larger than θ. Proof. For contradiction, assume there are more than edges longer than θ. The length of p is given by X i pi = X pi + pi ≤θ X pi >θ pi ≥ X pi ≤θ pi + x θ x θ > x θ thus a contradiction to the assumption that the path length is at most x. Lemma 1 assures that after pruning all edges larger than θ (from the graph), there are at most xθ Connected Comx θ ponents (CCs) {Cj }j=1 in SOPT . In addition, it holds that P xθ j=1 |Cj | = αn, and all the edges inside any connected component Cj are shorter than θ. Next, we (lower) bound the total gain of RDFS. Say that RDFS starts at a reward in component Cj . Then, since all edges in Cj are shorter than θ, it collects either all the rewards in Cj , or at least x/2θ rewards. Thus, RDFS collects Approximating Optimal Discounted TSP Using Local Policies  Ω min{|Cj |, xθ } rewards. Notice that the first random step leads RDFS to a vertex in |C | CC Cj with probability nj . If more than half of rewards x are in CCs s.t |Cj | ≥ θ , then x RDFS ≥γ dmax θ X |Cj | j=1 ≥γ dmax n xo · min |Cj |, n θ X j:|Cj |≥ x θ αx |Cj | x · ≥ γ dmax . n θ 2θ If more than half of rewards in SOPT are in CCs such that |Cj | ≤ xθ , let s be the number of such CCs and notice that s ≤ xθ . We get that: Step 3. Finally, we consider the general case where OPT may collect its value in a segment of length larger than x. Notice that the value which OPT collects from rewards that follow the first log2 (n) segments of length x in its tour is at most 1 (since γ log2 (n)·x = n1 ). This means that there exists at least one segment of length x in which OPT collects at least logOPT of its value. Combining this with the analysis 2 (n) in the previous step, the proof is complete. 3.4. NN with a Random Ascent (RA) x θ X |Cj | n xo RDFS = γ · min |Cj |, n θ j=1   s X |Cj |2 s dmax  1 X dmax 2 ≥γ ≥ γ |Cj | n n s j=1 x dmax 1 This guarantees that with probability log(n) our guess for n0 will be off of its true value by a factor of at most 2. With this guess we will work with an approximation√ of θ which is off of its true value by a factor of at most 2. These approximations degrade our bounds by a factor of log(n). j:|Cj |≤ θ  2 s θα2 n s dmax  1 X γ |Cj | ≥ γ dmax . ≥ s j=1 4x Jensen n By setting θ = √xn we guarantee that the value of RDFS is √ at least γ dmax α2 n/4. Since dmax − dmin ≤ x, √ RDFS γ dmax α2 n/4 αγ x α √ = √ , ≥ ≥ d OPT γ min αn 4 n 2 n where the last inequality follows from the triangle inequality. Step 2. Assume that OPT gets its value from n0 < n rewards that it collects in a segment of length x (and from all other rewards OPT collects a negligible value). Recall that the NN-RDFS policy is either NN with probability 0.5 or RDFS with probability 0.5. By picking the single reward closest to the starting point, NN gets at least 1/n0 of the value of OPT. Otherwise, with probability n0 /n, RDFS starts with one of the n0 rewards picked by OPT and then, by the analysis of step 1, if it sets θ = √xn0 , RDFS collects √1 of the value collected by OPT (we use Step (1) with 2 n0 α = 1). It follows that √ NN-RDFS 1 1 1 n0 1 1 n0 ≥ · 0+ · · √ = 0+ . 0 OPT 2 n 2 n 2 n 2n 4n 2 This lower bound is smallest when n0 ≈ n 3 , in which case NN-RDFS collect Ω(n−2/3 ) of OPT. First, notice that since n0 is not known to NN-RDFS, it has to be guessed in order to choose θ. This is done by setting n0 at random from n0 = n/2i , i ∼ Uniform{1, 2, ..., log2 (n)}. We now describe the NN-RA policy (Algorithm 3). Similar in spirit to NN-RDFS, the policy performs NN with probability 0.5 and local policy which we call RA with probability 0.5. RA starts at a random node, s1 , sorts the nodes in increasing order of their distance from s1 and then visits all other nodes in this order. The algorithm is simple to implement, as it does not require guessing any parameters (like θ which RDFS has to guess). However, this comes at the cost of a worse worst case bound. Algorithm 3 NN with RA Input: Graph G, with n nodes, and s0 the first node Flip a coin, if outcome = heads then # Perform RA Visit a node at random, denote it by s1 Sort the remaining nodes by increasing distances from s1 , call this permutation π Visit the nodes in π in increasing order else Execute NN end if The performance guarantees for the NN-RA method are given in Theorem 6. The analysis follows the same steps as the proof of the NN-RDFS algorithm. We emphasize that here, the pruning parameter θ is only used for analysis purposes and is not part of the algorithm. Consequently, we see only one logarithmic factor in the performance bound of Theorem 6 in contrast with two in Theorem 5. Theorem 6 (NN-RA Performance) For a discounted reward traveling salesman graph with n nodes and for any discount factor γ:    2  n− 3  , if OPT = Ω(n). Ω log(n)  NN-RA   ≥ 3  OPT n− 4  , otherwise. Ω log(n) Approximating Optimal Discounted TSP Using Local Policies 4. Simulations In this section, we evaluate and compare the performance of deterministic and stochastic local policies by measuring the (cumulative discounted) reward achieved by each algorithm on different RD-TSP instances as a function of n, the number of the rewards, with n ∈ {100, 200, 400, 600, 800, 1000}. For each n, we set γ = 1− n1 and place the rewards such that OPT can collect almost all of them within a constant discount (i.e., OP T ≈ αn). We always place the initial state s0 at the origin, i.e., s0 = (0, 0). We define x = log γ1 (2), and ` = 0.01x denotes a short distance. Next, we describe five scenarios (Figure 1, ordered from left to right) that we considered for evaluation. For each of these graph types, we generate Nmaps = 10 different graphs, and report the reward achieved by each algorithm on average over the Nmaps graphs (Figure 1, Top), and in the worst-case (the minimal among the Nmaps scores) (Figure 1, Bottom). As some of our algorithms are stochastic, we report average results, i.e., for each graph we run each algorithm Nalg = 100 times and report the average score. Finally, we provide a visualization of the different graphs and the tours taken by the different algorithms, which helps in understanding our numerical results. (1) Random Cities. For a vanilla TSP with n rewards randomly distributed on a 2D plane, it is known that the NN algorithm yields a tour which is 25% longer than optimal on average (Johnson & McGeoch, 1997). We used a similar input to compare our algorithms for RD-TSP, specifically, we Discounted Cumulative Reward Random Cities Line generated a graph with n rewards ri ∼ (U(0, x), U(0, x)), where U is the uniform distribution. Inspecting Figure 1, we can see that the NN algorithm performs the best both on the average and in the worst case. This observation suggests that when the rewards are distributed at random, selecting the nearest reward is the most reasonable thing to do. In addition, we can see that NNRDFS performs the best among the stochastic policies (as predicted by our theoretical results). On the other hand, the RA policy performs the worst among stochastic policies. This happens because sorting the rewards by their distances from s1 , introduces an undesired “zig-zag” behavior while collecting rewards at equal distance from s1 . (2) Line. This graph demonstrates a scenario where greedy algorithms like NN and R-NN are likely to fail. The rewards are located in three different groups; each contains n/3 of the rewards. In group 1, the rewards are located in a cluster left to the origin ri ∼ (U[−θ/3 − `, −θ/3 + `], N(0, `)), while in group 2 they are located in a cluster right to the origin ri ∼ (U[θ/3 − 3`, θ/3 − 2`], N(0, `)) but a bit closer than group 1 (θ = √xn ). Group 3 is also located to the right, but the rewards are placed in increasing distances, such that the i-th reward is located at (θ/3)2i . Inspecting the results, we can see that NN and R-NN indeed perform the worst. To understand this, consider the tour that each algorithm takes. NN goes to group 2, then 3 then 1 (and loses a lot from going to 3). The stochastic tours depend on the choice of s1 . If it belongs to group 1, they Random Clusters Circle Rural vs. Urban Average Results Min Results Figure 1. Evaluation of deterministic and stochastic local policies over different RD-TSPs. The cumulative discounted reward of each policy is reported for the average and worst case scenarios. Approximating Optimal Discounted TSP Using Local Policies collect group1 then 2 then 3, from left to right, and perform relatively the same. If it belongs to group 3, they will first collect the rewards to the left of s1 in ascending order and then come back to collect the remaining rewards to the right, performing relatively the same. However, if s1 is in group 2, then NN-RDFS, NN-RA will visit group 1 before going to 3, while R-NN is tempted to go to group 3 before going to 1 (and loses a lot from doing it). (3) Random Clusters. This graph demonstrates the advantage of stochastic policies. We first randomly place k = 10 cluster centers cj , j = 1, . . . , k on a circle of radius x (the centers are at a small random (Guassian) distance from the circle) Then to draw a reward ri we first draw a cluster center cj uniformly and then draw ri such that ri ∼ (U[cjx − 10`, cjx + 10`], U[cjy − 10`, cjy + 10`]). This scenario is motivated by maze navigation problems, where collectible rewards are located at rooms (clusters) while in between rooms there are fewer rewards to collect. Inspecting the results, we can see that NN-RDFS and R-NN perform the best, in particular in the worst case scenario. The reason for this is that NN picks the nearest reward, and most of its value comes from rewards collected at this cluster. On the other hand, the stochastic algorithms visit larger clusters first with higher probability and achieve higher value by doing so. √ (4) Circle. In this graph, there are n circles, all centered at the origin, and the radii of the ith circle is ρi = √xn · (1 + √ 1 i √ n rewards are placed at 4 n ) On each circle we place equal distances. Here, NN-RDFS performs the best among all policies since it collects rewards closer to s1 first. The greedy algorithms, on the other hand, are “tempted” to collect rewards that take them to the outer circles which results in lower values. (5) Rural vs. Urban. Here, the rewards are sampled from a mixture of two normal distributions. Half of the rewards are located in a “city”, i.e., their position is a Gaussian random variable with a small standard deviation s.t. ri ∼ (N(x, `), N(0, `)). The other half is located in a “village”, i.e., their position is a Gaussian random variable s.t. ri ∼ (N(−x, 10x), N(0, 10x)). In this graph, we can see that in the worst case scenario, the stochastic policies perform much better than NN. This happens because NN is mistakenly choosing rewards that take it to remote places in the rural area, while the stochastic algorithms remain near the city with high probability and collect its rewards. 4.1. Visualization We now present a visualization of the tours taken by the different algorithms. Note that in order to distinguish between the algorithms qualitatively, we compare the NN algorithm with the stochastic algorithms RDFS and RA (and not of NN-RDFS and NN-RA, i.e., without balancing them with NN). All the graphs we present have n = 800 rewards, displayed on a 2D grid using gray dots. For each graph type, we present a single graph sampled from the appropriate distribution. On top of it, we display the tours taken by the different algorithms, such that each row corresponds to a single algorithm. For the stochastic algorithms, we present the best (Figure 2) and the worst tours (Figure 3), among 20 different runs (for NN we display the same tour since it is deterministic). Finally, for better interpretability, we only display the first n/k rewards of each tour, in which the policy collects most of its value, with k = 8 ( n/k = 100 ) unless mentioned otherwise. Discussion. Random cities: Inspecting the worst tours, we can see that the stochastic tours are longer than the NN tour due to the distance to the first reward (that is drawn at random). In addition, we observe a “zig-zag” behavior in the tour of NN-RA while collecting rewards at equal distanced from s1 (but which are not close to each other), which causes it to perform the worst in this scenario. The best tours exhibit similar behavior, but in this case, s1 is located closer to s0 . Line: Recall that in this graph, the rewards are located in three different groups; each containing n/3 of the rewards. Ordered from left to right, group 1 is a cluster of rewards located left to the origin at distance θ/3. Group 2 is also a cluster, but located to the right, in a slightly shorter distance from the origin than group 1. Group 3 is also located to the right, but the rewards are placed in increasing distances, such that the i-th reward is located at (θ/3)2i . For visualization purposes, we added a small variance in the locations of the rewards at groups 1 and 2 and rescaled the axes. The two vertical lines of rewards represent these two groups, while we cropped the graph such that only the first few rewards in group 3 are observed. Finally, we chose k = 2, such the first half of the tour is displayed, and we can see the first two groups visited in each tour. Examining the best tours (Figure 2), we can see that NN first visits group 2, but is then tempted going right, to group 3, which harms its performance. On the other hand, the stochastic algorithms are staying closer to the origin and collect the rewards in groups 1 and 2 first. In the case of the worst case tours, the algorithms perform relatively the same. Random Clusters: Here, we can see that NN first visits the cluster nearest to the origin. The nearest cluster is not necessarily the largest one, and in practice, NN collects all the rewards in this cluster and traverses between the remaining clusters which result in lower performance. On Approximating Optimal Discounted TSP Using Local Policies Figure 2. Visualization of the best tours, taken by deterministic and stochastic local policies over different RD-TSPs. Figure 3. Visualization of the worst tours, taken by deterministic and stochastic local policies over different RD-TSPs. Approximating Optimal Discounted TSP Using Local Policies the other hand, since the stochastic algorithms are selecting the first reward at random, with high probability the reach larger clusters and achieve higher performance. Circles: In this scenario, the distance between adjacent rewards on the same circle is longer than the distance between adjacent rewards on two consecutive circles. Examining the tours, we can see that indeed NN and R-NN are taking tours that lead them to the outer circles. On the other hand, RDFS and RA are staying closer to the origin. Such local behavior is beneficial for RDFS, which achieves the best performance in this scenario. However, while RA performs well in the best case, its performance is much worse than the other algorithms in the worst case. Hence, its average performance is the worst in this scenario. Rural vs. Urban: In this graph, we have a large rural area, where the city is located near the origin which is hard to visualize. To improve the visualization here, we chose k = 2, such the first half of the tour is displayed. Since half of the rewards belong to the city, choosing k = 2 ensures that any tour that is reaching the city only the first segment of the tour (until the tour reaches the city) will be displayed. Looking at the best tours, we can see that NN is taking the longest tour before it reaches the city, while the stochastic algorithms reach it earlier. By doing so, the stochastic algorithms can collect many rewards by traversing short distances and therefore perform much better in this scenario. 5. Related work Pre-defined rules for option selection are used in several studies. Karlsson (1997) suggested a policy that chooses greedily withPrespect to the sum of the local Qvalues a∗ = argmaxa i Qi (s, a). Humphrys (1996) suggested to choose the option with the highest local Q-value a∗ = argmaxa,i Qi (s, a) (NN). Such greedy combination of local policies that were optimized separately may not necessarily perform well. Barreto et al. (2017) considered a transfer framework similar to ours (Definition 1), but did not focus on collectible reward decomposition (Definition 2). Instead, they proposed a framework where the rewards are linear in some reward features Ri (s, a) = wiT φ(s, a).5 Similar to (Humphrys, 1996), they suggested using NN as the pre-defined rule for option selection (but referred to it as GPI). In addition, the authors provided performance guarantees for using GPI in the form of additive (based on regret) error bounds but did not provide impossibility results. In contrast, we prove multiplicative performance guarantees for NN, as well as for three stochastic policies. We also proved, for the first time, impossibility results for such local option selection methods. 5 Notice that Definition 2 is a special case of the framework considered by Barreto et al. A different approach to tackle these challenges is Multitask learning, in which we optimize the options in parallel with the policy over options (Russell & Zimdars, 2003; Sprague & Ballard, 2003; van Seijen et al., 2017). One method that achieves that goal is the local SARSA algorithm (Russell & Zimdars, 2003; Sprague & Ballard, 2003). Similar to (Karlsson, 1997), a Q function is learned locally for each option (concerning a local reward). However, here the local Q functions are learnt on-policy (using SARSA)P with respect to the policy over options π(s) = argmaxa i Qi (s, a), instead of being learned off-policy with Q learning. Russell & Zimdars (2003) showed that if the policy over options is being updated in parallel with the local SARSA updates, then the local SARSA algorithm is promised to converge to the optimal value function. 6. Conclusions In this work, we provided theoretical guarantees for reward decomposition in deterministic MDPs, which allows one to learn many policies in parallel and combine them into a composite solution efficiently and safely. In particular, we focused on approximate solutions that are local, and therefore, easy to implement and do not require many computational resources. Local deterministic policies, like Nearest Neighbor, are being used in practice for hierarchical reinforcement learning. Our study provides an important theoretical guarantee on the reward collected by three such policies, as well as impossibility results for any local policy. These policies outperform NN in the worst case scenario; we evaluated them in the average, and worst case scenarios, suggesting when is it better to use each one. Approximating Optimal Discounted TSP Using Local Policies References Bacon, Pierre-Luc, Harb, Jean, and Precup, Doina. The optioncritic architecture. pp. 1726–1734, 2017. Barreto, Andre, Munos, Remi, Schaul, Tom, and Silver, David. Successor features for transfer in reinforcement learning. Advances in Neural Information Processing Systems, 2017. Barto, Andrew G and Mahadevan, Sridhar. Recent advances in hierarchical reinforcement learning. Discrete Event Dynamic Systems, 13(4):341–379, 2003. Mann, Timothy Arthur, Mannor, Shie, and Precup, Doina. Approximate value iteration with temporally extended actions. Journal of Artificial Intelligence Research (JAIR), 53, pp. 375–438, 2015. Mnih, Volodymyr, Kavukcuoglu, Koray, Silver, David, Rusu, Andrei A, Veness, Joel, Bellemare, Marc G, Graves, Alex, Riedmiller, Martin, Fidjeland, Andreas K, Ostrovski, Georg, et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015. Beattie, Charles, Leibo, Joel Z, Teplyashin, Denis, Ward, Tom, Wainwright, Marcus, Küttler, Heinrich, Lefrancq, Andrew, Green, Simon, Valdés, Vı́ctor, Sadik, Amir, et al. Deepmind lab. arXiv preprint arXiv:1612.03801, 2016. Moravčı́k, Matej, Schmid, Martin, Burch, Neil, Lisỳ, Viliam, Morrill, Dustin, Bard, Nolan, Davis, Trevor, Waugh, Kevin, Johanson, Michael, and Bowling, Michael. Deepstack: Expert-level artificial intelligence in heads-up no-limit poker. Science, 356 (6337):508–513, 2017. Bellemare, Marc G, Naddaf, Yavar, Veness, Joel, and Bowling, Michael. The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research (JAIR), 47, pp. 253–279, 2013. Oh, Junhyuk, Guo, Xiaoxiao, Lee, Honglak, Lewis, Richard L, and Singh, Satinder. Action-conditional video prediction using deep networks in atari games. In Advances in Neural Information Processing Systems, pp. 2863–2871, 2015. Bellman, Richard. Dynamic programming treatment of the travelling salesman problem. Journal of the ACM (JACM), 9(1): 61–63, 1962. Oh, Junhyuk, Singh, Satinder, Lee, Honglak, and Kohli, Pushmeet. Zero-shot task generalization with multi-task deep reinforcement learning. Proceedings of the 34th International Conference on Machine Learning (ICML-17), 2017. Blum, Avrim, Chawla, Shuchi, Karger, David R, Lane, Terran, Meyerson, Adam, and Minkoff, Maria. Approximation algorithms for orienteering and discounted-reward tsp. SIAM Journal on Computing, 37(2):653–670, 2007. Rosenkrantz, Daniel J, Stearns, Richard E, and Lewis, II, Philip M. An analysis of several heuristics for the traveling salesman problem. SIAM journal on computing, 6(3):563–581, 1977. Farbstein, Boaz and Levin, Asaf. Discounted reward tsp. Algorithmica, pp. 1–24, 2016. Russell, Stuart J and Zimdars, Andrew. Q-decomposition for reinforcement learning agents. pp. 656–663, 2003. Held, Michael and Karp, Richard M. A dynamic programming approach to sequencing problems. Journal of the Society for Industrial and Applied Mathematics, 10(1):196–210, 1962. Silver, David, Huang, Aja, Maddison, Chris J, Guez, Arthur, Sifre, Laurent, Van Den Driessche, George, Schrittwieser, Julian, Antonoglou, Ioannis, Panneershelvam, Veda, Lanctot, Marc, et al. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016. Higgins, Irina, Pal, Arka, Rusu, Andrei A, Matthey, Loic, Burgess, Christopher P, Pritzel, Alexander, Botvinick, Matthew, Blundell, Charles, and Lerchner, Alexander. Darla: Improving zeroshot transfer in reinforcement learning. Proceedings of the 34th International Conference on Machine Learning (ICML17), 2017. Humphrys, Mark. Action selection methods using reinforcement learning. From Animals to Animats, 4:135–144, 1996. Johnson, David S and McGeoch, Lyle A. The traveling salesman problem: A case study in local optimization. Local search in combinatorial optimization, 1997. Karlsson, Jonas. Learning to solve multiple goals. PhD thesis, 1997. Kempka, Michał, Wydmuch, Marek, Runc, Grzegorz, Toczek, Jakub, and Jaśkowski, Wojciech. Vizdoom: A doom-based ai research platform for visual reinforcement learning. In Computational Intelligence and Games (CIG), 2016 IEEE Conference on, pp. 1–8. IEEE, 2016. Kulkarni, Tejas D, Narasimhan, Karthik, Saeedi, Ardavan, and Tenenbaum, Josh. Hierarchical deep reinforcement learning: Integrating temporal abstraction and intrinsic motivation. In Advances in Neural Information Processing Systems, pp. 3675– 3683, 2016. Silver, David, Hubert, Thomas, Schrittwieser, Julian, Antonoglou, Ioannis, Lai, Matthew, Guez, Arthur, Lanctot, Marc, Sifre, Laurent, Kumaran, Dharshan, Graepel, Thore, et al. Mastering chess and shogi by self-play with a general reinforcement learning algorithm. arXiv preprint arXiv:1712.01815, 2017. Sprague, Nathan and Ballard, Dana. Multiple-goal reinforcement learning with modular sarsa (0). 2003. Sutton, Richard S, Precup, Doina, and Singh, Satinder. Between mdps and semi-mdps: A framework for temporal abstraction in reinforcement learning. Artificial intelligence, 112(1-2):181– 211, 1999. Teh, Yee, Bapst, Victor, Pascanu, Razvan, Heess, Nicolas, Quan, John, Kirkpatrick, James, Czarnecki, Wojciech M, and Hadsell, Raia. Distral: Robust multitask reinforcement learning. In Advances in Neural Information Processing Systems, pp. 4497– 4507, 2017. Tesauro, Gerald. Temporal difference learning and td-gammon. Communications of the ACM, 38(3):58–69, 1995. Tessler, Chen, Givony, Shahar, Zahavy, Tom, Mankowitz, Daniel J, and Mannor, Shie. A deep hierarchical approach to lifelong learning in minecraft. AAAI, 2017. Approximating Optimal Discounted TSP Using Local Policies van Seijen, Harm, Fatemi, Mehdi, Romoff, Joshua, Laroche, Romain, Barnes, Tavian, and Tsang, Jeffrey. Hybrid reward architecture for reinforcement learning. Advances in Neural Information Processing Systems, 2017. Vezhnevets, Alexander Sasha, Osindero, Simon, Schaul, Tom, Heess, Nicolas, Jaderberg, Max, Silver, David, and Kavukcuoglu, Koray. Feudal networks for hierarchical reinforcement learning. Proceedings of the 34th International Conference on Machine Learning (ICML-17), 2017. Yao, Andrew Chi-Chin. Probabilistic computations: Toward a unified measure of complexity. In Proceedings of the 18th Annual Symposium on Foundations of Computer Science, pp. 222–227. IEEE Computer Society, 1977. Approximating Optimal Discounted TSP Using Local Policies 7. NN performance γ = 1 − n1 we have that Theorem (NN Performance) For any discounted reward traveling salesman graph with n nodes, and any discount factor γ: NN 1 ≥ . OPT n Proof. Denote by i∗ the nearest reward to the origin s0 , and by d0,i∗ the distance from the origin to i∗ . The distance from s0 to the first reward collected by OPT is at least d0,i∗ . Thus, if o0 = s0 , o1 , . . . , on−1 are the rewards ordered in the order by which OPT collects them we get that OPT = n−1 X γ Pj t=0 n 4 ≤ P n2 −1 i=0 γi = 1−γ n/2 1−γ ≤ n 2 and thus P n2 P n2 −1 i (2i−1)d + γ nd+1 i=0 γ D-Local i=1 γ = P n2 P n2 −1 i −1 2d+ n OPT d (2i−1)d 2 γ i=0 γ + γ i=1 γ P n2 (2i−1)d nd+1 + 0.5nγ i=1 γ ≤ P n2 −1 i d γ i=0 γ P n2 n i 2 i=1 0.25 + 0.54 n 6 24 . = ≤ P n −1 ≤ P n2 −1 i n 2 i 0.5 i=0 γ i=0 γ 8.2. Stochastic Local Policies Theorem (Impossibility Results for Stochastic Local Policies) For each stochastic local policy S-Local, there exists a graph with n nodes and a discount factor γ = 1 − √1n such that dot ,ot+1 j=0 ≤ γ d0,i∗ 1+ n−1 X γ Pj t=1 S-Local 8 ≤ √ . OPT n ! dot ,ot+1 ≤ nγ d0,i∗ j=1 On the other hand, the NN heuristic chooses i∗ in the first round, thus, its cumulative reward is at least γ d0,i∗ and we get that NN γ d0,i∗ 1 ≥ = . OPT n nγ d0,i∗ Proof. We consider a family of graphs, G, each of which consists of a star with a central vertex and n leaves. The starting vertex is the central vertex, and there is a reward at each leaf. The length of each edge is d, where d is chosen such that γ d = 12 . Each graph in √ G corresponds to a subset of n leaves which we pairwise connect to form a clique. √ P√n−1 i Since γ = 1 − √1n , we have that i=0 γ ≥ 2n , and therefore √ n−1 OPT = γ d ≥ 0.5 X i γ +γ i=0 √ n−1 8. Impossibility results 8.1. Deterministic Local Policies X √ 2d+ n−1 √ n− n X γ (2i−1)d i=1 √ γ i ≥ 0.25 n. i=0 Theorem (Impossibility Results for Deterministic Local Policies) For any deterministic local policy D-Local, there exists a graph with n nodes and a discount factor γ = 1 − n1 such that 24 D-Local ≤ . OPT n On the other hand, local policy at the central vertex cannot distinguish among the rewards and therefore for every graph in G it picks the first reward from the same distribution. The policy continues to choose rewards √ from the same distribution until it hits the first reward from the n-size clique. Each graph of the family G corresponds to a different subset of n/2 of the leaves which we connect (pairwise) by edges of length 1. (The other n2 leaves are only connected to the central vertex.) While at the central vertex, local policy cannot distinguish among the n rewards (they all at the same distance from the origin), and therefore its choice is the same for all graphs in G. (The following choice is also the same and so on, as long as it does not hit one of the n/2 special rewards.) To argue formally that every S-Local policy has small expected reward on a graph from G, we use Yao’s principle (Yao, 1977) and consider the expected reward of a D-Local policy on the uniform distribution over G. √ Let p1 = n/n √ be the probability that D-Local picks its first vertex from the n-size clique. √ Assuming that the first vertex is not in the clique, let p2 = n/(n − 1) be the probability that the second vertex is from the clique, and let p3 , p4 , . . . be defined similarly. When D-local picks a vertex in√the clique then its reward (without the cumulative discount) is O( n). However, each time D-Local misses the clique then it collects a single reward but suffers a discount of γ 2d = 1/4. Neglecting the rewards collected until it hits the clique, the total value of D-Local is  √  O p1 + (1 − p1 )γ 2d p2 + (1 − p1 )(1 − p2 )γ 4d p3 . . . n It follows that, for any given policy, there exists a graph in G such that the adjacent n/2 rewards are visited last. Finally, since √ Since pi ≤ 2/ n for 1 ≤ i ≤ n/2 this value is O(1) Proof. Consider a family of graphs, G, each of which consists of a star with a central vertex and n leaves. The starting vertex is the central vertex, and there is a reward at each leaf. The length of each edge is d, where d is chosen s.t. γ d = 12 . Approximating Optimal Discounted TSP Using Local Policies 9. NN-RA where the last inequality follows from the triangle inequality. Theorem (NN-RA Performance) For any RD-TSP instance with n nodes and for any discount factor γ    −2  n 3  , if OPT = Ω(n). Ω log(n) NN-RA   ≥ 3 −  OPT n 4  , otherwise. Ω log(n) Step 2. Assume that OPT gets its value from n0 < n rewards that it collects in a segment of length x (and from all other rewards OPT collects a negligible value). Recall that the NN-RA policy is either NN with probability 0.5 or RA with probability 0.5. By picking the single reward closest to the starting point, NN gets at least 1/n0 of the value of OPT. Otherwise, with probability n0 /n, RA starts with one of the n0 rewards picked by OPT and then, by the analysis of step 1, if it sets θ = (n0x)2/3 , RA collects 2(n01)2/3 of the value collected by OPT (we use Step (1) with α = 1). It follows that Proof. Step 1. Assume that OPT collects a set SOPT of αn rewards for some 0 ≤ α ≤ 1, in a segment p of length x = log 1 (2) (i.e. x γ is the distance from the first reward to the last reward – it does not include the distance from the starting point to the first reward). Let dmin , dmax the shortest and longest distances from s0 to a reward in SOPT respectively. By the triangle inequality, dmax − dmin ≤ x. We further assume that OP T ≤ O(γ dmin αn) (i.e., That is the value that OPT collects from rewards which are not in SOPT is negligible). Let θ be a threshold that we will fix below, and denote by {Cj } the CCs of SOPT that are created by deleting edges longer than θ among vertices of SOPT . By Lemma 1, we have at most x/θ CC. Assume that RA starts at a vertex of a component Cj , such that |Cj | = k. Since the diameter of Cj is at most (|Cj | − 1)θ then it collects its first k vertices (including s1 ) within a total distance P of 2 ki=2 (i − 1)θ ≤ k2 θ. So if k2 θ ≤ x then it collects at least |Cj | rewards before traveling a total distance of x, and if p k2 θ > x it collects at least b x/θc rewards. (We shall omit the floor function for in the sequal.) It follows that RA collects  p brevity Ω min{|Cj |, xθ } rewards. Notice that the first random step |C | leads RDFS to a vertex in CC Cj with probability nj . If more p than half of rewards are in CCs s.t |Cj | ≥ xθ , then x r   θ X |Cj | x · min |Cj |, n θ j=1 r r X |Cj | α x x ≥γ dmax · ≥ γ dmax . θ 2 θ √x n RA ≥γ dmax j:|Cj |≥ x RA = γ ≥γ dmax θ n X |Cj | xo · min |Cj |, n θ j=1 X j:|Cj |≤ s dmax ≥ γ n Jensen √x |Cj |2 s ≥ γ dmax n n s 1X |Cj |2 s j=1 ! θ s 1X |Cj | s j=1 !2 ≥ γ dmax θα2 n . 4x x By setting θ = n2/3 we guarantee that the value of RA is at least dmax 2 1/3 γ α n /4. Since dmax − dmin ≤ x, γ dmax α2 n1/3 /4 RA αγ x α ≥ ≥ = , d OPT γ min αn 4n2/3 2n2/3 3 This lower bound is smallest when n0 ≈ n 4 , in which case NN-RA collect Ω(n−3/4 ) of OPT. Step 3. By the same arguments from Step 3 in the analysis of NN-RDFS, it follows that    −2  n 3  , if OPT = Ω(n). Ω log(n) NN-RA   ≥ −3  OPT n 4  , otherwise. Ω log(n) 10. R-NN We now analyze the performance guarantees of the R-NN method. The analysis is conducted in two steps. In the first step, we assume that OPT achieved a value of Ω(n0 ) by collecting n0 rewards and consider the case that n0 = αn. The second step considers the more general case and analyzes the performance of NN-Random for the worst value of n0 . We emphasize that unlike the previous two algorithms, we do not assume this time that OPT collects its rewards at a segment of length x6 . Theorem (R-NN) For a discounted reward traveling salesman graph with n nodes: θ If than half of rewards in SOPT are in CCs such that |Cj | ≤ pmore x , let s be the number of such CCs and notice that s ≤ xθ . We θ get that: dmax (n0 )1/3 NN-RA 1 1 1 1 n0 1 ≥ · 0 + · · . = + 0 0 2/3 OPT 2 n 2 n 2(n ) 2n 4n R-NN ≥Ω OPT  log(n) n  Proof. Step 1. Assume OPT collects Ω (n) rewards. √ Define x = log 1 (2) and θ = √xn (here we can replace the n by any γ fractional power of n, this will not affect the asymptotics of the result) and denote by {Cj } the CCs that are obtained by pruning edges longer than θ. We define a CC to be large if it contains more √ than log(n) rewards. Observe that since there are at most n CCs (Lemma 1), at least one large CC exists. Lemma 2 Assume that s1 is in a large component C. Let p be the path covered by NN starting from s1 until it reaches si in a large component. Let d be the length of p and let r1 be the number of rewards collected by NN in p (including the last reward in p which is back in a large component, but not including s1 ). Note that r1 ≥ 1. Then d ≤ (2r1 − 1)θ. 6 Therefore, we do not perform a third step like we did in the analysis of the previous methods. Approximating Optimal Discounted TSP Using Local Policies Proof. Let pi be the prefix of p that ends at the ith reward on p (i ≤ r1 ) and let di be the length of pi . Let `i be the distance from the ith reward on p to the (i + 1)th reward on p. Since when NN is at the ith reward on p, the neighbor of s1 in C is at distance at most di + θ from this reward we have that `i ≤ di + θ. Thus, di+1 ≤ 2di + θ (with the initial condition d0 = 0). The solution to this recurrence is di = (2i − 1)θ. Lemma 3 For k < log(n), we have that after k visits of R-NN in large CCs, for any s in a large CC there exists an unvisited reward at distance shorter than (k + 1)θ from s. Proof. Let s be a reward in a large component C. We have collected at most k rewards from C. Therefore, there exists a reward s0 ∈ C which we have not collected at distance at most (k + 1)θ from s. Lemma 2 and Lemma 3 imply the following corollary. Corollary 6.1 Assume that k < log(n), and let p be the path of NN from its kth reward in a large CC to its (k + 1)st reward in a large connected component. Let d denote the length of p and rk be the number of rewards on p (excluding the first and including the last). Then d ≤ (2rk − 1)(k + 1)θ ≤ 2rk +1 kθ. The following lemma concludes the analysis of this step. Lemma 4 Let p be the prefix of R-NN of length x. Let k be the number of segments on p of R-NN that connect rewards in large CCs and contain internally rewards in small CCs. For 1 ≤ i ≤ k, let ri be the number of rewards R-NN collects in the ith segment. P Then ki=1 ri = Ω(log n). (We assume that p splits exactly into k segments, but in fact the last segment may be incomplete, this requires a minor adjustment in the proof.) Proof. since ∀i, ri ≥ 1, then if k ≥ log(n) the lemma follows. So assume that s < log n. By Corollary 6.1 we have that x≤ k X 2ri +1 iθ ≤ 2rmax +2 k2 θ . (1) i=1 √ where rmax √ = argmax {ri }ki=1 . Since θ = x/ n, Equation (1) rmax +2 2 implies that n ≤ 2 k and since k ≤ log n we get that √ n ≤ 2rmax +2 log2 (n). Taking logs the lemma follows. Lemma 4 guarantees that once at s1 ∈ Cj , R-NN collects Ω(log(n)) rewards before traversing a distance of x. Next, notice that the chance that s1 (as defined in Algorithm 3) belongs to one √ of the large CCs is p = n− nnlog(n) , which is larger than 1/2 for n ≥ 256. Finally, similar to NN-RDFS, assume that the value of OPT is greater than a constant fraction of n, i.e., OP T ≥ n/2α . This means that OPT must have collected the first n/2α+1 rewards after traversing a distance of at most x̃ = (α + 1) x, 7 , and denote this fraction of the rewards by SOPT . Further denote by dmin , dmax the 7 To see this, recall that after traversing a distance of x̃, OPT achieved less than n/2α+1 . Since it already traversed x̃ it can only achieve less than n/2α+1 from the remaining rewards, thus a contraction with the assumption that it achieved more than n/2α . shortest and longest distances from s0 to SOPT respectively. By the triangle inequality, dmax − dmin ≤ x̃; therefore, with a constant 1 , we get that s1 ∈ SOPT . By taking expectation probability of 2α+1 over the first random pick, it follows that log(n) R-NN 1 γ dmax log(n) ≥ α+1 = α+1 = Ω OPT 2 γ dmin n 4 n  log(n) n  . Step 2. Similar to the analysis of NN-RDFS, we now assume that OPT collects its value from n0 < n rewards that it collects in a segment of length x (and from all other rewards OPT collects a negligible value). Recall that the R-NN is either NN with probability 0.5 or a random pick with probability 0.5 followed by NN. By picking the single reward closets to the starting point, NN gets at least 1/n0 of the value of OPT. Notice, that we do not need to assume anything about the length of the tour that OPT takes to collect the n0 rewards (since we didn’t use it in Step 1). It follows that: log(n0 ) R-NN 1 1 1 n0 log(n0 ) 1 ≥ · 0 + · · = + OPT 2 n 2 n n0 2n0 n Thus, in the worst case scenario, n0 log(n0) ≈ n, which implies n that n0 = Θ( log(n) ). Therefore R-NN = Ω log(n) . OPT n Approximating Optimal Discounted TSP Using Local Policies 11. Exact solutions for the RD-TSP possible states in which the optimal policy could be. We now present a variation of the Held-Karp algorithm for the RDTSP. Note that similar to the TSP, C({S, k}) denotes the length of the tour visiting all the cities in S, with k being the last one (for TSP this is the length of the shortest tour). However, our formulation required the definition of an additional recursive quantity, V ({S, k}), that accounts for the value function (the discounted sum of rewards) of the shortest path. Using this notation, we observe that Held-Karp is identical to doing tabular Q-learning on SMDP Ms . Since Held-Karp is known to have exponential complexity, it follows that solving Ms using SMDP algorithms is also of exponential complexity. Since we were able to classify a state space of polynomial size which contains all states of the optimal policy, then we can describe a dynamic programming scheme (Algorithm 5) that finds the optimal policy. The algorithm computes a table V , where V (`, r, →) is the maximum value we can get by collecting all rewards 1, . . . ` and r, . . . , n starting from r − 1, and V (`, r, ←) is defined analogously starting from ` + 1. The algorithm first initializes the entries of V where either ` = 0 or r = n + 1. These entries correspond to the cases where all the rewards to the left (right) of the agent have been collected. (In these cases the agent continues to collect all remaining rewards one by one in their order.) It then iterates over t, a counter over the number of rewards that are left to collect. For each value of t, we define S as all the combinations of partitioning these t rewards to the right and the left of the agent. We fill V by increasing value of t. To fill an entry V (`, r, ←) such that ` + (n + 1 − r) = t we take the largest among 1) the value to collect ` and then the rewards 1, . . . ` − 1 and r, . . . , n, appropriately discounted and 2) the value to collect r and then 1, . . . ` and r + 1, . . . , n. We fill V (`, r, →) analogously. Algorithm 4 The Held-Karp for the TSP (blue) and RDTSP (black) Input: Graph G, with n nodes for k := 2 to n do C({k}, k) := d1,k C({k}, k) := d1,k V ({k}, k) := γ d1,k end for for s := 2 to n − 1 do for all S ⊆ {2, ..., n}, |S| = s do for all k ∈ S do C(S, k) = minm6=k,m∈s [C(S \ {k}, m) + dm,k ] Q(S, k, a) = [V (S \{k}, a)+γ C(S\{k},a) ·γ da,k ] a∗ = arg maxa6=k,a∈s Q(S, k, a) C(S, k) = C(S \ {k}, a∗ ) + da∗ ,k V (S, k) = Q(S, k, a∗ ) end for end for end for opt := mink6=1 [C({2, ..., n}, k) + dk,1 ] opt := maxk6=1 [V ({2, . . . , n}, k) + γ C({2,...,n},k)+dk,1 return (opt) 11.1. Exact solutions for simple geometries We now provide exact, polynomial time solutions based on dynamic programming for simple geometries, like a line and a star. We note that such solutions (exact and polynomial) cannot be derived for general geometries. 11.1.1. DYNAMIC PROGRAMMING ON A LINE (1D) Given an RD-TSP instance, such that all the rewards are located on a single line (denoted by the integers 1, . . . , n from left to right), it is easy to see that an optimal policy collects, at any time, either the nearest reward to the right or the left of its current location. Thus, at any time the set of rewards that it has already collected lie in a continuous interval between the first uncollected reward to the left of the origin, denoted `, and the first uncollected reward to the right of the origin denoted r. The action to take next is either to collect ` or to collect r. It follows that the state of the optimal agent is uniquely defined by a triple (`, r, c), where c is the current location of the agent. Observe that c ∈ {` + 1, r − 1} and therefore there are O(n2 ) The optimal value for starting position j is 1 + V (j − 1, j + 1, →). Note that the Algorithm computes the value function; to get the policy, one has merely to track the argmax at each maximization step. 11.1.2. DYNAMIC PROGRAMMING ON A d- STAR We consider an RD-TSP instance, such that all the rewards are located on a d-star, i.e., all the rewards are connected to a central connection point via one of d lines and there are ni rewards along the ith line. We denote the rewards on the ith line by mij ∈ {1, .., ni }, ordered from the origin to the end of the line, and focus on the case where the agent starts at the origin.8 It is easy to see that an optimal policy collects, at any time, the uncollected reward that is nearest to the origin along one of the d lines. Thus, at any time the set of rewards that it has already collected lie in d continuous intervals between the origin and the first uncollected reward along each line, denoted by `¯ = {`i }di=1 . The action to take next is to collect one of these nearest uncollected rewards. It follows that the state of the optimal agent is uniquely defined by a ¯ c), where c is the current location of the agent. Observe tuple (`, that c ∈ {mi`i −1 }di=1 and therefore there are O(dnd ) possible states in which the optimal policy could be. Since we were able to classify a state space of polynomial size which contains all states of the optimal policy then we can describe a dynamic programming scheme (Algorithm 7) that finds the optimal policy. The algorithm computes a table V , where ¯ c) is the maximum value we can get by collecting all reV (`, wards {mi`i , . . . , mini }di=1 starting from c. The algorithm first initializes the entries of V where all `i = ni + 1 except for exactly one entry. These entries correspond to the cases where all the rewards have been collected, except in one line segment (in these cases the agent continues to collect all remaining rewards one by one in their order.) It then iterates over t, a counter over the number of rewards that are left to collect. For each value of t, we define S as all the combinations of partitioning these t rewards among d lines. We fill V by increasing value of ¯ c) such that P li = n − t we take the t. To fill an entry V (`, largest among the values for collecting `i and then the rewards 8 The more general case is solved by applying Algorithm 5 until the origin is reached followed by Algorithm 7 Approximating Optimal Discounted TSP Using Local Policies Algorithm 6 Optimal solution for the RD-TSP on a line. The rewards are denoted by 1, . . . , n from left to right. We denote by di,j the distance between reward i and reward j. We denote by V (`, r, →) the maximum value we can get by collecting all rewards 1, . . . , ` and r, . . . , n starting from reward r − 1. Similarly, we denote by V (`, r, ←) maximum value we can get by collecting all rewards 1, . . . , ` and r, . . . , n starting from ` + 1. If the leftmost (rightmost) reward was collected we define ` = 0 (r = n + 1). Input: Graph G, with n nodes Init V (·, ·, ·) = 0 for t = 1, n do V (` = t, n +  1, →) := γ dt,n ·  Pj P2 1 + j=t γ i=0 di,i−1   Pn−1 Pj V (0, r = t, ←) := γ d1,t · 1 + j=t γ i=0 di,i+1 end for for t = 2, ..n − 1 do S = {(i, n + 1 − j)|i + j = t} for (`, r) ∈ S do if V (`, r, ←) = 0 then ( γ d`,`+1 [1 + V (` − 1, r, ←)] V (`, r, ←) = max γ d`+1,r [1 + V (`, r + 1, →)] end if if V (`, r, →) = 0 then ( γ d`,r−1 [1 + V (` − 1, r, ←)] V (`, r, →) = max γ dr−1,r [1 + V (`, r + 1, →)] end if end for end for m1`1 . . . m1n1 , . . . , mi`i +1 , . . . mini , . . . , md`d , . . . , mdnd appropriately discounted. Note that the Algorithm computes the value function; to get the policy, one has merely to track the argmax at each maximization step. Algorithm 8 Optimal solution for the RD-TSP on a d-star. We denote by ni the amount of rewards there is to collect on the ith line, and denote by mij ∈ {1, .., ni } the rewards along this line, from the center of the star to the end of that line. We denote by dmti ,mkj the distance between reward i on line t and reward j on line k. The first uncollected reward along each line is denoted by `i , and the maximum value we can get by collecting all the remaining rewards m1`1 . . . m1n1 , . . . , md`d , . . . , md nd starting from reward c is defined by V (`¯ = {`i }di=1 , c). If all the rewards were collected on line i we define `i = ni + 1. Input: Graph G, with n rewards. Init V (·, ·) = 0 for i ∈ {1, .., d} do for `i ∈ {1, .., ni } do for c ∈ {m1n1 , .., mi`i −1 , mdnd } do V + 1, .., li , nd + 1, c) =  (n1P Pj d k,k+1 1 + j=`i γ k=0 end for end for end for for t = 2, ..n − 1 do ¯ i ∈ {1, .., ni } P `i = n − t} S = {`|` for `¯ ∈ S, c ∈ {mi`i −1 }di=1 do ¯ c) = 0 then if V (`, A = {mij |j = `i , j ≤ ni } for a ∈ A do γ dc,mi `i ( ¯ c) V (`, ¯   V (`, c) = max γ dc,a 1 + V (`¯ + ea , a) end for end if end for end for ·
2cs.AI
Entanglement-assisted quantum MDS codes from constacyclic codes with large minimum distance Liangdong Lua,b,† , Wenping Maa , Ruihu Lib , Yuena Mab , Yang Liub , Hao Caoa arXiv:1803.04168v1 [cs.IT] 12 Mar 2018 a. State Key Laboratory of Integrated Services Networks, Xidian University, Xi’an, Shaanxi, 710051, China, (email: † [email protected], ) b. College of Science, Air Force Engineering University, Xi’an, Shaanxi, 710051, China Abstract The entanglement-assisted (EA) formalism allows arbitrary classical linear codes to transform into entanglement-assisted quantum error correcting codes (EAQECCs) by using pre-shared entanglement between the sender and the receiver. In this work, we propose a decomposition of the defining set of constacyclic codes. Using this method, we construct four classes of q-ary entanglement-assisted quantum MDS (EAQMDS) codes based on classical constacyclic MDS codes by exploiting less pre-shared maximally entangled states. We show that a class of q-ary EAQMDS have minimum distance upper limit greater than 3q − 1. Some of them have much larger minimum distance than the known quantum MDS (QMDS) codes of the same length. Most of these q-ary EAQMDS codes are new in the sense that their parameters are not covered by the codes available in the literature. Index terms: Entanglement-assisted quantum error correcting codes (EAQECCs), MDS codes, cyclotomic cosets, constacyclic code. 1 Introduction Quantum error-correcting codes play an important role in quantum information processing and quantum computation [1–4]. The stabilizer formalism allows standard quantum codes to be constructed from dual-containing (or self-orthogonal) classical codes [3]. However, the dual-containing condition forms a barrier in the development of quantum coding theory. In Ref. [5], Brun et al. proposed a entanglement-assisted (EA) stabilizer formalism 1 which shown that non-dual-containing classical codes can be used to construct EAQECCs if shared entanglement is available between the sender and receiver. Let q be a prime power. A q-ary [[n, k, d; c]] EAQECC that encodes k information qubits into n channel qubits with the help of c pairs of maximally-entangled Bell states (ebits) can correct up to ⌊ d−1 2 ⌋ errors, where d is the minimum distance of the code. If c = 0, then it is called a q-ary standard [[n, k, d]] quantum code Q. We denote a q-ary [[n, k, d; c]] EAQECC by [[n, k, d; c]]q . Currently, many works have focused on the construction of EAQECCs based on classical linear codes, see [6–16]. As in classical coding theory, one of the central tasks in quantum coding theory is to construct quantum codes and EA-quantum codes with the best possible minimum distance. Theory 1.1 [5, 8]. ( EA-Quantum Singleton Bound) An [[n, k, d; c]]q EAQECC satisfies n + c − k ≥ 2(d − 1), where 0 ≤ c ≤ n − 1. A EAQECC achieving this bound is called a EA-quantum maximum-distance-separable (EAQMDS) code. If c = 0, then this bound is Quantum Singleton Bound, and a code achieving the bound is called quantum maximum-distance-separable (QMDS) code. Just as in the classical linear codes, QMDS codes and EAQMDS codes form an important family of quantum codes. Constructing QMDS codes and EAQMDS codes had become a central topic for quantum error-correction codes. Many classes of QMDS codes have been constructed by different methods, in particular, the constructions of obtained from constacyclic codes or negacyclic codes containing their Hermitian dual over Fq2 [17–25]. According to the MDS conjecture in [4], the maximal-distance-separable (MDS) code cannot exceed q 2 + 1. Therefore, for larger distance than q + 1 of code length n ≤ q 2 + 1, one need to construct a EA-quantum MDS code. The following Proposition is one of the most frequently used construction methods. Proposition 1.2 [5, 7]. If C= [n, k, d]q2 is a classical code over Fq2 and H is its parity check matrix, then C ⊥h EA stabilizes an [[n, 2k − n + c, d; c]]q EAQECC, where c =rank(HH † ) is the number of maximally entangled states required and H † is the conjugate matrix of H over Fq2 . In resent years, many scholars have constructed several entanglement-assisted quantum codes with good parameters in [5, 7]. 2 Li et al. [26,27] proposed the concept about a decomposition of the defining set of cyclic codes, and they construct some good entanglement-assisted quantum codes with the help of this concept [26]. Lu and Li [14]constructed some families of entanglement-assisted quantum codes from primitive quaternary BCH codes. In this paper, we construct several families of EA-quantum MDS codes with length n from classical constacyclic codes. From [18], Chen points that dual-containing constacylic codes over Fq2 exist only when the order r is a divisor of q + 1. We construct EA-quantum MDS codes from constacyclic codes with r = q + 1 or negacyclic codes with r = 2, respectively. More precisely, Our main contribution on new q-ary quantum MDS codes is as follows: (1) [[q 2 + 1, q 2 − 2d + 7, d; 4]]q , where q be an odd prime power, q + 3 ≤ d ≤ 3q − 1 is even. 2 2 (2) [[ q 10+1 , q 10+1 − 2d + 3, d; 1]]q , where q be an odd prime power of the form 10m + 3, 2 ≤ d ≤ 6m + 2 is even. 2 2 (3) [[ q 10+1 , q 10+1 − 2d + 3, d; 1]]q , where q be an odd prime power of the form 10m + 7, 2 ≤ d ≤ 6m + 4 is even. (4) [[ q 2 −1 h of q + 1 and ,q 2 −1 h −2d+3, d; 1]]q , where q be (q+1)(h+3) q+1 − 1. h +1 ≤ d ≤ 2h an odd prime power, h ∈ {3, 5, 7} is a divisor In construction (1), we obtain some EA-quantum MDS codes with the minimal distance upper limit greater than 3q + 2 by consuming four pre-shared maximally entangled states. In construction (2)-(3), consumed only one pair of maximally entangled states, each EAquantum MDS code has half length comparing with the standard QMDS code of the same minimum distance constructed. In construction (4), consuming only one pair of maximally entangled states, we obtain a family of EA-quantum MDS codes with the minimal distance larger than the standard quantum MDS codes in Ref. [18]. Comparing the parameters with all known EA-quantum MDS codes, we find that these quantum MDS codes are new in the sense that their parameters are not covered by the codes available in the literature. The paper is organized as follows. In Section 2, basic notations and results about EAquantum codes and constacyclic codes are provided. The concept of a decomposition of the defining set of constacyclic codes is stated, which is extended from the decomposition of the defining set of cyclic codes proposed by Li [26]. In Section 3, we give some new classes of EA-quantum MDS codes. The conclusion is given in Section 4. 3 2 Preliminaries In this section, we review some basic results on constacyclic codes, BCH codes, and EAQECCs for the purpose of this paper. For details on BCH codes and constacyclic codes can be found in standard textbook on coding theory [28, 29], and for EAQECCs please see Refs. [5–15]. Let p be a prime number and q a power of p, ie., q = pl for some l > 0. Fq2 denotes the finite field with q 2 elements. For any α ∈ Fq2 , the conjugation of α is denoted by α = αq . Given two vectors x = (x1 , x2 , · · · , xn ) and y = (y1 , y2 , · · · , yn ) ∈ Fqn2 , their P Hermitian inner product is defined as (x, y)h = xi y i = x1 y 1 + x2 y 2 + · · · + xn y n . For a linear code C over Fq2 of length n, the Hermitian dual code C ⊥h is defined as C ⊥h = {x ∈ Fqn2 |(x, y)h = 0, ∀y∈ C}. If C ⊆ C ⊥h , then C is called a Hermitian dual containing code, and C ⊥h is called a Hermitian self-orthogonal code. We now recall some results about classical constacyclic codes, negacyclic code and cyclic codes. For any vector (c0 , c1 , · · · , cn−1 ) ∈ Fqn2 , a q 2 -ary linear code C of length n is called η-constacyclic if it is invariant under the η-constacyclic shift of Fqn2 : (c0 , c1 , · · · , cn−1 ) → (ηcn−1 , c0 , · · · , cn−2 ), where η is a nonzero element of Fq2 . Moreover, if η = 1, then C is called a cyclic code ; and if η = −1, then C is called a negacyclic code. For a constacyclic code C, each codeword c = (c0 , c1 , · · · , cn−1 ) is customarily represented in its polynomial form: c(x) = c0 + c1 x + · · · + cn−1 xn−1 , and the code C is in turn identified with the set of all polynomial representations of its codewords. The proper context for studying constacyclic codes is the residue class ring Rn = Fq [x]/(xn − η). xc(x) corresponds to a constacyclic shift of c(x) in the ring Rn . As we all know, a linear code C of length n over Fq2 is constacyclic if and only if C is an ideal of the quotient ring Rn = Fq [x]/(xn − η). It follows that C is generated by monic factors of (xn − η), i.e., C = hf (x)i and f (x)|(xn − η). The f (x) is called the generator polynomial of Cn . Let η ∈ Fq2 be a primitive rth root of unity. Let gcd(n, q) = 1, then there exists a primitive rn-th root of unity ω in some extension field field of Fq2 such that ω n = η. Qn−1 Hence, xn − η = i=0 (x − ω 1+ir ). Let Ω = {1 + ir|0 ≤ i ≤ n − 1}. For each j ∈ Ω, let Cj be the q 2 -cyclotomic coset modulo rn containing j. Let C be an η-constacyclic code of length n over Fq2 with generator polynomial g(x). The set T = {j ∈ Ω|g(ω j ) = 0} is 4 called the defining set of C. Let s be an integer with 0 ≤ s < rn, the q 2 -cyclotomic coset modulo rn that contains s is defined by the set Cs = {s, sq 2 , sq 2·2 , · · · , sq 2(k−1) } (mod rn), where k is the smallest positive integer such that xq 2k ≡ x (mod rn). We can see that the defining set T is a union of some q 2 -cyclotomic cosets module rn and dim(C) = n − |T |. Lemma 2.1 [22]. Let C be a q 2 -ary constacyclic code of length n with defining set T . T Then C contains its Hermitian dual code if and only if T T −q = ∅, where T −q denotes the set T −q = {−qz(mod rn)|z ∈ T }. Let C be a constacyclic code with a defining set T = S Cs . Denoting T −q = {rn − s∈S qs|s ∈ T }, then we can deduce that the defining set of C ⊥h is T ⊥h =Zn \T −q , see Ref. [22]. Since there is a striking similarity between cyclic codes and constacyclic code, we give a correspondence defining of skew aymmetric and skew asymmetric as follows. A cyclotomic coset Cs is skew symmetric if rn − qs mod rn ∈ Cs ; and otherwise is skew asymmetric otherwise. Skew asymmetric cosets Cs and Crn−qs come in pair, we use (Cs , Crn−qs ) to denote such a pair. The following results on q 2 -cyclotomic cosets, dual containing constacyclic codes are bases of our discussion. Lemma 2.2 [22, 27]. Let r be a positive divisor of q + 1 and η ∈ Fq∗2 be of order r. Let C be a η-constacyclic code of length n over Fq2 with defining set T , then C ⊥h ⊆ C if and only if one of the following holds: (1) T ∩T −q = ∅, where T −q = {rn − qs | s ∈ T }. (2) If i, j, k ∈ T , then Ci is not a skew asymmetric coset and (Cj , Ck ) is not a skew asymmetric cosets pair. According to Lemma 2.2, C ⊥h ⊆ C can be described by the relationship of its cyclotomic coset Cs . However, a defining set T of a non-dual-containing (or non-self-orthogonal) classical codes is T ∩T −q 6= ∅. We gave a definition for decomposition of the defining set of cyclic codes in [14, 26] and Chen et al. defined a decomposition of the defining set of negacyclic codes in [33]. In order to construct EA-quantum MDS codes for larger distance than q + 1 of code length n ≤ q 2 + 1, we introduce a fundamental definition of decomposition of the defining set of constacyclic codes. Definition 2.1 Let η ∈ Fq2 be a primitive rth root of unity. Let C be a η-constacyclic code of length n with defining set T . Denote Tss = T ∩T −q and Tsas = T \Tss , where T −q = {rn − qx|x ∈ T } and r is a factor of q + 1. T = Tss ∪ Tsas is called decomposition 5 of the defining set of C. To determine Tss and Tsas , we give the following lemma to characterize them. Lemma 2.3 [22, 27]. Let r be a positive divisor of q + 1 and η ∈ Fq∗2 be of order r. Let gcd(q, n) = 1, ordrn (q 2 ) = m, 0 ≤ x, y, z ≤ n − 1. 2t+1 (mod (1) Cx is skew symmetric if and only if there is a t ≤ ⌊ m 2 ⌋ such that x ≡ xq rn). (2) If Cy 6= Cz , (Cy , Cz ) form a skew asymmetric pair if and only if there is a t ≤ ⌊ m 2⌋ such that y ≡ zq 2t+1 (mod n) or z ≡ yq 2t+1 (mod rn). Using the decomposition of a defining set T of a constacyclic code C, one can give a decomposition of C ⊥h as follow. Lemma 2.4 Let C be a constacyclic code with defining set T , T = Tss ∪ Tsas be decomposition of T . Denote the constacyclic codes C R and C E with defining set Tsas and ⊥h ⊥h ⊥h h Tss , respectively. Then C ⊥ R ⊆ C R , C E ∩ C E = {0}, C R ⊂ C E , C R ∩ C E = C and C R + ⊥h . h C⊥ E =C Using the decomposition of a defining set T , one can calculate the number of needed ebits with a algebra method. Lemma 2.5. Let T be a defining set of a constacyclic code C, T = Tss ∪ Tsas be decomposition of T . Using C ⊥h as EA stabilizer, the optimal number of needed ebits is c =| Tss |. Proof. According to Definition 2.1, we denote the defining sets of constacyclic codes C1 and C2 into Tss and Tsas , respectively.  The parity check matrix of C1 and C2 are H1  H1  be the parity check matrix of C. Then, and H2 , respectively. Let H =  H2  HH † =  H1 H1† H1 H2† H2 H1† H2 H2†  . Since H2 is the parity check matrix of C2 with defining set of Tsas , H2 H2† = 0. Because of C1⊥h ⊆ C2 , H1 H2† = 0 and H2 H1† = 0. Therefore,  HH † =  H1 H1† 0 0 0  . According to [5–7], one obtain that c = rank(HH † ) = rank(H1 H1† ). Since H1 is the parity check matrix of C2 with defining set of Tss , H1 is a full-rank matrix. Hence, 6 c = rank(H1 H1† ) = |Tss |. Lemma 2.6 [30]. (The BCH bound for Constacyclic Codes) Let C be an η-constacyclic code of length n over Fq2 , where η is a primitive rth root of unity. Let ω be a primitive rn-th root of unity in an extension field of Fq2 such that ω n = η. Assume the generator polynomial of C has roots that include the set {ω 1+ri |i1 ≤ i ≤ i1 + d − 2}. Then the minimum distance of C is at least d. Theorem 2.7. Let C be an [n, k, d]q2 Constacyclic code with defining set T , and the decomposition of T be T = Tss ∪ Tsas . Then C ⊥h EA stabilizes an q-ary [[n, n − 2|T | + |Tss |, d ≥ δ; |Tss |]] EAQECC. Proof. The dimension of C is k = n − |T |. From Proposition 1.2, Lemma 2.5 and Lemma 2.6, we know C ⊥h EA stabilizes an q-ary EAQECC with parameters [[n, 2k − n + c, d; c]] = [[n, n − 2|T | + |Tss |, d; |Tss |]]. 3 New EA-quantum MDS Codes In this section, we consider η-constacyclic codes over Fq2 of length n to construct EAquantum codes. To do this, we give a sufficient condition for a decomposition of the defining set of η-constacyclic codes over Fq2 of length n which do not contain their Hermitian duals. First, we compute q 2 -cyclotomic cosets modulo rn where r = q +1(constacyclic codes) or r = 2(negacyclic codes). 3.1 EA-quantum MDS Codes of Lenght n = q 2 + 1 with q ≡ 1 mod 4 In this subsection, we construct some classes of q-ary EA-quantum MDS Codes of Length n = q 2 + 1, where q be a prime or an odd prime power of the form q ≡ 1 mod 4. First, Let r = 2, we consider negacyclic codes over Fq2 of length n to construct EA-quantum codes. To do this, we give a decomposition of defining set sufficient condition for negacyclic codes over Fq2 of length n which do not contain their Hermitian duals. Lemma 3.1 [21]: Let n = q 2 + 1 with q ≡ 1 mod 4, and s = n 2. Then the q 2 −cyclotomic cosets modulo 2n containing odd integers form 1 to 2n are Cs = {s} and Cs+2i = {s + 2i, s − 2i} for 1 ≤ i ≤ s − 1. Lemma 3.2: Let n = q 2 + 1 with q ≡ 1 mod 4, q ≥ 5 and s = n2 . If C is a q 2 S ary negacyclic code of length n with define set T = ki=0 Cs+2i , where 0 ≤ k ≤ 3q−3 2 , 7 and the decomposition of a defining set T = Tss S Tsas , then (Cs+q+1 , Cs−q+1 ) forms a skew-asymmetric pair. And   0, if 0 ≤ k ≤ q−1 ; 2 |Tss | =  4, if q+1 ≤ k ≤ 3q−3 . 2 2 2 Proof: (i) Since q ≡ 1 mod 4, 4|(q − 1), then −(s + q + 1)q ≡ −[ q−1 4 · 2(q + 1) + 2(q 2 + 1) + q 2 + q] ≡ 2n − ( q 2 +1 2 + q 2 + q) ≡ 2q 2 + 2 − ( q 2 +1 2 + q 2 + q) ≡ q 2 +1 2 1 4 · −q+1 ≡ s − q + 1 mod 2n. Hence, (Cs+q+1 , Cs−q+1 ) forms a skew-asymmetric pair. S (ii)Let T = ki=0 Cs+2i , where 0 ≤ k ≤ 3q−3 2 . According to Ref. [21], one can obtain that for 0 ≤ k ≤ For q+1 2 ≤k q−1 ⊥h ⊆ C. It means that |T | = 0 if 0 ≤ k ≤ q−1 . ss 2 , C 2 3q−3 ≤ 2 , since (Cs+q+1 , Cs−q+1 ) forms a skew-asymmetric pair, Tss com- prises the set {Cs+q+1 , Cs−q+1 } at least. According to the concept about a decomposition of the defining set T , one obtain that Tsas = T \Tss . In order to testify |Tss | = 4 if q+1 2 ≤ k ≤ 3q−1 2 , from Definition 2.1 and Lemma 2.2, we need to testify that there is no skew symmetric cyclotomic coset, and any two cyclotomic coset do not form a skew asymmetric pair in Tsas . Let I = {s + 2i|0 ≤ i ≤ 3q−3 2 }, I1 = {s + 2i|0 ≤ i ≤ q−1 2 }, I2 = {s + 2i| q+3 2 ≤i≤ 3q−3 2 }. Then I = I1 ∪ I2 ∪ {s + q + 1}. Only we need to testy that for ∀x ∈ I1 ∪ I2 , then −qx (mod 2n)6∈ I1 ∪ I2 . Let x, y ∈ I1 ∪ I2 . From Lemma 2.3, Cx is not a skew symmetric cyclotomic coset, and any Cx , Cy do not form a skew asymmetric pair if and only if x + yq 6≡ 0 mod 2n. q−1 q−1 2 4 (2n) < 4 (2n)+ q + 1 = s(q + 1) ≤ x+ yq ≤ (s + q − 1)(q + 1) = q−1 2 s(q + 1) + (q − 1)(q + 1) = q−1 4 (2n) + 2q < ( 4 + 1)(2n). q−1 2 2 For x, y ∈ I2 , then ( q−1 4 +1)(2n) < 4 (2n)+q +1+(q +1)+4(q +1) = (s+q +3)(q + 2 1) ≤ x + yq ≤ (s + 3q − 3)(q + 1) = s(q + 1) + (3q − 3)(q + 1) = q−1 4 (2n) + 4(q + 1) − 6 < ( q−1 4 + 2)(2n). q−1 2 For x ∈ I1 ∪ I2 and y ∈ I2 , then ( q−1 4 + 1)(2n) < 4 (2n) + 2(q + 1) + 3q − 1 = For x, y ∈ I1 , then s + (s + q + 3)(q + 1) ≤ x + yq ≤ (s + 3q − 3)(q + 1) = s(q + 1) + (3q − 3)(q + 1) = q−1 4 (2n) + 4(q 2 + 1) − 6 < ( q−1 4 + 2)(2n). Hence, there is no skew symmetric cyclotomic cosets, and any two cyclotomic coset do not form a skew asymmetric pair. That implies that Tss = {Cs+q+1 , Cs−q+1 } = {s + q + 1, s − q − 1, s − q + 1, s + q − 1} and |Tss | = 4 for S T = ki=0 Cs+2i , where 0 ≤ k ≤ 3q−3 2 . 8 q+1 2 ≤k≤ 3q−3 2 , when the defining set Theory 3.3: Let q be an odd prime power and q ≡ 1 mod 4. Then there exists a q-ary [[q 2 + 1, q 2 − 2d + 7, d; 4]]- EA-quantum MDS codes, where q + 3 ≤ d ≤ 3q − 1 is even. Proof: Consider the negacyclic codes over Fq2 of length n = q 2 + 1 with defining set S for q be an odd prime power and q ≡ 1 mod 4. T = ki=0 Cs+2i , where 0 ≤ k ≤ 3q−3 2 q+1 2 By Lemma 3.2 (i), there is c = |Tss | = 4 if ≤k ≤ 3q−3 2 . Since every q 2 -cyclotomic coset Cx = {x, n − x} and x must be odd number, we can obtain that T consists of 2k + 1 integers {n − (s + 2k), · · · , n − s − 2, s, s + 2, · · · , s + 2k}. It implies that C has minimum distance at least 2k + 2. Hence, C is a q 2 -ary negacyclic code with parameters [n, n − 2(k + 1) + 1, ≥ 2k + 2]. Combining Theory 2.7 with EA-quantum Singleton bound, we can obtain a EA-quantum MDS code with parameters [[q 2 + 1, q 2 − 2d + 7, d; 4]]q , where q + 3 ≤ d ≤ 3q − 1 is even. Table 1 EA-quantum MDS codes with length n = q 2 + 1 for q ≡ 1 mod 4 3.2 q n [[n, k, d; 4]]q d d in [33] 9 82 [[82, 88 − 2d, d; 4]]9 12 ≤ d ≤ 26 is even 14 ≤ d ≤ 18 is even 13 170 [[170, 176 − 2d, d; 4]]13 16 ≤ d ≤ 38 is even 18 ≤ d ≤ 26 is even 17 290 [[290, 296 − 2d, d; 4]]17 20 ≤ d ≤ 50 is even 22 ≤ d ≤ 34 is even 19 362 [[362, 368 − 2d, d; 4]]19 22 ≤ d ≤ 56 is even 24 ≤ d ≤ 38 is even 25 626 [[626, 632 − 2d, d; 4]]25 28 ≤ d ≤ 74 is even 30 ≤ d ≤ 50 is even 29 842 [[842, 848 − 2d, d; 4]]29 32 ≤ d ≤ 86 is even 34 ≤ d ≤ 58 is even EA-quantum MDS Codes of Lenght n = q 2 + 1 with q ≡ 3 mod 4 In this subsection, we construct some classes of q-ary EA-quantum MDS Codes of Length n = q 2 + 1, where q be a prime or an odd prime power of the form q ≡ 3 mod 4. Let r = q + 1 and η ∈ Fq2 be a primitive rth root of unity. To do this, we give a decomposition of defining set sufficient condition for η-constacyclic codes over Fq2 of length n which do not contain their Hermitian duals. Lemma 3.4 [22]: Let n = q 2 + 1 with q ≡ 3 mod 4, and s = n 2. Then the q 2 −cyclotomic cosets modulo (q + 1)n containing odd integers form 1 to (q + 1)n are Cs = {s} and Cs+(q+1)i = {s + (q + 1)i, s − (q + 1)i} for 1 ≤ i ≤ s − 1. Lemma 3.5: Let n = q 2 + 1 with q ≡ 3 mod 4, q ≥ 7 and s = n2 . If C is a q 2 -ary S constacyclic code of length n with define set T = ki=0 Cs+(q+1)i , where 0 ≤ k ≤ 3q−3 2 , 9 and the decomposition of a defining set T = Tss S Tsas , then (Cs+ (q+1) (q+1) , Cs+ (q−1) (q+1) ) 2 2 forms a skew asymmetric pair. And   0, if 0 ≤ k ≤ q−1 ; 2 |Tss | =  4, if q+1 ≤ k ≤ 3(q−1) . 2 2 Proof: Let r = q + 1. (a) From Lemma 3.4, C1 = Cs− (q−1) (q+1) = Cs+ (q−1) (q+1) , Since −(s + −[r(q 2 + 1) + r q−1 2 − (q 2 +1) 2 ] ≡ −( q 2 −1 2 2 − q 2 +1 2 ) 2 q−1 2 }, I2 = {s+ri| q+3 2 ≤i ≡ ≡ 1 mod rn, (Cs+ (q+1) (q+1) , Cs+ (q−1) (q+1) ) 2 forms a skew-asymmetric pair. S (b) Let T = ki=0 Cs+ri , where 0 ≤ k ≤ I1 = {s+ri|0 ≤ i ≤ (q+1)2 2 )q 3q−3 2 . Let I 3q−3 ≤ 2 }. Then 2 = {s + ri|0 ≤ i ≤ 3q−3 2 }, I = I1 ∪I2 ∪{Cs+ (q−1) (q+1) }. 2 Only we need to testy that for ∀x ∈ I1 ∪ I2 , then −qx (mod rn)6∈ I1 ∪ I2 . Let x, y ∈ I1 ∪ I2 . From Lemma 2.3, Cx is not a skew symmetric cyclotomic coset, and any Cx , Cy do not form a skew asymmetric pair if and only if x + yq 6≡ 0 mod rn. 2 For x, y ∈ I1 , then s(q + 1) ≤ x + yq ≤ [s + (q + 1) q−1 2 ](q + 1) = rq < rn. Tss = ∅ and |Tss | = 0 if 0 ≤ k ≤ For q+1 2 ≤ k ≤ q−1 2 . 3(q−1) , 2 since (Cs+ (q+1) (q+1) , Cs+ (q−1) (q+1) ) forms a skew-asymmetric 2 2 pair, Tss comprises the set {Cs+ (q+1) (q+1) , Cs+ (q−1) (q+1) } at least. According to the concept 2 2 about a decomposition of the defining set T , one obtain that Tsas = T \Tss . In order to testify |Tss | = 4 if q+1 2 ≤k≤ 3q−3 2 , from Definition 2.1 and Lemma 2.2, we need to testify that there is no skew symmetric cyclotomic coset, and any two cyclotomic coset do not form a skew asymmetric pair in Tsas . For x, y ∈ I2 , then rn < r(q 2 +1)+r(2q +1) = [s+ q+3 2 (q +1)](q +1) ≤ x+yq ≤ [s+(q + 1) 3(q−1) 2 ](q + 1) = rn < 2q 2 +3q+1 r 2 = 3(q−1) q 2 +1 (q + 1)r = 2rq 2 < 2(rn). For x ∈ I1 ∪ I2 and y ∈ I2 , then 2 r+ 2 3(q−1) 2 s + [s + (q + 1) q+3 2 ]q ≤ x + yq ≤ [s + 2 (q + 1)](q + 1) = 2rq < 2(rn). Hence, there is no skew symmetric cyclotomic cosets, and any two cyclotomic coset do not form a skew asymmetric pair. That implies that Tss = {Cs+ (q+1) (q+1) , Cs+ (q−1) (q+1) } = 2 2 (q+1) (q−1) (q−1) (q+1) 2 (q + 1), s − 2 (q + 1), s + 2 (q + 1), s − 2 (q + 1)} and |Tss | S 3(q−1) q+1 , when the defining set T = ki=0 Cs+ri , where 0 ≤ k ≤ 3q−3 2 ≤k ≤ 2 2 . {s + = 4 for Theory 3.6: Let q be an odd prime power and q ≡ 3 mod 4. Then there exists a q-ary [[q 2 + 1, q 2 − 2d + 7, d; 4]]- EA-quantum MDS codes, where q + 3 ≤ d ≤ 3q − 1 is even. 10 Proof: Consider the constacyclic codes over Fq2 of length n = q 2 + 1 with defining S for q be an odd prime power and q ≡ 3 mod 4. set T = ki=0 Cs+ri , where 0 ≤ k ≤ 3q−3 2 By Lemma 3.5, there is c = |Tss | = 4 if q+1 2 ≤k≤ 3(q−1) . 2 Since every q 2 -cyclotomic coset Cs+x = {s + x, s − x} and s + x must be odd number, we can obtain that T consists of 2k + 1 integers {s − rk, · · · , s − r, s, s + r, · · · , s + rk} (mod rn). It implies that C has minimum distance at least 2k + 2. Hence, C is a q 2 -ary constacyclic code with parameters [n, n − 2(k + 1) + 1, ≥ 2k + 2]. Combining Theory 2.7 with EA-quantum Singleton bound, we can obtain a EA-quantum MDS code with parameters [[q 2 + 1, q 2 − 2d + 7, d; 4]]q , where q + 3 ≤ d ≤ 3q − 1 is even. Table 2 EA-quantum MDS codes with length n = q 2 + 1 for q ≡ 3 mod 4 3.3 q n [[n, k, d; 4]]q d 7 50 [[50, 56 − 2d, d; 4]]7 10 ≤ d ≤ 20 is even 11 122 [[122, 128 − 2d, d; 4]]11 14 ≤ d ≤ 32 is even 19 362 [[362, 368 − 2d, d; 4]]19 22 ≤ d ≤ 56 is even 23 530 [[530, 536 − 2d, d; 4]]23 26 ≤ d ≤ 68 is even 31 962 [[962, 968 − 2d, d; 4]]31 34 ≤ d ≤ 92 is even 43 1850 [[1850, 1856 − 2d, d; 4]]43 46 ≤ d ≤ 128 is even EA-quantum MDS Codes of Lenght n = q 2 +1 10 In this subsection, we construct new q-ary EA-quantum MDS Codes of Length n = q 2 +1 10 , where q be an odd prime power of the form q = 10m + 3 or q = 10m + 7. Let r = 2 and η ∈ Fq2 be a primitive rth root of unity. To do this, we give the well-known result of q 2 −cyclotomic coset Ci modulo 2n for each i ∈ Z2n . Lemma 3.7: Let n = q 2 +1 10 . Then the q 2 -cyclotomic coset Ci modulo 2n containing all odd integers form 1 to 2n are Cn = {n} and C2i−1 = {2i − 1, 2i + 1} for 1 ≤ i ≤ n−1 2 . 2 +1 . If C is a q 2 -ary Lemma 3.8: Let q be and prime power of form 10m + 3 and n = q 10 S negacyclic code of length n with define set T = ki=0 Cn+2i , where 0 ≤ k ≤ 3m, and the S decomposition of a defining set T = Tss Tsas , then (i) Cn is skew symmetric. q 2 +1 (ii) |Tss | = 1, if 0 ≤ k ≤ 3m. 2 2 q +1 q +1 Proof: Since −nq ≡ −( 5 · q−1 2 − 10 ) ≡ 10 ≡ n mod 2n, Cn is skew symmetric. S Let T = ki=0 Cn+2i , where 0 ≤ k ≤ 3m. Since Cn is skew symmetric, Tss comprises the set {Cn } at least. According to the concept about a decomposition of the defining 11 set T , one obtain that Tsas = T \Tss . In order to testify |Tss | = 1 if 0 ≤ k ≤ 3m, from Definition 2.1 and Lemma 2.2, we need to testify that there is no skew symmetric cyclotomic coset, and any two cyclotomic coset do not form a skew asymmetric pair in Tsas . Let I = {n + 2i|1 ≤ i ≤ 3m}. Only we need to testy that for ∀x ∈ I, −qx (mod 2n)6∈ I and Tss = {Cn }. That implies that if x, y ∈ I, from Lemma 2.3, Cx is not a skew symmetric cyclotomic coset, and any Cx , Cy do not form a skew asymmetric pair if and only if x + yq 6≡ 0 mod 2n. Divide I into three parts I1 = [n + 2, n + 2m], I2 = [n + 2m + 2, n + 4m] and I3 = [n + 4m + 2, n + 6m]. q = 10m + 3 and n = If x, y ∈ I1 , then q+1 2 q 2 +1 10 = 10m2 + 6m + 1. · 2n < (n + 2)(q + 1) ≤ x + yq ≤ (n + 2m)(q + 1) = (q + 1)n + q+1 20m2 + 8m < ( q+1 2 + 1) · 2n. If x, y ∈ I2 , then ( 2 + 1) · 2n < q+1 2 · 2n + 20m2 + 28m + 8 = (n + 2m + 2)(q + 1) ≤ x + yq ≤ (n + 4m)(q + 1) = (q + 1)n + 40m2 + 16m < ( q+1 2 + 2) · 2n. If x, y ∈ I3 , then ( q+1 2 + 2) · 2n < q+1 2 · 2n + 40m2 + 36m + 8 = (n + 4m + 2)(q + 1) ≤ x + yq ≤ (n + 6m)(q + 1) = (q + 1)n + 60m2 + 24m < ( q+1 2 + 3) · 2n. Hence, there is no skew symmetric cyclotomic cosets, and any two cyclotomic coset do not form a skew asymmetric pair in T \ {Cn }. That implies that Tss = {Cn } and |Tss | = 1 S for 0 ≤ k ≤ 3m, when the defining set T = ki=0 Cn+2i , where 0 ≤ k ≤ 3m. Theory 3.9: Let q be an odd prime power of form 10m + 3, where m is a positive integer. Let n = q 2 +1 10 . 2 2 +1 q +1 Then there exists a q-ary [[ q 10 , 10 − 2d + 3, d; 1]]- EA-quantum MDS codes, where 2 ≤ d ≤ 6m + 2 is even. 2 +1 with defining set Proof: Consider the negacyclic codes over Fq2 of length n = q 10 Sk T = i=0 Cn+2i , where 0 ≤ k ≤ 3m for q be an odd prime power with the form of 10m + 3. By Lemma 3.8, there is c = |Tss | = 1 if 0 ≤ k ≤ 3m. Since every q 2 -cyclotomic coset Cn+x = {n + x, n − x}, 0 ≤ x ≤ n − 1 and n + x must be odd number, we can obtain that T consists of 2k + 1 integers {n − 2k, · · · , n − 2, n, n + 2, · · · , n + 2k}. It implies that C has minimum distance at least 2k + 2. Hence, C is a q 2 -ary negacyclic code with parameters [n, n − 2(k + 1) + 1, ≥ 2k + 2]. Combining Theory 2.7 with EA-quantum Singleton bound, 2 2 we can obtain a EA-quantum MDS code with parameters [[ q 10+1 , q 10+1 − 2d + 3, d; 1]]q , where 2 ≤ d ≤ 6m + 2 is even. Recently, based on constacyclic codes, zhang et al. [24] constructed a family of new q-ary quantum MDS codes with parameters [[ q 12 2 +1 5 ,q 2 +1 5 − 2d + 2, d]]q , where q = 10m + 3 and 2 ≤ d ≤ 6m + 2 is even. However, there is no paper about the construction of q-ary 2 2 +1 q +1 , 10 − 2d + 2, d]]q with the minimal distance quantum MDS codes with parameters [[ q 10 of 2 ≤ d ≤ 6m + 2 is even. From Theory 3.8, with the help of 1 ebit, we construct a family 2 2 +1 q +1 , 10 − 2d + 3, d; 1]], where of new q−ary EA-quantum MDS codes with parameters [[ q 10 q = 10m + 3 and 2 ≤ d ≤ 6m + 2. Table 4 EAQMDS codes with n = q 2 +1 10 for q = 10m + 3 q [[n, k, d; 1]]q d 13 [[17, 20 − 2d, d; 1]]13 2 ≤ d ≤ 8 is even 23 [[53, 56 − 2d, d; 1]]23 2 ≤ d ≤ 14 is even 43 [[185, 188 − 2d, d; 1]]43 2 ≤ d ≤ 26 is even 53 [[281, 284 − 2d, d; 1]]53 2 ≤ d ≤ 32 is even 2 Lemma 3.10: Let q be and prime power of form 10m + 7 and n = q 10+1 . If C is a S q 2 -ary negacyclic code of length n with define set T = ki=0 Cn+2i , where 0 ≤ k ≤ 3m + 1, S and the decomposition of a defining set T = Tss Tsas , then (i) Cn is skew symmetric. (ii) |Tss | = 1, if 2 ≤ δ ≤ 6m + 4. 2 q 2 +1 2 q +1 q +1 Proof: Since −nq ≡ −( 5 · q−1 2 − 10 ) ≡ 10 ≡ n mod 2n, Cn is skew symmetric. S Let T = ki=0 Cn+2i , where 0 ≤ k ≤ 3m+1. Since Cn is skew symmetric, Tss comprises the set {Cn } at least. According to the concept about a decomposition of the defining set T , one obtain that Tsas = T \Tss . In order to testify |Tss | = 1 if 0 ≤ k ≤ 3m + 1, from Definition 2.1 and Lemma 2.2, we need to testify that there is no skew symmetric cyclotomic coset, and any two cyclotomic coset do not form a skew asymmetric pair in Tsas . Let I = {n + 2i|1 ≤ i ≤ 3m + 1}. Only we need to testy that for ∀x ∈ I, −qx (mod 2n)6∈ I and Tss = {Cn }. That implies that if x, y ∈ I, from Lemma 2.3, Cx is not a skew symmetric cyclotomic coset, and any Cx , Cy do not form a skew asymmetric pair if and only if x + yq 6≡ 0 mod 2n. Divide I into three parts I1 = [n + 2, n + 2m], I2 = [n + 2m + 2, n + 4m + 2] and I3 = [n + 4m + 4, n + 6m + 2]. q = 10m + 7 and n = If x, y ∈ I1 , then q+1 2 q 2 +1 10 = 10m2 + 14m + 5. · 2n < (n + 2)(q + 1) ≤ x + yq ≤ (n + 2m)(q + 1) = (q + 1)n + q+1 20m2 + 16m < ( q+1 2 + 1)·2n. If x, y ∈ I2 , then ( 2 + 1)·2n < (n+2m+2)(q+1) ≤ x+yq ≤ (n+4m+2)(q+1) = If x, y ∈ I3 , then ( q+1 2 + 2) · 2n < q+1 2 q+1 2 2 ·2n + 20m + 36m + 16 = (q+1)n+40m2 +52m+16 < ( q+1 2 +2)·2n. · 2n + 2n + 16m + 12 = (n + 4m + 4)(q + 1) ≤ x + yq ≤ (n + 6m + 2)(q + 1) = (q + 1)n + 3n − 16m − 14 < ( q+1 2 + 3) · 2n. 13 Hence, there is no skew symmetric cyclotomic cosets, and any two cyclotomic coset do not form a skew asymmetric pair in T \ {Cn }. That implies that Tss = {Cn } and |Tss | = 1 S for 0 ≤ k ≤ 3m + 1, when the defining set T = ki=0 Cn+2i , where 0 ≤ k ≤ 3m + 1. Theory 3.11: Let q be an odd prime power of form 10m + 7, where m is a positive integer. Let n = q 2 +1 10 . 2 2 +1 q +1 Then there exists a q-ary [[ q 10 , 10 − 2d + 3, d; 1]]- EA-quantum MDS codes, where 2 ≤ d ≤ 6m + 4 is even. 2 +1 with defining set Proof: Consider the negacyclic codes over Fq2 of length n = q 10 Sk T = i=0 Cn+2i , where 0 ≤ k ≤ 3m + 1 for q be an odd prime power with the form of 10m + 7. By Lemma 3.11, there is c = |Tss | = 1 if 0 ≤ k ≤ 3m + 1. Since every q 2 -cyclotomic coset Cn+x = {n + x, n − x}, 0 ≤ x ≤ n − 1 and n + x must be odd number, we can obtain that T consists of 2k + 1 integers {n − 2k, · · · , n − 2, n, n + 2, · · · , n + 2k}. It implies that C has minimum distance at least 2k + 2. Hence, C is a q 2 -ary negacyclic code with parameters [n, n − 2(k + 1) + 1, ≥ 2k + 2]. Combining Theory 2.7 with EAquantum Singleton bound, we can obtain a EA-quantum MDS code with parameters 2 2 +1 q +1 , 10 − 2d + 3, d; 1]]q , where 2 ≤ d ≤ 6m + 4 is even. [[ q 10 Table 5 EAQMDS codes with n = 3.4 q 2 +1 10 for q = 10m + 7 q [[n, k, d; 1]]q d 17 [[29, 32 − 2d, d; 1]]17 2 ≤ d ≤ 10 is even 27 [[73, 76 − 2d, d; 1]]27 2 ≤ d ≤ 16 is even 37 [[137, 140 − 2d, d; 1]]37 2 ≤ d ≤ 22 is even 47 [[221, 224 − 2d, d; 1]]47 2 ≤ d ≤ 28 is even EA-quantum MDS Codes of Lenght n = q 2 −1 h with h ∈ {3, 5, 7} Let h ∈ {3, 5, 7}, q be an odd prime power with h|(q + 1). Suppose n = q 2 −1 h and r = h. Let η ∈ Fq2 be a primitive r th root of unity. Since rn = q 2 − 1 clearly, every q 2 -cyclotomic coset modulo rn contains exactly one element. In [18], using Hermitian construction and η-constacyclic codes, Chen et al. have constructed a family of new quantum MDS codes with parameters [[ q 2 −1 h ,q 2 −1 h − 2d + 2, d]]q , where q is an odd prime power with h|(q + 1), h ∈ {3, 5, 7} and 2 ≤ d ≤ (q+1)(h+1) 2h − 1. It is very hard to enlarge the minimal distance for this type quantum MDS codes. In this subsection, adding one ebit, we construct a new family of a family of new EA-quantum MDS codes with parameters 14 [[ q 2 −1 h ,q 2 −1 h − 2d + 3, d; 1]]q , where (q+1)(h−1) 2h +1≤d≤ (q+1)(h+3) 2h − 1. 2 Lemma 3.12: Let q is an odd prime power with h|(q + 1), h ∈ {3, 5, 7} and n = q h−1 . S If C is a q 2 -ary constacyclic code of length n with define set T = ki= (h−3)(q+1) {C1+hi }, 2h S where (h−3)(q+1) ≤ k ≤ q − 2, and the decomposition of a defining set T = T Tsas , then ss 2h (i) C1+( (h−1)(q+1) −1)h is skew symmetric. 2h − 1)h = Proof: (i) Since 1 + ( (h−1)(q+1) 2h 2 q ≡ −[ h−1 − (h−1)(q−1) 2 2 (q − 1) − (h−1)(q−1) ] 2 C1+( (h−1)(q+1) −1)h is skew symmetric. 2h S (ii) Let T = ki= (h−3)(q+1) {C1+hi }, where 2h ≡ (h−1)(q+1) −1 ≤ k ≤ q −2. 2h (h−1)(q−1) (h−1)(q+1) and −[1 + ( − 1)h]q ≡ 2 2h (h−1)(q−1) ≡ 1 + ( (h−1)(q+1) − 1)h mod hn, 2 2h (ii) |Tss | = 1, if (h−3)(q+1) 2h ≤ k ≤ q−2. Since C1+( (h−1)(q+1) −1)h 2h is skew symmetric, Tss comprises the set {C1+( (h−1)(q+1) −1)h } at least. According to the 2h concept about a decomposition of the defining set T , one obtain that Tsas = T \Tss . In order to testify |Tss | = 1 if (h−1)(q+1) 2h − 1 ≤ i ≤ q − 2, from Definition 2.1 and Lemma 2.2, we need to testify that there is no skew symmetric cyclotomic coset, and any two cyclotomic coset do not form a skew asymmetric pair in Tsas . ≤ i ≤ q − 2} \ (1 + ( (h−1)(q+1) − 1)h) and r = h. Only we Let I = {1 + hi| (h−3)(q+1) 2h 2h need to testy that for ∀x ∈ I, −qx (mod rn)6∈ I and Tss = {C1+( (h−1)(q+1) −1)h }. That 2h implies that if x, y ∈ I, from Lemma 2.3, Cx is not a skew symmetric cyclotomic coset, and any Cx , Cy do not form a skew asymmetric pair if and only if x + yq 6≡ 0 mod 2n. (h−3)(q+1) h, 1 + ( (h−1)(q+1) − 2)h], I2 = [1 + 2h 2h )h, 1 + ( (h+1)(q+1) − 2)h] and I3 = [1 + ( (h+1)(q+1) − 1)h, 1 + (q − 2)h]. rn = ( (h−1)(q+1) 2h 2h 2h (h−3)(q+1) 2 h)(q + 1) ≤ hn = q 2 − 1. If x, y ∈ I1 , then (h−3) 2 (q − 1) + (h − 2)(q + 1) = (1 + 2h 2 − 2)h)(q + 1) = (h−1) x + yq ≤ (1 + ( (h−1)(q+1) 2h 2 (q − 1) − h(q + 1). Since h ∈ {3, 5, 7}, (h−3) = {0, 1, 2}, and (h−1) = {1, 2, 3}. Hence, for x, y ∈ I1 , ln < x + yq < (l + 1)n, where 2 2 Divide I into three parts I1 = [1 + 0 ≤ l ≤ 2 is a integer. If x, y ∈ I2 , then h−1 2 2 (q −2)h)(q +1) = ( (h+1)(q+1) 2h and (h+1) 2 − 1) + h(q + 1) = (1 + ( (h−1)(q+1) )h)(q + 1) ≤ x + yq ≤ (1 + 2h h+1 2 2 (q −1)−(h−2)(q +1). Since h ∈ {3, 5, 7}, (h−1) 2 = {1, 2, 3}, = {2, 3, 4}. Hence, for x, y ∈ I2 , ln < x + yq < (l + 1)n, where 1 ≤ l ≤ 3 is a integer. For h = 3, if x, y ∈ I3 , then 2n < h+1 2 2 (q −1)+2(q +1) = (1+( (h+1)(q+1) −1)h)(q +1) ≤ 2h x + yq ≤ (1 + (q − 2)h)(q + 1) = h(q 2 − 1) − (h − 1)(q + 1) < 3n. − 1)h, 1 + ( (h+3)(q+1) − 2)h] For h = 5, Divide I3 into two parts I3 = [1 + ( (h+1)(q+1) 2h 2h ′ − 1)h, 1 + (q − 2)h]. and I3 = [1 + ( (h+3)(q+1) 2h ′′ 15 ′ if x, y ∈ I3 , then 3n < h+1 2 2 (q − 1) + 2(q + 1) − 2)h)(q + 1) = (1 + ( (h+3)(q+1) 2h h+3 2 2 (q = (1 + ( (h+1)(q+1) − 1)h)(q + 1) ≤ x + yq ≤ 2h ′′ − 1) − (h − 2)(q + 1) < 4n; if x, y ∈ I3 , 4n < −2)h)(q+1) ≤ x+yq ≤ (1+(q−2)h)(q+1) = h(q 2 −1)−(h−1)(q+1) < 5n. (1+( (h+3)(q+1) 2h − 1)h, 1 + ( (h+3)(q+1) − 2)h], For h = 7, Divide I3 into two parts I3a = [1 + ( (h+1)(q+1) 2h 2h I3b = [1+( (h+3)(q+1) −1)h, 1+( (h+5)(q+1) −2)h] and I3c = [1+( (h+5)(q+1) −1)h, 1+(q −2)h]. 2h 2h 2h If x, y ∈ I3a , 4n < − 1) + 2(q + 1) = (1 + ( (h+1)(q+1) − 1)h)h)(q + 1) ≤ x + 2h h+1 2 2 (q yq ≤ (1 + ( (h+3)(q+1) − 2)h)(q + 1) = 2h h+3 2 2 (q − 1) − (h − 2)(q + 1) < 5n; if x, y ∈ I3b , − 2)h)(q + 1) ≤ x + yq ≤ (1 + ( (h+5)(q+1) − 2)h)(q + 1) = 5n < (1 + ( (h+3)(q+1) 2h 2h h+5 2 2 (q − − 1)h)(q + 1) ≤ x + yq ≤ 1) − (h − 2)(q + 1) < 6n. if x, y ∈ I3c , 6n < (1 + ( (h+3)(q+1) 2h (1 + (q − 2)h)(q + 1) = h(q 2 − 1) − (h − 1)(q + 1) < 7n. Hence, there is no skew symmetric cyclotomic cosets, and any two cyclotomic coset do not form a skew asymmetric pair in T \ {C1+( (h−1)(q+1) −1)h }. That implies that 2h − 1 ≤ i ≤ q − 2, when the defining Tss = {C1+( (h−1)(q+1) −1)h } and |Tss | = 1 for (h−1)(q+1) 2h 2h Sk set T = i= (h−3)(q+1) {C1+hi }, where (h−3)(q+1) ≤ k ≤ q − 2. 2h 2h Theory 3.13: Let q is an odd prime power with h|(q + 1), h ∈ {3, 5, 7} and n = Then there exists a q-ary [[ q 2 −1 h ,q 2 −1 h − 2d + 3, d; 1]], where q+1 h q 2 −1 h . (h+3)(q+1) − 1. 2h q 2 −1 with definh +1≤d≤ Proof: Consider the constacyclic codes over Fq2 of length n = S ≤ k ≤ q − 2, h ∈ {3, 5, 7} and ing set T = ki= (h−3)(q+1) {C1+hi }, where (h−3)(q+1) 2h 2h h|(q + 1) for q be an odd prime power. By Lemma 3.12, there is c = |Tss | = 1 if (h−1)(q+1) 2h − 1 ≤ k ≤ q − 2. Since every q 2 -cyclotomic coset has one element which must be odd number, we can obtain that T consists of (i − (h−3)(q+1) 2h + 1) integers )h, 1 + ( (h−3)(q+1) + 1)h, 1 + ( (h−3)(q+1) + 2)h · · · , 1 + kh}. It implies {1 + ( (h−3)(q+1) 2h 2h 2h that C has minimum distance at least (i − stacyclic code with parameters [n, n − 2(i − (h−3)(q+1) 2h (h−3)(q+1) 2h + 2). Hence, C is a q 2 -ary con+ 1) + 1, ≥ (i − (h−3)(q+1) 2h + 2)]. Combining Theory 2.7 with EA-quantum Singleton bound, we can obtain a EA-quantum MDS code with parameters [[ q 2 −1 h ,q 2 −1 h − 2d+ 3, d; 1]]q , where Table 6 EAQMDS codes with n = 16 q+1 h +1 q 2 −1 h ≤d≤ (h+3)(q+1) − 1. 2h 4 q h [[n, k, d; 1]]q d 11 3 [[40, 43 − 2d, d; 1]]11 5 ≤ d ≤ 11 17 3 [[96, 99 − 2d, d; 1]]17 7 ≤ d ≤ 17 19 5 [[72, 75 − 2d, d; 1]]19 5 ≤ d ≤ 15 29 5 [[168, 171 − 2d, d; 1]]29 7 ≤ d ≤ 23 13 7 [[24, 27 − 2d, d; 1]]13 3≤d≤9 41 7 [[240, 243 − 2d, d; 1]]41 7 ≤ d ≤ 29 SUMMARY In this paper, we propose a concept about a decomposition of the defining set of constacyclic codes. Using this concept, we construct four classes of q-ary entanglement-assisted quantum MDS (EAQMDS) codes based on classical constacyclic MDS codes by exploiting less pre-shared maximally entangled states. In Table 7, we list the q-ary entanglement-assisted quantum MDS codes constructed in this paper. By consuming four pre-shared maximally entangled states, we obtain the EA-quantum MDS codes of n = q 2 + 1 with the minimal distance upper limit greater than 3q − 1 (even). These EA-quantum MDS codes are improved the parameters of codes in Ref. [33]. Consumed only one pair of maximally entangled states, each EA-quantum MDS code of length n = q 2 +1 10 has half length comparing with the standard QMDS code of the same minimum distance constructed in Ref. [18]. Moreover, consuming only one pair of maximally entangled states, we obtain a family of EA-quantum MDS codes of q 2 −1 h with the minimal distance larger than the standard quantum MDS codes in Ref. [18]. Comparing the parameters with all known q-ary EA-quantum MDS codes, we find that these quantum MDS codes are new in the sense that their parameters are not covered by the codes available in the literature. Table 7 New parameters of EAQMDS codes 17 Class length [[n, k, d; c]]q Distance 1 q2 + 1 [[q 2 + 1, q 2 − 2d + 7, d; 4]]q q + 3 ≤ d ≤ 3q − 1 is even 2 q 2 +1 10 ,q = 10m + 3 [[ q 10+1 , q 10+1 − 2d + 3, d; 1]]q 3 q 2 +1 10 ,q = 10m + 7 [[ q 10+1 , q 10+1 − 2d + 3, d; 1]]q 4 n= q 2 −1 h [[ q 2 2 2 2 2 −1 h ,q 2 −1 h − 2d + 3, d; 1]]q 2 ≤ d ≤ 4m + 2 is even 2 ≤ d ≤ 4m + 4 is even q+1 h +1≤d≤ (h+3)(q+1) 2h −1 h ∈ {3, 5, 7} Remark: This paper was submitted to the Journal ”Designs, Codes and Cryptography” on Jan. 11, 2018. Acknowledgment This work is supported by the National Key R&D Program of China under Grant No. 2017YFB0802400, the National Natural Science Foundation of China under grant No. 61373171, the National Natural Science Foundation of China under Grant No.11471011, the Natural Science Foundation of Shaanxi under Grant No.2017JQ1032, 111 Project under grant No.B08038, and the Key Project of Science Research of Anhui Province(Grant No.KJ2017A519). References [1] P. Shor, Scheme for reducing decoherence in quantum memory. Phys. Rev. A 52, 2493C2496, 1995. [2] A. Calderbank, P. Shor, Good quantum error-correcting codes exist. Phys. Rev. A 54, 1098-1105, 1996. [3] A. Calderbank, E. Rains, P. Shor, and N. A. Sloane, Quantum error correction via codes over GF(4), IEEE Trans. Inf. Theory, vol. 44, no. 4, pp. 1369-1387, 1998. [4] A. Ketkar, A. Klappenecker, S. Kumar and P. Sarvepalli, Nonbinary stabilizer codes over finite fields. IEEE Trans. Inform. Theory, vol. 52, no. 11, p. 4892-4914, 2006. 18 [5] T. Brun, I. Devetak, and M. Hsieh, Correcting quantum errors with entanglement. Science 314, 436-439 (2006) [6] M. Hsieh, I. Devetak, and T. Brun, General entanglement-assisted quantum errorcorrecting codes. Phys.Rev, A, 76, 062313, 2007. [7] M. Wilde, and T. Brun, Optimal entanglement formulas for entanglement-assisted quantum coding. Phys. Rev, A, 77, 064302, 2008. [8] C. Lai, and T. Brun, Entanglement Increases the Error-Correcting Ability of Quantum Error-Correcting Codes. Phys. Rev, A, 88, 012320, 2013. [9] C. Lai, T. Brun, and M. Wilde, Duality in Entanglement-assisted quantum error correction. IEEE Trans Inf. Theory 59, 4020-4024 (2013). [10] C. Lai, T. Brun and M. Wilde, Dualities and Identities for Entanglement-Assisted Quantum Codes. Quantum Inf. Process, 13:957C990 (2014). [11] M. Hsieh,W. Yen, and L. Hsu, High performance entanglement-assisted quantum LDPC codes need little entanglement, IEEE Trans. Inf. Theory, 57, 1761-1769, 2011. [12] Y. Fujiwara, D. Clark, P. Vandendriessche, M. De Boeck, and V. D. Tonchev, Entanglementassisted quantum low-density parity-check codes, Phys. Rev. A, vol. 82, no. 4, p. 042338, 2010. [13] M. Wilde, M. Hsieh, and Z. Babar, Entanglement-assisted quantum turbo codes, IEEE Trans. Inf. Theory, 60, 1203-1222, 2014. [14] L. Lu and R. Li, Entanglement-assisted quantum codes constructed from primitive quaternary BCH codes, Int. J. Quantum Inf., 12, 1450015, 2014. [15] L. Lu, R. Li, L. Guo, Q, Fu, Maximal entanglement entanglement-assisted quantum codes constructed from linear codes, Quantum Inf. Process, 12, 1450015, 2015. [16] L. Guo and R. Li, Linear Plotkin bound for entanglement-assisted quantum codes. Phys. Rev. A 87, 032309 (2013). [17] R. Li and Z. Xu, Construction of [[n, n.4, 3]]q quantum MDS codes for odd prime power q, Phys. Rev. A, 82, 052316-1-052316-4, 2010. 19 [18] B. Chen, S. Ling, and G. Zhang, Application of constacyclic codes to quantum MDS codes. IEEE Trans. Inform. Theory, 61, 1474-1484, 2015. [19] X. He, L. Xu, and H. Chen, New q-ary quantum MDS codes with distances bigger than 2q . Quantum Inf. Process., 15, 2745-2758, 2016. [20] L. Jin, H. Kan and J. Wen, Quantum MDS codes with relatively large minimum distance from Hermitian self-orthogonal codes. Des. Codes Cryptogr., 2016, doi:10.1007/s10623-016-0281-9. [21] X. Kai, and S. Zhu, New quantum MDS codes from negacyclic codes. IEEE Trans. Inform. Theory, 59, 1193-1197, 2013. [22] X. Kai, S. Zhu and P. Li, Constacyclic codes and some new quantum MDS codes. IEEE Trans. Inform. Theory, 60, 2080-2086, 2014. [23] L. Wang, and S. Zhu, New quantum MDS codes derived from constacyclic codes. Quantum Inf. Process., 14, 881-889, 2015. [24] T. Zhang, and Ge, G, Some New Classes of Quantum MDS Codes From Constacyclic Codes. IEEE Trans. Inform. Theory, 61, 5224-5228, 2015. [25] Zhang, T. and Ge, G, Quantum MDS code with large minimum distance. Des. Codes Cryptogr., doi:10.1007/s10623-016-0245-0. 2016 [26] R. Li, F. Zuo, Y. Liu, A study of skew asymmetric q2-cyclotomic coset and its application. J. Air Force Eng. Univ. (Nat. Sci. Ed.) 12(1), 87C89, 2011. (in Chinese) [27] R. Li, F. Zuo, Y. Liu and Z. Xu. Hermitian dual containing BCH codes and Construction of new quantum codes. Quantum Inf. and Comp, 13, 0021-0035, 2013. [28] F. J. Macwilliams and N. J. A. Sloane, The Theory of Error-Correcting Codes, Amsterdam, the Netherlands: North-Holland, 1977. [29] W. C. Huffman and V. Pless, Fundamentals of Error-Correcting Codes, Cambridge University Press, Cambridge, 2003. [30] Y. Yang and W. Cai, On self-dual constacyclic codes over finite fields, Designs, Codes Cryptogr., vol. 74, no. 2, pp. 355-364, 2013. 20 [31] J. Qian, L. Zhang, Entanglement-assisted quantum codes from arbitrary binary linear codes. Des. Codes Cryptogr., 77(1), 193-202, 2015. [32] J. Qian, L. Zhang, On MDS linear complementary dual codes and entanglementassisted quantum codes, Des. Codes Cryptogr., 2017, doi:10.1007/s10623-017-0413-x. [33] J. Chen, Y. Huang, C. Feng, R. Chen, Entanglement-assisted quantum MDS codes constructed from negacyclic codes, Quantum Inf. Process., 2017, doi: 10.1007/s11128017-1750-4. 21
7cs.IT
Finding Large Set Covers Faster via the Representation Method∗ Jesper Nederlof1 1 Department of Mathematics and Computer Science, Eindhoven University of Technology, The Netherlands. [email protected] arXiv:1608.03439v1 [cs.DS] 11 Aug 2016 Abstract The worst-case fastest known algorithm for the Set Cover problem on universes with n elements still essentially is the simple O∗ (2n )-time dynamic programming algorithm, and no non-trivial consequences of an O∗ (1.01n )-time algorithm are known. Motivated by this chasm, we study the following natural question: Which instances of Set Cover can we solve faster than the simple dynamic programming algorithm? Specifically, we give a Monte Carlo algorithm that determines 4 the existence of a set cover of size σn in O∗ (2(1−Ω(σ ))n ) time. Our approach is also applicable to Set Cover instances with exponentially many sets: By reducing the task of finding the chromatic number χ(G) of a given n-vertex graph G to Set Cover in the natural way, we show there is an 4 O∗ (2(1−Ω(σ ))n )-time randomized algorithm that given integer s = σn, outputs NO if χ(G) > s and YES with constant probability if χ(G) ≤ s − 1. On a high level, our results are inspired by the ‘representation method’ of Howgrave-Graham and Joux [EUROCRYPT’10] and obtained by only evaluating a randomly sampled subset of the table entries of a dynamic programming algorithm. 1998 ACM Subject Classification G.2.2 Graph Algorithms, Hypergraphs Keywords and phrases Set Cover, Exact Exponential Algorithms, Fine-Grained Complexity Digital Object Identifier 10.4230/LIPIcs.ESA.2016.[79] 1 Introduction The Set Cover problem is, after determining satisfiability of CNF formulas or Boolean circuits, one of the canonical NP-complete problems. It not only directly models many applications in practical settings, but also algorithms for it routinely are used as tools for theoretical algorithmic results (e.g., [17]). It is a problem ‘whose study has led to the development of fundamental techniques for the entire field’ of approximation algorithms.1 However, the exact exponential time complexity of Set Cover is still somewhat mysterious: We know algorithms need to use super-polynomial time assuming P 6= N P and (denoting n for the universe size) O∗ (2Ω(n) ) time assuming the Exponential Time Hypothesis, but how large the exponential should be is not clear. In particular, no non-trivial consequences of an O∗ (1.01n )-time algorithm are currently known. Even though it is one of the canonical NP-complete problems, the amount of studies of exact algorithms for Set Cover pales in comparison with the amount of literature on exact algorithms for CNF-Sat: Many works focus on finding O∗ (cn )-time algorithms for c < 2 ∗ 1 Funded by the NWO VENI project 639.021.438. This work was partly done while the author was visiting the Simons Institute for the Theory of Computing during the program ‘Fine-Grained Complexity and Algorithm Design’ in the fall of 2015. As the Wikipedia page on Set Cover quotes the textbook by Vazirani [34, p15]. © Jesper Nederlof; licensed under Creative Commons License CC-BY 24th Annual European Symposium on Algorithms (ESA 2016). Editors: Piotr Sankowski and Christos Zaroliagis; Article No. [79]; pp. [79]:1–[79]:20 Leibniz International Proceedings in Informatics Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany [79]:2 Finding Large Set Covers Faster via the Representation Method for CNF-Sat on n-variable CNF-formulas in special cases such as, among others, bounded clause width [33, 17, 12], bounded clause density [11, 26] or few projections [28, 31, 32]. Improved exponential time algorithms for special cases of problems other than CNF-Sat were also studied for e.g. Graph Coloring or Traveling Salesman on graphs bounded degree/average degree [8, 9, 15, 20]. In this paper we are interested in the exponential time complexity of Set Cover, and study which properties are sufficient to have improved exponential time algorithms. Our interest in finding faster exponential time algorithms for Set Cover does not only stem from it being a canonical NP-complete problem, but also from its unclear relation with CNF-Sat. Intriguingly, on one hand Set Cover has some similarities with the CNFSat: 1. Both problems take an (annotated) hypergraph as input 2. The improvability of the worst-case complexity of CNF-Sat is essentially equivalent to the improvability of the worst-case complexity of Hitting Set [14], which is just a reparametrization2 of Set Cover. But, on the other hand the problems are quite different to our understanding: 1. Most algorithms for Set Cover use dynamic programming or some variant of inclusion exclusion, while most algorithms for CNF-Sat are based on branching 2. No connection between the exponential time complexities of both problems is known (see [14]). One hope would be that a better understanding of the exact complexity of Set Cover might shed more light on this unclarity. Moreover, Cygan et al. [14] also show that if we would like to improve the run time O∗ (f (k)) of several parameterized algorithms to O∗ (f (k)1−Ω(1) ), we first need to find an O∗ (2(1−Ω(1))n )-time algorithm for Set Cover. These parameterized algorithms include the classic algorithm for Subset Sum, as well as more recent algorithms for Connected Vertex Cover and Steiner Tree. Relevant previous work The algorithmic results on Set Cover that are the most relevant to our work are as follows: The folklore dynamic programming algorithm runs in O∗ (2n ) time. A notable special case of Set Cover that can be solved in O∗ (2(1−Ω(1))n ) time is due 1 to Koivisto [29]: He gives an algorithm that runs in time O∗ (2(1− O(r) )n )-time algorithm if all sets are at most of size r. Björklund et al. [10] show that the problem can be solved in 2n poly(n) time (which is faster if the number of sets is exponentially large in n). Björklund et al. [7] give a randomized algorithm that assumes all sets are of size q and determines whether there exist p pairwise disjoint sets in O∗ (2(1−)pq ) time where  > 0 depends on q. Our Main Results We investigate what are sufficient structural properties of instances of Set Cover, and the closely related Set Partition (in which the picked sets need to be disjoint), problems to be solvable in time significantly faster than the currently known algorithms. We will outline our main results now: I Theorem 1.1. There is a Monte Carlo algorithm that takes an instance of Set Cover on n elements and m sets and an integer s as input and determines whether there exists a 4 set cover of size s in O(2(1−Ω(σ ))n m) time, where σ = s/n. We remark that this generalizes the result of Koivisto [29] in the sense that it solves a larger class of instances in O∗ (2(1−Ω(1))n ) time: If all set sizes are bounded by a constant r, a set partition needs to consist of at least n/r sets and Theorem 1.1 applies with σ = 1/r 2 One way of stating Hitting Set in this context, is that we have an instance of the Set Cover problem but aim to find an O∗ (2(1−Ω(1))m ) time algorithm, where m denotes the number of sets. Jesper Nederlof [79]:3 (although this gives a slower algorithm than Koivisto’s in this special case). Moreover, it seems hard to extend the approach of Koivisto to our more general setting. The second result demonstrates that our techniques are also applicable to Set Cover instances with exponentially many sets, a canonical example of which being graph coloring: I Theorem 1.2. There is a randomized algorithm that given graph G and integer s = σn, in 4 O∗ (2(1−Ω(σ ))n ) time outputs yes with constant probability, if χ(G) < s, and no, if χ(G) > s. Representation method for Set Cover We feel the main technique used in this paper is equally interesting as the result, and will therefore elaborate on its origin here. Our technique is on a high level inspired by the following simple observation ingeniously used by HowgraveGraham and Joux [24]: Suppose S ⊆ 2[m] is a set of solutions implicitly given and we seek  [m] for a solution X ∈ S with |X| = s by listing all sets of s/2 and performing pairwise checks to see which two combine to an element of S. Then we can restrict our search in various s pairs guiding us to X. In [24] and all subsequent ways since there will be as many as s/2 works (including [3, 4, 1, 2]), this idea was used to speed up ‘meet-in-the-middle attacks’ (also called ‘birthday attacks’ [27, Chapter 6]). We will refer to uses of this idea as the ‘representation method’ since it crucially relies on the fact that X has many representations as pairs. To indicate the power of this technique in the context of Set Cover and Set Partition we show that without changes it already gives an O∗ (20.3399m )-time Monte Carlo algorithm for the Set Partition problem with m sets, and even for a more general linear satisfiability problem on m variables. For the latter problem this improves the O∗ (2m/2 ) time algorithm based on the meet-in-the-middle attack that was the fastest known before. At first sight the representation method seemed to be inherently only useful for improving algorithms based on the meet-in-the-middle attack. However, the main conceptual contribution of this work is to show that it is also useful in other settings, or at least for improving the dynamic programming algorithm for the Set Cover and Set Partition problems if the solution size is large. On a high level, we show this as follows in the case of Set Partition:3 for a subset W of the elements of the Set Partition instance, define T [W ] to be the minimum number of disjoint sets needed to cover all elements of W . Stated slightly oversimplified, we argue that  if a minimal set partition of size s is large, we have s that T [W ] + T [[n] \ W ] = s for s/2 sets W with |W | close to n/2. To relate this to later sections, let us remark we refer to such a set W as a witness halve. Subsequently, we exploit the presence of many witness halves by using a dynamic programming algorithm that samples a set of the subsets with size close to n/2 and only evaluates table entries from this sample plus the table entries required to compute the table entries from the sample. Organization This paper is organized as follows: In Section 2, we recall preliminaries and introduce notation. In Section 3, we discuss new observations and basic results that we feel are useful for developing a better understanding of the complexity of Set Cover with respect to several structural properties of instances. In Section 4 we formally present the notion of witness halves and prepare tools for exploiting the existence of many witness halves. In Section 5 we prove our main results and in Section 6 we suggest further research. 3 The algorithm for Set Cover actually reduces to Set Partition. ESA 2016 [79]:4 Finding Large Set Covers Faster via the Representation Method 2 Preliminaries and Notation For a real number x, |x| denotes the absolute value of x. For a Boolean predicate p, we let [p] denote 1 if p is true and 0 otherwise. On the other hand, if p is an integer we let [p] denote {1, . . . , p}. As usual, N denotes all positive integers. Running times of algorithms are often stated using O∗ (·) notation which suppresses factors polynomial in the input size. To avoid superscript, we sometimes use exp(x) to denote ex . We denote lg for the base-2 logarithm. If G = (V, E) and v ∈ V we denote N (v) = {w ∈ V : (v, w) ∈ E} and for X ⊆ V S we extend this notation to N (X) = v∈X N (v). For reals a, b > 0 we let a ± b denote the interval [a − b, a + b]. A false positive (negative) of an algorithm is an instance on which it incorrectly outputs YES (respectively, NO). In this work we call an algorithm Monte Carlo if it has no false positives and if any instance is a false negative with probability at most 1/4. We denote vectors with boldface for clarity. For a real number x ∈ [0, 1], h(x) = −x lg x − (1 − x) lg(1 − x) denotes the binary entropy of x, where 0 lg 0 should be thought of as 0. It is well known that ab ≤ 2h(a/b)b (and this can for example be proved using Stirling’s approximation). It is easy to see from the definition that h(·) is symmetric in the sense that h(x) = h(1 − x). I Lemma 2.1. The following can be verified using standard calculus: 1. h(1/2 − x) = h(1/2 + x) ≤ 1 − x2 for all x ∈ (0, 1/2), 2. h(x) ≤ x lg(4/x) for all x ∈ (0, 1), 3. (1 − 1/n)n ≤ 1/e. Ps I Lemma 2.2 (Hoeffding bound [22]). If X1 , . . . , Xs are independent, Y = i=1 Xi and  2 ai ≤ Xi ≤ bi for i = 1, . . . , s then Pr[|Y − E[Y ]| ≥ t] ≤ 2 · exp Ps −2t . (b −a )2 i=1 i i Set Cover / Set Partition In the Set Cover problem we are given a bipartite graph ˙ E) (where F and U shorthand ‘Family’ and ‘Universe’ respectively), together G = (F ∪U, with an integer s and the task is to determine whether there exists a solution S ⊆ F such that N (S) = U and |S| ≤ s. In the Set Partition problem we are given the same input as in the Set Cover problem, but we are set to determine whether there exists S ⊆ F with N (S) = U , |S| = s and additionally N (f ) ∩ N (f 0 ) = ∅ for every f, f 0 ∈ S with f 6= f 0 . We will refer to solutions of both problems as set covers and set partitions. Throughout this paper, we let n, m respectively denote |U | and |F |, and refer to instances of Set Cover or Set Partition as (n, m, s)-instances to quantify their parameters. Since this work concerns Set Cover or Set Partition with large solutions we record the following basic observation that follows by constructing for each4 c-tuple t = (f1 , . . . , fc ) ∈ F c St of sets in the original instance a set f t with N (f t ) = i=1 fi in the output instance: I Observation 2.3 ([14]). There is a polynomial time algorithm that takes a constant c ≥ 1 dividing s, and a (n, m, s)-instance of Set Cover (resp. Set Partition) as input and outputs an equivalent (n, mc , s/c)-instance of Set Cover (resp. Set Partition). Often it will be useful dispense with linear sized sets. To this end, the following can be achieved by simply iterating over all f ∈ F with |N (f )| ≥ n and checking for each such set whether there is a solution containing it using the 2n poly(n) algorithm for Set Cover [10]. 4 For Set Partition only do this for c-tuples (f1 , . . . , fc ) with N (fi ) disjoint. Jesper Nederlof [79]:5 I Observation 2.4. There is an algorithm that, given a real number  > 0, takes an (n, m, s)instance of Set Cover as input and outputs an equivalent (n, m0 , s)-instance with m0 ≤ m satisfying |N (f )| ≤ n for every f ∈ F . The algorithm runs in O(m2(1−)n poly(n)) time. As we will see in Theorem 3.4, it makes a difference in the Set Partition problem whether empty sets are allowed since we need to find a set partition of size exactly s. To exclude such sets, we will simply say that an instance is ‘without empty sets’. 3 Observations and Basic Results on Set Cover and Set Partition. To improve our understanding of which properties of instances of Set Cover and Set Partition allow faster algorithms, and which techniques are useful for obtaining such faster algorithms, we will record some observations and basic results in this section. To stress that the proof techniques in this section are not our main technical contribution, we postpone all proofs to Appendix A. We prefer to state our results in terms of Set Cover because it is slightly more natural and common, but since Set Partition often is easier to deal with for our purposes we will sometimes use the following easy reduction, all of whose steps are contained in [14]: I Theorem 3.1. There is an algorithm that, given a real 0 <  < 1/2, takes an (n, m, s)instance of Set Cover as input and outputs an equivalent (n, m0 , s)-instance of Set Partition with m0 ≤ m2n sets in time O(m2(1−)n ). For completeness, we show that in fact Set Cover and Set Partition are equivalent with respect to being solvable in time O∗ (2(1−Ω(1))n ). This was never stated in print to the best of our knowledge, but the proof uses standard ideas and is found in Appendix A.2. I Theorem 3.2. For some  > 0 there is an O∗ (2(1−)n ) time algorithm for Set Cover if 0 and only if for some 0 > 0 there is an O∗ (2(1− )n ) time algorithm for Set Partition. The following natural result is a rather direct consequence of a paper by Koivisto [29]. It reveals some more similarity with the k-CNF-Sat problem: Koivisto shows5 that for 1 maximum set size r, Set Cover can be solved in O∗ (2(1−Ω( r ))n ) which is analogous to 1 k-CNF-Sat being in O∗ (2(1−Ω( k ))n ) time [33, 17, 12], and similarly the following result is 1 the counterpart of O∗ (2(1−Ω( δ ))n )-time algorithms for CNF-formula’s of density δ (i.e. at most δn clauses) [11, 26]. Again, this result was never explicitly stated in print to the best of our knowledge, and therefore is proved in Appendix A.3. I Theorem 3.3. There is an algorithm solving (n, m, s)-instances of Set Cover or Set n Partition in time m · poly(n)2n− O(lg(m/n)) . Relevant to our work is the following subtlety on solution sizes in Set Partition. It shows that for Set Partition with empty sets, finding large solutions is as hard as the general case. The proof is postponed to Appendix A.4. I Theorem 3.4. Suppose there exist 0 < 1 , 2 < 1/2 and an algorithm solving (n, m, 1 n)instances of Set Partition in time O∗ (2(1−2 )n ). Then there exists an O∗ (2(1−2 /2)n )-time algorithm for Set Partition. 5 Koivisto only showed this for Set Partition, but the straightforward reductions in this section carry this result over to Set Cover. ESA 2016 [79]:6 Finding Large Set Covers Faster via the Representation Method Finally, it is insightful to see how well the representation method performs on the Set Partition problem with few sets (e.g., we consider running times of the O∗ (2m ), where m is the number of sets). A straightforward approach of the meet-in-the-middle attack leads directly to an O∗ (2m/2 ) time algorithm. We show that the representation method combined with the analysis of [2, 1] in fact solves the more general Linear Sat problem. In Linear Sat one is given an integer t, matrix A ∈ Zn×m , and vectors b ∈ Zn2 and ω ∈ Nm . The task 2 m is to find x ∈ Z2 satisfying Ax ≡ b and ω · x ≤ t. I Theorem 3.5. There is an O∗ (20.3399m )-time Monte Carlo algorithm solving Linear Sat. To our best knowledge no O∗ (2(0.5−Ω(1))m )-time algorithm for Linear Sat was known ˙ E), we can determine before. We get as a corollary that, given a bipartite graph G = (F ∪U, the smallest size of a set partition in time O∗ (20.3399m ). We take this as a first signal that the representation method is useful for solving Set Partition (and Set Cover) for instances with small universe. To see this consequence, note we can reduce this problem to Linear Sat as follows: For every f ∈ F add the incidence vector of N (f ) as a column to A, and set the cost ωi of picking this column to be n|N (f )| + 1. Then the minimum of ω · x subject to Ax ≡ 1 will be n2 + s where s is the number of sets in a minimum set partition. Let us remark that [16, Page 130] solves (a counting version) of Set Partition in time O∗ (1.2561m ) = O∗ (20.329m ), and Drori and Peleg [18] solve the problem in O∗ (20.3212m ) time,6 so by no means our algorithm is the fastest in this setting. However, both use sophisticated branching and we find it intriguing that the representation method does work quite well even for the seemingly more general Linear Sat problem. 4 Exploiting the Presence of Many Witness β-halves. For convenience we will work with Set Partition in this section; the results straightforwardly extend to Set Cover but we will not need this in the subsequent section. I Definition 4.1. Given an (n, m, s) instance of Set Partition, a subset W ⊆ U is said to be a witness β-halve if |W | ∈ ( 12 ± β)n and there exist disjoint subsets S1 , S2 ⊆ F such that P N (S1 ∪ S2 ) = U , f ∈S1 ∪S2 |N (f )| = n, N (S1 ) = W , N (S2 ) = U \ W and |S1 | + |S2 | = s. Note that this is similar to the intuitive definition outlined in Section 1, except that we require |W | ∈ ( 21 ± β)|U | and we adjusted the definition to the Set Partition problem. Since S1 ∪S2 is a set partition of size s we see that if a witness β-halve exists, we automatically have a yes instance. In this section we will give randomized algorithms that solve promise-variants of Set Partition with the promise that, if the instance is a yes-instance, there will be an exponential number of witness halves that are sufficiently balanced (i.e. of size close to n/2). In the first subsection we outline the basic algorithm and in the second subsection we show how tools from the literature can be combined with our approach to also give a faster algorithm if the number of sets is exponential in n. 6 We attempted to find any more recent faster algorithm, but did not find this. Though, we would not be surprised if using more recent tools in branching algorithms as [19] one should be able to more significantly outperform our algorithm for Set Partition. Jesper Nederlof 4.1 [79]:7 The basic algorithm I Theorem 4.2. There exists an algorithm A1 that takes an (n, m, s)-instance of Set √ Partition and real numbers β, ζ > 0 satisfying 2 β ≤ ζ < 1/4 as input, runs in time 4 2(1−(ζ/2) )n poly(n)m, and has the following property: If there exist at least Ω(2ζn ) witness β-halves it returns yes with at least constant probability, and if there does not exist a set partition of size s it returns no. Note that the theorem does not guarantee anything on Algorithm A1 if a partition of s sets exists and there are only few witness halves, but we will address this later. A high level description of the Algorithm A1 is given in Figure 1: √ Algorithm A1(G = (F ∪˙ U, E), s, ζ, β). Assumes 2 β ≤ ζ < 1/4 Output: An estimate of whether there exists a set partition of size s. 1: for integer l satisfying b(1/2 − β)nc < l < d(1/2 + β)ne do   2: Sample W ⊆ Ul by including every set of Ul with probability 2−ζn . 3: For every W ∈ W and i ∈ [n], compute ci (W ) and ci (U \ W ). 4: if ∃i ∈ [n] : ci (W ) ∧ cs−i (U \ W ) then return yes. 5: return no. Figure 1 High level description of the Algorithm implementing Theorem 4.2. Here, we define ci (W ) to be true if and only if there exists S1 ⊆ F with |S1 | = i, N (S1 ) = W , and for every f, f 0 ∈ S1 with f 6= f 0 , N (f ) ∩ N (f 0 ) = ∅. Given a set family W, we denote ↓W = {X : ∃W ∈ W ∧ X ⊆ W } for the down-closure of W. The following lemma concerns the sub-routine invoked in Algorithm 1 and can be proved via known dynamic programming techniques, and is postponed to Appendix A.6. I Lemma 4.3. There exists an algorithm that given a bipartite graph G = (F ∪˙ U, E) and W ⊆ 2U with |U | = n and |F | = m, computes ci (W ) for all W ∈ W and i ∈ [n] in O(poly(n)|↓W|m) time. Thus, for further preparation of the proof of Theorem 4.2, we need to analyze the maximum size of the (down/up)-closure of W in Algorithm A1 in Figure 1. p I Lemma 4.4. Let ζ > 0, β (which may be negative) be real numbers satisfying 2 |β| ≤ ζ <  4 U 1/4 and |U | = n. Suppose W ⊆ (1/2+β)n with |W| ≤ 2(1−ζ)n . Then |↓W| ≤ n2(1−(ζ/2) )n . Proof. Let λ ≤ β and wλ = |{W ∈ ↓W : |W | = λn}|, so |↓W| ≤ n · maxλ wλ . Then we have the following upper bounds:       λ n (1/2 + β)n (1−ζ)+h 1/2+β (1/2+β) n h(λ)n wλ ≤ ≤2 , wλ ≤ |W| ≤2 . λn λn To see the second upper bound, note that any set W ∈ W can have at most (1/2+β)n λn subsets of size λn. Thus, we see that |↓W|/n is upper bounded by 2f (ζ,β)n , where f (ζ, β) =   max min h(λ), (1 − ζ) + h λ≤1/2+β λ 1/2 + β    (1/2 + β) . ESA 2016 [79]:8 Finding Large Set Covers Faster via the Representation Method The remainder of the proof is therefore devoted to upper bounding f (ζ, β). We establish this by evaluating both terms of the minimum, setting λ to be λ0 = (1 − ζ 2 )(1/2 + β). First note that by our assumption λ0 = (1 − ζ 2 )(1/2 + β) = 1/2 − ζ 2 /2 + β − ζ 2 β ≤ 1/2 − ζ 2 /2 + ζ 2 /4 − ζ 2 β 0 λ /(1/2 + β) = 1 − ζ 2 < 1/2, > 1/2. since h(x) for x < 1/2, h(λ) ≤ h(λ0 ) for λ ≤ λ0 . Similarly,   is 0increasing   Therefore, λ λ h 1/2+β is at most h 1/2+β for λ ≥ λ0 , and we may upper bound f (ζ, β) by the maximum of the two terms of the minimum in f (ζ, β) obtained by setting λ = λ0 . For the first term of the minimum, note that by Lemma 2.1, Item 1: h(λ0 ) = h((1 − ζ 2 )(1/2 + β)) ≤ 1 − (1/2 − (1 − ζ 2 )(1/2 + β))2 2 = 1 − ζ 2 /2 − β + βζ 2 2 ≤ 1 − ζ 2 /2 − β ≤ 1 − (ζ 2 /4)2 = 1 − (ζ/2)4 . For the second term we have   λ0 1−ζ +h (1/2 + β) = 1 − ζ + h(1 − ζ 2 )(1/2 + β) 1/2 + β = 1 − ζ + h(ζ 2 )(1/2 + β)   4 33 2 ≤ 1 − ζ + ζ lg ζ 2 64 ≤ 1 − ζ + ζ 32 · by Lemma 2.1, Item 2 β< 1 64 by assumption   ζ lg ζ42 ≤ 32 33 64 ≤ 1 − ζ/10. note for the penultimate inequality that ζ lg( ζ42 ) is monotone increasing for 0 ≤ ζ ≤ 1/4 and substituting ζ = 1/4 in this expression thus upper bounds it with 3/2. J Now we are ready to wrap up this section with the proof of Theorem 4.2: Proof of Theorem 4.2. We can implement Line 3 by invoking the algorithm of Lemma 4.3 with both |W| and W 0 = {W : [n]\W ∈ W}. This will take time O(poly(n)(|↓W |+|↓W 0 |)m). This is clearly the bottleneck of the algorithm, so it remains to upper bound  (the expectation  n of) |↓W | + |↓W 0 | by applying Lemma 4.4. To do this, note that W ⊆ nl , W 0 ⊆ n−l , and p ≤ ζ by assumption so indeed we have that (1/2 − β)n ≤ l, n − l ≤ (1/2 + β)n. Also, 2 |β|  −ζn n Lemma 4.4 applies. Then on expectation |W| ≤ (1/2+β)n 2 ≤ 2(1−ζ)n , and thus the 7 running time indeed is as claimed. For the correctness, it is easily checked that the algorithm never returns false positives. Moreover, if there exist at least Ω(2ζn ) witness β-halves then for some l in the loop of Line 1, there are at least Ω(2ζn /n) witness halves of size l. Thus in this iteration we see by Lemma 2.1, Part 3 that  Ω(2ζn /n) 1 Pr[@ witness halve W ∈ W] ≤ 1 − ζn ≤ e−1/n . 2 7 (4.1) Due to the sampling in Line 2, we actually only get an upper bound on the expectation of the running time, but by Markov’s inequality we can simply ignore iterations where W exceeds twice the expectation. Jesper Nederlof [79]:9 and if a witness halve W ∈ W exists the algorithm returns yes since ci (W ) ∧ cs−i (U \ W ) holds for some i by the definition of witness halve. Therefore, if we perform n independent trials of Algorithm A1 it return yes with probability at least 1 − 1/e. J 4.2 Improvement in the case with exponentially many input sets. In this section we show that under some mild conditions, the existence of many witness halves can also be exploited in the presence of exponentially many sets. This largely builds upon machinery developed by Björklund et al. [10, 8]. To state our result as general as possible we assume the sets are given via an oracle so our running can be sublinear in the input if the number of sets is close to 2n . I Theorem 4.5. There exists an algorithm that, given oracle access to an (n, m, s)-instance √ of Set Partition and real numbers β, ζ > 0 satisfying 2 β ≤ ζ < 1/4, runs in time 4 2(1−(ζ/2) )n poly(n)T and has the following property: if there exist at least Ω(2ζn ) witness β-halves, it outputs yes with constant probability and if there does not exist a set partition of size s it outputs no. Here the oracle algorithm accepts X ⊆ U as input, and decides whether there exists f ∈ F with N (f ) = X in time T . The proof of Theorem 4.5 is identical to the proof of Theorem 4.2 (and therefore omitted), except that here we use the following lemma instead of Lemma 4.3: I Lemma 4.6. There exists an algorithm that, given W ⊆ 2U and oracle access to a bipartite graph G = (F ∪˙ U, E), computes the values ci (W ) for all W ∈ W in O(T |↓W| poly(n)) time. Here the oracle algorithm accepts X ⊆ U as input, and decides whether there exists f ∈ F with N (f ) = X in time T . This lemma mainly reiterates previous work developed by Björklund et al. [10, 8], but since they did not prove this lemma as such we include a proof here in Appendix A.7. 5 Finding Large Set Covers Faster In this section we will use the tools of the previous sections to prove our main results, Theorems 1.1 and 1.2. We first connect the existence of large solutions to the existence of many witness halves in the following lemma: I Lemma 5.1. If an (n, m, s)-instance of Set Partition has no empty sets and satisfies s ≥ σ0 n and |N (f )| ≤ σ04 n/8 for every f ∈ F , there is a solution if and only if there exist at least 2σ0 n /4 witness (σ02 /4)-halves. Proof. Note that the backward direction is trivial since by definition the existence of a witness halve implies the existence of a solution. For the other direction, suppose S = {f1 , . . . , fs } is a set partition, and denote di = |N (fi )|. Suppose S 0 ⊆ S is obtained by including every element of S with probability 1/2 in S 0 . since N (fi ) ∩ N (fj ) = ∅ for i 6= j, we have that the random variable |N (S 0 )| is a sum of s independent random variables that equal 0 and di with probability 1/2. By the Hoeffding bound (Lemma 2.2) we see that  2 4   2 4  h i −n σ0 /8 −n σ0 /8 0 0 2 Pr |N (S )| − E[|N (S )|] ≥ nσ0 /4 ≤ 2 · exp P < 34 , ≤ 2 · exp 2 n2 σ04 /8 e∈S de ESA 2016 [79]:10 Finding Large Set Covers Faster via the Representation Method P where the second inequality follows from de ≤ σ04 n/8 and e∈S de = n. So for at least 2|S| /4 ≥ 2σ0 n /4 subsets S 0 ⊆ S we have that |N (S 0 )| ∈ ( 21 ± σ02 /4)n. Thus, since for each such S 0 , N (S 0 ) determines S 0 and thus gives rise to a distinct witness halve, there are at least 2σ0 n /4 witness (σ02 /4)-halves. J Now we are ready to prove the first main theorem, which we recall here for convenience. I Theorem 1.1 (restated). There is a Monte Carlo algorithm that takes an instance of Set Cover on n elements and m sets and an integer s as input and determines whether there 4 exists a set cover of size s in O(2(1−Ω(σ ))n m) time, where σ = s/n. Proof. The algorithm implementing Theorem 1.1 is given in Figure 2. Algorithm A2(G = (F ∪˙ U, E), σ). 1: Ensure |N (f )| ≤ σ 4 n/1000 using Observation 2.4. 2: for every integer s satisfying bσn/2c ≤ s ≤ σn do 3: Create an (n, m0 , s)-instance ((F 0 ∪˙ U, E), s) of Set Partition where F 0 is constructed by adding a vertex f 0 with N (f 0 ) = X for all f ∈ F, X ⊆ N (f ). 4: Let σ0 = s/n. 5: if A1((F 0 ∪˙ U, E), s, σ0 , σ02 /4) = yes then return yes. U 6: Pick an arbitrary subset X ∈ n/2 . 7: Find the sizes l and r of the smallest set covers in the instances induced by elements X and respectively elements U \X in O(2n/2 poly(n)m) time with standard dynamic programming. 8: if l + r <= σn then return yes else return no. Figure 2 Algorithm for Set Cover large solutions (implementing Theorem 1.1). We first focus on the correctness of this algorithm. It is clear that the algorithm never returns false positives on Line 5 since Algorithm A1 also has this property. If yes is returned on Line 8 it is also clear there exists a solution. Now suppose that a set cover S of size at most σn exists. First suppose σn/2 ≤ |S| ≤ σn. We consider s = |S| in some iteration of the loop on Line 2. Notice that now in Line 3 we have reduced the problem to a yes-instance of Set Partition without empty sets satisfying |N (f )| ≤ σ 4 n/1000 for every f ∈ F . Therefore Lemma 5.1 applies with σ0 ≥ σ/2 and we see there are at least 2σ0 n /4 witness (σ02 /4)-halves. Thus, we can apply Theorem 4.2 with ζ = σ0 and β = σ02 /4 to find the set S with constant probability, since β ≤ (ζ/2)2 . Now suppose |S| ≤ σn/2. Then picking every element in S twice is a solution (as a multiset), and it implies that for every X ⊆ U the sizes of the smallest set covers l and r (as defined in the algorithm) satisfy l + r ≤ σn. Thus Lines 6-8 find such a set and the algorithm returns yes. 4 For the running time, Line 1 takes at most O(2(1−σ /1000)n poly(n)m) due to Observation 2.4. For Line 5, due to Theorem 4.2 this runs in time O(2(1−(ζ/2) 4 )n poly(n)m0 ) = O(2(1−(σ0 /2) ≤ O(2 )n (1−(σ/4) )n 4 = O(2(1+σ ≤ O(2 4 4 poly(n)2σ poly(n)2 (1/1000−1/4 ))n (1−Ω(σ 4 ))n 4 4 4 n/1000 σ n/1000 m) m) poly(n)m) poly(n)m). Jesper Nederlof [79]:11 J as claimed in the theorem statement. As a more direct consequence of the tools of the previous section we also get the following result for Set Partition: I Theorem 5.2. There exists a Monte Carlo algorithm for Set Partition that, given oracle access to an (n, m, σn)-instance satisfying 0 < |N (f )| ≤ σ 4 n/8 for every f ∈ F , runs 4 in 2(1−Ω(σ ))n poly(n)T time. Here the oracle algorithm accepts X ⊆ U as input, and decides whether there exists f ∈ F with N (f ) = X in time T . Proof. Lemma 5.1 implies the instance is a YES-instance if and only if there exist 2σn /4 witness (σ 2 /4)-halves. Thus Theorem 4.5 implies the theorem statement. J Note that this theorem also implies an O((m + 2(1−Ω(σ ))n ) poly(n)) time algorithm for Set Partition where the sets are given explicitly because we can construct a binary search tree after which we can implement the oracle to run in T = n query time. We remark that it would be interesting to see whether the assumption |N (f )| ≤ σ 2 n/4 is needed, but removing this assumption seems to require more ideas than the ones from this work: For example if the solution has three sets of size 3n/10 there will be no witness halve that is sufficiently balanced, and alternatively using Observation 2.4 seems to be too slow. However, if we settle for a additive 1-approximation we can deal with this issue in a simple way and have as a particular consequence the second result mentioned in the beginning of this paper: 4 I Theorem 1.2 (restated). There is a randomized algorithm that given graph G and integer 4 s = σn, in O∗ (2(1−Ω(σ ))n ) time outputs yes with constant probability, if χ(G) < s, and no, if χ(G) > s. Proof. Let G = (V, E) and define a Set Partition instance where for every independent set I ⊆ V of G there is an element f ∈ F with N (f ) = I. It is easy to see that this instance of Set Partition has a solution of size s if and only if χ(G) ≤ s.  Check in σ4nn/8 time whether G has an independent set of size σ 4 n/8. If such an independent set is found, remove this set from the graph and return yes if the obtained graph has a (k − 1)-coloring and no otherwise. Using the O∗ (2n ) time algorithm by Björklund et 4 al. [10] in the second step, this procedure clearly runs in time O∗ (2(1−Ω(σ ))n ), and always finds a coloring using at most one more color than the minimum number of colors if a large enough independent set exists. On the other hand, if the maximum independent set of G is of size at most σ 4 n/8, we may apply Theorem 5.2 with T = poly(n) since it can be verified in polynomial time whether a given X ⊆ V is an independent set, and the theorem follows. J 6 Directions for Further Research In this section, we relate the work presented to some notorious open problems. The obvious open question is to determine the exact complexity of the Set Cover problem: I Open Problem 1. Can Set Cover be solved in time O∗ ((2 − Ω(1))n )? This question was already stated at several places. It is known that if a version of Set Cover where the number of solutions modulo 2 is counted can be solved in (2 − Ω(1))n the Strong Exponential Time Hypothesis fails. We refer to [14], for more details. ESA 2016 [79]:12 Finding Large Set Covers Faster via the Representation Method Less ambitiously, it is natural to wonder whether our dependency on σ can be improved. Our algorithm and analysis seem loose, but we feel the gain of a sharpening this analysis does not outweigh the technical effort currently: For a better dependence, we need both a better bound in Lemma 4.4 and to reduce the set sizes more efficiently than in Observation 2.4. As further research we suggest to find a different algorithmic way to deal with the case where many witness halves are unbalanced. But this alone will not suffice to give linear dependence in σ since in Lemma 4.4 we do not expect to get linear dependence on ζ even if β = 0. It would also be interesting to see which other instances of Set Cover can be solved in O∗ ((2 − Ω(1))n ) time. One that might be worthwhile studying is whether this includes instances with optimal set covers in which the sum of the set sizes is at least (1 + Ω(1))n; one may hope to find exponentially many (balanced) witness halves here as well. In [14], the authors also give a reduction from Subset Sum to Set Partition. The exact complexity of Subset Sum with small integers is also something we explicitly like to state as open problem here, especially since the O∗ (t) time algorithm (where t is the target integer) is perhaps one of the most famous exponential time algorithms: I Open Problem 2. Can Subset Sum with target t be solved in time O∗ (t1−Ω(1) ), or can we exclude the existence of such an assuming the Strong Exponential Time Hypothesis? Note this question was before asked in [25] by the present author. It would be interesting to study the complexity of Subset Sum in a similar vein as we did in this paper: are there some special properties allowing a faster algorithm? For example, a faster algorithm for instances of high ‘density’ (e.g., n/ lg t) may be used for improving an algorithm of Horowitz&Sahni [23]. Note that here the ‘density’ of a Subset Sum instance is the inverse of what one would expect when relating to the definition of density of k-CNF formula. Another question that has already open for a while is I Open Problem 3. Can Graph Coloring be solved in time O∗ (2(1−Ω(1))n )? Could the techniques of this paper be used to make progress towards resolving this question? While our algorithm seems to benefit from the existence of many optimal colorings, in an interesting paper Björklund [5] actually shows that the existence of few optimal colorings can be exploited in graphs of small pathwidth. Related to this is also the Hamiltonicity problem. In our current understanding this problem becomes easier when there is a promise that there are few Hamiltonian cycles (see [6], but also e.g. [13] allows derandomizations of known probabilistic algorithms in this case), so a natural approach would be to deal explicitly with instances with many solutions by sampling dynamic programming table in a similar vein as done in this paper. Acknowledgements The author would like to thank Per Austrin, Petteri Kaski, Mikko Koivisto for their collaborations resulting in [1, 2] that mostly inspired this work, Karl Bringmann for discussions on applications of [2] to Set Partition, and anonymous reviewers for their useful comments. References 1 Per Austrin, Petteri Kaski, Mikko Koivisto, and Jesper Nederlof. Subset sum in the absence of concentration. In Ernst W. Mayr and Nicolas Ollinger, editors, 32nd International Symposium on Theoretical Aspects of Computer Science, STACS 2015, March 4-7, 2015, Garching, Germany, volume 30 of LIPIcs, pages 48–61. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2015. Jesper Nederlof 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 [79]:13 Per Austrin, Petteri Kaski, Mikko Koivisto, and Jesper Nederlof. Dense subset sum may be the hardest. In Nicolas Ollinger and Heribert Vollmer, editors, 33rd Symposium on Theoretical Aspects of Computer Science, STACS 2016, February 17-20, 2016, Orléans, France, volume 47 of LIPIcs, pages 13:1–13:14. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2016. Anja Becker, Jean-Sébastien Coron, and Antoine Joux. Improved generic algorithms for hard knapsacks. In Kenneth G. Paterson, editor, Advances in Cryptology - EUROCRYPT 2011 - 30th Annual International Conference on the Theory and Applications of Cryptographic Techniques, Tallinn, Estonia, May 15-19, 2011. Proceedings, volume 6632 of Lecture Notes in Computer Science, pages 364–385. Springer, 2011. Anja Becker, Antoine Joux, Alexander May, and Alexander Meurer. Decoding random binary linear codes in 2 n/20: How 1 + 1 = 0 improves information set decoding. In EUROCRYPT, volume 7237 of Lecture Notes in Computer Science, pages 520–536. Springer, 2012. Talk at http://www.iacr.org/cryptodb/data/paper.php?pubkey=24271. Andreas Björklund. Uniquely coloring graphs over path decompositions. CoRR, abs/1504.03670, 2015. Andreas Björklund, Holger Dell, and Thore Husfeldt. The parity of set systems under random restrictions with applications to exponential time problems. In Magnús M. Halldórsson, Kazuo Iwama, Naoki Kobayashi, and Bettina Speckmann, editors, Automata, Languages, and Programming - 42nd International Colloquium, ICALP 2015, Kyoto, Japan, July 610, 2015, Proceedings, Part I, volume 9134 of Lecture Notes in Computer Science, pages 231–242. Springer, 2015. Andreas Björklund, Thore Husfeldt, Petteri Kaski, and Mikko Koivisto. Narrow sieves for parameterized paths and packings. CoRR, abs/1007.1161, 2010. Andreas Björklund, Thore Husfeldt, Petteri Kaski, and Mikko Koivisto. Trimmed moebius inversion and graphs of bounded degree. Theory Comput. Syst., 47(3):637–654, 2010. Andreas Björklund, Thore Husfeldt, Petteri Kaski, and Mikko Koivisto. The traveling salesman problem in bounded degree graphs. ACM Transactions on Algorithms, 8(2):18, 2012. Andreas Björklund, Thore Husfeldt, and Mikko Koivisto. Set partitioning via inclusionexclusion. SIAM J. Comput., 39(2):546–563, 2009. Chris Calabro, Russell Impagliazzo, and Ramamohan Paturi. A duality between clause width and clause density for SAT. In 21st Annual IEEE Conference on Computational Complexity (CCC 2006), 16-20 July 2006, Prague, Czech Republic, pages 252–260. IEEE Computer Society, 2006. Timothy M. Chan and Ryan Williams. Deterministic apsp, orthogonal vectors, and more: Quickly derandomizing razborov-smolensky. In Robert Krauthgamer, editor, Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2016, Arlington, VA, USA, January 10-12, 2016, pages 1246–1255. SIAM, 2016. Suresh Chari, Pankaj Rohatgi, and Aravind Srinivasan. Randomness-optimal unique element isolation with applications to perfect matching and related problems. SIAM J. Comput., 24(5):1036–1050, 1995. Marek Cygan, Holger Dell, Daniel Lokshtanov, Dániel Marx, Jesper Nederlof, Yoshio Okamoto, Ramamohan Paturi, Saket Saurabh, and Magnus Wahlström. On problems as hard as CNF-SAT. In Proceedings of the 27th Conference on Computational Complexity, CCC 2012, Porto, Portugal, June 26-29, 2012, pages 74–84. IEEE, 2012. Marek Cygan and Marcin Pilipczuk. Faster exponential-time algorithms in graphs of bounded average degree. Inf. Comput., 243:75–85, 2015. Vilhelm Dahllöf. Exact Algorithms for Exact Satisfiability Problems. PhD thesis, Linköping University, TCSLAB, The Institute of Technology, 2006. ESA 2016 [79]:14 Finding Large Set Covers Faster via the Representation Method 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 Evgeny Dantsin, Andreas Goerdt, Edward A Hirsch, Ravi Kannan, Jon Kleinberg, Christos Papadimitriou, Prabhakar Raghavan, and Uwe Schöning. A deterministic (2 − 2/(k + 1))n algorithm for k-SAT based on local search. Theoretical Computer Science, 289(1):69 – 83, 2002. Limor Drori and David Peleg. Faster exact solutions for some NP-hard problems. Theoretical Computer Science, 287(2):473 – 499, 2002. Algorithms. Fedor V. Fomin, Fabrizio Grandoni, and Dieter Kratsch. A measure & conquer approach for the analysis of exact algorithms. J. ACM, 56(5), 2009. Alexander Golovnev, Alexander S. Kulikov, and Ivan Mihajlin. Families with infants: A general approach to solve hard partition problems. In Javier Esparza, Pierre Fraigniaud, Thore Husfeldt, and Elias Koutsoupias, editors, Automata, Languages, and Programming - 41st International Colloquium, ICALP 2014, Copenhagen, Denmark, July 8-11, 2014, Proceedings, Part I, volume 8572 of Lecture Notes in Computer Science, pages 551–562. Springer, 2014. Godfrey Hardy and Srinivasa Ramanujan. Asymptotic formulaæ in combinatory analysis. Proceedings of the London Mathematical Society, s2-17(1):75–115, 1918. Wassily Hoeffding. Probability inequalities for sums of bounded random variables. Journal of the American Statistical Association, 58(301):13–30, 1963. Ellis Horowitz and Sartaj Sahni. Computing partitions with applications to the knapsack problem. J. ACM, 21(2):277–292, 1974. Nick Howgrave-Graham and Antoine Joux. New generic algorithms for hard knapsacks. In Henri Gilbert, editor, Advances in Cryptology - EUROCRYPT 2010, 29th Annual International Conference on the Theory and Applications of Cryptographic Techniques, French Riviera, May 30 - June 3, 2010. Proceedings, volume 6110 of Lecture Notes in Computer Science, pages 235–256. Springer, 2010. Thore Husfeldt, Ramamohan Paturi, Gregory B. Sorkin, and Ryan Williams. Exponential Algorithms: Algorithms and Complexity Beyond Polynomial Time (Dagstuhl Seminar 13331). Dagstuhl Reports, 3(8):40–72, 2013. Russell Impagliazzo, Shachar Lovett, Ramamohan Paturi, and Stefan Schneider. 0-1 integer linear programming with a linear number of constraints. CoRR, abs/1401.5512, 2014. Antoine Joux. Algorithmic Cryptanalysis. Chapman & Hall/CRC, 1st edition, 2009. Petteri Kaski, Mikko Koivisto, and Jesper Nederlof. Homomorphic hashing for sparse coefficient extraction. In Dimitrios M. Thilikos and Gerhard J. Woeginger, editors, Parameterized and Exact Computation - 7th International Symposium, IPEC 2012, Ljubljana, Slovenia, September 12-14, 2012. Proceedings, volume 7535 of Lecture Notes in Computer Science, pages 147–158. Springer, 2012. Mikko Koivisto. Partitioning into sets of bounded cardinality. In Jianer Chen and Fedor V. Fomin, editors, Parameterized and Exact Computation, 4th International Workshop, IWPEC 2009, Copenhagen, Denmark, September 10-11, 2009, Revised Selected Papers, volume 5917 of Lecture Notes in Computer Science, pages 258–263. Springer, 2009. R. J. McEliece. A Public-Key Cryptosystem Based On Algebraic Coding Theory. Deep Space Network Progress Report, 44:114–116, January 1978. Daniel Paulusma, Friedrich Slivovsky, and Stefan Szeider. Model counting for cnf formulas of bounded modular treewidth. Algorithmica, pages 1–27, 2015. Sigve Hortemo Sæther, Jan Arne Telle, and Martin Vatshelle. Solving maxsat and #sat on structured CNF formulas. In Carsten Sinz and Uwe Egly, editors, Theory and Applications of Satisfiability Testing - SAT 2014, Vienna, Austria, July 14-17, 2014. Proceedings, volume 8561 of Lecture Notes in Computer Science, pages 16–31. Springer, 2014. Uwe Schöning. A probabilistic algorithm for k -SAT based on limited local search and restart. Algorithmica, 32(4):615–623, 2002. Jesper Nederlof 34 [79]:15 Vijay V. Vazirani. Approximation Algorithms. Springer-Verlag New York, Inc., New York, NY, USA, 2001. A Missing Proofs A.1 Proof of Theorem 3.1 Given an (n, m, s)-instance of Set Cover, use Observation 2.4 to assure that for every f ∈ F , |N (f )| ≤ n. It is easy to see that this instance is a YES-instance if and only if the Set Partition instance (G0 , s) is a YES-instance, where G0 is the bipartite graph with U on one side and F 0 on the other with for every f ∈ F and X ⊆ N (f ) a vertex in F 0 with neighborhood X. A.2 Proof of Theorem 3.2 0 For the backward direction, an O(mc 2(1− )n ) algorithm for Set Partition concatenated to the reduction of Theorem 3.1 with  = 0 /(2c) gives a  c  0 0 0 O m2 n/(2c) 2(1− )n + m2(1−)n ≤ O(mc 2(1− /2)n + m2(1−)n ) time algorithm for Set Cover. For the forward direction, given an (n, m, s)-instance of Set Partition first use Observation 2.3 with c = 2/ to obtain an (n, m2/ , s/2) instance of Set Partition (without loss of generality, we may assume s divides 2/ by decreasing  and adding at most 2/ sets and elements). Denoting s0 = s/2, now iterate over all unordered integer partitions ˙ 0 ), E 0 ), s0 ) a1 +. . .+as0 = n, and for each such partition solve a Set Cover instance (((F 0 ∪U constructed as follows from the Set Partition instance with the assumed algorithm: For every i = 1, . . . , s0 add an element ei to U 0 For every f ∈ F and i such that |N (f )| = ai add a set f 0 to F 0 with N (f 0 ) = f ∪ {ei }. Note that possibly ai = aj for i 6= j we make at least two copies of every set f ∈ F with |N (f )| = ai . It is easy to see that this new Set Cover instance is a YES-instance if and only if in the original Set Partition instance there is a set partition of size s0 : each set only contains one element from {e1 , . . . , es0 } so a Set Cover of size s0 needs to correspond Ps0 with f1 , . . . , fs0 ∈ F satisfying i=1 |N (fi )| = n. √ By a result of Hardy and Ramanujan [21], there are at most 2O( n) unordered integer partitions and hence concatenating this reduction with the assumed O(mc 2(1−)n ) time Set Cover algorithm leads to an √ O(2O( n) 0 mc2/ 2(1−)(n+s ) ) ≤ O∗ (2(1−)(1+/2)n ) ≤ O∗ (2(1−/2)n ) time algorithm for Set Partition. A.3 Proof of Theorem 3.3 I Theorem A.1 ([29]). Given an instance n-element set N , an integer r and a family F of subsets of N each of cardinality at most r, the partitions of N into apgive number of members of F can be counted in time |F|2λr n poly(n), where λr = (2r − 2)/ (2r − 1)2 − 2 ln 2. Note that the above result also immediately implies an algorithm deciding Set Cover in time 2r |F|2λr n poly(n) since we may reduce this Set Cover instance to a Set Partition ESA 2016 [79]:16 Finding Large Set Covers Faster via the Representation Method instance by adding all subsets of the given sets. It is easily seen that for r ≥ 3, λr is sandwiched as p 1 1/2 ≤ λr ≤ (2r − 2)/ (2r − 1.5)2 = 1 − . O(r) The algorithm implementing the theorem is given in Figure 3. ˙ E), s, r). Algorithm A3(G = (F ∪U, 1: if ∃f ∈ F : |N (f )| ≥ r then 2: return A3(G[F \ {f } ∪ U ], s, r) ∨ A2(G[F \ {f } ∪ U \ N (f )], s − 1, r). 3: else 4: Use the algorithm of Theorem A.1. Figure 3 Algorithm for Set Cover or Set Partition with few sets (implementing Theorem 3.3). We may bound the running time of Algorithm A3 by analyzing the branching tree of recursive calls where an execution reaching Line 4 represents a leaf. Note that the depth of the branching tree is at most m and if we refer to the second recursive call of Line 2 as the right branch, on any path from the root to a leaf there are at most n/r right branches since in each such recursion step we decrease the size of U in one call with at least r. Thus the  number of paths from the root to a leaf in this tree with i right branches is at most mi . For each such a branch the size of U has become n − ir and since Line 4 is reached all sets are of size at most r, Theorem 3.3 applies and thus we spend m · poly(n)2λr (n−ir) per leaf of this type. Thus, denoting µ = m/n, the total running time can be written as m  m   X m λr (n−ir) 1 λr n m · poly(n) 2 = m · poly(n)2 1 + λr r 2 i i=1   n ≤ m · poly(n) 2λr exp µ/2r/2 , where the equality follows from the binomial theorem and the inequality uses (1 + n1 )n ≤ e. Denoting µ = m/n and setting r = d4 lg(µ)e, we see that when taking asymptotics for growing µ (which is allowed since we may assume the running time complexity is monotone increasing with µ), the running time becomes  n n 1 m · poly(n) 21− O(lg µ) +lg(e)/µ = m · poly(n)2n− O(lg(m/n)) . A.4 Proof of Theorem 3.4 Given an arbitrary instance of Set Partition consisting of G = (F ∪˙ U, E) and integer s, first construct an equivalent (n, m0 , s0 ) instance using Observation 2.3 with s0 ≤ min{1 , 2 /2}n. Then add vertices e1 , . . . , es0 to U and replace every f ∈ F with s0 copies f1 , . . . , fs0 where N (fi ) = N (f )∪ei . It is easy to see that in this Set Partition instance all set partitions are of size s0 and it has a set partition of size s0 if and only if the original instance has one of size s. Thus, to transform the new instance into an equivalent (n + s0 , m0 , 1 (n + s0 ))-instance, simply add sufficiently many isolated vertices to F . Running the assumed algorithm on this 0 instance results thus in an O∗ (2(1−2 )(n+s ) ) = O∗ (2(1−2 )(1+2 /2)n ) = O∗ (2(1−2 /2)n ) time algorithm. Jesper Nederlof A.5 [79]:17 Proof of Theorem 3.5 Recall from the main text that the Linear Sat problem is defined as follows: given an integer t, matrix A ∈ Zn×m and vectors b ∈ Zn2 and ω ∈ Nm the task is to find x ∈ Zm 2 2 satisfying Ax ≡ b and ω · x ≤ t. We let a1 , . . . , am ∈ Zn2 denote the column vectors of A, and denote wt(x) for the Hamming weight of a vector x ∈ Zm 2 . The algorithm implementing the theorem is outlined in Algorithm 4. The algorithm assumes the Hamming weight of the vector x minimizing minimum ωx is at most 2m/3 (which is in order to facilitate the running time analysis). We can assume this since if the minimum set of columns summing to b consists of more than 2m/3 columns, the rank of A is at least 2m/3 (as the solution needs to consist of linearly independent columns) and we can solve the problem in O∗ (2m/3 ) time by finding a column basis, iterating over all subsets of columns not in the basis and for each such compute the unique way to extend it to a set of columns summing to b if it exists (this is a standard technique called ‘Information Set Decoding’ introduced in [30]). Note, to ensure the Hamming weight is at most 2m/3, we cannot simply assume the solution is at most size at most m/2 by looking for the complement otherwise since this gives a maximization problem. Running Time First note that Line 16 and Line 17 can be implemented in time 2s2 m. Line 18 can be implemented with linear delay by elementary methods, and thus list2 runs in O(m(2s2 + |list2(A, b, s2 )|))  time. For Algorithm list1, Line 11 and Line 12, we see that E[|list2(A, b, s1 /2)|] = s1m/2 2−s1 , because any vector from Zm 2 of weight s1 /2 will be included with probability 2−s1 in the output. Thus these lines run in expected time O(m(2s1 /2 + m −s1 )). s1 /2 2 Similarly, the for-loop at Line 13 will take on expectation |P|2−s1 iterations, where P is 2 the set of pairs (x, y) ∈ (Zm 2 ) such that wt(x) = wt(y) = s1 /2 and Ax + Ay ≡ b. Here we s1 divide by 2 since this is the probability that such pair is in L × R because for this it needs to satisfy HAx ≡ bL . For algorithm A4 at iteration s, Line 3 and Line 5 thus take expected time      m O m 2s1 /2 + 2−s1 + E[|P|2−s1 ] , s1 /2 2 where s1 denotes s/2. Note that E[|P|] ≤ s1m/2 /2s since any pair (x, y) in P also needs to satisfy Ax + Ay ≡ b. Denoting σ = s/m and using s2 = s1 /2 = s/4 we can thus upper bound the running time with   O m2m maxσ≤2/3 {σ/4,h(σ/4)−σ/2,2h(σ/4)−3σ/2} . It is easily verified by standard calculus or a Mathematica computation that the exponent in this expression is maximized for σ = 0.4444 where it is at most O(m20.3399m ). Note that Line 4 and Line 6 can be implemented in time O(m|L|) and O(m|R|) with standard data structures and similarly we can efficiently enumerate all pairs in the for-loop at Line 7. Also note that in this for-loop the number of enumerated pairs is at most |L0 | + |R0 | since all vectors Ax for x ∈ L0 are different. Thus the algorithm indeed runs in the claimed running time. Correctness For correctness, first note that whenever yes is returned on Line 8 this is clearly correct. Now suppose the instance of Linear Sat is a YES-instance. Then there ESA 2016 [79]:18 Finding Large Set Covers Faster via the Representation Method Assumes A ∈ Zn×m Algorithm A4(A, b, t, ω) , b ∈ Zn2 , t ∈ Z, ω ∈ Nm 2 m Output: yes, if ∃x ∈ Z2 such that wt(x) ≤ 2m/3 is a multiple of 4, Ax ≡ b and ω · x ≤ t. 1: for 1 ≤ s ≤ 2m/3 such that s is a multiple of 4 do 2: Pick H ∈ Zs×n , bL ∈ Zs2 uniformly at random. 2 3: Construct L = list1(HA, bL , s/2). 4: Construct L0 such that ∀x ∈ L : ∃!x0 ∈ L0 with ω · x0 ≤ ω · x and Ax ≡ Ax0 . 5: Construct R = list1(HA, b − bL , s/2). 6: Construct R0 such that ∀y ∈ R : ∃!y 0 ∈ R0 with ω · y 0 ≤ ω · y and Ay ≡ Ay 0 . 7: for all (x, y) ∈ L0 × R0 such that Ax + Ay ≡ b do 8: if ω · x + ω · y ≤ t then return yes 9: return no. 1 ×m 1 Assumes A ∈ Z2s , b ∈ Z2s Algorithm list1(A, b, s1 ) 2 2 , s1 ∈ Z m Output: O ⊆ Zm 2 such that ∀x ∈ Z2 : Ax ≡ b ∧ wt(x) = s1 , Pr[x ∈ O] ≥ Ω(1) . s1 ×2s1 s1 10: Pick H ∈ Z2 , bL ∈ Z2 uniformly at random. 11: Construct L = list2(HA, bL , s1 /2). 12: Construct R = list2(HA, b − bL , s1 /2). 13: for all (x, y) ∈ L × R such that Ax + Ay ≡ b do 14: if x + y ∈ {0, 1}n then add x + y to O if it is not in yet. 15: return O. Assumes A ∈ Z2s2 ×m , b ∈ Zs22 , s2 ∈ Z Algorithm list2(A, b, s2 ) Output: {x ∈ Zm 2 : Ax ≡ b ∧ wt(x) = s2 }. s 16: Construct a graph D with a vertex vi,y for every y ∈ Z22 . s2 17: For every i = 1, . . . , m and y ∈ Z2 , add arcs from vi,y to vi−1,y and to vi−1,y−ai . 18: Enumerate the set P of all paths in D from vm,b to v(0, 0) with s2 arcs of the second type in O(m|P|) time. 19: Initiate S = ∅ 20: for every path P in P do 21: Let x ∈ Zm 2 be vector with xi = 1 if and only if (vi,y , vi,y−ai ) ∈ P for some y. 22: add x to S. 23: return S. Figure 4 An O∗ (20.3399n ) time algorithm for Linear Sat. Jesper Nederlof [79]:19 exists a minimum weight x ∈ Zm 2 satisfying Ax ≡ b and ω · x ≤ t. Let us consider the iteration of the for-loop on Line 1 where s = wt(x). Since x has minimum weight, the columns i of A for which xi = 1 need to be linearly independent since otherwise leaving out a linear combination would result in a smaller weight  vector x. This means that if we s denote Z = {Ay : y ⊆ x ∧ wt(y) = s/2}, then |Z| = s/2 , where y ⊆ x denotes yi ≤ xi for every coordinate i. For v ∈ Zs2 , let f (v) = |{z ∈ Z : Hz = v}| be an indicator function. We see that   X X Pr[H(z1 − z2 ) = 0] ≤ |Z| + |Z|2 2−s ≤ 2|Z|. f (v)2  = E v∈Zs2 z1 ,z2 ∈Z 2 P By Markov’s inequality we thus see that Pr[ x∈Zs f (x)2 ≤ 4|Z|] ≥ 1/2 over the choice of H. 2 Conditioned on this, the Cauchy-Schwarz inequality implies that the number of x ∈ Zs2 such that f (x) > 0 is at least |Z|2 /(2|Z|) = |Z|/2. When this happens, we have with probability √ s at least s/2 /2s = Ω(1/ s) that bL ∈ Z. If bL ∈ Z, we can apply the same reasoning to √ show that list1(HA, bL , s/2) contains y with probability Ω(1/ s) and list1(HA, b − bL , s/2) √ contains x − y with probability Ω(1/ s), and the pair will be considered in the loop on Line 7. Since the latter two probabilities are independent we see that if a solution exists we find it with probability at least Ω(n−1.5 ), and thus we may repeat the algorithm O(n1.5 ) times to ensure it finds a solution with constant probability if it exists. A.6 Proof of Lemma 4.3 Let F = {f1 , . . . , fm } be arbitrarily ordered. For integers i ∈ [n], j ∈ [m] and X ⊆ U define cij [X] to be true if and only if there exists S1 ⊆ {f1 , . . . , fj } such that |S1 | = i, N (S1 ) = X and for every f, f 0 ∈ S1 with f 6= f 0 , N (f ) ∩ N (f 0 ) = ∅. We see that ci0 [X] is true if and only if i = 0 and X = ∅, and for j > 0 we have i cj [X] = (ci−1 j−1 [X \ N (fj )] ∧ N (fj ) ⊆ X) ∨ cj−1 [X]. The values ci (W ) for W ∈ W and i ∈ [n] can be read off from cji (W ), and because for computing cij [X] we only need the entries cij [Y ] where Y ⊆ X, we can restrict our computation to computing cij (X) for X ∈ ↓W, and the runtime bound follows. A.7 Proof of Lemma 4.6 Define fx (X) to be true if and only if ∃f ∈ F : N (f ) = X and |X| = x. Note we can, within the claimed time bound, create a table storing the values fx (X) for all X ∈ ↓W using the P oracle. Now define gx (Y ) = Y ⊆X fx (X). Let U = {u1 , . . . , un } and define X gxj (X) = fx (Y ). X∩{u1 ,...,uj }⊆Y ⊆X Then we see that gxn (X) = fx (X), and for j < n we can compute gxj using gxj (X) = [uj ∈ X]gxj+1 (X \ {uj }) + gxj+1 (X). (A.1) Thus, by straightforward dynamic programming using (A.1) we can compute gx (X) = gx0 (X) for every x and X ∈ ↓W in O(n2 |↓W|) time. Next, define X hx,i (X) = gx1 (X)gx2 (X) · · · gxi (X), x1 +x2 +...+xi =x ESA 2016 [79]:20 Finding Large Set Covers Faster via the Representation Method Pi or equivalently, hx,i is the number of tuples f1 , . . . , fi such that N (fi ) ⊆ X and l=1 |N (fi )| = x. Note that for a fixed X, given the entries gx (X) for every x ≤ n, we can compute hx,i (X) in time poly(n) using standard dynamic programming or the Fast Fourier Transformation. Denote ( ) i X i 0 cx,i (X) = (f1 , . . . , fi ) ∈ F : N ({f1 , . . . , fi }) = X ∧ |N (fl )| = x , l=1 which is easily seen to be the number i-tuples of disjoint sets whose union is exactly X P if x = |X|. Note that hx,i (X) = Y ⊆X c0x,i (Y ) since every i-tuple counted in hx,i (X) is counted once in a c0x,i (Y ) where Y is the union of the i-tuple. Then, since any non-empty set has equally many even-sized as odd-sized subsets (e.g., by inclusion exclusion), we see that   X X  c0x,i (X) = (−1)|Z|  c0x,i (Y ) Y ⊆X Z⊆X\Y  = X (−1)|Z|  Z⊆X (A.2)  X Y ⊆X\Z c0x,i (Y ) = X (−1)|Z| hx,i (X \ Z). Z⊆X Then, define hjx,i (X) = X (−1)|Z| hx,i (X \ Z), X∩{u1 ,...,uj }⊆Z⊆X and similarly to (A.1), we see that hnx,i (X) = hx,i (X) and for j < n we can compute hjx,i using j+1 hjx,i [X] = −[ui ∈ X]hj+1 x,i (X \ {ui }) + hx,i (X). (A.3) Consequently, we can use Equation A.3 for a straightforward dynamic programming algorithm to determine h0|X|,i [X] which equals c0x,i (X) by (A.2), and it is easy to see that ci (X) is true if and only if c|X|,i (X) > 0.
8cs.DS
Hierarchical Hyperlink Prediction for the WWW arXiv:1611.09084v1 [cs.DS] 28 Nov 2016 Garcia-Gasulla, D. & Ayguadé, E. & Labarta, J. Barcelona Supercomputing Center (BSC) [email protected] & Cortés, U. Universitat Politècnica de Catalunya - BarcelonaTECH & Suzumura, T. IBM T.J. Watson Research Center Barcelona Supercomputing Center (BSC) November 29, 2016 Abstract The hyperlink prediction task, that of proposing new links between webpages, can be used to improve search engines, expand the visibility of web pages, and increase the connectivity and navigability of the web. Hyperlink prediction is typically performed on webgraphs composed by thousands or millions of vertices, where on average each webpage contains less than fifty links. Algorithms processing graphs so large and sparse require to be both scalable and precise, a challenging combination. Similaritybased algorithms are among the most scalable solutions within the link prediction field, due to their parallel nature and computational simplicity. These algorithms independently explore the nearby topological features of every missing link from the graph in order to determine its likelihood. Unfortunately, the precision of similarity-based algorithms is limited, which has prevented their broad application so far. In this work we explore the performance of similarity-based algorithms for the particular problem of hyperlink prediction on large webgraphs, and propose a novel method which assumes the existence of hierarchical properties. We evaluate this new approach on several webgraphs and compare its performance with that of the current best similarity-based algorithms. Its remarkable performance leads us to argue on the applicability of the proposal, identifying several use cases of hyperlink prediction. We also describes the approach we took for the computation of large-scale graphs from the perspective of high-performance computing, providing details on the implementation and parallelization of code. Web Mining, Link Prediction, Large Scale Graph Mining 1 Introduction Link prediction is a link mining task focused on discovering new edges within a graph. Typically, we first observe the relationships (or links) between some pairs of entities in a network (or graph) and then we try to predict unobserved 1 links [35]. Link prediction methods are relevant for the field of Artificial Intelligence (AI) as for any Knowledge Base (KB) represented as a graph these methods may produce new knowledge in the form of relations. In most cases link prediction methods use one of two different types of graph data for learning [32, 37]: the graph structure (i.e., its topology) and/or the properties of the graph components (e.g., node attributes). Combinations of both have also been proposed [3]. Solutions based on the properties of components must rely on a previous study of those properties (e.g., which attributes are present, which is their type, their relevance, etc.). These solutions are therefore only applicable to domains where those properties hold, and can hardly be generalized. On the other hand, solutions based on structural information of the graph are easily applicable to most domains as this information is found in all graphs and their interpretation is broadly shared (i.e., all graphs have a topology and its interpretation is the connectivity of elements). In this article we will focus on link prediction methods based on structural data. The problem of link prediction is commonly focused on undirected, unweighted graphs. Whereas plain edges have some universal semantics (i.e., the existence of a relationship between two entities), additional properties such as directionality and weights may have varying interpretations (e.g., what is the difference between the origin and the destination of an edge? what is the impact of weights?). Link prediction methods taking into account one or more of these additional properties must make certain assumptions regarding their semantics, assumptions which will limit the domains to which they can be applied. However, if the semantics of these properties are shared by a large set of domains, the development of specific link prediction methods with these assumptions may be beneficial; these methods may increase the quality of the results while still being applicable to a relevant set of domains. 1.1 Hyperlink Prediction The World Wide Web naturally represents a webgraph: a large, directed graph composed by web pages connected through hyperlinks. In contrast with traditional instance-attribute based data sets, webgraphs define highly distributed and inter-connected structures [11, 43], often referred to as networks. Networks became popular data structures with the explosion of the Internet and motivated a change on how data was being processed by fields like Data Mining (DM) and Machine Learning (ML), as these fields had to shift their focus towards finding inter-entity patterns (i.e., entity-entity relations). Graph-based data mining [20], Statistical Relational Learning [18], Link Mining [32], Network or Link Analysis [19], Network Science [29] or Structural Mining [8] are all names used to identify these knowledge discovery tools which share the same precept: exploiting structural properties of high-dimensional, inter-connected data sets to learn about the relational patterns of its entities. For the sake of simplicity, we generally refer to this field of science with the term graph mining. A defining contribution to the graph mining field was the identification of certain problems which had not been successfully tackled by traditional DM and ML techniques [17]. Tasks like community detection [22], frequent subgraph discovery [21] and link-based object ranking [41]. These were to become the main target of graph mining algorithms. One of those tasks, the link prediction (LP) 2 problem of finding new relations in a high-dimensional data set (e.g., adding new edges to a given graph), is one with a particularly wide range of applications. Through LP one can automatically produce new knowledge (edges) within a domain (a graph) using its own language (the topology of vertices and edges), avoiding any interpretation step. This infrequent and enabling feature should have made of LP a technique widely applied to domains that could be represented as a graph. Unfortunately scalable LP methods are often either imprecise or non-generalizable, problems that have constrained their popularization so far. A prototypical domain of application of LP is the WWW webgraph, a large and sparse graph growing continuously in a distributed manner. Any algorithm to be applied to large graphs such as webgraphs needs to be scalable. In fact, their size often motivates the use of High Performance Computing (HPC) tools and infrastructure, even when using scalable methods. The challenges arising from the computation of large networks has been noticed by the HPC community, as demonstrates the popularization of the Graph500 benchmark to evaluate on how fast can supercomputers process large-scale graphs [48], or the development of graph-specific parallel programming models [33]. The collaboration between graph mining and HPC seems therefore to be inevitable in the near future, as both fields may benefit from it: graph mining by obtaining the means for its goals and HPC by obtaining a purpose for its capabilities. In this context, the work here presented provides the following contributions: • WWW: Discuss the importance of hierarchical properties for the topological organization of the WWW. Study the performance of four different LP scores for the specific problem of hyperlink prediction, including one that exploits hierarchical properties. • Graph mining: Propose an improved version of INFerence, a score for directed link prediction based on hierarchies, to process informal graphs not explicitly hierarchical. Compare the results with the current best similarity-based algorithms. Study the distinct types of local LP algorithms, and proposed a combined solution to significantly improve performance. • HPC: Present a feasible and applicable use case of large-scale graph mining. Discuss the scalability and parallelism of similarity-based LP algorithms and how these can be extended to distributed memory setting for the computation of large-scale graphs. The rest of the paper is structured as follows: the first two sections survey related work. §2 summarizes the current state-of-the-art on LP, and contains a brief survey on previous attempts to perform LP on WWW data. §3 reviews the role of hierarchies in the WWW according to previous work. Once the related work has been presented, in §4 we introduce and formalize the HybridINF score, a novel LP algorithm for the task of hyperlink prediction. We describe our experiments in §5, including the webgraphs used, the evaluation methodologies and the analysis of results. From those results we derive various possible applications in §6. We discuss computational details in §7, describing our implementation and parallelization approach, as well as its extension to a distributed memory setting. Finally, we present our conclusions in the context of both LP and the WWW in §8, and motivate our future work in §9. 3 2 Link Prediction Background LP algorithms are often classified in three families [32]: statistical relational models (SRM), maximum likelihood algorithms (MLA) and similarity-based algorithms (SBA). SRM may include probabilistic methods. Probabilistic SRM build a joint probability distribution representing the graph based on its edges, and estimate the likelihood of edges not found in the graph through inference [18]. These are frequently based on Markov Networks or Bayesian Networks. Non-probabilistic SRM are often based on tensor factorization, a useful approach for predicting edges in heterogeneous networks (i.e., composed by more than one type of relation) [40]. These have also been used in combination with probabilistic models [46]. The second type of LP algorithms, MLA, assume a given structure within the graph (e.g., a hierarchy, a set of communities, etc.) and fit a graph model through maximum likelihood estimation. Based on the model, MLA calculate the likelihood of non-existing edges. MLA provide relevant insight into the composition of the graph (i.e., how is its topology defined and why), information that can be used for other purposes beyond LP. An example of MLA is the Hierarchical Random Graph [7] which builds a dendrogram model representing a hierarchical abstraction of the graph, and obtains the connection probabilities that most accurately represents the graph hierarchically. The third type of LP methods, SBA, compute a similarity score for each pair of vertices given their topological neighbourhood. Each edge is evaluated on its own, thus potentially in parallel, without previously computing a graph model (unlike SRM and MLA). To score an edge, SBA focus on the paths between the pair of vertices. As a result SBA can be categorized in three classes according to the maximum path length they explore: local if they consider only paths of length 2, global if they consider paths without length constrain, and quasi-local if they consider limited path lengths larger than 2. Global SBA effectively build a model of the whole graph, with the consequent computational cost and poor scalability [32]. Quasi-local SBA are a compromise between between efficacy and efficiency, and are often based on the random walk model [30, 29] or on the number of paths between the pair of vertices [31]. Beyond the previous classification of LP methods, there are more heterogeneous solutions. In [3], the LP process is decomposed in two steps: a first macro-processing step obtains clusters in the graph satisfying certain structural properties, while a second micro-processing step tries to find new links on those clusters based on attribute properties. This work is particularly appropriate for heterogeneous networks that grow over time. The same bi-scale approach is taken in [12], where the mesoscale and the microscale components are combined through a statistical inference model. Significantly, the microscale component is based on the Resource Allocation algorithm [51]. A different solution is proposed in [29], where authors identify 12 topological features of directed weighted links which are used as input to a set of traditional DM classifiers. To make the approach feasible, link features are obtained at a local level, considering only the sub-graph at 2 to 4 hops away. Results indicate that their proposed algorithm (HPLP), which includes CN as feature, implements an ensemble of classifiers through bagging and random forests, and performs undersampling for correcting the class imbalance, outperforms SBA on both directed and an undirected weighted graphs. 4 2.1 Local similarity-based algorithms Local SBA are among the most scalable approaches to LP, as these methods only need to explore the direct neighbourhood of a given pair of vertices to estimate the likelihood of a edge between them. Since this estimation can be done in parallel (it does not depend on the likelihood of other edges) these methods scale very efficiently and hardly waste any resources. From the perspective of similarity scores, local are the most essential of SBA, as all quasi-local originate from a local one: when the number of steps is set to one, quasi-local SBA are equivalent to some local SBA [15]. This motivates the research on better local SBA from which better quasi-local SBA can be derived. For the remaining of this paper we will focus on SBA methods. Similarity-based algorithms were first compared in [28]. Nine algorithms were tested on five different scientific co-authorship graphs in the field of physics and compared with a random predictor as baseline. Although no method clearly outperformed the rest in all datasets, three methods consistently achieved the best results; local algorithms Adamic/Adar and Common Neighbours, and global algorithm Katz [23]. In [37] the same results were obtained, with Adamic/Adar and Common Neighbours achieving the best results among local algorithms. In [51] a new local algorithm called Resource Allocation was proposed and compared with other local similarity-based algorithms. Testing on six different datasets showed once again that Adamic/Adar and Common Neighbours provide the best results among local algorithms, but it also showed how Resource Allocation was capable of improving them both. The Common Neighbours (CN) algorithm [39] computes the similarity sx,y between two vertices x and y as the size of the intersection of their neighbours. Formally, let Γ(x) be the set of neighbours of x Definition 1 sCN x,y = |Γ(x) ∩ Γ(y)| The Adamic/Adar (AA) algorithm [2] follows the same idea as CN, but it also considers the rareness of edges. To do so, shared neighbours are weighted by their own degree and the score becomes Definition 2 X sAA x,y = z∈Γ(x)∩Γ(y) 1 log(|Γ(z)|) The Resource Allocation (RA) algorithm [51] is motivated by the resource allocation process of networks. In the algorithm’s simpler implementation, each vertex distributes a single resource unit evenly among its neighbours. In this case, the similarity between vertices x and y becomes the amount of resource obtained by y from x Definition 3 sRA x,y = X z∈Γ(x)∩Γ(y) 5 1 |Γ(z)| 2.2 Link Prediction for the WWW The first application that comes into mind when considering the application of LP to the WWW is to discover missing hyperlinks. However, there are few contributions which specifically focus on the application of SBA for such purpose. In contrast, other similar problems like predicting relations in social network graphs, have been thoroughly studied using a wide variety of methodologies [2, 28, 29, 38, 45]. Webgraphs have been occasionally used to evaluate LP scores, but always in combination with other graphs types, and never as an independent case of study. When webgraphs have been used for LP, only relatively small graphs have been computed (a few thousand vertices), typically due to the complexity of the models being used [31, 47]. An exception to the lack of attention to hyperlink prediction is the Wikipedia webgraph, composed by Wikipedia articles and the hyperlinks among those. The encyclopedic information contained in the vertices of this graph has served as a motivation to researchers, who have used it to enrich the learning process. As a result most LP methods used on the Wikipedia webgraph are heterogeneous combinations of ML methods that can be neither scaled to large graphs, nor generalized to other webgraphs. In [1] authors propose an algorithm combining clustering, natural language processing and information retrieval for finding missing hyperlinks among Wikipedia pages. With a more specific target in mind, in [36] authors use supervised learning to train a disambiguation classifier, in order to choose the appropriate target for article hyperlinks. Multiple contributions also exist from the semantic web community, typically using RDF data to implement link inference methods [25]. 3 Hierarchies in the WWW Hierarchies are the most widely used knowledge organization structure. They can be found in domains such as the human brain, metabolic networks, terrorist groups, protein interactions, food webs, social networks, and many others. Within the field of link prediction hierarchies have been acknowledged as a relevant source of information due to its power to describe knowledge [44]. However, while link prediction methods are predominantly focused on undirected structures for the sake of simplicity, hierarchies are necessarily directed. This structural divergence complicates the use of hierarchies for traditional link prediction, and is the main reason why this combination have not been further exploited so far. One of the main contributions of our work is the adaptation and application of a LP score based on hierarchies to the problem of hyperlink prediction. By doing so we are implicitly assuming that the WWW topology contains or is partly defined by hierarchical properties. That assumption has been discussed, empirically evaluated and exploited in the past, as presented in this section. What is novel in our contribution is its successful application to the use case of hyperlink discovery. As we will see in §5, we consider hierarchies at a much smaller scale than usual, thus enabling novel hierarchical properties and applications in the context of the WWW. The topology of routing units [42], and the WWW requests and traffic [9] were the first aspects of the Internet shown to have hierarchical properties. The 6 webgraph topology itself was found to compose a self-similar structure defined by a set of nested entities in [11], indicating a natural hierarchical characterization of the structure of the web. This idea was extended in [43], where authors showed how some key properties found in webgraphs and other real networks, such as being scale-free and having a high degree of clustering, could be satisfied through a hierarchical organization. In their work, Ravasz and Barbási propose a hierarchical network model fitting such networks, to be composed by small, dense groups, recursively combining to compose larger and sparser groups, in a fractal-like structure. The importance of hierarchies to model the structure of real networks was explored through its application to generative models: models built to produce artificial, large scale networks mimicking the topological properties of real networks. The work of Leskovec et al. [27] is of particular interest, as authors define generative models satisfying complex properties never captured before (e.g., densification, shrinking diameters and a heavy-tailed out-degree). The most complex model proposed by Leskovec et al., the forest fire model (FFM), builds a network by iteratively adding one new vertex x to a graph. In this model, one randomly provides an entry point to x in the form of a vertex x will point to through a first out-going edge (i.e., its ambassador). In a second step, a random number of out-going edges are added to x, to be chosen among the out-going and in-going edges of the ambassador (the former with a higher probability of being selected than the latter). This process is then repeated recursively for each new vertex that gets connected with x, with the restriction that each ambassador is considered only once. Notice that the FFM methodology contains no explicit hierarchy, and yet it generates data which is hierarchically structured. We will discuss how and why this happens in §8, by comparing FFM with our proposed LP algorithm. 4 The Hybrid-INF Score The local SBA obtaining the best predictive results so far are all based on the accumulated number of shared neighbours (e.g., CN). Additionally, some of these scores weight the evidence provided by each neighbour by its own degree, thus increasing the importance of rare edges (e.g., RA, AA) [28, 37, 51]. CN and AA have been shown to achieve the best results among local SBA in citation and social networks [28, 37], while RA was capable of achieving similar results to those of CN and AA in several domains, and even to improve them in a webgraph of political blogs [51]. To emphasize the importance of these type of scores, let us remark that the current best quasi-local SBA scores are based on one either CN, RA or AA [15], and most alternative solutions also include them [3, 12, 29]. This illustrates the potential benefits of obtaining better local SBA. Given the scarce information available to local SBA (i.e., the topological neighbourhood around a pair of vertices) it seems difficult to find scores which are significantly more precise than the current best. As one of the main contributions of this article we argue that in the presence of large amounts of data, implicit graph models may naturally exist, and that these models can be used to significantly improve the performance of LP. A model assumed to exist and exploited for learning has the benefit of providing additional knowledge on the structure of the network, while not adding any complexity (e.g., building the 7 model). As discussed in §3, the WWW webgraph is strongly related with hierarchical organizations, which leads us to believe that the hyperlink prediction process could be improved by the assumption of a hierarchical model. Hierarchies are at the core of INFerence (INF [15]), a LP score assuming hierarchical properties in the topology to guide the LP process. INF was designed for mining explicitly hierarchical graphs, implementing properties like edge transitivity. The original definition of INF however seems inappropriate for graphs without an explicit hierarchical structure, as INF assumes the computed graph implicitly implements a hierarchy. Based on this assumption INF defines a score for each ordered pair of vertices measuring its hierarchical evidence, providing higher scores to hierarchically coherent edges. So far INF has been shown to be a very good predictor for graphs containing explicitly hierarchical relations, such as the ontology is-a relation and the hyponymy/hyperonymy linguistic relation of WordNet [15]. Given a directed graph G = (N, E), and a given vertex x ∈ N , the original definition of INF refers to the vertices linked by an out-going edge from x as the ancestors of x (A(x)), and to the vertices linked by an in-going edge from x as the descendants of x (D(x)). Definition 4 ∀x, y ∈ N : y ∈ A(x) ↔ x → y ∈ E ∀x, y ∈ N : y ∈ D(x) ↔ y → x ∈ E Based on those sets INF defines two sub-scores named deductive sub-score (DED) and inductive sub-score (IND), in coherency with a specialization/generalization sense between descendants/ancestors. The DED sub-score follows a top-down reasoning process from the abstract to the specific, resembling that of a weighted deductive inference: if my four grandparents are mortal, I will probably be mortal too. In this case information of a vertex (me → mortal) is obtained from the ancestors of that vertex (me → grandparent → mortal), in proportion with the number of times the relation is satisfied (e.g., four out of four if I do not have any living grandparent). Hence, the more of our generalizations share a property, the more certain we can be about that property applying to ourselves. IND on the other hand follows a bottom-up reasoning process from the specific to the generic, resembling that of a weighted inductive inference: if most of an author’s publications are meticulous, the author will most likely be meticulous. In this example, the information of a vertex (author → meticulous) is obtained from the descendants of that vertex (publication → meticulous, publication → author), in proportion with the number of times the relation is satisfied. Hence, the more frequently our specializations share a property, the more certain we can be about that property applying to ourselves. See Figure 1 for a graphical representation of these processes. INF is then defined as DED + IN D, as in Definition §5. Definition 5 F sIN x→y = |A(x) ∩ D(y)| |D(x) ∩ D(y)| + |A(x)| |D(x)| The INF algorithm as defined in Definition §5 maximizes the predictions of links according to a natural sense of hierarchy. Anti-hierarchical evidence, such as the one provided by the common neighbour Z in the setting X ← 8 Figure 1: Dashed edge represents link inference. On the left: graphic representation of the top-down DEDuctive process for estimating link likelihood. On the right: graphic representation of the bottom-up INDuctive process for estimating link likelihood. On both cases C nodes represent the evidence considered by each sub-score. Z ← Y , is not considered when evaluating the score of the edge X → Y . This is potentially beneficial for domains which are implicitly hierarchical, such as webgraphs. However, INF is a purely proportional score, which makes it unsuited for real world data. 4.1 Proportional and Accumulative Scores INF was originally defined to target formal graphs [15] where relations have a high reliability due to an expert validation process (e.g., WordNet), or because they originate from formal properties (e.g., ontologies). To exploit this fact, INF was entirely based on the proportion of evidence (see Definition §5) as in formal graphs a single relation is as reliable as many, regardless of size (e.g., one out of one vs ten out of ten). Unfortunately, proportional evidence is not equally reliable when working with informal graphs (e.g., webgraphs, social networks, costumer-item networks, etc.), where edges often contain errors, outliers or imbalanced data. In this setting, considering a one out of one evidence set as a certain scenario would be precipitous and prone to error. An alternative to proportional evidence is accumulative evidence, the most frequent approach used by LP algorithms. Proportional scores, such as INF or Jaccard’s coefficient [28], weight the evidence of edges according to their local context, and provide a normalized similarity for each edge regardless of their degree. This makes proportional scores unbiased towards edges among highdegree vertices. Accumulative scores on the other hand measure the absolute amount of evidence, ignoring the local context (e.g., source vertex degree). In these scores, edges are evaluated and ranked from a graph-wide perspective, which benefits predictions around high-degree vertices. Previous results have shown that in general, accumulative scores perform better than proportional scores: the top three scores found in the bibliography are accumulative (i.e., CN, AA and RA). This can be explained by the importance of the preferential attachment process (i.e., the rich get richer), which accumulative scores implicitly satisfy, and proportional scores avoid. But the generally poor results of proportional scores is also caused by the volatility of predictions among low-degree vertices, which proportional scores do not neglect, and which are often less reliable due to arbitrariness. 9 4.2 From INF to a Hybrid Score To adapt the INF score to informal domains such as webgraphs, we extend it to consider both proportional and accumulative evidence. Although most current solutions are purely accumulative, we build a hybrid solution with the goal of combining the certainty of proportional scores with the reliability of accumulative ones. To do so we first normalize evidence given the local context (proportionally, through a division) and then weight that proportion through the absolute size it was based on (accumulatively, through a logarithm). By using a logarithmic function to weight the score, the accumulative evidence dominates the scores of low degree vertices, while the proportional evidence remains more important for high degree vertices. The IN F LOG variant is defined in Definition §6. Definition 6 F sIN x→y LOG = |A(x) ∩ D(y)| |D(x) ∩ D(y)| ∗ log(|A(x)|) + ∗ log(|D(x)|) |A(x)| |D(x)| The original INF definitions combines the evidence provided by DED and IND through an addition (IN F = DED + IN D), considering both of them equally reliable. However, through a preliminary evaluation we found that DED consistently achieves higher precisions than IND on most domains. Our hypothesis to understand this behaviour is two-fold. For formal graphs, the deductive reasoning process in which DED is based may be more reliable than the inductive reasoning process of IND. While for informal graphs, this behaviour may be caused by the fact that a vertex typically has a higher responsibility on defining the set of vertices it points to, while the set of vertices that point to it are often imposed onto it (e.g., webmasters rarely choose who links to their site). A higher reliability of the outgoing edges of a vertex would make DED more precise at defining the vertex itself, as those are the edges used by DED. Even though DED is more reliable than IND, the contribution of IND to the INF score is still relevant, as it is capable of detecting a type of relational evidence DED does not consider. We solve this problem by adding a multiplying factor to DED (IN F kD = k ∗ DED + IN D), as seen in Definition §7 Definition 7 F sIN x→y LOG kD =k∗ |D(x) ∩ D(y)| |A(x) ∩ D(y)| ∗ log(|A(x)|) + ∗ log(|D(x)|) |A(x)| |D(x)| We sampled the k value for several domains, and found the optimal value to be between 1 and 3 for most graphs. To provide a consistent score evaluation in §5, we set k = 2 in all our tests. This implementation of the Hybrid-INF score is identified as IN F LOG 2D. 5 Experiments Next we evaluate the INF LOG 2D algorithm for the hyperlink prediction problem within its same family of algorithms (local SBA scores). Although the hyperlink prediction problem is a directed link prediction problem, we use CN, AA and RA (all undirected link prediction scores) as baseline because these are the 10 current state-of-the-art in LP, and because there are no analogous directed local SBA scores in the bibliography. We compare the performance of INF LOG 2D, CN, AA and RA, on six webgraphs obtained from different sources: the webgraph of the University of Notre Dame domain in 1999 (webND) [4], the webgraph of Stanford and Stanford-Berkley in 2002 (webSB) [24], a webgraph provided by Google for its 2002 programming contest (webGL), a domain level webgraph of the Erdős WebGraph project from 2015 (erdős), and the webgraphs of the Chinese encyclopedias Baidu and Hudong [26]. Sizes of all graphs are shown in Table 1. All graphs used are publicly available through the referenced sources. Graph webND webSB webGL erdős hudong baidu Vertices 325,729 685,230 875,713 1,817,390 1,984,484 2,141,300 Edges 1,497,134 7,600,595 5,105,039 16,391,889 14,869,484 17,794,839 Table 1: Size of computed webgraphs. For its evaluation, the LP problem is reduced to a binary classification problem. In this context one has a set of correct instances (edges missing from the graph known to be correct) and a set of wrong instances (the rest of missing edges) and the goal of the predictors is to classify both sets of edges as well as possible. The most frequently used metrics to evaluate the predictors performance in this context are the Receiver Operating Characteristic (ROC) curve and the Precision-Recall (PR) curve. The ROC curve sets the True Positive Rate (TPR) against the False Positive Rate (FPR), making this metric unbiased towards entities of any class regardless of the size of classes. Unfortunately, their consideration of errors can result in mistakenly optimistic interpretations [10, 49], as ROC curves represent missclassifications relative to the number of errors that could be made. In domains where the negative class is very large and one can make millions or even billions of errors, showing mistakes as relative to the negative class size (i.e., FPR) may hide their actual magnitude and complicate a realistic assessment of predictive performance. Furthermore, in large and highly imbalanced domains most of the ROC curve becomes irrelevant in practice, as it represents inapplicable precisions below 1% [16]. Precision-recall (PR) curves are an alternative to ROC curves, which show precision on the y axis and recall on the x axis. PR curves do not show the number of correct classifications for the negative class, and instead represent miss-classifications relative to the number of predictions made. This allows for a straight-forward idea of the actual predictive quality in absolute terms, and makes the whole curve relevant regardless of the problem size. In fact, ROC and PR curves are strongly related, as a curve dominates another (it is above it) in the ROC space if and only if it also dominates it in PR space [10]. The main difference between ROC and PR curves is on how errors are represented, but this difference has a huge visual impact. Consider for example how the two curves represent a random classifier, which always performs poorly in a large 11 Webgraph webND webSB webGL erdős hudong baidu Positive edges 133,279 756,937 494,982 1,543,090 1,446,760 1,701,330 Negative edges 95,939 million 466,034 million 741,978 million 2,970,460 million 3,786,991 million 4,370,982 million Class imbalance 1:719,835 1:615,684 1:1,498,999 1:1,925,007 1:2,617,566 1:2,569,155 Table 2: For each graph, number of correct edges and number of incorrect edges being computed, and positive:negative class imbalance in the evaluated set. and highly imbalanced data set. The ROC curve always represents the random classifier as a straight line between points (0, 0) and (1, 1), regardless of class imbalance, with all better than random classifiers represented as lines above that diagonal. The more demanding PR curves on the other hand represent random classifiers in imbalanced data sets a flat line on the x axis, as their precision in imbalanced settings is always close to zero. Due to the previously outlined motives and the huge class imbalance found in our graphs (see Table 2), we will use the Area Under the Curve measure of the PR curve (AUPR) to compare the various scores evaluated. Let us remark this is the recommended approach in this setting [49], even though it has not been fully assimilated by the community yet. To calculate the PR curve one needs a source graph (on which we wish to predict edges) and a test set (edges missing from the source graph known to be correct). To build the test set we randomly split the edges of each graph 90%-10%, and use the 90% as source graph and the remaining 10% as test set. The size of test sets for each evaluated graph are shown in Table 2. Vertices becoming disconnected in the source graph after removing the test edges were not computed nor considered for score evaluation. A frequently used methodology to build test sets is 10-fold cross validation. However, due to the size of the graphs being used this is not necessary. The law of large numbers will make any significant portion (e.g., 10%) of a large domain tend towards a stable sample, thus making a single run a representative and accurate sample of the performance [16]. We build all PR curves exhaustively, as all possible edges in the graph are computed. Approximate and potentially dishonest methodologies such as test sampling [49] were avoided. Computing all possible edges for graphs the size of the ones used here equals to the evaluation of billions of edges (see Table 2), and any proposed solution needs to be scalable in order to be feasible. SBA fit this requirement perfectly, as they can be parallelized with maximum computational efficiency. See §7 for details on the implementation and parallelization of the solution. 5.1 Results Analysis The PR curves of all four scores on all six graphs can be seen in Figure 2. INF LOG 2D achieves the best predictive performance on all graphs by a large margin. The huge improvement in precision obtained by INF LOG 2D (i.e., much higher y axis values on its curve), results in an increase of up to one 12 Figure 2: PR curves for all webgraphs. Recall in x axis, precision in y axis. Erdős, hudong and, baidu curves are zoomed in for clarity. order of magnitude in the AUPR over the current state-of-the-art (see Table 3). Significantly, INF LOG 2D achieves precisions of 100% at a very small recall, even when the other scores never reach 50% precisions. As seen in these results, INF LOG 2D can recommend thousands of edges (those with higher reliability) making very few mistakes in the process. The leap in performance of INF LOG 2D, and its consistence within webgraphs, also stresses the importance of hierarchical properties for hyperlink prediction, and how much it can be gained by integrating implicit network models within predictive algorithms. 5.2 Hybrid Improvement INF LOG 2D is a hybrid score which outperforms the best accumulative scores on different six webgraphs (see Table 3). So far, purely accumulative scores had obtained the best results on almost every informal graph evaluated, which makes these results all the more important. To validate the importance of the hybrid approach also when compared to purely proportional scores, we tested a well13 webND webSB webGL erdős hudong baidu AA 0.31679 0.02218 0.08961 0.01933 0.00555 0.00285 CN 0.31855 0.01669 0.06227 0.01826 0.00743 0.00176 RA 0.21178 0.05491 0.10035 0.02220 0.00223 0.00308 INF LOG 2D 0.52640 0.45156 0.49210 0.07299 0.03424 0.00528 Improvement 65.24% (CN) 722.36% (RA) 390.38% (RA) 228.78% (RA) 360.83% (CN) 71.42% (RA) Table 3: AUC obtained by each score on the PR curves shown in Figure 2. Also, improvement over best accumulative score in percentage of PR-AUC. Webgraph webND webSB webGL erdős hudong baidu INF PR-AUC 0.09966 0.10530 0.12826 0.00698 0.00402 0.00065 Accum. PR-AUC 0.31855 0.05491 0.10035 0.02220 0.00743 0.00308 Hybrid Improve. 428% 328% 283% 945% 751% 712% Table 4: For each graph tested, PR-AUC of the basic INF score, PR-AUC of the top accumulative score, and percentage of PR-AUC improvement of INF LOG 2D over INF known proportional score on the same six webgraphs, the Jaccard’s coefficient score [28]. Where Γ(x) is the set of neighbours of x, Jaccard’s coefficient is defined as Definition 8 sJaccard = x,y |Γ(x) ∩ Γ(y)| |Γ(x) ∪ Γ(y)| Our results showed that Jaccard’s performance was incomparably worse on all webgraphs; when plotted together with the other scores, Jaccard’s PR curve was a flat line on the x axis on all graphs. Jaccard’s is not shown in Tables 3 or §4 because its AUPR was always zero given the five decimals used. Consistently with previous research, we find proportional scores to be very imprecise on their own. However, for the first time we show that proportional scores can help improve significantly the more precise accumulative scores when integrated. To explore the impact of making hierarchical assumptions on webgraphs, we now consider the performance of the basic INF score. INF is a purely proportional score assuming a hierarchical model in the graph, and it actually outperforms the best accumulative scores on two of the six webgraphs (see Table §4). INF is incomparably better than the also proportional Jaccard’s coefficient, which shows how the handicaps of using a proportional approach can be overcome by exploiting implicit data models found in the graph. Finally, let us compare the performance of INF LOG 2D with that of INF. As defined in §4.2, INF LOG 2D is a hybrid version of the purely proportional score INF. Their comparison is therefore a reliable test on the benefits of turning a proportional score into a hybrid score. In that regard, Table 4 shows that INF LOG 2D consistently improves the performance of INF by two orders of magnitude in the AUPR measure for all graphs, regardless of INF achieving competitive results on it or not. 14 6 Applications On large graphs, LP is frequently imprecise, does not scale or cannot be generalized. Three reasons which have constrained its application. SBA are clearly generalizable, as they can be applied to virtually any graph composed by vertices and directed edges1 . SBA are clearly scalable, as they can be perfectly parallelized to compute graphs of any size. Unfortunately, current SBA are rather imprecise, hardly reaching 50% precisions on graphs close to 1 million vertices (see Figure 2). webND webSB webGL erdős hudong baidu AA +66% +1935% +449% +277% +516% +85% CN +65% +2605% +690% +299% +360% +200% RA +148% +722% +390% +228% +1435% +71% Table 5: Percentage of PR-AUC improvement achieved by INF LOG 2D. In this context we have shown how, by considering inherent topological features (e.g., hierarchies) and by combining different approaches (e.g., proportional and accumulative into hybrid scores), one can overcome the limitation of imprecision. The results obtained by INF LOG 2D are a huge improvement over the previous state-of-the-art (see Table 5 for details). But more importantly, these results show how SBA can reach precisions between 1 and 0.9, enabling the reliable discovery of tens of thousands of hyperlinks in a straightforward fashion (i.e., without building a model of the graph). From these unprecedented results many applications can be derived. • Web search engines. Current search engines are composed by a wide variety of interacting metrics, which together produce a complete ranking of web page relevance. The measure of hierarchical similarity between webpages provided by INF LOG 2D may represent a different sort of evidence, and could be used to enrich the ranking of web pages from a different perspective, once the utility of the INF LOG 2D score to characterize webpages within a webgraph has been validated here. Scores such as INF LOG 2D could be combined with algorithms like Page Rank by spreading relevance not only to those webs directly connected, but also to those that LP algorithms estimate as potential neighbours with high reliability. • Web connectivity. Hierarchical properties of the WWW spontaneously emerge at a global level (see §3), but at a local level things are rather chaotic. Each webmaster must find appropriate webpages to link to, in a domain with billions of websites. As humans cannot be aware of every single webpage online, a hyperlink recommender could be useful for web masters to find relevant web pages to link to. This tool could improve 1 Depending on the score. CN, AA and RA do not require edge directionality. INF LOG 2D does. 15 Graph webND webSB webGL erdőss hudong baidu Vertices Edges Edges to evaluate Computation time 325,729 1,497,134 9,593.78M 20 seconds 685,230 7,600,595 466,028M 21 minutes 875,713 5,105,039 741,973M 50 seconds 1,817,390 16,391,889 2,970,460M 92 minutes 1,984,484 14,869,484 3,786,970M 49 minutes 2,141,300 17,794,839 4,370,980M 71 minutes Table 6: Size of computed webgraphs, missing edges to be evaluated and time spent evaluating them. Computational context is defined in §7.2 both the connectivity and coherency of the WWW, or of a specific web domain, as well as significantly enrich directory webs. • Bottom-up taxonomy building. Taxonomies are frequently used in online shops and encyclopedias (among others) to organize content. These taxonomies are often defined by external experts, requiring a continuous and expensive manual post-process of data mapping (e.g., fitting web pages and articles to taxonomy entities). According to our results it seems feasible to develop an automatic taxonomy building system, proposing a taxonomy of web pages based on their interrelations, similarly to what was proposed by Clauset et al.[7]. Such a taxonomy would have the benefit of originating from the data, making it necessarily relevant for the domain in question. It would also be easily updated. A taxonomy like this can be used to optimize the commercial organization of an online shop, for example by considering user navigation paths as source for the LP algorithms. 7 Implementation In the tests performed in this paper, the LP algorithms compute billions of edges for each processed graph (see Table 6). Computing these edges sequentially, one by one, is clearly impractical, which makes High Performance Computing (HPC) parallelism necessary for the feasibility of our work. In §7.1 we review the algorithmic design and code parallelization we developed to maximize efficiency. In §7.2 we consider the different computational contexts in which large scale graph mining problems can be set, and how we tackle this particular problem on each of them. 7.1 Parallelization Large-scale graphs and their need for parallel computing reinforce the importance of SBA, as these are extremely efficient parallel algorithms. The main challenge in the implementation of a graph processing algorithm lies in the high-dimensionality of graphs, which easily translates into data dependencies. Dependencies determine execution order constrains among portions of code and imply synchronization points, as one portion of code must wait for another portion to be executed first. Through the existence of dependencies, threads see their work flow halted as they wait for other threads. In essence, dependencies 16 define bottlenecks in the parallel execution of code, and reduce the efficiency of computational resources usage. A related concept within the field of parallel computing is that of embarrassingly parallel problems [14]. This notion applies to algorithms that can be parallelized without the definition of significant dependencies. These are therefore problems which can achieve a huge efficiency through parallelization, as there will be almost no idle resources in their computation. Embarrassingly parallel problems are capable of decreasing computational time almost linearly with the number of computing units available, as the various threads must not endure waiting times. As said in §2, one of the key features of similarity based LP algorithms is that the score of each edge can be calculated independently from the rest. This particularity gains a huge relevance now as it allow us to define LP as an embarrassingly parallel problem. Fully testing a LP algorithm on a graph equals to calculate the similarity of all possible ordered pairs of vertices (i.e., of each possible directed edge). Since each similarity can be calculated independently, we can evaluate them all simultaneously without dependencies. Considering the huge number of edges to test (at times in the order of billions), the code parallelization design defines the efficiency of the algorithm, and eventually, the size of the graph it can process. Our algorithmic design is divided into two parallel sections. On the first one we calculate the similarity of each possible edge in parallel, storing the results obtained for the edges originating on each vertex separately (in the n1 partial scores data structure). However, the score of each edge is not computed at once. Instead, evidence is accumulated as we find paths (n2 vertices) leading from a given target (n1 vertices) to a given goal (n3 vertices). An overview of this code can be seen in Algorithm 1. Notice that the iterations of the outermost loop (Line 3) can be computed in parallel without dependencies, with the only exception of the storage of results. Between the first and second part of the code we perform a reduction, combining the results obtained by all vertices. Since we are interested in evaluating the graph-wide performance, we need to know the number of true positive and false positive predictions achieved at every distinct threshold throughout the graph. This will allow us to simplify the second part of the code, and define it without dependencies. The second part of the code calculates the performance of LP algorithms. Given the total number of true positive and false positive predictions found at each distinct similarity value, we can calculate the points composing the curves discussed in §5 through an aggregation process. This task is also embarrassingly parallel, as the performance at each threshold (i.e., each point within the PR curves) can be calculated independently from the rest of thresholds. It is important to parallelize this task, as the number of distinct similarity values in large graphs can be also large (up to millions of values). An overview of this second section of code can be seen in Algorithm 2. Notice how the outermost loop can be parallelized with the only dependency of writing the results. 7.2 Computational Setting In this paper we computed graphs up to two million vertices. Graphs of this size are already challenging to process exhaustively due to the total number of 17 ALGORITHM 1: Code skeleton for similarity evaluation of all edges in a graph Input: A graph G. Output: Link prediction scores for all vertices in G, stored in a map structure. //Map to store all scores < sourceId, < targetId, score >> map < int, map < int, f loat >> graph scores; for vertex n1 in G do //Map to store partial scores < targetId, partial score > map < int, f loat > n1 partial scores; for vertex n2 neighbor of n1 do for vertex n3 neighbor of n2 do //If new target, initialize partial score if !n3 in n1 partial scores then n1 partial scores.insert(n3, initial score); end //Else, update partial score else n1 partial scores(n3).update; end end end //Score of edge n1 → x has been computed for all x in G graph scores.push back(n1 partial scores); end potential edges in the graph. Nevertheless, LP and the hyperlink prediction problem should target much larger graphs. Recent research is moving in that direction, as shows the work of [50] capable of training a factorization model for a graph with 50 million vertices. Regardless, the size of some of the most interesting graphs to process remains several orders of magnitude larger than that. For example, a webgraph from 2014 [34] covered 1.7 billion web pages connected by 64 billion hyperlinks. Large-scale graphs such as these represent the kind of problem for which SBA is the only feasible solution so far. A main concern when targeting large-scale graphs are computational resources. For medium sized graphs, such as the ones in §5.1, one can use a shared memory context, where all computational units have direct access to a centralized memory space. This approach assumes that the graph data can be stored into a single memory location. An assumption that can hardly be satisfied as graphs grow. Furthermore, even if we manage to store a large-scale graph into a single memory location, the number of computing units (i.e., cores) available for parallel execution will be physically constrained. To solve this limitation, HPC researchers use a distributed memory setting, where memory is split among several locations. In distributed memory, each memory location is directly accessible only by a subset of all computing units, but data from other locations can be fetched when necessary through communication channels. Using a distributed memory context entails several additional problems, such as how to distribute work among computational resources while achieving balance, and how to split input data among locations as to minimize communications. Problems that remain open issues, but that need to be addressed if we want to process large-scale graphs. See Figure 3 for a graphical representation of both paradigms. 18 ALGORITHM 2: Code skeleton for full graph performance evaluation Input: The list of all distinct similarities found in a graph, together with their corresponding true and false positive predictions. Output: List of performance rates, one per distinct similarity value //Structure to store the results obtained at all thresholds vector < pair < int, int >> f ull results; for similarity value sim1 in graph do int true pos sim1 = 0; int f alse pos sim1 = 0; for similarity value sim2 in graph do if sim2 >= sim1 then true pos sim1+ = true pos sim2; f alse pos sim1+ = f alse pos sim2; end end f ull results.push back(sim1, true pos sim1, f alse pos sim1); end Figure 3: On the left, structure of a shared memory architecture. On the right, structure of a distributed memory architecture. The code used for the tests presented in this paper was parallelized using the OpenMP shared memory model API [5]. This API provides a set of compiler directives that extend the C/C++ and FORTRAN base languages. OpenMP directives define how code is to be parallelized and how data is to be shared. We chose OpenMP because it is portable, scalable, flexible and the de-facto standard. Within the first section of code, the parallelization was done on the most external loop (Line 3 of Listing 1), effectively distributing its iterations among different threads. This design guarantees that all similarities n → X of a given vertex n (i.e., a full iteration of the outermost loop) are calculated by a single thread, thus avoiding any dependencies. The second section of the code was also parallelized on the most external loop (Line 3 of Listing 2). This design guarantees that each possible threshold (i.e., each point within the PR and ROC curves) was calculated by a unique thread, avoiding any dependencies. When parallelizing a loop through OpenMP one must decide how to distribute iterations among the team of running threads. Of the different ways of splitting iterations, we found that the most efficient for our problem was a dynamic scheduling, splitting iterations in chunks of pre-determined size which get 19 assigned to threads as these request it. It is key to define chunk sizes according to the problem size in order to minimize imbalances. If the chunk size is too large, at the end of the computation one or more threads will remain idle for a long time while the rest of threads finish their last chunks. If chunks are too small the constant scheduling of threads, as these finish chunks and requesting for more, may slow down the whole process. In the case of LP, the larger and denser the graph, the smaller the chunks must be, as iterations in a large graph will be more time consuming, thus increasing the possibilities of imbalance. For our code we found the optimal chunk size to be between 100 and 2000. For our tests we used the MareNostrum supercomputer, provided by BSC. We used a single Intel SandyBridge-EP E5-2670/1600 8-core at 2.6 GHz, with a limit of 28.1 GB of RAM. This translates into 8 parallel threads. The time spent for computing each graph is at most one hour and a half, as shown in Table 6. Those times include computing the four scores (CN, RA, AA and INF LOG 2D) for all the missing edges of each graph. Time does not include graph loading time, curve building time, and writing of results. 7.2.1 Distributed Memory Local similarity-based scores like the ones we evaluated here can compute graphs with a few million vertices in a shared memory environment. Eventually though, we will be interested in working with larger graphs for which a distributed memory is needed even by local methods. Consider for example the webgraph defined by Internet with 3.5 billion web pages, or a brain connectome graph composed by billions of neurons. For this kind of data sets the only feasible solution nowadays is distributed memory. And not only because of space requirement, but also because of the time complexity. Hundreds of cores computing in parallel will be needed to mine those graphs, and the number of computing cores accessing a single shared memory space is rarely over a few dozens. To run LP methods on a distributed memory environment using the same algorithmic design and parallelization described here we can use the OmpSs programming model [13], which supports OpenMP like directives and has been adapted to work on clusters with distributed memory [6]. Even though in the OmpSs version the graph data has to be communicated among computing entities (as the graph data is distributed among locations), our preliminary results show no relevant overhead added by this communication. This is so because in our LP algorithmic design it is easy to predict which edge will be evaluated next, and therefore which graph data will be needed next by each thread (i.e., which vertices and neighbours must be brought to memory). Thanks to this foreseeability, data can then sent before it is needed, thus avoiding idle threads and the consequent communication overhead. From a computational point of view this means that LP can scale almost linearly on distributed memory contexts. 8 Conclusions We have introduced a novel method which assumes the existence of hierarchical properties to improve the task of hyperlink prediction. Our first conclusion is that, according to the results shown in §5.1, the task of hyperlink prediction can be significantly improved through the consideration of hierarchical prop- 20 erties. In our tests the INF LOG 2D hierarchical score outperformed all nonhierarchical scores (CN, RA and AA) on six different webgraphs, doubling its AUPR measure in the worse case. The size and variety of the graphs used, and the thoroughness of the evaluation methodology guarantee the consistency of these results. Results that align with the previous work discussed in §3, providing further evidence on the importance of hierarchies for defining the topology of webgraphs. From a practical point of view, our main conclusion is that, through the leap in performance obtained shown here we can now predict thousands of hyperlinks in a webgraph with almost perfect precision, and in a scalable manner. As seen in Figure 2, INF LOG 2D achieves precisions close to 100% through its most certain predictions. According to our results, by considering hierarchies hyperlink prediction becomes a feasible problem. This immediately enables multiple interesting cases of application, including but not limited to: increase and improve the connectivity of web pages, optimize the navigability of web sites, tune search engines results through web similarity analysis, and refine product recommendation through item page linkage. From a general WWW perspective, this work continues the analysis of the relation between hierarchies and the WWW. A step in a different direction than most contributions, typically focused on defining generative models of real networks. Generative models and the problem of LP are strongly related, since generative models must produce new edges within a given graph. To illustrate that let us consider the FFM, briefly described in §3. The link adding process of FFM for out-going edges is in fact a particular case of the DED sub-score (see Figure 1). This alone indicates that INF LOG 2D and FFM are based on close hierarchical principles. There are nevertheless huge differences between them. FFM does not use a vertex in-edges to determine its out-edges, like INF LOG 2D does through the IND sub-score. INF LOG 2D calculates and rates all edges based on a similarity score, while the FFM randomly accepts and rejects edges as it does not seek faithfulness at a vertex level; it seeks topological coherency at a graph level. Finally, the FFM explores edges far away from the ambassador vertex though various iterations, thanks to its computational simplicity. INF LOG 2D performs only a one step exploration (i.e., its a local score), although building a quasi-local version of INF LOG 2D is one of our main lines of future work (see §9). To sum up, while the FFM and INF LOG 2D share a set of precepts, each model uses those for a different purpose: the FFM uses them to define a large, coherent topology model at graph scale, while INF LOG 2D uses them to define a high confidence and exhaustive edge likelihood score applicable at vertex level. In that regard, the good results achieved by both methods on their respective fields partly support the assumptions of the other. In this paper we also reached an interesting conclusion for the LP field in general. In our analysis of results we defined a simple categorization of SBA based on how these considered evidence: proportionally or accumulatively. Our proposed score INF LOG 2D is actually a hybrid, combining both approaches. While so far results indicate that accumulative solutions are more competitive than proportional ones, the good results of INF LOG 2D open the door to the consideration of hybrid solutions. This is of relevance for the LP field, as it motivates the integration of proportional features into current accumulative scores, with great potential benefits. From our code implementation and parallelization we derive conclusions for 21 the graph mining and HPC communities. As discussed in §5, SBA can be defined as parallel processes without dependencies. Thus, any amount of resources made available to SBA will be used efficiently. Significantly, this feature is consistent in a shared memory or distributed memory settings through the use of OpenMP and OmpSs. This opens the door to the computation of LP to graphs of arbitrary size. 9 Future Work From the LP perspective, the goal is to develop more precise prediction scores. Hybrid SBA are a promising family of LP methods that needs to be thoroughly studied. Also, the good performance of INF LOG 2D motivates the design of quasi-local scores based on it, which may achieve even more precise predictions. INF LOG 2D assumes a hierarchical structure, and thus works well on domains which satisfy this model to some degree. Other scores which assume (but do not compute) different underlying models (e.g., communities) should also be explored. Our final line of future work regards large-scale graphs. Our current implementation allows us to compute arbitrarily large graphs with OmpSs. Thanks to that we intend to develop an Internet-wide hyperlink recommender. For that purpose there are questions arising from a distributed memory context that must be considered, such as: How to split the graph data among different physical locations? Which data is to be allocated on each location? When must data be transfered? How is the code parallelized given these new restrictions? By solving this issues we intend to conclude the argument that we are starting with this work: that large-scale link prediction is not a field with a bright future, but instead one with a challenging present. Acknowledgements This work is partially supported by the Joint Study Agreement no. W156463 under the IBM/BSC Deep Learning Center agreement, by the Spanish Government through Programa Severo Ochoa (SEV-2015-0493), by the Spanish Ministry of Science and Technology through TIN2015-65316-P project and by the Generalitat de Catalunya (contracts 2014-SGR-1051). References [1] Sisay Fissaha Adafre and Maarten de Rijke. “Discovering missing links in Wikipedia”. In: Proceedings of the 3rd international workshop on Link discovery. ACM. 2005, pp. 90–97. [2] Lada A Adamic and Eytan Adar. “Friends and neighbors on the web”. In: Social networks 25.3 (2003), pp. 211–230. [3] Charu C. Aggarwal, Yan Xie, and Philip S. Yu. “A framework for dynamic link prediction in heterogeneous networks”. In: Statistical Analysis and Data Mining (2013). issn: 1932-1872. doi: 10.1002/sam.11198. 22 [4] Réka Albert, Hawoong Jeong, and Albert-László Barabási. “Internet: Diameter of the world-wide web”. In: Nature 401.6749 (1999), pp. 130–131. [5] OpenMP ARB. OpenMP Application Program Interface, v.4.0. http:// www.openmp.org/mp-documents/OpenMP4.0.0.pdf. 2013. [6] Javier Bueno et al. “Productive cluster programming with OmpSs”. In: Euro-Par 2011 Parallel Processing. Springer, 2011, pp. 555–566. [7] Aaron Clauset, Cristopher Moore, and Mark EJ Newman. “Hierarchical structure and the prediction of missing links in networks”. In: Nature 453.7191 (2008), pp. 98–101. [8] Diane J Cook et al. “Structural mining of molecular biology data”. In: Engineering in Medicine and Biology Magazine, IEEE 20.4 (2001), pp. 67– 74. [9] Mark E Crovella and Azer Bestavros. “Self-similarity in World Wide Web traffic: evidence and possible causes”. In: Networking, IEEE/ACM Transactions on 5.6 (1997), pp. 835–846. [10] Jesse Davis and Mark Goadrich. “The relationship between PrecisionRecall and ROC curves”. In: Proceedings of the 23rd international conference on Machine learning. ACM. 2006, pp. 233–240. [11] Stephen Dill et al. “Self-similarity in the web”. In: ACM Transactions on Internet Technology (TOIT) 2.3 (2002), pp. 205–223. [12] Enming Dong et al. “Bi-scale link prediction on networks”. In: Chaos, Solitons and Fractals 78 (2015), pp. 140–147. [13] Alejandro Duran et al. “OmpSs: A proposal for programming heterogeneous multi-core architectures”. In: Parallel Processing Letters 21.02 (2011), pp. 173–193. [14] Ian Foster. Designing and building parallel programs. Addison Wesley Publishing Company, 1995. [15] Dario Garcia-Gasulla and Ulises Cortés. “Link Prediction in Very Large Directed Graphs: Exploiting Hierarchical Properties in Parallel”. In: 3rd Workshop on Knowledge Discovery and Data Mining Meets Linked Open Data - 11th Extended Semantic Web Conference. 2014. [16] Dario Garcia-Gasulla et al. “Evaluating Link Prediction on Large Graphs”. In: Proceedings of the 18th International Conference of the Catalan Association of Artificial Intelligence. 2015. [17] Lise Getoor and Christopher P Diehl. “Link mining: a survey”. In: ACM SIGKDD Explorations Newsletter 7.2 (2005), pp. 3–12. [18] Lise Getoor and Ben Taskar. Introduction to statistical relational learning. MIT press, 2007. [19] Roger Guimerà and Marta Sales-Pardo. “Missing and spurious interactions and the reconstruction of complex networks”. In: Proceedings of the National Academy of Sciences 106.52 (2009), pp. 22073–22078. [20] Lawrence B Holder and Diane J Cook. “Graph-Based Data Mining.” In: Encyclopedia of data warehousing and mining 2 (2009), pp. 943–949. 23 [21] Akihiro Inokuchi, Takashi Washio, and Hiroshi Motoda. “An apriori-based algorithm for mining frequent substructures from graph data”. In: Principles of Data Mining and Knowledge Discovery. Springer, 2000, pp. 13– 23. [22] Brian Karrer and Mark EJ Newman. “Stochastic blockmodels and community structure in networks”. In: Physical Review E 83.1 (2011), p. 016107. [23] Leo Katz. “A new status index derived from sociometric analysis”. In: Psychometrika 18.1 (1953), pp. 39–43. [24] Ashraf Khalil and Yong Liu. “Experiments with PageRank computation”. In: Indiana University, Department Computer Science. URL: http://www. cs. indiana. edu/ akhalil/Papers/pageRank. pdf (2004). [25] Markus Krötzsch, Denny Vr Denny Vrandecic, and Max Völkel. “Wikipedia and the semantic web-the missing links”. In: Proceedings of Wikimania 2005. Citeseer. 2005. [26] Jérôme Kunegis. “KONECT: the Koblenz network collection”. In: Proceedings of the 22nd international conference on World Wide Web companion. International World Wide Web Conferences Steering Committee. 2013, pp. 1343–1350. [27] Jure Leskovec, Jon Kleinberg, and Christos Faloutsos. “Graphs over time: densification laws, shrinking diameters and possible explanations”. In: Proceedings of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining. ACM. 2005, pp. 177–187. [28] David Liben-Nowell and Jon Kleinberg. “The link-prediction problem for social networks”. In: Journal of the American society for information science and technology 58.7 (2007), pp. 1019–1031. [29] Ryan N Lichtenwalter, Jake T Lussier, and Nitesh V Chawla. “New perspectives and methods in link prediction”. In: Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM. 2010, pp. 243–252. [30] Weiping Liu and Linyuan Lü. “Link prediction based on local random walk”. In: EPL (Europhysics Letters) 89.5 (2010), p. 58007. [31] Linyuan Lü, Ci-Hang Jin, and Tao Zhou. “Similarity index based on local paths for link prediction of complex networks”. In: Physical Review E 80.4 (2009), p. 046122. [32] Linyuan Lü and Tao Zhou. “Link prediction in complex networks: A survey”. In: Physica A: Statistical Mechanics and its Applications 390.6 (2011), pp. 1150–1170. [33] Grzegorz Malewicz et al. “Pregel: a system for large-scale graph processing”. In: Proceedings of the 2010 ACM SIGMOD International Conference on Management of data. ACM. 2010, pp. 135–146. [34] Robert Meusel et al. “Graph Structure in the Web - Revisited”. In: 23rd International World Wide Web Conference (WWW2014). Seoul, Korea, Apr. 2014. 24 [35] Kurt Miller, Michael I. Jordan, and Thomas L. Griffiths. “Nonparametric Latent Feature Models for Link Prediction”. In: Advances in Neural Information Processing Systems 22. Ed. by Y. Bengio et al. Curran Associates, Inc., 2009, pp. 1276–1284. [36] David Milne and Ian H Witten. “Learning to link with wikipedia”. In: Proceedings of the 17th ACM conference on Information and knowledge management. ACM. 2008, pp. 509–518. [37] Tsuyoshi Murata and Sakiko Moriyasu. “Link prediction based on structural properties of online social networks”. In: New Generation Computing 26.3 (2008), pp. 245–257. [38] Tsuyoshi Murata and Sakiko Moriyasu. “Link prediction of social networks based on weighted proximity measures”. In: Web Intelligence, IEEE/WIC/ACM International Conference on. IEEE. 2007, pp. 85–88. [39] Mark EJ Newman. “Clustering and preferential attachment in growing networks”. In: Physical Review E 64.2 (2001), pp. 251021–251024. [40] Maximilian Nickel, Volker Tresp, and Hans-Peter Kriegel. “A three-way model for collective learning on multi-relational data”. In: Proceedings of the 28th international conference on machine learning (ICML-11). 2011, pp. 809–816. [41] Lawrence Page et al. “The PageRank citation ranking: bringing order to the web.” In: (1999). [42] Romualdo Pastor-Satorras, Alexei Vázquez, and Alessandro Vespignani. “Dynamical and correlation properties of the Internet”. In: Physical review letters 87.25 (2001), p. 258701. [43] Erzsébet Ravasz and Albert-László Barabási. “Hierarchical organization in complex networks”. In: Physical Review E 67.2 (2003), p. 026112. [44] Sid Redner. “Networks: teasing out the missing links”. In: Nature 453.7191 (2008), pp. 47–48. [45] Han Hee Song et al. “Scalable proximity estimation and link prediction in online social networks”. In: Proceedings of the 9th ACM SIGCOMM conference on Internet measurement conference. ACM. 2009, pp. 322–335. [46] Ilya Sutskever, Ruslan Salakhutdinov, and Joshua B Tenenbaum. “Modelling Relational Data using Bayesian Clustered Tensor Factorization.” In: NIPS. 2009, pp. 1821–1828. [47] Hanghang Tong, Christos Faloutsos, and Yehuda Koren. “Fast directionaware proximity for graph mining”. In: Proceedings of the 13th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM. 2007, pp. 747–756. [48] Koji Ueno and Toyotaro Suzumura. “Highly scalable graph search for the graph500 benchmark”. In: Proceedings of the 21st international symposium on High-Performance Parallel and Distributed Computing. ACM. 2012, pp. 149–160. [49] Yang Yang, Ryan N Lichtenwalter, and Nitesh V Chawla. “Evaluating link prediction methods”. In: Knowledge and Information Systems (2014), pp. 1–32. 25 [50] Hyokun Yun et al. “NOMAD: Non-locking, stOchastic Multi-machine algorithm for Asynchronous and Decentralized matrix completion”. In: arXiv preprint arXiv:1312.0193 (2013). [51] Tao Zhou, Linyuan Lü, and Yi-Cheng Zhang. “Predicting missing links via local information”. In: The European Physical Journal B 71.4 (2009), pp. 623–630. 26
8cs.DS
On a theorem of Blichfeldt Benjamin Sambale∗ arXiv:1606.02408v1 [math.GR] 8 Jun 2016 June 9, 2016 Abstract Let G be a permutation group on n < ∞ objects. Let f (g) be the number of fixed points of g ∈ G, and let {f (g) : 1 6= g ∈ G} = {f1 , . . . , fr }. In this expository note we give a character-free proof of a theorem of Blichfeldt which asserts that the order of G divides (n − f1 ) . . . (n − fr ). We also discuss the sharpness of this bound. Keywords: Blichfeldt’s Theorem, number of fixed points, permutation character AMS classification: 20B05 Let us consider a permutation group G on a finite set Ω consisting of n elements. By Lagrange’s Theorem applied to the symmetric group on Ω, it follows that the order |G| of G is a divisor of n!. In order to strengthen this divisibility relation we denote the number of fixed points of a subgroup H ≤ G on Ω by f (H). Moreover, let f (g) := f (hgi) for every g ∈ G. In 1895, Maillet [11] proved the following (see also Cameron’s book [4, p. 172]). Theorem 1 (Maillet). Let {f (H) : 1 6= H ≤ G} = {f1 , . . . , fr }. Then |G| divides (n − f1 ) . . . (n − fr ). Using the newly established character theory of finite groups, Blichfeldt [1] showed in 1904 that it suffices to consider cyclic subgroups H in Maillet’s Theorem (this was rediscovered by Kiyota [10]). Theorem 2 (Blichfeldt). Let {f (g) : 1 6= g ∈ G} = {f1 , . . . , fr }. Then |G| divides (n − f1 ) . . . (n − fr ). For the convenience of the reader we present the elegant argument which can be found in [4, Theorem 6.5]. Proof of Blichfeldt’s Theorem. Since f is the permutation character, the function ψ sending g ∈ G to (f (g) − f1 ) . . . (f (g) − fr ) is a generalized character of G (i. e. a difference of ordinary complex characters). From ( (n − f1 ) . . . (n − fr ) if g = 1, ψ(g) = 0 if g 6= 1 we conclude that ψ is a multiple of the regular character ρ of G. In particular, ρ(1) = |G| divides ψ(1) = (n − f1 ) . . . (n − fr ). It seems that no elementary proof (avoiding character theory) of Blichfeldt’s Theorem has been published so far. The aim of this note is to provide such a proof. Character-free proof of Blichfeldt’s Theorem. It suffices to show that 1 X (f (g) − f1 ) . . . (f (g) − fr ) ∈ Z, |G| g∈G ∗ Fachbereich Mathematik, TU Kaiserslautern, 67653 Kaiserslautern, Germany, [email protected] 1 since all summands with g 6= 1 vanish. Expanding the product we see that it is enough to prove Fk (G) := 1 X f (g)k ∈ Z |G| g∈G for k ≥ 0. Obviously, F0 (G) = 1. Arguing by induction on k we may assume that Fk−1 (H) ∈ Z for all H ≤ G. Let ∆1 , . . . , ∆s be the orbits of G on Ω, and let ωi ∈ ∆i for i = 1, . . . , s. For ω ∈ ∆i the stabilizers Gω and Gωi are conjugate in G. In particular, Fk−1 (Gω ) = Fk−1 (Gωi ). Recall that the orbit stabilizer theorem gives us |∆i | = |G : Gωi | for i = 1, . . . , s. This implies Fk (G) = = s 1 X X 1 X X 1 X |Gω |Fk−1 (Gω ) = f (g)k−1 = |Gω |Fk−1 (Gω ) |G| |G| |G| i=1 1 |G| ω∈Ω g∈Gω s X ω∈Ω |∆i ||Gωi |Fk−1 (Gωi ) = i=1 1 |G| ω∈∆i s X |G : Gωi ||Gωi |Fk−1 (Gωi ) = i=1 s X Fk−1 (Gωi ) ∈ Z. i=1 As a byproduct of the proof we observe that F1 (G) is the number of orbits of G. This is a well-known formula sometimes (inaccurately) called Burnside’s Lemma (see [13]). If there is only one orbit, the group is called transitive. In this case, F2 (G) is the rank of G, i. e. the number of orbits of any one-point stabilizer. It is known that Blichfeldt’s Theorem can be improved by considering only the fixed point numbers of non-trivial elements of prime power order. This can be seen as follows. Let Sp be a Sylow p-subgroup of G for every prime divisor p of |G|. Since {f (g) : 1 6= g ∈ Sp } ⊆ {f (g) : 1 6= g ∈ G has prime power order} = {f1 , . . . , fr }, Theorem 2 Q implies that |Sp | divides (n − f1 ) . . . (n − fr ) for every p. Since the orders |Sp | are pairwise coprime, also |G| = p |Sp | is a divisor of (n − f1 ) . . . (n − fr ). On the other hand, it does not suffice to take the fixed point numbers of the elements of prime order. An example is given by G = h(1, 2)(3, 4), (1, 3)(2, 4), (1, 2)(5, 6)i. This is a dihedral group of order 8 where every involution moves exactly four letters. Cameron-Kiyota [5] (and independently Chillag [6]) obtained another generalization of Theorem 2 where f is assumed to be any generalized character χ of G and n is replaced by its degree χ(1). A dual version for conjugacy classes instead of characters appeared in Chillag [7]. Numerous articles addressed the question of equality in Blichfeldt’s Theorem. Easy examples are given by the regular permutation groups. These are the transitive groups whose order coincides with the degree. In fact, by Cayley’s Theorem every finite group is a regular permutation group acting on itself by multiplication. A wider class of examples consists of the sharply k-transitive permutation groups G for 1 ≤ k ≤ n. Here, for every pair of tuples (α1 , . . . , αk ), (β1 , . . . , βk ) ∈ Ωk with αi 6= αj and βi 6= βj for all i 6= j there exists a unique g ∈ G such that αgi = βi for i = 1, . . . , k. Setting αi = βi for all i, we see that any non-trivial element of G fixes less than k points. Hence, {f (g) : 1 6= g ∈ G} ⊆ {0, 1, . . . , k − 1}. On the other hand, if (α1 , . . . , αk ) is fixed, then there are precisely n(n−1) . . . (n−k +1) choices for (β1 , . . . , βk ). It follows that |G| = n(n − 1) . . . (n − k + 1). Therefore, we have equality in Theorem 2. Note that sharply 1transitive and regular are the same thing. An interesting family of sharply 2-transitive groups comes from the affine groups Aff(1, pm ) = {ϕ : Fpm → Fpm | ∃a ∈ F× pm , b ∈ Fpm : ϕ(x) = ax + b ∀x ∈ Fpm } where Fpm is the field with pm elements. More generally, all sharply 2-transitive groups are Frobenius groups with abelian kernel. By definition, a Frobenius group G is transitive and satisfies {f (g) : 1 6= g ∈ G} = {0, 1}. The kernel K of G is the subset of fixed point free elements together with the identity. Frobenius Theorem asserts that K is a (normal) subgroup of G. For the sharply 2-transitive groups this can be proved in an elementary fashion (see [4, Exercise 1.16]), but so far no character-free proof of the full claim is known. The dihedral group h(1, 2, 3, 4, 5), (2, 5)(3, 4)i of order 10 illustrates that not every Frobenius group is sharply 2-transitive. A typical example of a sharply 3-transitive group is SL(2, 2m ) with its natural action on the set of onedimensional subspaces of F22m . We leave this claim as an exercise for the interested the reader. The sharply 2 k-transitive groups for k ∈ {2, 3} were eventually classified by Zassenhaus [16, 15] using near fields (see Passman’s book [14, Theorems 20.3 and 20.5]). On the other hand, there are not many sharply k-transitive groups when k is large. In fact, there is a classical theorem by Jordan [9] which was supplemented by Mathieu [12]. Theorem 3 (Jordan, Mathieu). The sharply k-transitive permutation groups with k ≥ 4 are given as follows: (i) the symmetric group of degree n ≥ 4 (k ∈ {n, n − 1}), (ii) the alternating group of degree n ≥ 6 (k = n − 2), (iii) the Mathieu group of degree 11 (k = 4), (iv) the Mathieu group of degree 12 (k = 5). We remark that the Mathieu groups of degree 11 and 12 are the smallest members of the sporadic simple groups. In accordance with these examples, permutation groups with equality in Theorem 2 are now called sharp permutation groups (this was coined by Ito-Kiyota [8]). Apart from the ones we have already seen, there are more examples. For instance, the symmetry group of a square acting on the four vertices has order 8 (again a dihedral group) and the non-trivial fixed point numbers are 0 and 2. Recently, Brozovic [3] gave a description of the primitive sharp permutation groups G such that {f (g) : 1 6= g ∈ G} = {0, k} for some k ≥ 1. Here, a permutation group is primitive if it is transitive and any one-point stabilizer is a maximal subgroup. The complete classification of the sharp permutation groups is widely open. Finally, we use the opportunity to mention a related result by Bochert [2] where the divisibility relation of |G| is replaced by an inequality. As usual ⌊x⌋ denotes the largest integer less than or equal to x ∈ R. Theorem 4 (Bochert). If G is primitive, then |G| ≤ n(n − 1) . . . (n − ⌊n/2⌋ + 1) unless G is the symmetric group or the alternating group of degree n. Acknowledgment This work is supported by the German Research Foundation (project SA 2864/1-1) and the Daimler and Benz Foundation (project 32-08/13). References [1] H. F. Blichfeldt, A theorem concerning the invariants of linear homogeneous groups, with some applications to substitution-groups, Trans. Amer. Math. Soc. 5 (1904), 461–466. [2] A. Bochert, Ueber die Zahl der verschiedenen Werthe, die eine Function gegebener Buchstaben durch Vertauschung derselben erlangen kann, Math. Ann. 33 (1889), 584–590. [3] D. P. Brozovic, The classification of primitive sharp permutation groups of type {0, k}, Comm. Algebra 42 (2014), 3028–3062. [4] P. J. Cameron, Permutation groups, London Mathematical Society Student Texts, Vol. 45, Cambridge University Press, Cambridge, 1999. [5] P. J. Cameron and M. Kiyota, Sharp characters of finite groups, J. Algebra 115 (1988), 125–143. [6] D. Chillag, Character values of finite groups as eigenvalues of nonnegative integer matrices, Proc. Amer. Math. Soc. 97 (1986), 565–567. [7] D. Chillag, On a congruence of Blichfeldt concerning the order of finite groups, Proc. Amer. Math. Soc. 136 (2008), 1961–1966. [8] T. Ito and M. Kiyota, Sharp permutation groups, J. Math. Soc. Japan 33 (1981), 435–444. 3 [9] C. Jordan, Sur la limite de transitivité des groupes non-alternées, Bull. Soc. Math. France 1 (1873), 40–71. [10] M. Kiyota, An inequality for finite permutation groups, J. Combin. Theory Ser. A 27 (1979), 119. [11] E. Maillet, Sur quelques propriétés des groupes de substitutions d’ordre donné, Ann. Fac. Sci. Toulouse Sci. Math. Sci. Phys. 9 (1895), 1–22. [12] É. Mathieu, Mémoire sur l’étude des fonctions de plusieurs quantités, sur la manière de les former et sur les substitutions qui les laissent invariables, J. Math. Pures Appl. 6 (1861), 241–323. [13] P. M. Neumann, A lemma that is not Burnside’s, Math. Sci. 4 (1979), 133–141. [14] D. S. Passman, Permutation groups, Dover Publications, Inc., Mineola, N.Y., 2012 (revised reprint of the 1968 original). [15] H. Zassenhaus, Kennzeichnung endlicher linearer Gruppen als Permutationsgruppen, Abh. Math. Sem. Univ. Hamburg 11 (1935), 17–40. [16] H. Zassenhaus, Über endliche Fastkörper, Abh. Math. Sem. Univ. Hamburg 11 (1935), 187–220. 4
4math.GR
Improved approximation algorithms for k-connected m-dominating set problems arXiv:1703.04230v1 [cs.DS] 13 Mar 2017 Zeev Nutov The Open University of Israel. [email protected]. Abstract. A graph is k-connected if it has k internally-disjoint paths between every pair of nodes. A subset S of nodes in a graph G is a kconnected set if the subgraph G[S] induced by S is k-connected; S is an m-dominating set if every v ∈ V \ S has at least m neighbors in S. If S is both k-connected and m-dominating then S is a k-connected m-dominating set, or (k, m)-cds for short. In the k-Connected mDominating Set ((k, m)-CDS) problem the goal is to find a minimum weight (k, m)-cds in a node-weighted graph. We consider the case m ≥ k and obtain the following approximation ratios. For unit disc-graphs we obtain ratio O(k ln k), improving the ratio O(k2 ln k) of [5,15]. For general graphs we obtain the first non-trivial approximation ratio O(k2 ln n). 1 Introduction A graph is k-connected if it has k internally disjoint paths between every pair of its nodes. A subset S of nodes in a graph G is a k-connected set if the subgraph G[S] induced by S is k-connected; S is an m-dominating set if every v ∈ V \ S has at least m neighbors in S. If S is both k-connected and m-dominating set then S is a k-connected m-dominating set, or (k, m)-cds for short. A graph is a unit-disk graph if its nodes can be located in in the Euclidean plane such that there is an edge between nodes u and v iff the Euclidean distance between u and v is at most 1. We consider the following problem for m ≥ k both in general graphs and in unit-disc graphs. k-Connected m-Dominating Set ((k, m)-CDS) Input: A graph G = (V, E) with node weights {wv : v ∈ V } and integers k, m. Output: A minimum weight (k, m)-cds S ⊆ V . The case k = 0 is the m-Dominating Set problem. Let αm denote the best known ratio for m-Dominating Set; currently αm = O(1) in unit-disc graphs [5] and αm = ln(∆ + m) + 1 < ln ∆ + 1.7 in general graphs [4], where ∆ is the maximum degree of the input graph. The (k, m)-CDS problem with m ≥ k was studied extensively. In recent papers Zhang, Zhou, Mo, and Du [15] and Fukunaga [5] obtained ratio O(k 2 ln k) for the problem in unit-disc graphs. For unit-disc graphs and k = 2 Zhang et al. [15] also obtained an improved ratio αm + 5. In a related paper Zhang et al. [16] obtained ratio O(k ln ∆) in general graphs with unit weights, mentionning that no non-trivial approximation algorithm for arbitrary weights is known. Let us say that a graph with a designated set T of terminals and a root node r is k-(T, r)-connected if it contains k internally-disjoint rt-paths for every t ∈ T . Our ratios for (k, m)-CDS are expressed in terms of αm and the best ratio for the following known problem: Rooted Subset k-Connectivity Input: A graph G = (V, E) with edge-costs/node-weights, a set T ⊆ V of terminals, a root node r ∈ V \ T , and an integer k. Output: A minimum cost/weight k-(T, r)-connected subgraph of G. Let βk and βk′ denote the best known ratios for the Rooted Subset kConnectivity problem with edge-costs and node-weights, respectively. Currently, βm = O(1) in unit-disc graphs [5], while in general graphs β2 = 2 [3], β3 = 6 32 [13], and βk = O(k ln k) for k ≥ 4 [11]. We also have βk′ = O(k 2 ln n) by [11] and the correction of Vakilian [14] to the algorithm and the analysis of [11]; see also [6]. Our main results are summarized in the following theorem. Theorem 1. Suppose that the m-Dominating Set problem admits ratio αm and that the Rooted Subset k-Connectivity problem admits ratios βk for edge-costs and βk′ for node-weights. Then (k, m)-CDS with m ≥ k admits ratios αm + βk′ + 2(k − 1) = O(k 2 ln n) for general graphs and αm + 5βk + 2(k − 1) = O(k ln k) for unit-disc graphs. Furthermore, (3, m)-CDS on unit-disc graphs admits ratio αm + 5β3 = αm + 33 31 . Our algorithm uses the main ideas as well as partial results from the papers of Zhang et al. [15] and Fukunaga [5]. Let us say that a graph G is k-T -connected if G contains k internally-disjoint paths between every pair of nodes in T . Both papers [15,5] consider unit-disc graphs and reduce the (k, m)-CDS problem with m ≥ k to the Subset k-Connectivity problem: given a graph with edge costs and a subset T of terminals, find a minimum cost k-T -connected subgraph. The problem admits a trivial ratio |T |2 for both edge-costs and node-weights, while | for |T | > k the best known ratios are |T|T|−k O(k ln k) = O(k 2 ln k) for edge-costs | and |T|T|−k O(k 2 ln n) = O(k 3 ln n) for node-weights [12]; see also [8]. In fact, these ratios are derived by applying O(k) times the algorithm for the Rooted Subset k-Connectivity problem. The main reason for our improvement over the ratios of [15,5] is a reduction to the easier Rooted Subset k-Connectivity problem. For small values of k we present a refined reduction, but for unit disc graphs and k = 2 the performance of our algorithm and that of [15] coincide, since for k = 2 and edge-costs both Subset k-Connectivity and Rooted Subset k-Connectivity admit ratio 2 [3]. 2 Proof of Theorem 1 For an arbitrary graph H = (U, F ) and u, v ∈ U let κH (u, v) denote the maximum number of internally disjoint uv-paths in H. We say that H is k-inconnected to r if H is k-(U \ {r}, r)-connected, namely, if κH (v, r) ≥ k every v ∈ U \ {r}. For A ⊆ U let ΓH (A) denote the set of neigbors of A in H. The proof of the following known statement can be found in [7], see also [1,2]; part (i) of the lemma relies on the Maders Undirected Critical Cycle Theorem [9]. Lemma 1. Let Hr be k-in-connected to r and let R = ΓHr (r). (i) The graph H = Hr \ {r} can be made k-connected by adding a set J of new edges on R; furthermore, if J is inclusionwise-minimal then J is a forest. (ii) Suppose that |R| = k. If k = 2, 3 then Hr is k-connected. Note that an inclusionwise-minimal edge set J as in Lemma 1(i) can be computed in polynomial time, by starting with J being a clique on R and repeatedly removing from J an edge e if H ∪ (J \ e) remains k-connected. A reason why the case m ≥ k is easier is given in the following lemma. Lemma 2. If a graph H = (V, E) has a k-dominating set T such that H is k-T -connected then H is k-connected. Proof. By a known characterization of k-connected graphs, it is sufficient to show that |V \ (A ∪ B)| ≥ k holds for any subpartition A, B of V such that E has no edge between A and B. If both A∩T, B ∩T are non-empty, this is so since H is k-T -connected. Otherwise, if say A ∩ T = ∅, then since T is a k-dominating set we have |ΓH (A)| ≥ k, and the result follows. ⊓ ⊔ Finally, we will need the following known fact, c.f. [11]. Lemma 3. Given a pair s, t of nodes in a node-weighted graph G, the problem of finding a minimum weight node set Pst such that G[Pst ] has k internallydisjoint st-paths admits a 2-approximation algorithm. For arbitrary k, we will show that the following algorithm achieves the desired approximation ratio. Algorithm 1: (G = (V, E), w, m ≥ k) 1 2 3 4 5 compute an αm -approximate m-dominating set T construct a graph Gr by adding to G a new node r connected to a set R ⊆ T of k nodes by a set Fr = {rv : v ∈ R} of new edges compute a βk′ -approximate node set S ⊆ V \ T such that the subgraph Hr of Gr induced by T ∪ S ∪ {r} is k-(T, r)-connected let H = H \ {r} = G[T ∪ S] and let J be a forest of new edges on R as in Lemma 1(i) such that the graph H ∪ J is k-connected for every uv ∈ J find a 2-approximate node set Puv such that [ G[T ∪ S ∪ Puv ] has k internally-disjoint uv-paths; let P = Puv uv∈J 6 return T ∪ S ∪ P We now prove that the solution computed is feasible. Lemma 4. The computed solution is feasible, namely, at the end of the algorithm T ∪ S ∪ P is a (k, m)-cds. Proof. Since T is an m-dominating set, so is any superset of T . Thus the node set T ∪ S ∪ P returned by the algorithm is an m-dominating set. It remains to prove that T ∪ S ∪ P is a k-connected set. We first prove that the graph Hr computed at step 3 is k-in-connected to r. By Menger’s Theorem, κHr (v, r) ≥ k iff for all A ⊆ T ∪ S with v ∈ A |ΓHr \R (A)| + |A ∩ R| ≥ k . (1) Let ∅ = 6 A ⊆ T ∪ S. If A ∩ T 6= ∅ then (1) holds since Hr is k-(T, r)-connected. If A ∩ S 6= ∅ then |ΓHr \R (A)| ≥ m ≥ k, since T is an m-dominating set and thus every node in A ∩ S has at least m neighbors in T . In both cases, (1) holds, hence Hr is k-in-connected to r. The graph H ∪ J is k-connected, which implies that the graph G[T ∪ S ∪ P ] is (T ∪ S)-k-connected and thus T -k-connected. Furthermore, T is a k-dominating set, since m ≥ k. Applying Lemma 2 on the graph G[T ∪ S ∪ P ] we get that this graphs is k-connected, as required. ⊓ ⊔ Lemma 5. Algorithm 1 has ratio αm + βk′ + 2(k − 1). Proof. Let S ∗ be an optimal solution to (k, m)-CDS. Clearly, w(T ) ≤ αm w(S ∗ ) ≤ βk′ w(S ∗ ). We claim that w(S) ≤ βk′ w(S ∗ \ T ). For this note that S ∗ \ T is a feasible solution to the problem considered at step 3 of the algorithm, while S is a βk′ -approximate solution. For the same reason, for each uv ∈ J the set S ∗ \(T ∪S) is a feasible solution to the problem considered at step 5, while the set Puv computed is a 2-approximate solution; thus w(Puv ) ≤ 2w(S ∗ \ (T ∪ S)). Finally, note that |J| ≤ k − 1, and thus w(P ) ≤ 2(k − 1)w(S ∗ ). The lemma follows. ⊓ ⊔ This concludes the proof of the case of general k and general graphs. Let us now consider unit disc graphs. Then we use the following result of [15]. Theorem 2 (Zhang, Zhou, Mo, and Du [15]). Any k-connected unit-disc graph has a k-connected spanning subgraph of maximum degree at most 5 if k = 2, and at most 5k if k ≥ 3. Note that any k-connected graph has minimum degree k. Thus Theorem 2 implies that when searching for a k-connected subgraph in a unit disc graph, one can convert node-weights to edge-costs while invoking in the ratio only a factor of 5/2 in the case k = 2 and 5 in the case k ≥ 3. Specifically, given node weights {wv : v ∈ V } define edge-costs cuv = wu + wv . Then for any subgraph (S, F ) of G with maximum degree ∆ and minimum degree δ we have: δw(S) ≤ c(F ) ≤ ∆w(S) since wv ≥ 0 for all v ∈ V and since X X c(F ) = (wu + wv ) = dF (v)wv . uv∈E v∈V We may use this conversion in some steps of our algorithm, and specifically in step 3, which concludes the proof of the case of general k and unit-disc graphs. In the case k = 3 we use a result of Mader [10] that any edge-minimal k-connected graph has at least (k−1)n+2 nodes of degree k. At step 3 of the 2k−1 algorithm we “guess” such a node r and the 3 edges incident to r in some edgeminimal optimal solution, remove from G all other edges incident to r, and run step 3 while omitting steps 4 and 5. By Lemma 1(ii) the graph G[S ∪ T ] will be already 3-connected. References 1. V. Auletta, Y. Dinitz, Z. Nutov, and D. Parente. A 2-approximation algorithm for finding an optimum 3-vertex-connected spanning subgraph. J. of Algorithms, 32(1):21–30, 1999. 2. Y. Dinitz and Z. Nutov. A 3-approximation algorithm for finding optimum 4, 5vertex-connected spanning subgraphs. J. of Algorithms, 32(1):31–40, 1999. 3. L. Fleischer, K. Jain, and D. Williamson. Iterative rounding 2-approximation algorithms for minimum-cost vertex connectivity problems. J. Computer and System Sciences, 72(5):838–867, 2006. 4. K.-T. Förster. Approximating fault-tolerant domination in general graphs. In ANALCO, pages 25–32, 2013. 5. T. Fukunaga. Constant-approximation algorithms for highly connected multidominating sets in unit disk graphs. arXiv:1511.09156[cs.DS],. 6. T. Fukunaga. Spider covers for prize-collecting network activation problem. In SODA, pages 9–24, 2015. 7. G. Kortsarz and Z. Nutov. Approximating node connectivity problems via set covers. Algorithmica, 37:75–92, 2003. 8. B. Laekhanukit. An improved approximation algorithm for minimum-cost subset k-connectivity. In ICALP, pages 13–24, 2011. 9. W. Mader. Ecken vom grad n in minimalen n-fach zusammenhängenden graphen. Archive der Mathematik, 23:219224, 1972. 10. W. Mader. On vertices of degree n in minimally n-connected graphs and digraphs. Combinatorics, Paul Erdös is eighty, 2:423449, 1993. 11. Z. Nutov. Approximating minimum cost connectivity problems via uncrossable bifamilies. ACM Transactions on Algorithms, 9(1):1, 2012. 12. Z. Nutov. Approximating subset k-connectivity problems. J. Discrete Algorithms, 17:5159, 2012. 13. Z. Nutov. Improved approximation algorithms for min-cost connectivity augmentation problems. In CSR, pages 324–339, 2016. 14. A. Vakilian. Node-weighted prize-collecting survivable network design problems. Master’s thesis, University of Illinois at Urbana-Champaign, 2013. 15. Z. Zhang, J. Zhou, Y. Mo, and D.-Z. Du. Approximation algorithm for minimum weight fault-tolerant virtual backbone in unit disk graphs. IEEE/ACM Transactions on networking, 2016. To appear. 16. Z. Zhang, J. Zhou, Y. Mo, and D.-Z. Du. Performance-guaranteed approximation algorithm for fault-tolerant connected dominating set in wireless networks. In INFOCOM, pages 1–8, 2016.
8cs.DS
Linking and Cutting Spanning Trees Luís M. S. Russoa,∗, Andreia Sofia Teixeiraa , Alexandre P. Franciscoa a INESC-ID and the Department of Computer Science and Engineering, Instituto Superior Técnico, Universidade de Lisboa. arXiv:1801.06846v1 [cs.DS] 21 Jan 2018 Abstract We consider the problem of uniformly generating a spanning tree, of a connected undirected graph. This process is useful to compute statistics, namely for phylogenetic trees. We describe a Markov chain for producing these trees. For cycle graphs we prove that this approach significantly outperforms existing algorithms. For general graphs we obtain no analytical bounds, but experimental results show that the chain still converges quickly. This yields an efficient algorithm, also due to the use of proper fast data structures. To bound the mixing time of the chain we describe a coupling, which we analyse for cycle graphs and simulate for other graphs. Keywords: Spanning Tree, Uniform Generation, Markov Chain, Mixing Time, Link Cut Tree PACS: 02.10.Ox, 02.50.Ga, 02.50.Ng, 02.70.Uu 2010 MSC: 05C81, 05C85, 60J10, 60J22, 65C40, 68R10 1. Introduction A spanning tree A of an undirected connected graph G is a tree, i.e., a connected set of edges without cycles, that spans every vertex of G. Every vertex of G occurs in some edge of A. Figure 1 shows an example. The vertexes of the graph are represented by A circles, the set of vertexes is denoted by V . The edges of G are represented by dashed lines, the set of edges is represented by E. The edges of the spanning tree A are represented by thick grey lines. We also use V and E to mean respectively the size of the set V and the size of set E, i.e., the number of vertexes and the number of edges. In case the expression can be interpreted as a Figure 1: A Spanning tree A over a graph G. ∗ Corresponding author. Preprint submitted to arXiv.org January 23, 2018 set, instead of a number, we avoid the ambiguity by writing |V | and |E|, respectively. We aim to compute one of such spanning trees A, uniformly among all possible spanning trees. The number of these trees may vary significantly, from 1 to V V −2 , depending on the underlying graph (Borchardt, 1860; Cayley, 1889; Aigner et al., 2010, Chapter 22). Computing such a tree uniformly and efficiently is challenging for several reasons: the number of such trees is usually exponential; the structure of the resulting trees is largely heterogeneous, as the underlying graphs change. The contributions of this paper are the following: • We present a new algorithm, which given a graph G, generates a spanning tree of G uniformly at random. The algorithm uses the link-cut tree data structure to compute randomizing operations in O(log V ) amortized time per operation. Hence, the overall algorithm takes O(τ log V ) time to obtain an uniform spanning tree of G, where τ is the mixing time of a Markov chain that is dependent on G. Theorem 1 summarizes this result. • We propose a coupling to bound the mixing time τ . The analysis of the coupling yields a bound for cycle graphs, Theorem 2, and for graphs which consists of simple cycles connect by bridges or articulation points, Theorem 3. We also simulate this procedure experimentally to obtain bounds for other graphs. The link-cut tree data structure is also key in this process. Section 4.3 shows experimental results, including other classes of graphs. The structure of the paper is as follows. In Section 2 we introduce the problem and explain its subtle nature. In Section 3 we explain our approach and point out that using the link cut tree data structure is much faster than repeating DFS searches. In Section 4 we thoroughly justify our results, proving that the underlying Markov chain has the necessary properties and providing experimental results of our algorithm. In Section 5 we describe the related work concerning random spanning trees, link cut trees and mixing time of Markov chains. In Section 6 we present our conclusions. 2. The Challenge We start by describing an intuitive process for generating spanning trees, that does not obtain a uniform distribution. Therefore it produces some trees with a higher probability than others. This serves to illustrate that the problem is harder than it may seem at first glance. Moreover we explain why this process is biased, using a counting argument. 2 A simple procedure to build A consists in using a Union-Find data structure (Galler and Fisher, 1964), to guarantee that A does not contain a cycle. Note that these structures are strictly incremental, meaning that they can be used to detect cycles but can not be used to remove an edge from the cycle. Therefore the only possible action is to discard the edge that creates the cycle. Let us analyse a concrete example of the resulting distribution of spanning trees. We shall show that this distribution is not uniform. First generate a permutation p of E and then process the edges in this order. Each edge that does not produce a cycle is added to A, edges that would otherwise produce cycles are discarded and the procedure continues with the next edge in the permutation. Consider the complete graph on 4 vertexes, K4 , and focus on the probability of generating a star graph, centered at the vertex la- 2 beled 1. Figure 2 illustrates the star graph. The K4 graph has 6 edges, hence there are 6! = 720 different permutations. To produce the star graph, from one such permutation, it is neces- 1 sary that the edges (1, 2) and (1, 3) are selected before the edge (2, 3) appears, in general the edges (1, u) and (1, v) must occur 3 4 before (u, v). One permutation that generates the star graph is (1, 2), (1, 3), (2, 3), (1, 4), (2, 4), (3, 4). Now (2, 3) can be moved to Figure 2: A star graph on K4 , centhe right to any of 3 different locations so we know 4 sequences tered at 1. that generate the star graph. The same reasoning can be applied to (2, 4) which can be moved once to the right. In total we counted 8 different sequences that generate the star graph, centered at 1. For each of these sequences it is possible to permute the vertexes 2, 3, 4, amongst themselves. Hence multiplying the previous count by 3! = 6. In total we counted 48 = 8 × 6 sequences that generate the star graph, therefore the total probability of obtaining a star graph is 48/6! = 1/15. According to Cayley’s formula the probability to obtain the star graph centered at 1 should be 1/42 = 1/16. Hence too many sequences are generating the star graph centered at 1. In the next section we fix this bias by discarding some edge in the potential cycle, not necessarily the edge that creates it. 3. Main Idea To generate a uniform spanning tree start by generating an arbitrary spanning tree A. One way to obtain this tree is to compute a depth first search in G, in which case the necessary time is O(V +E). 3 In general we wish that the mixing time of our chain is much smaller than O(E), specially for dense graphs. This initial tree is only generated once, subsequent trees are obtained by the randomizing process. To randomize A repeat the next process several times. Choose and edge (u, v) from E, uniformly at random, and consider the set A ∪ {(u, v)}. If (u, v) already belongs to A the process stops, otherwise A ∪ {(u, v)} contains a cycle C. To complete the process choose and edge (u′, v ′ ) uniformly from C \ {(u, v)} and remove it. Hence at each step the set A is transformed into the set (A ∪ {(u, v)}) \ {(u′ , v ′ )}. An illustration of this process is shown in Figure 3. This edge swapping process can be adequately modeled by a Markov chain, where the states corresponds to different spanning trees and the transitions among states correspond to the process A we have just described. In Section 4.1 we study the ergodic properties of this chain. For now let us focus on which data structures can be used to compute the transition procedure efficiently. A simple solution to this problem would be to compute a depth first search (DFS) on A, starting at u and terminating whenever v was reached. This would allow us to identify C in O(V ) time, recall C that A contains exactly V − 1 elements. The edge (u′ , v ′) could u then be easily removed. Besides G the elements of A would also need to be represented with the adjacency list data structure. For v our purposes this approach is inefficient. This computation is central to our algorithm and its complexity becomes a factor in the overall performance. Hence we will now explain how to perform this operation in only O(log V ) amortized time with the link cut u u′ tree data structure. The link cut tree (LCT) is a data structure that can used to represent a forest of rooted trees. The representation is dynamic so that edges can be removed and added. Whenever an edge is removed the v′ v (A ∪ {(u, v)}) \ {(u′ , v ′)} original tree is cut in two. Adding an edge between two trees links Figure 3: Edge swap procedure, inthem. This structure was proposed by Sleator and Tarjan (1985). serting the edge (u, v) into the iniBoth the link and cut operations can be computed in O(log V ) amortized time. 4 tial tree A generates a cycle C. The edge (u′ , v ′ ) is removed from C. Algorithm 1 Edge swapping process 1: procedure EdgeSwap(A) ⊲ A is an LCT representation of the current spanning tree 2: (u, v) ← Chosen uniformly from E 3: if (u, v) ∈ / A then ⊲ O(log V ) time ⊲ Makes u the root of A 4: ReRoot(A, u) 5: D ← Access(A, v) 6: i ← Chosen uniformly from {1, . . . , |D|} 7: (u′ , v ′ ) ← Select(D, i) 8: Cut(A, u′ , v ′) 9: Link(A, u, v) 10: 11: ⊲ Obtains a representation of the path C \ {(u, v)} ⊲ Obtain the i-th edge from D. end if end procedure The LCT can only represent trees, therefore the edge swap procedure must first cut the edge (u′, v ′ ) and afterwards insert the edge (u, v) with the Link operation. The randomizing process needs to identify C and select (u′ , v ′ ) from it. The LCT can also compute this process in O(log V ) amortized time. The LCT works by partitioning the represented tree into disjoint paths. Each path is stored in an auxiliary data structure, so that any of its edges can be accessed efficiently in O(log V ) amortized time. To compute this process we force the path D = C \ {(u, v)} to become a disjoint path. This means that D will be completely stored in one auxiliar data structure. Hence it is possible to efficiently select and edge from it. Moreover the size of D can also be computed efficiently. The exact process, to force D into an auxiliar structure, is to make u the root of the represented tree and then access v. Algorithm 1 shows the pseudo-code of the edge swapping procedure. We can confirm, by inspection, that this process can be computed in the O(log V ) amortized time bound that is crucial for our main result. Theorem 1. If G is a graph and A is a spanning tree of G then a spanning tree A′ can be chosen uniformly from all spanning trees of G in O((V + τ ) log V ) time, where τ is the mixing time of an ergodic edge swapping Markov chain. In section 4.1 we prove that the process we described is indeed an ergodic Markov chain, thus establishing the result. We finish this section by pointing out a detail in Algorithm 1. In the comment of line 3 we point out that the property (u, v) ∈ / A must be checked in at most O(log V ) 5 time. This can be achieved in O(1) time by keeping an array of booleans indexed by E. Moreover it can also be achieved in O(log V ) amortized time by using the LCT data structure, essentially by delaying the verification until D is determined and verifying if |D| = 6 1. 4. The details 4.1. Ergodic Analysis In this section, we analyse the Markov chain Mt induced by the edge swapping process. It should be clear that this process has the Markov property because the probability of reaching a state depends only on the previous state. In other words the next spanning tree depends only on the current tree. To prove that our procedure is correct we must show that the stationary distribution is uniform for all states. Let us first establish that such a stationary distribution exists. Note that, for a given finite graph G, the number of spanning trees is also finite. More precisely for complete graphs Cayley’s formula yields V V −2 spanning trees. This value is an upper bound for other graphs, as all spanning trees of a certain graph are also spanning trees of the complete graph with the same number of vertexes. Therefore the chain is finite. If we show that it is irreducible and aperiodic, it follows that it is ergodic (Mitzenmacher and Upfal, 2005, Corollary 7.6) and therefore it has a stationary distribution (Mitzenmacher and Upfal, 2005, Theorem 7.7). The chain is aperiodic because self-loops may occur, i.e., transitions where the underlying state does not change. Such transitions occur when (u, v) is already in A, therefore their probability is at least (V − 1)/E, because there are V − 1 edges in a spanning tree A. To establish that the chain is irreducible it is enough to show that for any pair of states i and j there is a non-zero probability path from i to j. First note that the probability of any transition on the chain is at least 1/(EV ), because (u, v) is chosen out of E elements and (u′ , v ′ ) is chosen from C \ {(u, v)}, that contains at most V − 1 edges. To obtain a path from i to j let Ai and Aj represent the respective trees. We consider the following cases: • If i = j we use a self-loop transition. • Otherwise, when i 6= j, it is possible to choose (u, v) from Aj \ Ai , and (u′ , v ′ ) from (C \ {(u, v)}) ∩ (Ai \ Aj ) = C \ Aj ; note that the set equality follows from the assumption that (u, v) belongs to Aj . For the last property note that if no such (u′, v ′ ) existed then C ⊆ Aj , which is a contradiction because Aj is a tree and C is a cycle. As mentioned above, the 6 probability of this transition is at least 1/(EV ). After this step the resulting tree is not necessarily Aj , but it is closer to that tree. More precisely (Ai ∪ {(u, v)}) \ {(u′, v ′ )} is not necessarily Aj , however the set Aj \ ((Ai ∪ {(u, v)}) \ {(u′ , v ′)}) is smaller than the original Aj \ Ai . Its size decreases by 1 because the edge (u, v) exists on the second set but not on the first. Therefore this process can be iterated until the resulting set is empty and therefore the resulting tree coincides with Aj . The maximal size of Aj \ Ai is V − 1, because the size of Aj is at most V − 1. This value occurs when Ai and Aj do no share edges. Multiplying all the probabilities in the process of transforming Ai into Aj we obtain a total probability of at least 1/(EV )V −1 . Now that the stationary distribution is guaranteed to exist, we will show that it coincides with the uniform distribution by proving that the chain is time reversible (Mitzenmacher and Upfal, 2005, Theorem 7.10). We prove that for any pair of states i and j, with j 6= i, for which there exists a transition from i to j, with probability Pi,j , there exists, necessarily, a transition from j to i with probability Pj,i = Pi,j . If the transition from i to j exists it means that there are edges (u, v) and (u′ , v ′) such that (Ai ∪ {(u, v)}) \ {(u′ , v ′)} = Aj , where (u′ , v ′ ) belongs to the cycle C contained in Ai ∪ {(u, v)}. Hence we also have that (Aj ∪ {(u′, v ′ )}) \ {(u, v)} = Ai , which means that the tree Ai can be obtained from the tree Aj by adding the edge (u′ , v ′ ) and removing the edge (u, v). In other words, the process in Figure 3 is similar both top down or bottom up. This process is a valid transition in the edge-swap chain, where the cycle C is the same in both transitions, i.e., C is the cycle contained in Ai ∪ {(u, v)} and in Aj ∪ {(u′ , v ′ )}. Now we obtain our result by observing that Pi,j = 1/(E(C − 1)) = Pj,i . In the transition from i to j the factor 1/E comes from the choice of (u, v) and the factor 1/(C − 1) from the choice of (u′ , v ′). In the transition between j to i, the factor 1/E comes from the choice of (u′ , v ′ ) and the factor 1/(C − 1) from the choice of (u, v). Hence we established that the algorithm we propose correctly generates spanning trees uniformly, provided we can sample from the stationary distribution. Hence, we need to determine the mixing time of the chain, i.e., the number of edge swap operations that need to be performed on an initial tree until the distribution of the resulting trees is close enough to the stationary distribution. Before analyzing the mixing time of this chain we point out that it is possible to use a faster version of this chain by choosing (u, v) uniformly from E \A, instead of from E. This makes the chain faster but proving that it is aperiodic is trickier. In this chain we have that Pr(Mt+1 = i|Mt = i) = 0, 7 for any state i. We will now prove that Pr(Mt+s = i|Mt = i) 6= 0, for any state i and s > 1. It is enough to show for s = 2 and s = 3, all other values follow from the fact that the greatest common divisor of 2 and 3 is 1. For the case of s = 2 we use the time reverse property and the following deduction: Pr(Mt+2 = i|Mt = i) ≥ Pi,j Pj,i ≥ (1/EV )2 > 0. For the case of s = 3 we observe that the cycle C must contain at least 3 edges (u, v), (u′, v ′ ) and (u′′ , v ′′ ). To obtain Aj we insert (u, v) and remove (u′, v ′ ), now we move from this state to state Ak by inserting (u′′ , v ′′ ) and removing (u, v). Finally we move back to Ai by inserting (u′, v ′ ) and removing (u′′ , v ′′ ). Hence, for this case we have Pr(Mt+3 = i|Mt = i) ≥ (1/EV )3 > 0 4.2. A Coupling In this section, we focus on bounding the mixing time. We did not obtain general analytical bounds from existing analysis techniques, such as couplings (Levin and Peres, 2017; Mitzenmacher and Upfal, 2005), strong stopping times (Levin and Peres, 2017) and canonical paths (Sinclair, 1992). The coupling technique yielded a bound only for cycle graphs and moreover a simulation of the resulting coupling converges for ladder graphs. Before diving into the reasoning in this section, we first need a finer understanding of the cycles generated in our process. We consider a closed walk to be a sequence of vertexes v0 , . . . , vn = v0 , starting and ending at the same vertex, and such that any two consecutive vertexes vi and vi+1 are adjacent, in our case (vi , vi+1 ) ∈ A ∪ {(u, v)}. The cycles we consider are simple, in the sense that they consist of a set of edges for which a closed walk can be formed, that traverses all the edges in the cycle and moreover no vertex repetitions are allowed, except for the vertex v0 , which is only repeated at the end. Formally this can be stated as: if 0 ≤ i, j < n and i 6= j then vi 6= vj . The cycles that occur in our randomizing process are even more regular. A cordless cycle in a graph is a cycle such that no two vertices of the cycle are connected by an edge that does not itself belong to the cycle. The cycles we produce also have this property, otherwise if such a chord existed then it would form a cycle on our tree A, which is a contradiction. In fact a spanning tree over a graph can alternatively be defined as a set of edges such that for any pair of vertexes v and v ′ there is exactly one path linking v to v ′ . A coupling is an association between two copies of the same Markov chain Xt and Yt , in our case the edge swapping chain. The goal of a coupling is to make the two chains meet as fast as possible, i.e., obtain Xτ = Yτ , for a small value of τ . At this point we say that the chains have coalesced. 8 The two chains may share information and cooperate towards this goal. However, when analysed in isolation, each chain must be indistinguishable from the original chain Mt . Obtaining Xτ = Yτ with a high probability implies that at time τ the chain is well mixed. Precise statements of these claims are given in Section 5. We use the random variable Xt to represent the state of the first chain, at time t. The variable Yt represents the state of the second chain. We consider the chain Xt in state x and the chain Yt in state y. In one step the chain Xt will transition to the state x′ = Xt+1 and the chain Yt will transition to state y ′ = Yt+1 . The set Ax \ Ay contains the edges that are exclusive to Ax and likewise the set Ay \ Ax contains the edges that are exclusive to Ay . The number of such edges provides a distance d(x, y) = |Ax \ Ay | = |Ay \ Ax |, that measures how far apart are the two states. We refer to this distance as the edge distance. We define a coupling when d(x, y) ≤ 1, which can be extended for states x and y that are farther apart, by using the path coupling technique (Bubley and Dyer, 1997). To use the path coupling technique we cannot alter the behavior of the chain Xt as, in general, it is determined by the previous element in the path. We denote by ix the edge that gets added to Ax , and by ox the edge that gets removed from the corresponding cycle Cx ⊆ Ax ∪ {ix }, in case such a cycle exists. Likewise, iy represents the edge that is inserted into Ay and oy the edge that gets removed from the corresponding cycle Cy ⊆ Ay ∪ {iy }, in case such a cycle exists. The edge ix is chosen uniformly at random from E and ox is chosen uniformly at random from Cx \ {ix }. The edges iy and oy will be obtained by trying to mimic the chain Xt , but still exhibiting the same behavior as Mt . In this sense the information flows from Xt to Yt . Let us now analyse d(x, y). 4.2.1. d(x, y) = 0 If d(x, y) = 0 then x = y, which means that the corresponding trees are also equal, Ay = Ax . In this case Yt uses the same transition as Xt , by inserting ix , i.e., set iy = ix , and removing ox , i.e., set oy = ox . 4.2.2. d(x, y) = 1 If d(x, y) = 1 then the edges ex ∈ Ax \ Ay and ey ∈ Ay \ Ax exist and are distinct. We also need the following sets: I = Cx ∩ Cy , Ex = Cx \ I and Ey = Cy \ I. The set I represents the edges that are common to Cx and Cy . The set Ex represents the edges that are exclusive to Cx , from the cycle point of view. This should not be confused with ex which represents the edge that is exclusive to 9 ex ex ∈ Ex ex ix ∈ I ix ix ey ey ey ∈ Ey Figure 4: Schematic representation of the relations between Ex , Ey and I. Ax , i.e., from a tree point of view. Likewise, Ey represents the edges that are exclusive to Cy . Also we consider the cycle Ce as the cycle contained in Ax ∪ {ey }, which necessarily contains ex . The following Lemma describes the precise structure of these sets. Lemma 1. When iy = ix we either have Cx = Cy = I and therefore Ex = Ey = ∅ or Ex , Ey and I form simple paths and the following properties hold: • ex ∈ Ex , ey ∈ Ey , ix ∈ I • Ex ∩ Ey = ∅, Ex ∩ I = ∅, Ey ∩ I = ∅ • Ex ∪ I = Cx , Ey ∪ I = Cy , Ex ∪ Ey = Ce . Notice that in particular this means that, in the non-trivial case, Ex and Ey partition Ce . A schematic representation of this Lemma is shown in Figure 4. We have several different cases described below. Aside from the lucky cases 2 and 3, we will usually choose iy = ix , as Yt tries to copy Xt . Likewise, if possible, we would like to set oy = ox . When this is not possible we must choose oy ∈ Ey , ideally we would choose oy = ey , but we must be extra careful with this process to avoid loosing the behavior of Mt . To maintain this behaviour, we must sometimes choose oy ∈ Ey \ {ey }. Since Xt provides no information on this type of edges we use oy = sy chosen uniformly from this Ey \ {ey }, i.e., select from Cy but not ey nor edges that are also in Cx . There is a final twist to this choice, which makes the coupling non Markovian, i.e., it does not verify the conditions in Equations (1a) and (1b). We can choose oy = ey more often than would otherwise be permissible, by keeping track of how ey was determined. If ey was obtained deterministically, for example by the initial selection of x and y, then this is not possible. In general ey might 10 x y ix = ey ↓ oy = ey ↑ ox = ex ↓ iy = ex ↑ x′ = y y′ = x ex′ ey′ Figure 5: Case 2. be determined by the changes in Xt , in which case we want to take advantage of the underlying randomness. Therefore, we keep track of the random processes that occur. The exact information we store is a set of edges Uy ⊆ Ce \ {ex } such that ey ∈ Uy and moreover this set contains the edges that are equally likely to be ey . This information can be used to set oy = ey when sy ∈ Uy , however after such an action the information on Uy must be purged. To illustrate the possible cases we use Figures 5 to 15, where the edges drawn with double lines represent a generic path, that may contain several edges, or none at all. The precise cases are the following: 1. If the chain Xt loops (x′ = x), because ix ∈ Ax then Yt also loops and therefore y ′ = y. The set Uy does not change, i.e., set Uy′ = Uy . 2. If ix = ey and ox = ex then set iy = ex and oy = ey . In this case the chains do not coalesce, they swap states because x′ = y and y ′ = x, see Figure 5. Set Uy′ = Ce \ {ex′ }. 3. If ix = ey and ox 6= ex then set iy = ex and oy = ox . In this case the chains coalesce, i.e., x′ = y ′ , see Figure 6. When the chains coalesce the edges ex′ and ey′ no longer exist and the set Uy′ is no longer relevant. 11 y x ix = ey ↓ ey ex ↓ iy = ex oy = ox ↓ ox ↓ x′ = y ′ Figure 6: Case 3. 4. If ix 6= ey set iy = ix . We now have 3 sub-cases, which are further sub-divided. These cases depend on whether |Cx | = |Cy |, |Cx | < |Cy | or |Cx | > |Cy |. We start with |Cx | = |Cy | which is simpler and establishes the basic situations. When |Cx | < |Cy | or |Cx | > |Cy | we use some Bernoulli random variables to balance out probabilities and whenever possible reduce to the cases considered for |Cx | = |Cy |. When this is not possible we present the corresponding new situation. a) If |Cx | = |Cy | we have the following situations: i) If ox = ex then set oy = ey . In this case the chains coalesce, see Figure 7. ii) If ox ∈ I \ {ix } then set oy = ox . In this case the chains do not coalesce, in fact the exclusive edges remain unchanged, i.e., ex′ = ex and ey′ = ey , see Figures 8 and 9. When ox ∈ / Ce the set Ce′ remains equal to Ce and likewise Uy′ remains equal to Uy , see Figure 8. Otherwise when ox ∈ Ce the set Ce′ is different from Ce and we assign Uy′ = Uy ∩ Ce′ , see Figure 9. iii) If ox ∈ Ex \ {ex } then select sy uniformly from Ey \ {ey }. If sy ∈ Uy then set oy = ey , / Uy is see Figure 10. In this case set Uy′ = Ex \ {ex }. The alternative, when sy ∈ considered in the next case (4.a.iv). 12 x y ix ↓ iy = ix ↓ ox = ex ↓ oy = ey ↓ x′ = y ′ Figure 7: Case 4.a.i, case 4.b.i and case 4.c.i. y x ↑ ox ix ↓ ↑ oy = ox iy = ix ↓ ey ex y′ x′ ey′ = ey ex′ = ex Figure 8: Case 4.a.ii, case 4.b.ii and case 4.c.ii, when ox ∈ / Ce . 13 y x ey ex ix ↓ iy = ix ↓ ↓ ox ↓ oy = ox x′ y′ ey′ = ey ex′ = ex Figure 9: Case 4.a.ii, case 4.b.ii and case 4.c.ii, when ox ∈ Ce . y x ox ↑ ex iy = ix ↓ ix ↓ oy = ey ↓ y′ x′ ex′ = ex ey′ Figure 10: Case 4.a.iii, case 4.b.iv and case 4.c.iv, when B ′ is true. 14 y x ox ↑ ex iy = ix ↓ ix ↓ ey x′ ex′ = ex ↓ o y = sy y′ z′ ez ′ ey′ = ey Figure 11: Case 4.a.iv, case 4.b.iv and case 4.c.iv. iv) If ox ∈ Ex \ {ex } and sy ∈ / Uy , then set oy = sy . This case is shown in Figure 11. In this case the distance of the coupled states increases, i.e., d(x′ , y ′) = 2. Therefore we include a new state z ′ , in between x′ and y ′ and define ez ′ to be the edge in Az ′ \ Ax′ ; and ey′ the edge in Ay′ \ Az ′ ; and ex′ the edge in Ax′ \ Az ′ . The set Uz ′ should contain the edges that provide alternatives to ez ′ . In this case set Uz ′ = Ex \ {ex } and Uy′ = (Uy ∩ Ey ) \ {oy }. b) If |Cx | < |Cy | then Xt will choose ox ∈ I with a higher probability then what Yt should. Therefore we use a Bernoulli random variable B with a success probability p defined as follows: p= Cx − 1 Cy − 1 In Lemma 2 we prove that p properly balances the necessary probabilities, for now note that when |Cx | = |Cy | the expression for p yields p = 1. This is coherent with the following cases, because when B yields true we use the choices defined for |Cx | = |Cy |. The following situations are possible: i) If ox = ex then we reduce to the case 4.a.i, both when B yields true or when B 15 y x ey o y = sy ↑ iy = ix ↓ ix ↓ ox = ex ↓ y′ x′ ey′ = ey ex′ Figure 12: Case 4.b.i when B fails and sy ∈ / Uy . fails and sy ∈ Uy . Set oy = ey , see Figure 7. The new case occurs when B fails and sy ∈ / Uy , in this situation set oy = sy and Uy′ = (Uy ∩ Cy ) \ {oy }, see Figure 12. ii) If ox ∈ I \ {ix } then we reduce to the case 4.a.ii when B yields true. Set oy = ox , see Figures 8 and 9. When B fails and sy ∈ Uy we have a new situation. Set oy = ey and Uy′ = I \ {ix }. The chains preserve their distance, i.e., d(x′ , y ′) = 1, see Figure 13. The alternative, when sy ∈ / Uy is considered in the next case (4.b.iii). iii) If ox ∈ I \ {ix } and B fails and sy ∈ / Uy . We have a new situation, set oy = sy . The distance increases, d(x′ , y ′) = 2, see Figure 14. Set Uz ′ = I \ {ix } and Uy′ = (Uy ∩ Ey ) \ {oy }. iv) If ox ∈ Ex \ {ex } then if sy ∈ Uy use case 4.a.iii (Figure 10), otherwise, when sy ∈ / Uy , use case 4.a.iv (Figure 11). c) If |Cx | > |Cy | we have the following situations: i) If ox = ex then use case 4.a.i and set oy = ey , see Figure 7. The chains coalesce. ii) If ox ∈ I \ {ix } then use case 4.a.ii and set oy = ox , see Figures 8 and 9. 16 y x ↑ ox ix ↓ iy = ix ↓ ex oy = ey ↓ y′ x′ ey′ ex′ = ex Figure 13: Case 4.b.ii. y x iy = ix ↓ ↑ ox ix ↓ ex ey o y = sy ↓ x′ y′ z′ ez ′ ex′ = ex ey′ = ey Figure 14: Case 4.b.iii. 17 y x iy = ix ↓ ix ↓ ↑ o y = si ey ex ox ↓ x′ y′ z′ ex′ ey′ = ey ez ′ Figure 15: Case 4.c.iii, when B ∗ is true. iii) If ox ∈ Ex \ {ex } then we use a new Bernoulli random variable B ∗ with a success probability p∗ defined as follows:   (Cx − 1)(I − 1) 1 1 ∗ × − p = Cy − 1 Cx − 1 Ex − 1 In Lemma 2 we prove that B ∗ properly balances the necessary probabilities. For now, note that when |Cx | = |Cy | the expression for p∗ yields p∗ = 0, because 1/(Cy − 1) − 1/(Cx − 1) becomes 0. This is coherent because when B ∗ returns false we will use the choices defined for |Cx | = |Cy |. The case when B ∗ fails is considered in the next case (4.c.iv). If B ∗ is successful we have a new situation. Set oy = si , where si is chosen uniformly from I \ {iy } and see Figure 15. We have ey′ = ey , Uy′ = (Uy ∩ Ey ) \ {oy }, ez ′ = ox and Uz ′ = Ex \ {ex }. iv) If ox ∈ Ex \ {ex } and B ∗ fails we use another Bernoulli random variable B ′ with a success probability p′ defined as follows: p′ = 1 − (Cx − 1)(Ey − 1) (Cy − 1)(Ex − 1)(1 − p∗ ) 18 In Lemma 2 we prove that B ′ properly balances the necessary probabilities. In case B ′ yields true, use case 4.a.iii and set oy = ey , see Figure 10. Otherwise, if sy ∈ Uy , use case 4.a.iii (Figure 10) or, if sy ∈ / Uy , use case 4.a.iv (Figure 11). Notice that the case 4.a.ii applies when Cx = Cy , thus solving this situation as a particular case. This case is shown in Figure 8. It may even be the case that Cx = Cy and Cx and Ce are disjoint, i.e., Cx ∩ Ce = ∅. This case is not drawn. Formally, a coupling is Markovian when Equations (1a) and (1b) hold, where Zt is the coupling, which is defined as a pair of chains (Xt , Yt ). The chain Mt represents the original chain. Pr(Xt+1 = x′ | Zt = (x, y)) = Pr(Mt+1 = x′ | Mt = x) (1a) Pr(Yt+1 = y ′ | Zt = (x, y)) = Pr(Mt+1 = y ′ | Mt = y) (1b) To establish vital insight into the coupling structure we will start by studying it when it is Markovian. Lemma 2. When Uy = {ey } the process we described is a Markovian coupling. Proof. The coupling verifies Equation (1a), because we do not alter the behavior of the chain Xt . Hence the main part of the proof focus on Equation (1b). First let us prove that for any edge i ∈ E the probability that iy = i is 1/E, i.e., Pr(iy = i) = 1/E. The possibilities for iy are the following: • i ∈ Ay , this occurs only in case 1, when ix ∈ Ax . It may be that i = ey , this occurs when ix = ex in which case iy = ey = i and this is the only case where iy = ey . In this case Pr(iy = i) = Pr(ix = ex ) = 1/E. Otherwise i ∈ Ay ∩ Ax , in these cases iy = ix and therefore Pr(iy = i) = Pr(ix = i) = 1/E. • i = ex , this occurs in cases 2 and 3, i.e., when ix = ey , which is the decisive condition for this choice. Therefore Pr(iy = i) = Pr(ix = ey ) = 1/E. • i ∈ E \ Ay , this occurs in case 4. In this case iy = ix so again we have that Pr(iy = i) = Pr(ix = i) = 1/E. Before focusing on ox we will prove that the Bernoulli random variables are well defined, i.e., that the expression on the denominators are not 0 and that the values of p, p∗ , p′ are between 0 and 1. 19 • Analysis of B. We need to have Cy − 1 6= 0 for p to be well defined. Any cycle must contain at least 3 edges, therefore 3 ≤ Cy and hence 0 < 2 ≤ Cy − 1. This guarantees that the denominator is not 0. The same argument proves that 0 < Cx − 1, thus implying that 0 < p, as both expressions are positive. We also establish that p < 1 because of the hypothesis of case 4.b which guarantees Cx < Cy and therefore Cx − 1 < Cy − 1. • Analysis of B ∗ . As in seen the analysis of B we have that 0 < Cy − 1 and 0 < Cx − 1, therefore those denominators are not 0. Moreover we also need to prove that Ex − 1 6= 0. In general we have that 1 ≤ Ey , because ey ∈ Ey . Moreover, the hypothesis of case 4.c.iii is that Cy < Cx and therefore Ey < Ex , obtained by removing I from the both sides. This implies that 1 < Ex and therefore 0 < Ex − 1, thus establishing that the last denominator is also not 0. Let us now establish that 0 ≤ p∗ and p∗ < 1. Note that p∗ can be simplified to the expression (Cx − Cy )(I − 1)/((Cy − 1)(Ex − 1)), where all the expressions in parenthesis are non-negative, so 0 ≤ p∗ . For the second property we use the new expression for p∗ and simplify p∗ < 1 to (Ex −Ey )(I −1) < (Ex −1)(Cy −1). The deduction is straightforward using the equality Cx − Cy = Ex − Ey that is obtained by removing I from the left side. The properties Ex − Ey ≤ Ex − 1 and I − 1 < Cy − 1 establish the desired result. • Analysis of B ′ . We established, in the analysis of B, that Cy − 1 is non-zero. In the analysis of B ∗ we also established that Ex − 1 is non-zero, note that case 4.c.iv also assumes the hypothesis that Cy < Cx . Moreover in the analysis of B ∗ we also established that p∗ < 1, which implies that 0 < 1 − p∗ and therefore the last denominator is also non-zero. Let us also establish that 0 ≤ p′ and p′ ≤ 1. For the second property we instead prove that 0 ≤ 1 − p′ , where 1 − p′ = (Cx − 1)(Ey − 1)/((Cy − 1)(Ex − 1)(1 − p∗ )) and all of the expressions in parenthesis are non-negative. We use the following deduction of equivalent inequalities to 20 establish that 0 ≤ p′ : 0 ≤ p′ −p′ ≤ 0 1 − p′ ≤ 1 (Cx − 1)(Ey − 1) ≤ (Cy − 1)(Ex − 1)(1 − p∗ )   (Cx − Cy )(I − 1) (Cx − 1)(Ey − 1) ≤ (Cy − 1)(Ex − 1) 1 − (Cy − 1)(Ex − 1) (Cx − 1)(Ey − 1) ≤ (Cy − 1)(Ex − 1) − (Cx − Cy )(I − 1) (Ex − 1)(Ey − 1) + I(Ey − 1) ≤ (Ey − 1)(Ex − 1) + I(Ex − 1) − (Ex − Ey )(I − 1) I(Ey − 1) ≤ I(Ex − 1) − (Ex − Ey )(I − 1) I((Ey − 1) + Ex − Ey ) ≤ I(Ex − 1) + Ex − Ey I(Ex − 1) + Ey ≤ I(Ex − 1) + Ex Ey ≤ Ex Cy ≤ Cx This last inequality is part of the hypothesis of case 4.c.iv. Now let us focus on the edge ox . We wish to establish that for any o ∈ Cy \ {iy } we have that Pr(oy = o) = 1/(Cy − 1). We analyse this edge according to the following cases: 1. When the cycles are equal Cx = Cy . This involves cases 2 and 3. • o = ey , this occurs only in case 2 and it is determined by the fact that ox = ex , therefore Pr(oy = o) = Pr(ox = ex ) = 1/(Cx − 1) = 1/(Cy − 1). • o 6= ey , this occurs only in case 3 and it is determined by the fact that ox 6= ex , in this case oy = ox . Therefore Pr(oy = o) = Pr(ox = o) = 1/(Cx − 1) = 1/(Cy − 1). 2. When the cycles have the same size |Cx | = |Cy |, case 4.a. The possibilities for o are the following: • o = ey , this occurs only in the case 4.a.i. This case is determined by the fact that ox = ex . Therefore Pr(oy = o) = Pr(ox = ex ) = 1/(Cx − 1) = 1/(Cy − 1). Note that according to the Lemma’s hypothesis, case 4.a.iii never occurs. 21 • o ∈ I \ {iy }, this occurs only in case 4.a.ii. This case is determined by the fact that ox ∈ I \ {ix } and sets oy = ox = o. Therefore Pr(oy = o) = Pr(ox = o) = 1/(Cx − 1) = 1/(Cy − 1). • o ∈ Ey \ {ey }, this occurs only in case 4.a.iv. This case is determined by the fact that ox ∈ X \ {ex } and moreover sets oy = sy , which was uniformly selected from Ey \ {ey }. We have the following deduction where we use the fact that the events are independent and that |Cx | = |Cy | implies |Ex | = |Ey |: Pr(oy = o) = Pr(ox ∈ Ex \ {ex } and sy = o) = Pr(ox ∈ Ex \ {ex }) Pr(sy = o) = 1 Ex − 1 × Cx − 1 Ey − 1 = 1/(Cx − 1) = 1/(Cy − 1) 3. When Cx < Cy this involves case 4.b. The cases for o are the following: • o = ey , this occurs only in the case 4.b.i and when B is true. This case occurrs when ox = ex . We make the following deduction, that uses the fact that the events are independent and the success probability of B: Pr(oy = o) = Pr(ox = ex and B = true) = Pr(ox = ex ) Pr(B = true) = 1 Cx − 1 × Cx − 1 Cy − 1 = 1/(Cy − 1) • o ∈ I \ {iy }, this occurs only in case 4.b.ii and when B is true. This case is determined by the fact that ox ∈ I \ {ix } and sets oy = ox = o. We make the following deduction, 22 that uses the fact that the events are independent and the success probability of B: Pr(oy = o) = Pr(ox = o and B = true) = Pr(ox = o) Pr(B = true) = Cx − 1 1 × Cx − 1 Cy − 1 = 1/(Cy − 1) • o ∈ Ey \ {ey }, this occurs in case 4.b.iv, but also in cases 4.b.iii and 4.b.i when B is false. We have the following deduction, that uses event independence, the fact that the cases are disjoint events and the succes probability of B: Pr(oy = o) = Pr(4.b.iv or (4.b.iii and B = false) or (4.b.i and B = false)) = Pr(4.b.iv) + Pr(4.b.iii and B = false) + Pr(4.b.i and B = false) = Pr(ox ∈ Ex \ {ex } and sy = o) + Pr(ox ∈ I and B = false and sy = o) + Pr(ox = ex and B = false and sy = o) = Pr(ox ∈ Ex \ {ex }) Pr(sy = o) + Pr(ox ∈ I) Pr(B = false) Pr(sy = o) + Pr(ox = ex ) Pr(B = false) Pr(sy = o) = Pr(ox ∈ Ex \ {ex }) Pr(sy = o) + Pr(ox ∈ I ∪ {ex }) Pr(B = false) Pr(sy = o) = [Pr(ox ∈ Ex \ {ex }) + Pr(ox ∈ I ∪ {ex })(1 − Pr(B = true))] Pr(sy = o) = [Pr(ox ∈ Cx ) − Pr(ox ∈ I ∪ {ex }) Pr(B = true)] Pr(sy = o) = [1 − Pr(ox ∈ I ∪ {ex }) Pr(B = true)] Pr(sy = o)   I − 1 + 1 Cx − 1 = 1− Pr(sy = o) × Cx − 1 Cy − 1   I Pr(sy = o) = 1− Cy − 1 Cy − 1 − I 1 = × Cy − 1 Ey − 1 1 = Cy − 1 4. When Cx > Cy , this concerns case 4.c. The cases for o are the following: 23 • o = ey , this occurs in the case 4.c.i and case 4.c.iv when B ′ is true. We use the following deduction: Pr(oy = o) = Pr(4.c.i or (4.c.iv and B ′ = true)) = Pr(4.c.i) + Pr(4.c.iv and B ′ = true) = Pr(ox = ex ) + Pr(ox ∈ Ex \ {ex } and B ∗ = false and B ′ = true)   Ex − 1 (Cx − 1)(Ey − 1) 1 ∗ + (1 − p ) 1 − = Cx − 1 Cx − 1 (Cy − 1)(Ex − 1)(1 − p∗ )   1 (Cx − 1)(Ey − 1) Ex − 1 ∗ = 1−p − + Cx − 1 Cx − 1 (Cy − 1)(Ex − 1) 1 Ex − 1 Ex − 1 ∗ Ey − 1 = + − p − Cx − 1 Cx − 1 Cx − 1 Cy − 1   Ey − 1 Ex − 1 1 1 1 (I − 1) − + − − = Cx − 1 Cx − 1 Cy − 1 Cx − 1 Cy − 1 1 Ex − 1 I −1 I −1 Ey − 1 = + − + − Cx − 1 Cx − 1 Cy − 1 Cx − 1 Cy − 1 Ex + I − 1 Ey + I − 1 1 = − + Cx − 1 Cy − 1 Cy − 1 1 Cx − 1 Cy − 1 − + = Cx − 1 Cy − 1 Cy − 1 1 = Cy − 1 • o ∈ I \ {iy }, this occurs in case 4.c.ii and case 4.c.iii when B ∗ is true. We make the 24 following deduction Pr(oy = o) = Pr(4.c.ii or 4.c.iii) = Pr(4.c.ii) + Pr(4.c.iii) = Pr(ox = o) + Pr(ox ∈ Ex \ {ex } and B ∗ = true and si = o) = Pr(ox = o) + Pr(ox ∈ Ex \ {ex }) Pr(B ∗ = true) Pr(si = o)   1 (Cx − 1)(I − 1) Ex − 1 1 1 1 = × + × − × Cx − 1 Cx − 1 Cy − 1 Cx − 1 Ex − 1 I −1 1 1 1 = + − Cx − 1 Cy − 1 Cx − 1 1 = Cy − 1 • o ∈ Ey \ {ey }, this occurs in case 4.c.iv when B ′ is false. We have the following deduction: Pr(oy = o) = Pr(4.c.iv and B ′ = false) = Pr(ox ∈ Ex \ {ex } and B ∗ = false and B ′ = false and sy = o) = Pr(ox ∈ Ex \ {ex }) Pr(B ∗ = false) Pr(B ′ = false) Pr(sy = o) Ex − 1 1 (Cx − 1)(Ey − 1) (1 − p∗ ) × ∗ Cx − 1 (Cy − 1)(Ex − 1)(1 − p ) Ey − 1 1 = Cy − 1 = Lemma 3. The process we described is a non-Markovian coupling. Proof. In the context of a Markovian coupling we analyse the transition from y to y ′ given the information about x. In the non-Markovian case we will use less information about x. We assume that ey is a random variable and that x provides only ex and Uy and we know only that ey ∈ Uy and moreover that Pr(ey = e) = 1/Uy , for any e ∈ Uy and Pr(ey = e) = 0 otherwise. Then the chain Xt makes its move and provides information about ix and ox . Let us consider only the cases when Yt 25 then chooses iy = ix , because nothing changes in the cases where this does not happen. Now ix can be used to define Cx and Cy and, therefore, Ex and Ey . We focus our attention on Ey ∩ Uy because, except for the trivial cases, we must have ey ∈ (Ey ∩ Uy ). Hence, we instead alter our condition to Pr(ey = e) = 1/|Ey ∩ Uy |, for any e ∈ (Ey ∩ Uy ) and 0 otherwise. Note that this is a reasonable process because we established in Lemma 1 that, in the non-trivial case, Ex and Ey partition Ce and, therefore, because Uy ⊆ Ce \ {ex }, we have that Ex and Ey also partition Uy . This means that Uy ∩ I = ∅ and so we are not loosing any part of Uy in this process, we are only dividing it into cases. This process is also the reason why, even when sy ∈ / Uy we define Uy′ = (Uy ∩ Cy ) \ {oy } = (Uy ∩ Ey ) \ {oy }. Now the cases considered in Lemma 2 must be changed. Substitute the original cases of o = ey for o ∈ Uy ∩ Ey . Also, substitute the case o ∈ Ey \ {ey } by o ∈ Ey \ Uy . The other cases remain unaltered. Except for the first case, the previous deductions still apply. We will exemplify how the deduction changes for o ∈ Uy ∩ Ey . We consider only the situation when |Cx | = |Cy |. For the remaining situations, Cx < Cy and Cx > Cy , we use a general argument. Hence our precise assumptions are: o ∈ Uy ∩ Ey and |Cx | = |Cy |. Pr(oy = o) = Pr(ox ∈ Ex \ {ex } and sy ∈ Uy ) + Pr(ox = ex ) Pr(ey = o) = Pr(ox ∈ Ex \ {ex }) Pr(sy ∈ Uy ) + Pr(ox = ex ) Ex − 1 |Uy ∩ Ey | − 1 1 × + Cx − 1 Ey − 1 Cx − 1 |Uy ∩ Ey | = Cx − 1 |Uy ∩ Ey | = Cy − 1 = Which is correct according to our assumption. The general argument follows the above derivation. Whenever the Markovian coupling would produce oy = ey , we obtain 1/(Cy − 1) probability. Moreover, for every edge such that oy ∈ Ey \ {ey } produced by the Markovian coupling we obtain another 1/(Cy − 1) probability. This totals to |Uy ∩ Ey |/(Cy − 1) as desired. This also occurs in the cases when Cx < Cy and Cx > Cy , only the derivations become more cumbersome. Finally will argue why the property that Pr(ey = e) = 1/Uy when e ∈ Uy . The set Uy is initialised to contain only the edge ey , i.e., Uy = {ey }. As the coupling proceeds Uy′ is chosen to represent the 26 edges, from which ey′ was chosen, or is simply restricted if ey does not change. More precisely: in case 4.a.iii we chose ey′ = ox ; in case 4.b.iv we chose ez ′ = ox ; in case 4.b.ii we chose ey′ = ox . We obtained no general bounds on the coupling we presented, it may even be that such bounds are exponential even if the Markov chain has polynomial mixing time. In fact Kumar and Ramesh (1999) proved that this is the case for Markovian couplings of the Jerrum-Sinclair chain (Jerrum and Sinclair, 1989). Note that according to the classification of Kumar and Ramesh (1999) the coupling we present is considered as time-variant Markovian. Hence their result applies to type of coupling we are using, albeit we are considering different chains so it is not immediate that indeed there exist no polynomial Markovian couplings for the chain we presented. Cycle graphs are the only class of graphs for which we establish polynomial bounds, see Figure 18. Theorem 2. For any cycle graph G the mixing time τ of edge swap chain is O(V ) for the normal version of the chain and 1/ log4 (V − 1) for the fast version. Proof. For any two trees Ax and Ay we have a maximum distance of 1 edge, i.e., d(Ax , Ay ) ≤ 1. Hence our coupling applies directly. For the fast version, case 1 does not occur, because ix is chosen from E \ {Ax }. Hence, the only cases that might apply are cases 2 and 3. In first case, the chains preserve their distance and in the last case the distance is reduced to 0. Hence, E[d(X1 , Y1 )] = 1/(V − 1), which corresponds to the probability of case 2. Each step of the coupling is independent, which means we can use the previous result and Markov’s inequality to obtain Pr(d(Xτ , Yτ ) ≥ 1) ≤ 1/(V − 1)τ . Then, we use this probability in the coupling Lemma 4 to obtain a variation distance of 1/4, by solving the following equation: 1/(V − 1)τ = 1/4. For the slow version of the chain, case 1 applies most of the time, i.e., for V − 1 out of V choices of ix . It takes V − 1 steps for the standard chain to behave as the fast chain and, therefore, the time should be (V − 1)/ log4 (V − 1). This result is in stark contrast with the alternative algorithms, random walk and Wilson’s (see Section 5), which require O(V 2 ) time (Levin and Peres, 2017). More recent algorithms are also at least O(V 4/3 ) for this case. Moreover when a graph is a connect set of cycles connected by bridges or articulation points we can also establish a similar result. Figure 16 shows one such graph. 27 Figure 16: Cycles connected by bridges or articulation points. Theorem 3. For any graph G which consists of n simple cycles connect by bridges or articulation points, such that m is the size of the smallest cycle, then the mixing time τ of the fast edge swap chain is the following: log(4n) τ= log  n(m−1) n(m−1)−(m−2)  The mixing time for the slow version is obtained by using |E| instead of n in the previous expression. Proof. To obtain this result we use a path coupling argument. Then for two chains at distance 1   m−2 we have E[d(X1 , Y1)] ≤ 1 − n(m−1) . We assume that the different edge occurs in the largest cycle. In general the edges inserted and deleted do not alter this situation, hence the term 1. However with probability 1/n the chain Xt inserts an edge that creates the cycle where the diference occurs. In that case with probability (m − 2)/(m − 1) the chains coalesce. Hence applying path coupling the mixing time must verify the following equation: τ  1 m−2 ≤ n 1− n(m − 1) 4 For the slow version the chain choose the correct edge with probability 1/E instead of 1/n. 28 4.3. Experimental Results 4.3.1. Convergence Testing Before looking at the performance of the algorithm we started by testing the convergence of the edge swap chain. We estimate the variation distance after a varying number of iterations. The results are shown in Figures 19, 20, 21, 22, 23, 24, and 25. We now describe the structure of these figures. Consider for example Figure 21. The structure is the following: • The bottom left plot shows the graph properties, the number of vertexes V in the x axis and the number edges E on the y axis. For the dense case graph 0 has 10 vertexes and 45 edges. Moreover, graph 6 has 40 vertexes and 780 edges. These graph indexes are used in the remaining plots. • The top left plot show the number of iterations t of the chain in the x axis and the estimated variation distance on the y axis, for all the different graphs. • The top right plot is similar to the top left, but the x axis contains the number of iterations divided by (V 1.3 + E). Besides the data this plot also show a plot of ln(1/ε̂) for reference. • The bottom right plot is the same as the top right plot, using a logarithmic scale on the y axis. To avoid the plots from becoming excessively dense, we do not plot points for all experimental values, instead plot one point out of 3. However, the lines pass through all experimental points, even those that are not explicit. The variation distance between two distributions D1 and D2 on a countable state space S is given P by kD1 − D2 k = x∈S |D1 (x) − D2 (x)|/2. This is the real value of ε. However, the size of S quickly becomes larger than we can compute. Instead, we compute a simpler variation distance kD1 −D2 kd , where S is reduced from the set of all spanning trees of G to the set of integers from 0 to V −1, which correspond to the edge distance, defined in Section 4.2, of the generated tree A to a fixed random spanning tree R. More precisely, we generate 20 random trees, using a random walk algorithm described in Section 5. For each of these trees, we compute kπ − Mt kd , i.e., the simpler distance between the stationary distribution π and the distribution Mt obtained by computing t steps of the edge swapping chain. To obtain Mt , we start from a fixed initial tree A0 and execute our chain t 29 Figure 17: A ladder graph. Figure 18: A cycle graph. times. This process is repeated several times to obtain estimates for the corresponding probabilities. We keep two sets of estimates Mt and Mt′ and stop when kMt − Mt′ kd < 0.05. Moreover we only estimate values where kπ − Mt kd ≥ 0.1. We use the same criteria to estimate π, but in this case the trees are again generated by the random walk algorithm. The final value ε̂ is obtained as the maximum value obtained for the 20 trees. We generated dense graphs, sparse graphs and some in between graphs. The sparse graphs are ladder graphs; an illustration of these graphs is shown in Figure 17. The cycle graphs consist of a single cycle, as shown in Figure 18. The dense graphs are actually the complete graphs KV . We also generated other dense graphs labelled biK which consisted of two complete graphs connected by two edges. We also generated graphs based on the duplication model dmP. Let G0 = (V0 , E0 ) be an undirected and unweighted graph. Given 0 ≤ p ≤ 1, the partial duplication model builds a graph G = (V, E) by partial duplication as follows (Chung et al., 2003): start with G = G0 at time t = 1 and, at time t > 1, perform a duplication step: 1. Uniformly select a random vertex u of G. 2. Add a new vertex v and an edge (u, v). 30 0 50 100 1 1023456 0.75 ε̂ 2 0.5 0.25 3 150 0.1 0 250 300 t/(V 1.3 + E) 350 6543210 56 4 ln(1/ε̂) 65 4 3 1 0 200 2 1 5 6 1 2 3 4 3 5 4 65 4 6 6 1 6543210 0.75 t sparse 130 120 6 110 5 100 90 4 E 80 3 70 60 2 50 1 40 30 0 20 10 20 30 40 50 60 70 80 90 V 65 3 6 02 5 1 4 63 ln(1/ε̂) 43 0.5 2 1 6 5 4 ε̂ 0.25 3 0265 1 4 0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Figure 19: Estimation of variation distance as a function of the number of iterations for sparse graphs (see Section 4.3.1 for details). 3. For each neighbor w of u, add an edge (v, w) with probability p. The different values at the end of dmP, namely in Figures 23, 24 and 25, correspond to the choices of p. These graphs show the convergence of the Markov chain and moreover V 1.3 + E seems to be a reasonable bound for τ . Still, these results are not entirely binding. On the one hand the estimation of the variation distance groups several spanning trees into the same distance, which means that within a group the distribution might not be uniform, even if the global statistics are good. So the actual variation distance may be larger and the convergence might be slower. On the other hand we chose the exponent 1.3 experimentally by trying to force the data of the graphs to converge at the same point. The actual value may be smaller or larger. 4.3.2. Coupling Simulation As mentioned before, we obtained no general bounds on the coupling we presented. In fact, experimental simulation for the coupling does not converge for all classes of graphs. We obtained 31 0 50 100 150 200 1 6 12354 0.75 0 ε̂ t/(V 1.3 + E) 250 65432 1 ln(1/ε̂) 0 0.5 0.25 0.1 0 56 4 123 0 2 3 54 65 4 32 1 6 4 3 02 1 6543 21 0 0.75 t cycle 90 80 6 70 5 60 4 3 E 50 40 2 30 1 20 0 10 10 20 30 40 50 60 70 80 90 V 65 ln(1/ε̂) 0.5 65 ε̂ 0.25 43 2 1 65 4 0.1 0 0.1 0.2 0.3 0.4 0.5 3 0.6 0 0.7 0.8 Figure 20: Estimation of variation distance as a function of the number of iterations for cycle graphs (see Section 4.3.1 for details). 32 0 50 1 0123456 150 200 6543210 6 0.1 0 5 4 0 0.25 3 6 2 6 1 54 3 2 1 6 5 4 1 6543210 6 5 4 0.75 t 0.5 6 5 4 5 5 34 0 2 1 ln(1/ε̂) 3 6 2 54 ε̂ 1 0 0.25 6 3 1 0 10 15 20 25 30 35 40 45 V 1 6 dense 2 ln(1/ε̂) 54 3 2 0.5 900 800 700 600 E 500 400 300 200 100 0 t/(V 1.3 + E) 250 6 45 0.75 ε̂ 100 3 5 4 0.1 0 0.1 0.2 0.3 2 0.4 0.5 0.6 Figure 21: Estimation of variation distance as a function of the number of iterations for dense graphs (see Section 4.3.1 for details). 33 0 t/(V 1.3 + E) 50 100 150 200 250 300 350 400 450 500 1 0124356 124356 4 65 6 3 0 5 4 2 0.75 1 3 4 ε̂ 0.5 2 0 1 3 2 0.25 0 1 0 1 2 0.1 0 6543210 6546325461 6 5 ln(1/ε̂) 5340 6 2 1 65 43 6 4 5 6 3 2 4 5 4 t biK 1800 1600 6 1400 1200 5 E 1000 4 800 3 600 400 2 1 200 0 0 10 20 30 40 50 60 70 80 90 V 2 0 65 43 1 5 4 23 5 4 5 1 6543210 6546325461 563402 0.75 1 65 43 0.5 2 0 65 43 1 ε̂ 2 0.25 54 3 5 4 0 21 2 1 ln(1/ε̂) 1 2 0.1 0 0.1 0.2 0.3 0 0.4 0 2 1 0 0.5 0.6 0.7 Figure 22: Estimation of variation distance as a function of the number of iterations for biK graphs (see Section 4.3.1 for details). 34 0 1 586 7 2 1 0.75 3 4 ε̂ 0.5 0 ln(1/ε̂) 865 7 21 3 4 0 8 5 76 2 3 4 0.25 0.1 0 8 0 1 2 3 48 3 1 85 76 0.75 321 4 0.5 t 5 7 65 8 dmP05 100 90 80 70 60 E 50 40 30 20 10 0 t/(V 1.3 + E) 20 40 60 80 100 120 140 160 180 200 6 7 8 4 0 10 20 30 40 50 60 70 80 90 100 V ln(1/ε̂) 0 8 ε̂ 0.25 7 65 4 3 8 0.1 0 2 0.1 0.2 0.3 0.4 2 0.5 0.6 0.7 Figure 23: Estimation of variation distance as a function of the number of iterations for dmP graphs (see Section 4.3.1 for details). 35 0 50 100 150 200 t/(V 1.3 + E) 250 1 123546 0 654231 ln(1/ε̂) 0 0.75 ε̂ 0.5 23 0.25 54 6 6 1 0.1 0 54 1 1 654231 0 0.75 t 6 4 2 0 5 3 23 6 dmP25 110 100 90 80 E 70 60 50 40 30 20 54 5 0.5 0.25 10 15 20 25 30 35 40 45 V 54 23 1 1 54 ln(1/ε̂) 6 ε̂ 6 6 4 0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Figure 24: Estimation of variation distance as a function of the number of iterations for dmP graphs (see Section 4.3.1 for details). 36 0 50 100 150 200 t/(V 1.3 + E) 250 1 123546 654321 ln(1/ε̂) 0.75 ε̂ 0 0.5 45 0 6 3 0.25 12 0.1 0 54 3 1 654321 0.75 t 4 0.5 6 5 3 1 ε̂ 2 10 15 20 25 30 35 40 45 V 21 6 54 ln(1/ε̂) 0 6 54 3 0.25 2 1 0 5 54 6 dmP50 140 120 100 80 E 60 40 20 0 6 0.1 0 6 4 5 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Figure 25: Estimation of variation distance as a function of the number of iterations for dmP graphs (see Section 4.3.1 for details). 37 experimental convergence for cycle graph, as expected from Theorem 2, and for ladder graphs. For the remaining graphs we used an optimistic version of the coupling which always assumes that sy ∈ Uy and that B ∗ fails. With these assumptions, all the cases which increase the distance between states are eliminated and the coupling always converges. Note that this approach does not yield a sound coupling, but in practice we verified that this procedure obtained good experimental variation distance. Moreover, the variation distance estimation for these tests is not the simpler distance but the actual experimental variation distance, obtained by generating several experimental trees, such that in average each possible tree is obtained 100 times. The simulation of the path coupling proceeds by generating a path with e ln V steps, essentially selecting two trees at distance e ln V from each other. This path is obtained by computing e ln V steps of the fast chain. Recall that our implementation and all simulations use the fast version of the chain. The simulation ends once this path contracts to size ln V . Let t′ be the number of steps in this process. Once this point is obtained our estimate for mixing time is τ̂ = t′ ln V . In general, we wish to obtain τ̂ such that the probability that the two general chains Xt and Xt coalesce is at least 75%. Hence, we repeat this process 4 times and choose the second largest value of τ̂ as our estimate. Table 1 summarizes results for the experimental variation distance. The number of possible spanning trees for each graph was computed through the Kirchoff’s theorem. Then, we generated 100 times the number of possible trees, and we computed the variation distance. As stated above, we got good results for the variation distance, getting a median well below 25% for all tested graph topologies. We now present experimental results for larger graphs where we use the optimistic coupling. All experiments were conducted on a computer with an Intel(R) Xeon(R) CPU E5-2630 v3 @ 2.40GHz with 4 cores and 32GB of RAM. We present running times for different graph topologies and sizes in Figures 26, 27, 28, 29, 30, 31 and 32. Note beforehand that the coupling estimate needs only to be computed once for each graph. Once the estimate is known, we can generate as many spanning trees as we want. Although the edge swapping method is not always the faster compared with the random walk and the Wilson’s algorithm, it is competitive in practice for dmP and torus graphs, and it is faster for biK, cycle and sparse (ladder) graphs. As expected, it is less competitive for dense graphs. Hence, experimental results seem to point out that the edge swapping method is more competitive in practice for those instances that are harder for random walk based methods, namely biK and cycle graphs. The results for biK and dmP are of particular interest as most 38 Table 1: Variation distance (VD) for different graph topologies. Median and maximum VD computed over 5 runs for each network. Since dmP graphs are random, results for dmP were further computed over 5 different graphs for each size |V |. Graph |V | median VD max VD dense {5, 7} 0.060 0.194 biK {8, 10} 0.065 0.190 cycle {16, 20, 24} 0.001 0.004 sparse {10, 14, 20} 0.053 0.110 torus {9, 12} 0.094 0.383 dmP {8, 10, 12} 0.069 0.270 real networks seem to include these kind of topologies, i.e., they include communities and they are scale-free Chung and Lu (2006). 5. Related Work For a detailed exposure on probability on trees and networks see Lyons and Peres (2016, Chapter 4). As far as we know, the initial work on generating uniform spanning trees was by Aldous (1990b) and Broader (1989), which obtained spanning trees by performing a random walk on the underlying graph. The author also further studied the properties of such random trees (Aldous, 1990a), namely giving general closed formulas for the counting argument we presented in Section 2. In the random walk process a vertex v of G is chosen and at each step this vertex is swapped by an adjacent vertex, where all neighboring vertices are selected with equal probability. Each time a vertex is visited by the first time the corresponding edge is added to the growing spanning tree. The process ends when all vertexes of G get visited at least once. This amount of steps is known as the cover time of G. To obtain an algorithm that is faster than the cover time, Wilson (1996) proposed a different approach. A vertex r of G is initially chosen uniformly and the goal is to hit this specific vertex r from a second vertex, also chosen uniformly from G. This process is again a random walk, but with a loop erasure feature. Whenever the path from the second vertex intersects itself all the edges in the corresponding loop must be removed from the path. When the path eventually reaches r it becomes part of the spanning tree. The process then continues by choosing a third vertex and 39 Dense 9 Coupling estimate Edge swapping Random walk 8 Wilson’s algorithm 7 time (s) 6 5 4 3 2 1 0 0 × 100 2 × 106 4 × 106 6 × 106 8 × 106 1 × 107 1 × 107 1 × 107 E Figure 26: Running times for dense (fully connected) graphs averaged over five runs, including the running time for computing the optimistic coupling estimate, the running time for generating a spanning tree based on that estimate and on the edge swapping algorithm, the running time for generating a spanning tree through a random walk, and also the running time for Wilson’s algorithm. 40 BiK 9 Coupling estimate Edge swapping Random walk 8 Wilson’s algorithm 7 time (s) 6 5 4 3 2 1 0 0 × 100 1 × 106 2 × 106 3 × 106 4 × 106 5 × 106 6 × 106 7 × 106 E Figure 27: Running times for biK graphs averaged over five runs, including the running time for computing the optimistic coupling estimate, the running time for generating a spanning tree based on that estimate and on the edge swapping algorithm, the running time for generating a spanning tree through a random walk, and also the running time for Wilson’s algorithm. 41 Cycle 100 Coupling estimate Edge swapping Random walk 90 Wilson’s algorithm 80 70 time (s) 60 50 40 30 20 10 0 2 × 103 4 × 103 6 × 103 8 × 103 1 × 104 1 × 104 1 × 104 2 × 104 2 × 104 2 × 104 E Figure 28: Running times for cycle graphs averaged over five runs, including the running time for computing the optimistic coupling estimate, the running time for generating a spanning tree based on that estimate and on the edge swapping algorithm, the running time for generating a spanning tree through a random walk, and also the running time for Wilson’s algorithm. 42 Ladder 2000 Coupling estimate Edge swapping Random walk 1800 Wilson’s algorithm 1600 1400 time (s) 1200 1000 800 600 400 200 0 0 × 100 2 × 104 4 × 104 6 × 104 8 × 104 E 1 × 105 1 × 105 1 × 105 2 × 105 Figure 29: Running times for sparse (ladder) graphs averaged over five runs, including the running time for computing the optimistic coupling estimate, the running time for generating a spanning tree based on that estimate and on the edge swapping algorithm, the running time for generating a spanning tree through a random walk, and also the running time for Wilson’s algorithm. 43 Torus 10 Coupling estimate Edge swapping Random walk 9 Wilson’s algorithm 8 7 time (s) 6 5 4 3 2 1 0 0 × 100 2 × 103 4 × 103 6 × 103 8 × 103 1 × 104 1 × 104 1 × 104 2 × 104 2 × 104 2 × 104 E Figure 30: Running times for square torus graphs averaged over five runs, including the running time for computing the optimistic coupling estimate, the running time for generating a spanning tree based on that estimate and on the edge swapping algorithm, the running time for generating a spanning tree through a random walk, and also the running time for Wilson’s algorithm. 44 Torus 7 6 Coupling estimate Edge swapping Random walk Wilson’s algorithm time (s) 5 4 3 2 1 0 1 × 103 2 × 103 3 × 103 4 × 103 5 × 103 6 × 103 7 × 103 8 × 103 9 × 103 1 × 104 E Figure 31: Running times for rectangular torus graphs averaged over five runs, including the running time for computing the optimistic coupling estimate, the running time for generating a spanning tree based on that estimate and on the edge swapping algorithm, the running time for generating a spanning tree through a random walk, and also the running time for Wilson’s algorithm. 45 Duplication model 300 Coupling estimate Edge swapping Random walk Wilson’s algorithm 250 time (s) 200 150 100 50 0 0 × 100 5 × 104 1 × 105 2 × 105 2 × 105 2 × 105 3 × 105 4 × 105 4 × 105 5 × 105 5 × 105 6 × 105 E Figure 32: Running times for dmP graphs averaged over five runs, including the running time for computing the optimistic coupling estimate, the running time for generating a spanning tree based on that estimate and on the edge swapping algorithm, the running time for generating a spanning tree through a random walk, and also the running time for Wilson’s algorithm. 46 also computing a loop erasure path from it, but this time it is not necessary to hit r precisely, it is enough to hit any vertex on the branch that is already linked to r. The process continues by choosing random vertexes an computing loop erasure paths that hit the spanning tree that is already computed. We implemented the above algorithms, as they are accessible, although several theoretical results where obtained in recent years we are not aware of an implementation of such algorithms. We will now survey these results. Another approach to this problem relies on the Kirchoff’s Theorem Kirchhoff (1847) that counts the number of spanning trees by computing the determinant of a certain matrix, related to the graph G. This relation is researched by Guénoche (1983); Kulkarni (1990), which yielded an O(EV 3 ) time algorithm. This result was improved to O(V 2.373 ), by Colbourn, Day, and Nel (1989); Colbourn, Myrvold, and Neufeld (1996), where the exponent corresponds to the fastest algorithm to compute matrix multiplication. Improvements on the random walk approach where obtained by Kelner and Mądry (2009); Mądry (2011), culminating in an Õ(E o(1)+4/3 ) time algorithm by Mądry, Straszak, (2015), which relies on insight provided by the effective resistance metric. Interestingly, the initial work by Broader (1989) contains reference to the edge swapping chain we presented in this paper (Section 5, named the swap chain). The author mentions that the mixing time of this chain is E O(1) , albeit the details are omitted. As far as we can tell, this natural approach to the problem did not receive much attention precisely due to the lack of an efficient implementation. Even though link cut trees were known at the time (Sleator and Tarjan, 1981, 1985) their application to this problem was not established prior to this work. Their initial application was to network flows (Goldberg and Tarjan, 1989). We also found another reference to the edge swap in the work of Sinclair (1992). In the proposal of the canonical path technique the author mentions this particular chain as a motivating application for the canonical path technique, still the details are omitted and we were not able to obtain such an analysis. We considered the LCT version where the auxiliary trees are implemented with splay trees Sleator and Tarjan (1985), i.e., the auxiliary data structures we mentioned in Section 3 are splay trees. This means that in step 5 of Algorithm 1 all the vertexes involved in the path C \ {(u, v)} get stored in a splay tree. This path oriented approach of link cut trees makes them suitable for our goals, as opposed to other dynamic connectivity data structures such as Euler tour trees (Henzinger and King, 1995). Splay trees are self-adjusting binary search trees, therefore the vertexes are ordered in such a way 47 that the inorder traversal of the tree coincides with the sequence of the vertexes that are obtained by traversing C \ {(u, v)} from u to v. This also justifies why the size of this set can also be obtained in O(log V ) amortized time. Each node simply stores the size of its sub-tree and these values are efficiently updated during the splay process, which consists of a sequence of rotations. Moreover, these values can also be used to Select an edge from the path. By starting at the root and comparing the tree sizes to i we can determine if the first vertex of the desired edge is on the left sub-tree, on the root or on the right sub-tree. Likewise we can do the same for the second vertex of the edge in question. These operations splay the vertexes that they obtain and therefore the total time depends on the Splay operation. The precise total time of the Splay operation is O((V + 1) log n), however the V log V term does not accumulate over successive operations, thus yielding the bound of O((V + τ ) log V ) in Theorem 1. In general the V log V term should not be a bottleneck because for most graphs we should have τ > V . This is not always the case, if G consists of a single cycle then τ = 1, but V may be large. Figure 18 shows an example of such a graph. We finish this Section by reviewing the formal definitions of variational distance and mixing time τ (Mitzenmacher and Upfal, 2005). Definition 1. The variation distance between two distributions D1 and D2 on a countable space S is given by ||D1 − D2 || = X |D1 (x) − D2 (x)| x∈S 2 (2) Definition 2. Let π be the stationary distribution of a Markov chain with state space S. Let ptx represent the distribution of the state of the chain starting at state x after t steps. We define ∆x (t) = |ptx − π| (3) ∆(t) = max ∆s (t) (4) x∈S That is ∆x (t) is the variation distance between the stationary distribution and ptx and ∆(t) is the maximum of these values over all states x. We also define τx (ε) = min{t : ∆x (t) ≤ ε} (5) τ (ε) = max τx (ε) (6) x∈S When we refer only to the mixing time we mean τ (1/4). Finally the coupling Lemma justifies the coupling approach: 48 Lemma 4. Let Zt = (Xt , Yt) be a coupling for a Markov chain M on a state space S. Suppose that there exists a T such that, for every x, y ∈ S, Pr(XT 6= YT |X0 = x, Y0 = y) ≤ ε (7) Then τ (ε) ≤ T . That is, for any initial state, the variation distance between the distribution of the state of the chain after T steps and the stationary distribution is at most ε. If there is a distance d defined in S then the property XT 6= YT can be obtained using the condition d(XT , YT ) ≥ 1. For this condition we can use the Markovian inequality Pr(d(XT , YT ) ≥ 1) ≤ E[d(XT , YT )]. The path coupling technique (Bubley and Dyer, 1997) constructs a coupling by chaining several chains, such that the distance between then is 1. Therefore we obtain d(XT , YT ) = d(XT0 , XT1 ) + d(XT1 , XT2 ) + . . . + d(XTD−1 , XTD ) = 1 + 1 + . . . + 1, where XT = XT0 and YT = XTD . 6. Conclusions and Future Work In this paper we studied a new algorithm to obtain the spanning trees of a graph in an uniform way. The underlying Markov chain was initially sketched by Broader (1989) in the early study of this problem. We further extended this work by proving the necessary Markov chain properties and using the link cut tree data structure. This allows for a much faster implementation than repeating the DFS procedure. This may actually be the reason why this approach has gone largely unnoticed during this time. The main shortcoming of our approach is the lack of a general theoretical bound of the mixing time. Such a bound might be possible using new approaches as the insight into the resistance metric (Mądry, Straszak, and Tarnawski, 2015). Although a general analysis would be valuable we addressed this problem by simulating a coupling, both sound or optimistic. The lack of analysis is not a shortcoming of the algorithm itself, which is both practical and efficient. We implemented it and compared it against existing alternatives. The experimental results show that it is very competitive. A theoretical bound would still be valuable, specially if it is the case that this algorithm is more efficient than the alternatives. On the one hand computing the mixing time of the underlying chain is complex, time consuming and hard to analyse in theory. On the other hand the user of this process can fix a certain number of steps to execute. This is a very useful parameter, as it can be used to swap randomness for 49 time. Depending on the type of application the user may sacrifice the randomness of the underlying trees to obtain faster results or on the contrary spend some extra time to guarantee randomness. Existing algorithms do not provide such a possibility. As a final note, we point out that our approach can be generalized by assigning weights to the edges of the graph. The edge to be inserted can then be selected with a probability that corresponds to its weight, divided by the global sum of weights. Moreover the edge to remove from the cycle should be removed according to its weight. The probability should be its weight divided by the sum of the cycle weights. The ergodic analysis of Section 4.1 generalizes easily to this case, so this chain also generates spanning trees uniformly, albeit the analysis of the coupling of Section 4.2 might need some adjustments. A proper weight selection might obtain a faster mixing timer, possibly something similar to the resistance of the edge. Acknowledgements This work was funded in part by European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 690941 and national funds through Fundação para a Ciência e a Tecnologia (FCT) with reference UID/CEC/50021/2013. Bibliography References Aigner, M., Ziegler, G. M., Quarteroni, A., 2010. Proofs from the Book. Vol. 274. Springer. Aldous, D., 1990a. A random tree model associated with random graphs. Random Structures & Algorithms 1 (4), 383–402. Aldous, D. J., Nov. 1990b. The random walk construction of uniform spanning trees and uniform labelled trees. SIAM J. Discret. Math. 3 (4), 450–465. Borchardt, C. W., 1860. Über eine interpolationsformel für eine art symmetrischer functionen und über deren anwendung. Math. Abh. der Akademie der Wissenschaften zu Berlin, 1–20. Broader, A., 1989. Generating random spanning trees. In: IEEE Symposium on Fondations of Computer Science. pp. 442–447. 50 Bubley, R., Dyer, M., Oct 1997. Path coupling: A technique for proving rapid mixing in Markov chains. In: Proceedings 38th Annual Symposium on Foundations of Computer Science. pp. 223– 231. Cayley, A., 1889. A theorem on trees. Pure Appl. Math. 23, 376–378. Chung, F. R., Lu, L., 2006. Complex graphs and networks. No. 107. American Mathematical Soc. Chung, F. R. K., Lu, L., Dewey, T. G., Galas, D. J., 2003. Duplication models for biological networks. Journal of Computational Biology 10 (5), 677–687. URL https://doi.org/10.1089/106652703322539024 Colbourn, C. J., Day, R. P. J., Nel, L. D., 1989. Unranking and ranking spanning trees of a graph. J. Algorithms 10 (2), 271–286. Colbourn, C. J., Myrvold, W. J., Neufeld, E., 1996. Two algorithms for unranking arborescences. J. Algorithms 20 (2), 268–281. Galler, B. A., Fisher, M. J., May 1964. An improved equivalence algorithm. Commun. ACM 7 (5), 301–303. URL http://doi.acm.org/10.1145/364099.364331 Goldberg, A. V., Tarjan, R. E., Oct. 1989. Finding minimum-cost circulations by canceling negative cycles. J. ACM 36 (4), 873–886. Guénoche, A., 1983. Random spanning tree. Journal of Algorithms 4 (3), 214–220. Henzinger, M. R., King, V., 1995. Randomized dynamic graph algorithms with polylogarithmic time per operation. In: Proceedings of the twenty-seventh annual ACM symposium on Theory of computing. ACM, pp. 519–527. Jerrum, M., Sinclair, A., 1989. Approximating the permanent. SIAM J. Comput. 18 (6), 1149–1178. Kelner, J. A., Mądry, A., Oct 2009. Faster generation of random spanning trees. In: 2009 50th Annual IEEE Symposium on Foundations of Computer Science. pp. 13–21. Kirchhoff, G., 1847. Ueber die auflösung der gleichungen, auf welche man bei der untersuchung der linearen vertheilung galvanischer ströme geführt wird. Annalen der Physik 148 (12), 497–508. 51 Kulkarni, V., 1990. Generating random combinatorial objects. Journal of Algorithms 11 (2), 185 – 207. Kumar, V. S. A., Ramesh, H., 1999. Markovian coupling vs. conductance for the Jerrum-Sinclair chain. In: 40th Annual Symposium on Foundations of Computer Science. pp. 241–251. Levin, D. A., Peres, Y., 2017. Markov chains and mixing times. Vol. 107. American Mathematical Soc. Lyons, R., Peres, Y., 2016. Probability on trees and networks. Vol. 42. Cambridge University Press. Mądry, A., 2011. From graphs to matrices, and back: new techniques for graph algorithms. Ph.D. thesis, Massachusetts Institute of Technology. Mitzenmacher, M., Upfal, E., 2005. Probability and Computing: Randomized Algorithms and Probabilistic Analysis. Cambridge University Press, New York, NY, USA. Mądry, A., Straszak, D., Tarnawski, J., 2015. Fast generation of random spanning trees and the effective resistance metric. In: Indyk, P. (Ed.), Proceedings of the Twenty-Sixth Annual ACMSIAM Symposium on Discrete Algorithms, SODA 2015, San Diego, CA, USA, January 4-6, 2015. SIAM, pp. 2019–2036. Sinclair, A., 1992. Improved bounds for mixing rates of Markov chains and multicommodity flow. Combinatorics, probability and Computing 1 (4), 351–370. Sleator, D. D., Tarjan, R. E., 1981. A data structure for dynamic trees. In: Proceedings of the Thirteenth Annual ACM Symposium on Theory of Computing. STOC ’81. ACM, New York, NY, USA, pp. 114–122. Sleator, D. D., Tarjan, R. E., Jul. 1985. Self-adjusting binary search trees. J. ACM 32 (3), 652–686. Wilson, D. B., 1996. Generating random spanning trees more quickly than the cover time. In: Proceedings of the Twenty-eighth Annual ACM Symposium on Theory of Computing. STOC ’96. ACM, New York, NY, USA, pp. 296–303. 52
8cs.DS
Maximizing Diversity for Multimodal Optimization arXiv:1406.2539v1 [cs.NE] 10 Jun 2014 Fabrı́cio Olivetti de França CMCC, Federal University of ABC (UFABC) – Brazil [email protected] 1 Introduction Most multimodal optimization algorithms use the so called niching methods [1] in order to promote diversity during optimization, while others, like Artificial Immune Systems [2] try to find multiple solutions as its main objective. One of such algorithms, called dopt-aiNet [3], introduced the Line Distance that measures the distance between two solutions regarding their basis of attraction. In this short abstract I propose the use of the Line Distance measure as the main objective-function in order to locate multiple optima at once in a population. 1.1 Line Distance The Line Distance between two solutions x and y is calculated by taking the middle point z = (x + y)/2 and building three (n + 1)-dimensional vectors x′ = [x, f (x)], the distance between the point z ′ and the line formed by x′ and y ′ is then calculated. If both x and y are at the same optimum, the line formed between them will be very close to the function contour resulting in a small distance, but if they are at different optima, it will be proportional to f (x). 1.2 Maximizing Diversity The maximization of the Line Distance between two points will have its maximum when both points are at different optima, so if we maximize this objective, it is expected that the algorithm returns several different optima. This objectivefunction may be used as the main goal of the optization process or as a supporting operator (i.e., crossover, mutation) for any populational algorithm. To test this claim, we can use this simples algorithm: create an initial population P with a single solution drawn uniformly at random on the problem domain. Next, for a given number of iterations repeat two procedures: expansion of current solutions and suppression of similar solutions. During the expansion process, for each solution x, m new solutions are created by optimizing argmaxld(x + α.d/kkdkk, x), where d is a random direction drawn α uniformly at random from [−1, 1]n , α is the step to walk into the given direction, that can be found by any unidimensional search method, and ld(x, y) is the line distance between x and y. Each new solution is aggregated into the population if its distance to the original solution x is greater than a threshold σ. If for a given solution, all of its generated solutions are discarded, then this solution is moved to the local optima population LP as it is assumed that all the nearby optima around this solution was already discovered. At the suppression step, all the solutions of each of the populations are compared to each other, whenever two solutions have distance greater than σ, the worst of the two is discarded. Notice that the optimization process is indirectly performed by the distance function that has its maximum whenever one of the solutions is located at a local optima. 2 Experimental Results For a simple experiment it was chosen two well-known multimodal functions: Rastrigin and Grienwank, those functions were tested in ℜ2 and the parameters were m = 10 for both function and, σ = 0.9 for Rastrigin and σ = 0.1 for Grienwank. For 10 repetitions of 1, 000 iterations of this simple procedure could find 9 different local optima for Rastrigin and 12 for Griewank, in both cases the global optima was also found. The results are depicted in Fig. 1. 4 5 2 0 0 −2 −5 −4 −4 −2 0 2 (a) Rastrigin 4 −5 0 5 (b) Grienwank Fig. 1. Some of the optima found by maximizing diversity. 3 Conclusion In this extended abstract it is argued that the direct maximization of the diversity may be better suited for multimodal optimization algorithms. It is shown that by using the adequate diversity measure with a proper algorithm it is possible to find many different local optima. It must be noticed that some objectivefunctions may have an explosive growth on the number of local optima with the growth of dimension, so this procedure may be best suited as an evolutionary operator to promote diversity. References 1. Mahfoud, S.W.: Niching methods for genetic algorithms. Urbana 51(95001) (1995) 2. de França, F.O., Coelho, G.P., Castro, P.A., Von Zuben, F.J.: Conceptual and practical aspects of the ainet family of algorithms. International Journal of Natural Computing Research (IJNCR) 1(1) (2010) 1–35 3. de Franca, F.O., Von Zuben, F.J., de Castro, L.N.: An artificial immune network for multimodal function optimization on dynamic environments. In: Proceedings of the 2005 conference on Genetic and evolutionary computation, ACM (2005) 289–296
9cs.NE
Diffusion Adaptation Over Clustered Multitask Networks Based on the Affine Projection Algorithm Vinay Chakravarthi Gogineni1, Mrityunjoy Chakraborty2 Department of Electronics and Electrical Communication Engineering arXiv:1507.08566v4 [cs.DC] 1 Oct 2015 Indian Institute of Technology, Kharagpur, INDIA Phone: +91 − 3222 − 283512 E.Mail : 1 Fax: +91 − 3222 − 255303 [email protected], 2 [email protected] Abstract Distributed adaptive networks achieve better estimation performance by exploiting temporal and as well spatial diversity while consuming few resources. Recent works have studied the single task distributed estimation problem, in which the nodes estimate a single optimum parameter vector collaboratively. However, there are many important applications where the multiple vectors have to estimated simultaneously, in a collaborative manner. This paper presents multi-task diffusion strategies based on the Affine Projection Algorithm (APA), usage of APA makes the algorithm robust against the correlated input. The performance analysis of the proposed multi-task diffusion APA algorithm is studied in mean and mean square sense. And also a modified multi-task diffusion strategy is proposed that improves the performance in terms of convergence rate and steady state EMSE as well. Simulations are conducted to verify the analytical results. I. I NTRODUCTION Distributed adaptation over networks has emerged as an attractive and challenging research area with the advent of multiagent( wireless or wireline) networks. Recent results in the field can be found in [1]-[3]. In adaptive networks, the interconnected nodes continuously learn and adapt, as well as perform the assigned tasks such as parameter estimation from observations collected by the dispersed agents. Consider a connected network consisting of N nodes observing temporal data arising from different spatial sources with possibly different statistical profiles. The objective is to enable the nodes to estimate a parameter vector of interest, wopt from the observed data. In a centralized approach, the data or local estimates from all nodes would be conveyed to a central processor where they would be fused and the vector of parameters estimated. In order to reduce the requirement of powerful central processor and extensive amount of communications in a traditional centralized solution, a distributed solution is developed relying only on local data exchange and interactions between intermediate neighborhood nodes, while retaining the estimation accuracy of centralized solution. In distributed networks, the individual nodes share the computational burden so that communications are reduced as compared to the centralized network, and power and bandwidth usage are also there by reduced. Due to these merits, distributed estimation has received more attention recently and been widely 2 used in many applications, such as in precision agriculture, environmental monitoring, military surveillance, transportation and instrumentation. The mode of cooperation that is allowed among the nodes determines the efficiency of any distributed implementation. In incremental mode of cooperation, each node transfers information to its corresponding adjacent node in sequential manner using cyclic pattern of collaboration. This approach reduces communications between nodes and improves the network autonomy as compared the centralized solution. In practical wireless sensor networks, it may be more difficult to establish a cyclic pattern as required in the incremental mode of cooperation as the number of sensor nodes increase. On the other hand, in diffusion mode of cooperation, each node exchanges information with its neighborhood (i.e., the set of all its neighbors including itself), Nk as directed by the network topology. There exist several useful distributed strategies for sequential data processing over networks including consensus strategies [4]-[6], incremental strategies [7]-[9] and diffusion strategies [10]-[13]. Diffusion strategies exhibit superior stability and performance over consensus based algorithms [14]. The existing literature on distributed algorithms shows that most works focus primarily on the case where the nodes estimate a single optimum parameter vector collaboratively. We shall refer to problems of this type as single-task problems. However, many problems of interest happen to be multi-task oriented i.e., consider the general situation where there are connected clusters of nodes, and each cluster has a parameter vector to estimate. The estimation still needs to be performed cooperatively across the network because the data across the clusters may be correlated and, therefore, cooperation across clusters can be beneficial. This concept is relevant to the context of distributed estimation and adaptation over networks. Initial investigations along these lines for the traditional diffusion strategy appear in [15]-[19]. It is well known that in the case of a single adaptive filter, one major drawback of the LMS algorithm is its slow convergence rate for colored input signals and the APA algorithm is a better alternative to LMS is such an environment. For distributed networks, highly correlated inputs also deteriorate the performance of the multi-task diffusion-LMS (multi-task d-LMS) algorithm. In this paper we therefore focus on a new APA-based multi-task distributed learning scheme over networks to obtain a good compromise between convergence performance and computational cost and to analyze their performance in terms of mean-square error and convergence rate. II. N ETWORK M ODELS AND M ULTI TASK LEARNING Consider a network with N nodes deployed over a certain geographical area. At every time instant n, every node k has access to time realizations {dk (n), uk (n)} with dk (n) denoting a scalar zero mean reference signal and uk (n) is an L × 1 regression vector, uk (n) = [uk (n), uk (n − 1), ..., uk (n − L + 1)]T with covariance matrix Ru,k = E[uk (n)uTk (n)]. The data at node k is assumed to be related via the linear measurement model: dk (n) = uTk (n) w⋆k + ǫk (n) (1) where w⋆k is an unknown optimal parameter vector to be estimated at node k and ǫk (n) is an observation noise with variance ξ 0 which is assumed to be zero mean white noise and also independent of uk (n) for all k. Considering the number of parameter vectors to be estimated, which we shall refer to as the number of tasks, the distributed learning problem can be single-task or multi-task oriented. Therefore we distinguish among the following three types of networks, as illustrated by Fig. 1, depending 3 (a) (b) (c) Figure 1: Three types of networks. Through direct links, nodes can communicate with each other in one hop. (a) Single-task Network. (b) Multi-task network. (c) Clustered multi-task network on how the parameter vectors w⋆k are related across nodes: • Single-task networks: All nodes in the network have to estimate the same parameter vector w⋆k . That is, in this case we have that w⋆k = w⋆ , • ∀k ∈ 1, 2, ..., N (2) Multi-task networks: Each node k in the network has to determine its own optimum parameter vector, w⋆k . However, it is assumed that similarities and relationships exist among the parameters of neighboring nodes, which we denote by writing w⋆k ∼ w⋆l , if l ∈ Nk (3) The sign ∼ represents a similarity relationship in some sense, and its meaning will become clear soon once we introduce expression (8) and (9) further ahead. There are many situations in practice where the objective parameters are not identical across clusters but have inherent relationships. It is therefore beneficial to exploit these relationships to enhance performance. Here we focus on promoting the similarity of objective parameter vectors via their distance to each other. • Clustered Multi-task Networks: Nodes are grouped into Q clusters, and there is one task per cluster. The optimum parameter vectors are only constrained to be equal within each cluster. The optimum parameter vectors are only constrained to be equal within each cluster, but similarities between neighboring clusters are allowed to exist, namely, w⋆k = w⋆Cq , whenever k ∈ Cq w⋆Cp ∼ w⋆Cq , if (4) Cp , Cq are connected where p and q denote two cluster indexes. We say that two clusters Cp and Cp are connected if there exists at least one edge linking a node from one cluster to a node in the other cluster. One can observe that the single-task and multi-task networks are particular cases of the clustered multi-task network. In the case where all nodes are clustered together, the clustered multi-task network reduces to the single-task network. on the other 4 hand, in the case where each cluster only involves one node, the clustered multi-task network becomes a multi-task network. Building on the literature on diffusion strategies for single-task networks, we shall now generalize its usage and analysis for distributed learning over clustered multi-task networks. These results will also be applicable to multi-task networks by setting the number of clusters equal to the number of nodes. III. P ROBLEM F ORMULATION In clustered multitask networks the nodes that are grouped into cluster estimate the same coefficient vector. Thus, consider the cluster C(k) to which node k belongs. Under certain settings, in order to provide independence from the input data correlation statistics, we introduce normalized updates with respect to the input regressor at each node uk (n). A local cost function, Jk (wC(k) ), is associated with node k and it is assumed that the Hessian matrix of the cost function is positive semi-definite. The local cost function Jk (wC(k) ) is defined as Jk (wC(k) ) = E{| dk (n) − uTk (n)wCk 2 | } kuk (n)k (5) Depending on the application, there may be certain properties among the optimal vectors {w∗C1 , . . . , {w∗CQ }. This Mutual information among tasks could be used to improve the estimation accuracy. Among the possible options, a simple yet effective, Euclidian distance based regularizer was enforced in [18]. The squared Euclidean distance regularizer is given as ∆(wCk , wCl ) = kwCk − wCl k2 (6) To estimate the unknown parameter vectors {w∗C1 , . . . , {w∗CQ }, it was shown in [18] that the local cost (5) and the regularizer (6) can be combined at the level of each cluster. This formulation led to the following estimation problem defined in terms of Q Nash equilibrium problems [20], where each cluster Cj estimates w∗Cj by minimizing the regularized cost function JCj (wCj , w−Cj ): (P1 )     min JCj (wCj , w−Cj ) wCj P P  d (n)−uT  k (n)wCk 2  E{| k kuk (n)k | }+η with JCj (wCj , w−Cj ) = k∈Cj P (7) ρkl kwCk − wCl k2 k∈Cj l∈Nk \Cj for j = 1, . . . , Q. wCj (n) is the parameter vector associated with cluster Cj , η > 0 is a regularization parameter, and the symbol \ is the set difference. Note that we have kept the notation wCk in above equation to make the role of the regularization term clearer, even though we have wC(k) = wCj for all k in Cj . The notation w−Cj denotes the collection of weight vectors estimated by the other clusters, that is, w−Cj = {wCq ; q = 1, . . . , Q} − {wCj }. The non-negative coefficients ρkl aim at the adjusting the regularization strength. In [18]. The coefficients {ρkl } were chosen to satisfy the conditions:     ρkl > 0, if l ∈ Nk \ C(k),    X ρkl = 1, and ρkk ≥ 0,    l∈Nk \C(k)    ρkl = 0, otherwise (8) 5 We impose ρkl = 0 for all l ∈ / Nk \ C(k), since nodes belonging to the same cluster estimate the same parameter vector. The solution for the problem P1 requires that every node in the network should have the access to the statistical moments Ru,k and pud,k over its cluster, however, node k can only be assumed to have direct access to the information from its neighborhood Nk , which may include the nodes that are not part of the cluster C(k) Therefore, to enable a distributed solution that relies only on measured data from neighborhood, as mentioned in [18], [19] the cost function is relaxed into following form: X JC(k) (wk ) = clk E{| l∈Nk ∩C(k) + X dl (n) − uTl (n)wk 2 | }+η kuk (n)k X ρkl kwk − wl k2 l∈Nk \C(k) (9) blk kwk − wol k2 l∈Nk− ∩C(k) where the coefficients clk are non-negative and satisfy the conditions: N X clk = 1, and clk = 0 if k ∈ / Nl ∩ C(l) (10) k=1 and the coefficients blk are also non-negative. Following the same line of reasoning from [10], [11] in the single-task case, and extending the argument to problem (9) by using Nash-equilibrium properties [20], and by following same procedure mentioned in [10], [21] the following diffusion strategy of the adapt-then-combine (ATC) for clustered multi-task Normalized LMS (NLMS) is derived in distributed manner:   uk (n) T   ψ k (n + 1) = wk (n) + µ kε+uk (n)k2 [dk (n) − uk (n)wk (n)] + µk η    wk (n + 1) = P alk ψ l (n + 1) P ρkl (wl (n) − wk (n)) l∈Nk \C(k) (11) l∈Nk ∩C(k) By extending the above clustered multi-task diffusion strategy to data-reuse case, we can derive the following Affine projection algorithm (APA) [22] based clustered multi-task diffusion strategy:  −1   ψ k (n + 1) = wk (n) + µUTk (n) εI + Uk (n)UTk (n) [dk (n) − uk (n)wk (n)]       P ρkl (wl (n) − wk (n)) +µk η  l∈Nk \C(k)     P   alk ψ l (n + 1) wk (n + 1) = (12) l∈Nk ∩C(k) where η denotes a regularization parameter with small positive value, ε is employed to avoid the inversion of a rank deficient matrix Uk (n)UTk (n) and the input data matrix Uk (n), desired response vector dk (n) are given as follows      Uk (n) =     uk (n) uk (n − 1) .. . uk (n − P + 1)      ,    The clustered multi-task diffusion APA algorithm is given below:      dk (n) =     dk (n) dk (n − 1) .. . dk (n − P + 1)          (13) 6 Algorithm 1: Diffusion APA for clustered multi-task networks 0: Start wk (0) = 0 for all k, and repeat: −1 ψ k (n + 1) = wk (n) + µk UTk (n) εI + Uk (n)UTk (n) [dk (n) − Uk (n)wk (n)] X ρkl (wl (n) − wk (n)) + µk η wk (n + 1) = X (14) l∈Nk \C(k) alk ψ l (n + 1) l∈Nk ∩C(k) In a single-task network, there is a single cluster that consists of the entire set of nodes we get Nk ∩ C(k) = Nk and Nk \ C(k) = ∅ for all k, so that the expression (14) reduces to the diffusion adaptation strategy [10] as described in algorithm 2: Algorithm 2: Diffusion APA for single-task networks 0: Start wk (0) = 0 for all k, and repeat: −1 ψ k (n + 1) = wk (n) + µk UTk (n) εI + Uk (n)UTk (n) [dk (n) − Uk (n)wk (n)] X wk (n + 1) = alk ψ l (n + 1) (15) l∈Nk In the case of multi-task network where the size of each cluster is one, we have Nk ∩ C(k) = k and Nk \ C(k) = Nk− for all k. Then algorithm 1 degenerates into Algorithm 3. This is the instantaneous gradient counterpart of (14) for each node. Algorithm 3: Diffusion APA for multi-task networks 0: Start wk (0) = 0 for all k, and repeat: −1 wk (n + 1) = wk (n) + µk UTk (n) εI + Uk (n)UTk (n) [dk (n) − Uk (n)wk (n)] X + µk η ρkl (wl (n) − wk (n)) (16) l∈Nk− IV. M EAN -S QUARE E RROR P ERFORMANCE A NALYSIS A. Network Global Model The space-time structure of the algorithm leads to challenge in the performance analysis. To proceed, first, Let us define the global representations as ψ(n) = col{ψ 1 (n), ψ 2 (n), . . . , ψ N (n)}, w(n) = col{w1 (n), w2 (n), . . . , wN (n)} U(n) = diag{U1 (n), U2 (n), . . . , UN (n)}, d(n) = col{d1 (n), d2 (n), . . . , dN (n)} (17) where U(n) is an N M × LN block diagonal matrix. The LN × LN diagonal matrices D and η are defined as D = diag{µ1 IL , µ2 IL , . . . , µN IL } (18) η = diag{η1 IL , η2 IL , . . . , ηN IL } 7 to collect the local step-sizes and regularization parameters. From the linear model of the form (1), the global model at network level is obtained as d(n) = U(n)w⋆ + v(n) (19) where w⋆ (n) and v(n) are global optimal weight and noise vectors given as follows w⋆ (n) = col{w⋆1 (n), w⋆2 (n), . . . , w⋆N (n)} (20) v(n) = col{v1 (n), v2 (n), . . . , vN (n)} To facilitate analysis, the network topology is assumed to be static (i.e. alk (n) = alk ). This assumption does not compromise the algorithm derivation or its operation, and is used for analysis only. The analysis presented in [23] and [24] serves as the basis for this work. Using the above expressions, the global model of multi-task diffusion APA is therefore formulated as follows: h i  −1 w(n + 1) = A w(n) + D UT (n) εI + U(n)UT (n) [d(n) − U(n)w(n)] − Dη Q w(n) (21) where A = AT ⊗ IL (22) Q = ILN − P ⊗ IL with ⊗ denoting the Kronecker product, A is the N × N symmetric matrix that defines the network topology and P is the N × N asymmetric matrix that defines regularizer strength among the nodes with ρkk = 1 if Nk \ C(k). Now the objective is to study the performance behavior of the multi-task diffusion APA governed by the form (21). B. Mean Error Behavior Analysis The global error vector e(n) is related to the local error vectors ek (n) as e(n) = col{e1 (n), e2 (n), . . . , eN (n)} e (n) = w⋆ − w(n), the global weight error vector can be rewritten as By denoting w (23) i e(n) = [d(n) − U(n)w(n)] = U(n)e w(n) + v(n) = ea (n) + v(n) (24) ea (n) = U(n)e w(n) (25) where 8 Using these results the recursive update equation of global weight error vector can be written as h i  −1  −1  e (n + 1) = A w e (n) − DUT (n) εI + U(n)UT (n) e (n) − w⋆ w U(n)e w(n) − DUT (n) εI + U(n)UT (n) v(n) − Dη Q w h i  −1 e (n) − AI DUT (n) εI + U(n)UT (n) = A ILN − D Z(n) − D η Q w v(n) + A D η Q w⋆ (26)  −1 where Z(n) = UT (n) εI + U(n)UT (n) U(n). Taking the expectation E[·] on both sides, and using the statistical inde- pendence between wk (n) and Uk (n) (i.e., independence assumption), and recalling that vk (n) is zero-mean i.i.d and also independent of Uk (n) and thus of wk (n) we can write      e (n + 1) = A ILN − D E Z(n) − D η Q E[e w(n)] + A D η Q w⋆ E w  (27) Then, for any initial condition, in order to guarantee the stability of the multi-task diffusion APA strategy in the mean sense, the step size µk has to be chosen to satisfy    λmax A ILN − D Z − D η Q < 1 (28) where Z = E[Z(n)], and λmax (·) denotes the maximum eigen value of its argument matrix. Therefore, using the norm inequalities and recalling the fact that the combining matrix A is a left stochastic matrix ( i.e., block maximum norm is equal to one), we have     kA ILN − D Z − D η Q kb,∞ ≤ k ILN − D Z − D η Q kb,∞   ≤ k ILN − D Z − D η + D η P ⊗ IL kb,∞ (29) Let A be the an L × L matrix, then from Gershgorin circle theorem, we have: |λ − ai,i | ≤ X |ai,j | (30) j6=i Therefore, using the above result, and recalling the fact that P is a right stochastic matrix, a sufficient condition for (29) to hold is to choose µ such that 0 < µk < 2 maxk {λmax ( Zk )} + 2η (31) h i  −1 Uk (n) . Above result clearly shows that the mean stability limit of the clustered where Zk = E UTk (n) εI + Uk (n)UTk (n) multi-task diffusion APA is lower than the diffusion APA due to the presence of η. In steady-state i.e., as n → ∞ the asymptotic mean bias is given by "  #−1  e (n) = ILN − A ILN − D Z − D η Q lim E w A D η Q w⋆ n→∞  (32) 9 C. Mean-Square Error Behavior Analysis The recursive update equation of weight error vector can also be rewritten as follows: where  −1 e (n + 1) = G(n) w e (n) − A DUT (n) εI + U(n)UT (n) v(n) + r w h i G(n) = A ILN − D Z(n) − D η Q (33) (34) r = A D η Q w⋆ Using the standard independent assumption between Uk (n) and wk (n) and E[v(n)] = 0, the mean square of the weight error e (n + 1), weighted by any positive semi-definite matrix Σ that we are free to choose, satisfies the following relation: vector w    T        e (n) E G T (n) Σ r + rT Σ E G(n) E w e (n) + krk2Σ Eke w(n + 1)k2Σ = Eke w(n)k2EΣ′ + E vT (n)YΣ (n)v(n) + E w (35) where h i ′ EΣ = E G T (n) Σ G(n) and    = AT Σ A − E Z(n)] D AT ΣA − AT ΣA DE Z(n) − AT ΣS − ST ΣA       + E Z(n) D ATI ΣS + ST Σ AD E Z(n) + E UT (n)YΣ U(n) + ST Σ S  −1  −1 YΣ = εI + U(n)UT (n) U(n) D AT Σ A D UT (n) εI + U(n)UT (n) (36) (37) S = ADηQ In order to study the behavior of the multi-task diffusion APA algorithm, the following moments in (35) and (36) must be evaluated: h i E Z(n) D AT Σ A D Z(n) h i  −1  −1 E vT (n) εI + U(n)UT (n) U(n) D AT Σ A D UT (n) εI + U(n)UT (n) v(n) (38) To extract the matrix Σ from the expectation terms, a weighted variance relation is introduced by using L2 N 2 × 1 column vectors: σ = bvec{Σ} and ′ ′ σ = bvec{EΣ } (39) where bvec{·} denotes the block vector operator. In addition, bvec{·} is also used to recover the original matrix Σ from σ. One property of the bvec{·} operator when working with the block Kronecker product [26] is used in this work, namely, bvec{QΣP} = (PT ⊗b Q) σ (40) 10 where P ⊗b Q denotes the block Kronecker product [25], [26] of two block matrices. Using (40) to (36) after block vectorization, the following terms on the right side of (36) are given by  n o  bvec AT Σ A = AT ⊗b AT σ   n  o      ILN ⊗b D AT ⊗b AT σ bvec E Z(n) D AT Σ A = ILN ⊗b E Z(n)    n o    bvec AT Σ A D E[Z(n)] = E Z(n) ⊗b ILN D ⊗b ILN AT ⊗b AT σ n o n o bvec AT Σ S = bvec AT Σ A D η Q      = QT ⊗b ILN η ⊗b ILN D ⊗b ILN AT ⊗b AT σ n o n o bvec ST Σ A = bvec QT η D AT Σ A      = ILN ⊗b QT ILN ⊗b η ILN ⊗b D AT ⊗b AT σ n  o n  o   bvec E Z(n) D AT Σ S = bvec E Z(n) D AT Σ A D η Q        T Q ⊗b ILN η ⊗b ILN D ⊗b D AT ⊗b A σ = ILN ⊗b E Z(n) n n  o  o bvec ST Σ A D E Z(n) = bvec QT η DAT Σ A D E Z(n)       = E[Z(n)] ⊙ ILN ILN ⊗b QT ILN ⊗b η D ⊗b D AT ⊗b AT σ n o n  o bvec UT (n)YΣ U(n) = bvec E Z(n) D AT Σ A D Z(n)     = E Z(n) ⊗b Z(n) D ⊗b D AT ⊗b AT σ n o n o bvec ST Σ S = bvec QT η D AT Σ A D η Q      = QT ⊗b QT η ⊗b η D ⊗b D AT ⊗b AT σ (41) (42) (43) (44) (45) (46) (47) (48) (49) ′ Therefore, a linear relation between the corresponding vectors {σ, σ }is formulated by ′ σ = Fσ (50) 11 where F is an L2 N 2 × L2 N 2 matrix and given by       F=          ILN − ILN ⊗b Z ILN ⊗b D − Z ⊗b ILN D ⊗b ILN       − QT ⊗b ILN η ⊗b ILN D ⊗b ILN − ILN ⊗b QT ILN ⊗b η ILN ⊗b D     + ILN ⊗b Z QT ⊗b ILN η ⊗b ILN D ⊗b D      + Z ⊗b ILN ILN ⊗b QT ILN ⊗b η D ⊗b D + Π D ⊗b D    + QT ⊗ b QT η ⊗ b η D ⊗ b D         AT ⊗b AT      (51) h i where Π = E Z(n) ⊗b Z(n) Let Λv = E[v(n)vT (n)] denote a N M × N M diagonal matrix, whose entries are the noise 2 variances σv,k for k = 1, 2, · · · , N and given by 2 2 2 Λv = diag{σv,1 IM , σv,2 IM , · · · , σv,N IM } (52)   Using the independence assumption of noise signals, the term E vT (n)YΣ (n)v(n) can be written as     E vT (n) YΣ (n) v(n) = T r A D E[Φ] D AT Σ (53) = γT σ  −1  −1 where Φ = UT (n) εI + U(n)UT (n) Λv (n) εI + U(n)UT (n) U(n) and  γ = vec A DE[Φ] DT AT    = A ⊗ A D ⊗ D vec E[WT Λv W]     = A ⊗ A D ⊗ D E WT ⊗ WT γ v (54)  −1 with W = εI + U(n)UT (n) U(n) and γ v = vec{Λv }.  Finally, let us define the f r, E[e w(n)], σ as the last three terms on the right hand side of the (35), i.e,   T        e (n) E G T (n) Σ r + rT Σ E G(n) E w e (n) f r, E[e w(n)], σ = krk2Σ + E w (55) Each term can be evaluated as follows. Now, let us consider the term Ekrk2Σ , that can be written krk2Σ = =  oT n T σ bvec A D η Q w⋆ w⋆ QT η D AT (56) rTb σ where    n   T o rb = E A ⊗b A D ⊗b D η ⊗b η Q ⊗b Q bvec w⋆ w⋆ (57) 12 h i e T (n) G T (n) Σ r that can be simplified as follows: Consider the second term E w h i  h i  e T (n) G T (n)Σ r = T r E r w e T (n) G T (n) Σ E w = (58) αT1 (n) σ where      I ⊗ Q I ⊗ D I ⊗ η LN b LN b LN b        α1 (n) = A ⊗b A  I ⊗ Q − D ⊗ D I ⊗ η Z ⊗b ILN LN b b LN b      − D ⊗b D η ⊗b η Q ⊗b Q h i e (n) can be written as follows: In the same way, third term E rT Σ G(n) w  n o  T  bvec w⋆ E[e w (n)]   i  h i  h e (n) = T r E G(n) w e (n) rT Σ E rT Σ G(n) w (59) (60) = αT2 (n) σ where     α2 (n) = A ⊗b A       D ⊗b ILN η ⊗b ILN Q ⊗b ILN     − D ⊗b D η ⊗b ILN Q ⊗b ILN ILN ⊗b Z    − D ⊗b D η ⊗b η Q ⊗b Q   n   o ⋆ T  bvec E[e w (n)] w   (61) Therefore, the mean-square behavior of the multi-task diffusion APA algorithm is summarized as follows:  Eke w(n + 1)k2σ = Eke w(n)k2Fσ + γ T σ + f r, E[e w(n)], σ   = Eke w(n)k2Fσ + γ T σ + rTb + αT1 (n) + αT2 (n) σ (62) Therefore, the multi-task diffusion strategy presented in (14) is mean square stable if the matrix F is stable. Iterating the recursion (62) starting from n = 0, we get Eke w(n + 1)k2σ = Eke w(0)k2Fn+1 σ + γ T n X i=0 Fi σ + n X f r, E[e w(n − i)], Fi σ i=0  (63) e (0) = w⋆ − w(0). If the matrix F is stable then the first and second terms in the above equation with initial condition w converge to a finite value as n → ∞. Now, let us consider the third term on the RHS of the (63). We know that E[e w(n)] is uniformly bounded because (27) is a BIBO stable recursion with bounded driving term A D η Q w⋆ . Therefore, from (55)  f r, E[e w(n − i)], Fi σ can be written as    f r, E[e w(n − i)], Fi σ = rTb + αT1 (n − i) + αT2 (n − i) Fi σ (64) Provided that F is stable and there exist a matrix norm, denoted by k · kp such that kFkp = cp < 1. Applying this norm to f  and using the matrix norms and triangular inequality, we can write kf r, E[e w(n − i)], Fi σ k ≤ v cip , given v is a small positive constant. Therefore Eke w(n + 1)k2σ converges to a bounded value as n → ∞, and the algorithm is said to be mean square 13 stable. By selecting Σ = 1 N ILN Eke w(n + we can relate Eke w(n + 1)k2σ and Eke w(n)k2σ as follows: 1)k2σ = Eke w(n)k2σ − n−1 X T n + γ F σ − Eke w(0)k i 2 f r, E[e w(n − 1 − i)], F σ i=0  I(LN )2 −F Fn σ  + n X f r, E[e w(n − i)], Fi σ i=0  (65) we can rewrite the last two terms in the above equation as, n X i=0 X  n−1    f r, E[e w(n − i)], Fi σ − f r, E[e w(n − 1 − i)], Fi σ = rTb Fn σ + αT1 (n) + αT2 (n) + Γ(n) σ (66) i=0 where Γ(n) = n  X i=1 n−1   X αT1 (n − 1 − i) + αT2 (n − 1 − i) Fi σ αT1 (n − i) + αT2 (n − i) Fi σ − (67) i=0 Therefore, the recursion presented in (62) can be rewritten as, w(0)k2 w(n)k2σ + γ T Fn σ − Eke Eke w(n + 1)k2σ = Eke Γ(n + 1) = Γ(n)F + with Γ(0) = 01×(LN )2 . h I(LN )2 −F i  αT1 (n) + αT2 (n) [F − I(LN )2 ]  Fn σ   + rTb Fn σ + αT1 (n) + αT2 (n) + Γ(n) (68) Steady-state MSD of the multi-task diffusion APA strategy is given as follows lim Eke w(n)k2 n→∞  = γ T σ + f r, E[e w(∞)], σ I(LN )2 −F σ  (69) D. New Approach to Improve the Performance of Clustered multi-task diffusion APA Clustered multi-task diffusion strategy presented in (14) has mainly 2 drawbacks • At time instance n, assume that the node l exhibiting poor performance over the node k. The multi-task diffusion strategy forces the node k to learn from node l during the adaptation step where l ∈ Nk \ Ck . This affects the performance in transient state. • For all l ∈ Nk \ Ck we have w∗k ≃ w∗l i.e, only the underlying system is same. However, the multitask diffusion strategy forces the node k to learn from node l even in the steady state. This hampers the steady state performance of the algorithm. To address these problems a control variable called similarity measure, δkl (n) is introduced to control the regularizer term in the multi-task diffusion strategy. At each time instance n, node k has access to its neighborhood filter coefficient vectors. Since the node is learning from its neighborhood filter coefficient vectors, it is reasonable to check the similarity among the filter coefficient vectors. The similarity measure is calculated as follows δkl (n) =    1 2 1 + sign σk2 (n) − σkl (n) 2 (70) 14 2 where σk2 (n) and σlk (n) are estimated error variances and can be calculated as h i2 σk2 (n) = λ σk2 (n − 1) + (1 − λ) dk (n) − uTk (n)wk (n) h i2 2 2 σkl (n) = λ σkl (n − 1) + (1 − λ) dk (n) − uTk (n)wl (n) (71) for l ∈ Nk \ C(k) and λ is a positive constant with λ ∈ [0, 1]. 2 To explain, suppose that at index n, the node l performs better than node k, i.e., σkl (n) < σk2 (n). Then for node k the h i  2 similarity measure δkl (n) = 12 1 + sign σk2 (n) − σkl (n) = 1, which implies that node k would learn the weight information from node l by adding the difference of their current weight vectors, i.e., [wl (n) − wk (n)] as a correction term to its weight 2 update. On the other hand, suppose the node l does not perform better than node k, i.e., σkl (n) > σk2 (n). Then for node k the i h  2 (n) = 0, which implies that node k would neglect the weight vector similarity measure δkl (n) = 12 1 + sign σk2 (n) − σkl wl (n). Thus improves the convergence rate and steady state performance over the multi-task diffusion strategy presented in (14). Therefore, by taking the similarity measure, δkl (n) into account the modified clustered multi-task diffusion APA is given below −1 ψ k (n + 1) = wk (n) + µk UTk (n) εI + Uk (n)UTk (n) [dk (n) − UTk (n)wk (n)] X   ′ ρkl (n) wl (n) − wk (n) + µk η l∈Nk \C(k) wk (n + 1) = X (72) alk ψ l (n + 1) l∈Nk ∩C(k) ′ where ρkl (n) = ρkl δkl (n). Therefore, using the above expressions, the global model of modified multi-task diffusion APA is formulated as follows: h i  −1 w(n + 1) = A w(n) + D UT (n) εI + U(n)UT (n) e(n) − D η Qδ (n) w(n) (73) where Qδ (n) = Dδ (n) − Pδ (n) ⊗ IL (74) Dδ (n) = diag{δ1 (n)IL (n), δ2 (n)IL (n), · · · , δN (n)IL (n)} (75) and with δk (n) = X l∈Nk \Ck ′ ρkl (n) = X ρkl δkl (n) (76) l∈Nk \Ck the matrix Pδ (n) = P ⊙ δ(n) (’⊙’ indicates the Hadamard product) is the N × N asymmetric matrix that defines regularizer strength among the nodes with δ k (n) = 1 and Pδ,kk (n) = 1 if Nk \ C(k) is empty. The matrices A, D, U(n) are as same as 15 the matrices defined in the network model section. Now the objective is to study the performance behavior of the multi-task diffusion APA governed by the form (73). E. Mean Error Behavior Analysis e (n) = w⋆ − w(n) the recursive update equation of global weight error vector can be written as By denoting w h i  −1 e (n + 1) = A ILN − D Z(n) − D η Qδ (n) w e (n) − A DUT (n) εI + U(n)UT (n) w v(n) + A D η Qδ (n) w⋆ (77) Taking the expectation of both sides, in addition to the statistical independence between wk (n) and Uk (n) (i.e., independence assumption) we are assuming statistical independence between wk (n) and δkl (n) and also recalling that vk (n) is zero-mean i.i.d and also independent of Uk (n) and thus of wk (n) we can write     e (n + 1) = A ILN − D Z − D η Qδ E[e E w w(n)] + A D η Qδ w⋆ (78) where Z = E[Z(n)] and Qδ = E[Qδ (n)]. The quantity Qδ is given as follows:   Qδ = E Dδ (n) − Pδ (n) ⊗ IL = Dδ − Pδ ⊗ IL (79) where Dδ = E[Dδ ], Pδ = E[Pδ (n)] and Qδ,ij =                P ρi l E[δi l (n)] if i=j if i 6= j l∈Ni \Ci ρi j E[δi j (n)] 0 (80) otherwise Then, for any initial condition, in order to guarantee the stability of the modified multi-task diffusion APA strategy in the mean sense, if the step size chosen to satisfy    λmax A ILN − D Z − D η Qδ < 1 (81) Now using the same arguments that are used in II. B, we will have      λmax A ILN − D Z − D η Qδ ≤ k ILN − D Z − D η Dδ + D η Pδ ⊗ IL kb,∞ (82) From Gershgorin circle theorem, a sufficient condition for (81) to hold is to choose µk such that 0 < µk < where δ k = E[δ k (n)] = P 2 maxk {λmax (Zk )} + 2 η maxk (δ k ) (83) ρkl E[δkl (n)]. Recalling the fact that δkl (n) is equal to either 0 or 1, we can write 0 ≤ l∈Nk \Ck E[δkl (n)] ≤ 1 that imply δ k ≤ 1. Therefore, the presence of similarity measure, δkl (n) makes the modified multi-task diffusion strategy mean stability is better than the multi task diffusion strategy mentioned in (14) however, lower than the 16 diffusion APA due to the presence of η. In steady-state i.e., as n → ∞ the asymptotic mean bias is given by "  #−1   e (n) = ILN − A ILN − D Z − D η Qδ lim E w A D η Qδ w ⋆ n→∞ (84) F. Mean-Square Error Behavior Analysis The recursive update equation of the modified multi-task diffusion APA weight error vector can also be rewritten as where  −1 e (n + 1) = G δ (n)e w w(n) − A D UT (n) εI + U(n)UT (n) v(n) + rδ i h G δ (n) = A ILN − D Z(n) − Dη Qδ (85) (86) r δ = A D η Qδ w ⋆ In addition to standard independent assumption between Uk (n) and wk (n) and E[v(n)] = 0 that was taken in II.C, here e (n + 1), we assume statistical independence between δkl (n) and wk (n). Then the mean square of the weight error vector w weighted by any positive semi-definite matrix Σ that we are free to choose, satisfies the following relation:    T  T     e (n) E G δ (n) Σ rδ + rTδ Σ E G δ (n) E w e (n) + krδ k2Σ Eke w(n + 1)k2Σ = Eke w(n)k2EΣ′ + E vT (n)YΣ (n)v(n) + E w δ (87) where h i ′ EΣδ = E G Tδ (n) Σ G δ (n) and = AT Σ A − Z D AT Σ A − AT Σ A D Z − AT Σ E[S δ ] − E[S Tδ ] Σ A   + Z D AT Σ E[S δ ] + E[S Tδ ] Σ A D Z + E UT (n)YΣ U(n) + E[S Tδ Σ S δ ]  −1  −1 YΣ = εI + U(n)UT (n) U(n) D AT Σ A D UT (n) εI + U(n)UT (n) (88) (89) S δ = A D η Qδ Following the same procudre mentioned in II. C, to extract the matrix Σ from the expectation terms, a weighted variance relation is introduced by using L2 N 2 × 1 column vectors: σ = bvec{Σ} and σ δ = bvec{EΣδ } (90) with a linear relation between the corresponding vectors {σ, σ δ } σ δ = Fδ σ (91) 17 where Fδ is an L2 N 2 × L2 N 2 matrix and given by       Fδ =           ILN − ILN ⊗b Z ILN ⊗b D − Z ⊗b ILN D ⊗b ILN      T T − Qδ ⊗b ILN η ⊗b ILN D ⊗b ILN − ILN ⊗b Qδ ILN ⊗b η ILN ⊗b D     T + ILN ⊗b Z Qδ ⊗b ILN η ⊗b ILN D ⊗b D     T + Z ⊗b ILN ILN ⊗b Qδ ILN ⊗b η D ⊗b D + Π D ⊗b D    + E[QTδ ⊗b QTδ ] η ⊗b η D ⊗b D         AT ⊙ AT      (92)   where Π = E Z(n) ⊗b Z(n)   The noise term E vT (n)YΣ (n)v(n) = γ T σ.  Finally, let us define the f rδ , E[e w(n)], σ as the last three terms on the right hand side of the (35), i.e, where   T  T     e (n) E G δ (n) Σ rδ + rTδ Σ E Gδ (n) E w e (n) f rδ , E[e w(n)], σ = krδ k2Σ + E w   = rTb,δ + αTδ,1 + αTδ,2 σ     h n  T o rb,δ = E A ⊗b A D ⊗b D η ⊗b η E Qδ ⊗b Qδ bvec w⋆ w⋆  and    I ⊗ I ⊗ D I ⊗ η Q LN b LN b LN b δ        αδ,1 (n) = A ⊗b A   − D ⊗b D ILN ⊗b η ILN ⊗b Qδ Z ⊗b ILN     − D ⊗b D η ⊗b η Qδ ⊗b Qδ     αδ,2 (n) = A ⊗b A       D ⊗b ILN η ⊗b ILN Qδ ⊗b ILN     − D ⊗b D η ⊗b ILN Qδ ⊗b ILN ILN ⊗b Z     − D ⊗b D η ⊗b η E Qδ ⊗b Qδ (93) (94)   n o  T  bvec w⋆ E[e w (n)]     n   o ⋆ T  bvec E[e w (n)] w   (95) (96) Therefore, the mean-square behavior of the modified multi-task diffusion APA algorithm is summarized as follows: Eke w(n + 1)k2σ = Eke w(n)k2Fδ σ + γ T σ + f rδ , E[e w(n)], σ  (97) Therefore, the modified multi-task diffusion APA strategy presented in (72) is mean square stable if the matrix Fδ is stable. Iterating the recursion (97) starting from n = 0, we get Eke w(n + 1)k2σ = Eke w(0)k2Fn+1 σ + γ T δ n X i=0 Fiδ σ + n X i=0 f rδ , E[e w(n − i)], Fiδ σ  (98) e (0) = w⋆ − w(0). If the matrix Fδ is stable then the first and second terms in the above equation with initial condition w converge to a finite value as n → ∞. Now, let us consider the third term on the RHS of the (98). We know that E[e w(n)] is uniformly bounded because (78) is a BIBO stable recursion with bounded driving term A D η E[Qδ ] w⋆ . Therefore, from 18  (93) f rδ , E[e w(n − i)], Fiδ σ can be written as    f rδ , E[e w(n − i)], Fiδ σ = rTb,δ + αTδ,1 (n − i) + αTδ,2 (n − i) Fiδ σ (99) Provided that Fδ is stable and there exist a matrix norm, denoted by k · kp such that kFδ kp = cp,δ < 1. Applying this norm  to f and using the matrix norms and triangular inequality, we can write kf rδ , E[e w(n − i)], Fiδ σ k ≤ vcip , given v is a small positive constant. Therefore Eke w(n + 1)k2σ converges to a bounded value as n → ∞, and the algorithm is said to be mean square stable. By selecting Σ = 1 N ILN we can relate Eke w(n + 1)k2σ and Eke w(n)k2σ as follows: Eke w(n + 1)k2σ = Eke w(n)k2σ + γ T Fnδ σ − Eke w(0)k2 + n X f rδ , E[e w(n − i)], Fiδ  σ − i=0  I(LN )2 −Fδ Fn δ σ n−1 X f rδ , E[e w(n − 1 − i=0 i)], Fiδ σ  (100) we can rewrite the last two terms in the above equation as, n X i=0 X   n−1   f rδ , E[e w(n − i)], Fiδ σ − f rδ , E[e w(n − 1 − i)], Fiδ σ = rTb,δ Fn σ + αTδ,1 (n) + αTδ,2 (n) + Γδ (n) σ (101) i=0 where Γδ (n) = n−1 n    X X αT1,δ (n − 1 − i) + αT2,δ (n − 1 − i) Fiδ σ αT1,δ (n − i) + αT2,δ (n − i) Fiδ σ − (102) i=0 i=1 Therefore, the recursion presented in (97) can be rewritten as, w(0)k2 Eke w(n + 1)k2σ = Eke w(n)k2σ + γ T Fnδ σ − Eke  I(LN )2 −Fδ Fn δσ h i  Γδ (n + 1) = Γδ (n)F + αT1,δ (n) + αT2,δ (n) [Fδ − I(LN )2 ]   + rTb,δ Fnδ σ + αT1,δ (n) + αT2,δ (n) + Γδ (n) (103) with Γδ (0) = 01×(LN )2 . Steady-state MSD of the modified multi-task diffusion APA strategy is given as follows lim Eke w(n)k2 n→∞  = γ T σ + f rδ , E[e w(∞)], σ I(LN )2 −Fδ σ  (104) V. S IMULATION R ESULTS A network consists of 9 nodes with the topology shown in Fig. 2 was considered for simulations. The nodes were divided into 3 clusters: C1 = {1, 2, 3}, C2 = {4, 5, 6}, and C3 = {7, 8, 9}. First, for theoretical performance comparison purpose, we first considered randomly generated two dimension vectors of the form w∗Ck = w0 + δCk wCk where δC1 = 0.025, δC2 = −0.025 and δC3 = 0.015. The input regressors uk (n) were taken from zero mean, Gaussian distribution with correlation matrices Ru,k = IL and the observation noises were i. i. d zero-mean Gaussian random variables, independent of any other signals with noise variance σv2 = 0.001. The multi-task diffusion APA algorithm was run with different step sizes and regularization parameters. Regularization strength ρkl was set to ρkl = |Nk \ C(k)|−1 for l ∈ Nk \ C(k), and ρkl = 0 for any other l. This 19 C1 1 2 3 9 C3 4 8 7 5 C2 6 Figure 2: Network Topology settings usually leads to asymmetrical regularization weights. The coefficient matrix C was taken to be identity matrix and the combiner coefficients alk were set according to Metropolis rule. Simulations were carried out to illustrate the performance of several learning strategies: 1) the non-cooperative APA algorithm, 2) the multi-task algorithm (Algorithm 3), and 3) the clustered multi-task algorithm (Algorithm 1). The non-cooperative algorithm was obtained by assigning a cluster to each node and setting η = 0. The multi-task algorithm was obtained by assigning a cluster to each node and setting η 6= 0. Note that algorithm 2 was not considered for comparison since it is a single-task estimation method. Normalized MSD was taken as the performance parametric to compare the diffusion strategies. Projection order was taken to be 4 and the initial taps were chosen to be zero. 0 Simulated MSD Theoritical MSD −5 Normalized MSD (dB) −10 −15 −20 µ = 0.2, η = 0.0001 −25 µ = 0.2, η = 1 µ = 0.1, η = 0.0001 −30 −35 0 10 20 30 40 50 60 70 80 90 100 Iterations (n) Figure 3: Comparison of transient NMSD for different step-sizes and regularization parameters. Secondly, the modified multi-task diffusion APA is compared with multi-task diffusion APA. For that, randomly generated coefficient vectors of the form w∗Ck = w0 + δCk wCk with L = 256 taps length were chosen as δC1 = 0.025, δC2 = −0.025 and 20 δC3 = 0.015. The input signal vectors were taken from zero mean, Gaussian distribution with correlation statistics as shown in the Fig. 4, and the observation noises were i. i. d zero-mean Gaussian random variables, independent of any other signals with noise variances as shown in the Fig. 5. 1 AR(1) pole, α k 0.8 0.6 0.4 0.2 0 1 2 3 4 5 6 7 8 9 8 9 Node number (k) Figure 4: Input signal Statistics. 0.02 2 σv, k 0.015 0.01 0.005 0 1 2 3 4 5 6 7 Node number (k) Figure 5: Noise statistics. Projection order was taken to be 8 and the initial taps were chosen to be zero. The step-size and regularization parameters (µ, η) were adjusted to compare the steady state MSD and convergence rate properly. Simulation results were obtained by averaging 50 Monte-Carlo runs. The learning curves of diffusion strategies were presented in Fig. 6. It can be observed that the performance of the non-cooperative strategy was poor as nodes do not collaborate for additional benefit. In the case of multi-task diffusion strategy the performance is improved over non-cooperative strategy due to regularization between nodes. The cluster information in addition to regularization among nodes in the clustered multi task results in better performance over the non-cooperative and multi task diffusion strategies. The extra information information in the regularization among nodes results in great improvement in the performance of modified clustered multi-task diffusion strategy. VI. C ONCLUSIONS In this paper, we presented the diffusion APA strategies which are suitable for multi-task networks and also robust against the correlated input conditions. The performance analysis of the proposed multi-task diffusion APA is presented in mean and mean square sense. By introducing similarity measure, the modified multi-task diffusion APA algorithms is proposed to achieve the improved performance over the multi-task diffusion strategies existed in literature. R EFERENCES [1] A. H .Sayed, S.-Y. Tu, J. Chen, X. Zhao and Z. J. Towfic, “Diffusion strategies for adaptation and learning over networks,” in IEEE Signal Process. Mag., vol. 30, no. 3, pp. 155-171, May 2013. 21 0 Non− cooperative APA −4 Multi−task diffusion APA Clustered multi−task diffusion APA Modified clustered multi−task diffusion APA Normalized MSD (dB) −12 −20 −28 −36 −44 −52 0 1,000 2,000 3,000 4,000 Iterations (n) Figure 6: Comparison of Modified Multi-task diffusion APA with multi-task diffusion APA [2] A. H .Sayed, R. Chellappa and S. Theodoridis “Diffusion adaptation over networks,” in Academic Press Library in Signal Processing, Eds. Amesterdam, The Netherlands: Elsevier, pp. 322-454, 2014. [3] A. H .Sayed, “Adaptive networks,” in Proc. of IEEE, vol. 102, no. 4, pp. 460-497, April 2014. [4] L. Xiao and S. Boyd, “Fast linear iterations for distributed averaging,” in Syst. Control Lett., vol. 53, no. 9, pp. 65-78, Sep. 2004. [5] P. Braca, S. Marano and V. Matta, “Running consensus in wireless sensor networks,” in in Proc. FUSION, Cologne, Germany, pp. 1-6, 2008. [6] S. Kar and J. M. F. Moura, “Distributed consensus algorithms in sensor networks with imperfect communication: Link failures and channel noise,” in IEEE Trans. Signal Process., vol. 57, no. 1, pp. 355-369, Jan. 2009. [7] D. P. Bertsekas, “A New Class of Incremental Gradient Methods for Least Squares Problems,” in SIAM Journal on Optimiz., vol. 7, no. 4, pp. 913-926, Nov. 1997. [8] C. G. Lopes and A. H. Sayed, “Incremental adaptive strategies over distributed networks,” in IEEE Trans. Signal Process., vol. 55, no. 8, pp. 4064-4077, Aug. 2007. [9] Leilei Li, J. A. Chambers, C. G. Lopes, and A. H. Sayed, “Distributed estimation over an adaptive incremental network based on the Affine Projection Algorithm,” in IEEE Trans. Signal Process., vol. 58, no. 1, pp. 151-164, 2010. [10] C. G. Lopes and A. H. Sayed, “Diffusion least-mean squares over adaptive networks: Formulation and performance analysis,” in IEEE Trans. Signal Process., vol. 56, no. 7, pp. 3122-3136, Jul. 2008. [11] F. S. Cattivelli and A. H. Sayed, “Diffusion LMS stretagies for distributed estimation,” in IEEE Trans. Signal Process., vol. 58, no. 3, pp. 1035-1048, May. 2010. [12] Leilei Li and J. A. Chambers, “Distributed adaptive estimation based on the APA algorithm over diffusion networks with changing topology,” in Proc. IEEE SSP, pp.757-760, Aug. 2009. [13] J. Chen and A. H. Sayed, “Diffusion adaptation strategies for distributed optimization and learning over networks,” in IEEE Trans. Signal Process., vol. 60, no. 8, pp. 4289-4305, Aug. 2012. [14] S.-Y. Tu and A. H. Sayed, “Diffusion strategies outperform consensus strategies for distributed estimation over adaptive networks,” in IEEE Trans. Signal Process., vol. 60, no. 12, pp. 6217-6234, Dec. 2012. 22 [15] X. ZHAO and A. H. Sayed, “Clustering via diffusion adaptation over networks,” in Proc. CIP, Parador de Baiona, Spain, pp. 1-6, May 2012. [16] N. Bogdanovic, J. Platta-Chaves and K. Berberidis, “Distributed incremental-based LMS for node specific parameter estimation over adaptive networks,” in Proc. IEEE ICASSP, Vancouver, BC, Canada, pp. 5425-5429, May 2013. [17] J. Chen, C. Richard and A. H. Sayed, “Diffusion LMS for clustered multitask networks,” in Proc. IEEE ICASSP, Florence, Italy, pp. 5487-5491, May 2014. [18] J. Chen, C. Richard and A. H. Sayed, “Multitask diffusion adaptation over networks,” in IEEE Trans. Signal Process., vol. 62, no. 16, pp. 4129-4144, Aug. 2014. [19] J. Chen, C. Richard and A. H. Sayed, “Diffusion LMS over multitask networks,” in to appear in IEEE Trans. Signal Process., Aug. 2015. [20] T. Basar and G. J. Olsder, “Dynamic noncooperative game theory,” in IEEE Trans. Signal Process., London, Academic Press, 2nd edition, 1995. [21] A. H. Sayed, Fundamentals of adaptive filtering, John Wiley and Sons, 2003. [22] K. Ozeki and T. Umeda, “An adaptive filtering algorithm using an orthogonal projection to an affine subspace and its properties,” in Transactions IECE Japan, vol. 67-A, no. 5, pp. 19-27, 1984. [23] H-C. Shin and A. H. Sayed, “Mean-square performance of a family of ane projection algorithms,” in IEEE Trans. Acoustics, Speech, Signal Process., vol. 52, no. 1, pp. 90-102, January, 2004. [24] N. R. Yousef and A. H. Sayed, “A unified approach to the steady- state and tracking analyzes of adaptive filters,” in IEEE Trans. Acoustics, Speech, Signal Process., vol. 49, no. 2, pp. 314-324, February, 2001. [25] R. H. Koning, H. Neudecker and T. Wansbeek, “Block Kronecker products and the vecb operator,” Economics Dept., Institute of Economics Research, Univ of Groningen, Groningen, The Netherlands, Research Memo. No. 351, 1990. [26] D. S. Tracy and R. P. Singh, “A new matrix product and its applications in partitioned matrix differentiation,” in Statistica Neerlandica., vol. 26, no. NR.4, pp. 143-157, 1972. [27] S. Haykin, Adaptive filter theory, 3rd , Prentice Hall. [28] B. Farhang-Boroujeny, Adaptive filters, John Wiley and Sons. [29] D. G. Manolakis, V. K. Ingle, and S. M. Kogon, Statistical and adaptive signal processing, McGraw-HILL.
3cs.SY
arXiv:1712.09687v1 [cs.NE] 27 Dec 2017 Combining Representation Learning with Logic for Language Processing Tim Rocktäschel A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy of University College London. Department of Computer Science University College London December 29, 2017 To Paula, Emily, Sabine, and Lutz. Acknowledgements I am deeply grateful to my supervisor and mentor Sebastian Riedel. He always has been a great source of inspiration and supportive and encouraging in all matters. I am particularly amazed by the trust he put into me from the first time we met, the freedom he gave me to pursue ambitious ideas, his contagious optimistic attitude, and the many opportunities he presented to me. There is absolutely no way I could have wished for a better supervisor. I would like to thank my second supervisors Thore Graepel and Daniel Tarlow for their feedback, as well as Sameer Singh whose collaboration and guidance made my start into the Ph.D. very smooth, motivating, and fun. Furthermore, I thank Edward Grefenstette, Karl Moritz Hermann, Thomas Kociský and Phil Blunsom, who I was fortunate to work with during my internship at DeepMind in Summer 2015. I am thankful for the guidance by Thomas Demeester, Andreas Vlachos, Pasquale Minervini, Pontus Stenetorp, Isabelle Augenstein, and Jason Naradowsky during their time at the University College London Machine Reading lab. In addition, thanks to Dirk Weissenborn for many fruitful discussions. I would also like to thank Ulf Leser, Philippe Thomas, and Roman Klinger for preparing me well for the Ph.D. during my studies at Humboldt-Universität zu Berlin. I had the pleasure to work with brilliant students at University College London. Thanks to Michal Daniluk, Luke Hewitt, Ben Eisner, Vladyslav Kolesnyk, Avishkar Bhoopchand, and Nayen Pankhania for their hard work and trust. Ph.D. life would not have been much fun without my lab mates Matko Bosnjak, George Spithourakis, Johannes Welbl, Marzieh Saeidi, and Ivan Sanchez. Thanks to Matko, Dirk, Pasquale, Thomas, Johannes, and Sebastian for feedback on this thesis. Furthermore, I thank my examiners Luke Dickens and Charles Sutton for their extremely helpful in-depth comments and corrections of this thesis. Thank you, Federation Coffee in Brixton, for making the best coffee in the world. Many thanks to Microsoft Research for supporting my work through its Ph.D. Scholarship Programme, and to Google Research for a Ph.D. Fellowship in Natural 6 Language Processing. Thanks to their generous support, as well as the funding from the University College London Computer Science Department, I was able to travel to all conferences that I wanted to attend. I am greatly indebted and thankful to my parents Sabine and Lutz. They sparked my interest in science, but more importantly, they ensured that I had a fantastic childhood. I always felt loved, supported, and well protected from the many troubles in life. Furthermore, I would like to thank Christa, Ulrike, Tillmann, Hella, Hans, Gretel, and Walter for their unconditional support over the last years. Lastly, thanks to the two wonderful women in my life, Paula and Emily. Thank you, Paula, for keeping up with my ups and downs during the Ph.D., your love, motivation, and support. Thank you for giving me a family, and for making us feel at home wherever we are. Emily, you are the greatest wonder and joy in my life. Declaration I, Tim Rocktäschel confirm that the work presented in this thesis is my own. Where information has been derived from other sources, I confirm that this has been indicated in the thesis. T IM ROCKT ÄSCHEL Abstract The current state-of-the-art in many natural language processing and automated knowledge base completion tasks is held by representation learning methods which learn distributed vector representations of symbols via gradient-based optimization. They require little or no hand-crafted features, thus avoiding the need for most preprocessing steps and task-specific assumptions. However, in many cases representation learning requires a large amount of annotated training data to generalize well to unseen data. Such labeled training data is provided by human annotators who often use formal logic as the language for specifying annotations. This thesis investigates different combinations of representation learning methods with logic for reducing the need for annotated training data, and for improving generalization. We introduce a mapping of function-free first-order logic rules to loss functions that we combine with neural link prediction models. Using this method, logical prior knowledge is directly embedded in vector representations of predicates and constants. We find that this method learns accurate predicate representations for which no or little training data is available, while at the same time generalizing to other predicates not explicitly stated in rules. However, this method relies on grounding first-order logic rules, which does not scale to large rule sets. To overcome this limitation, we propose a scalable method for embedding implications in a vector space by only regularizing predicate representations. Subsequently, we explore a tighter integration of representation learning and logical deduction. We introduce an end-to-end differentiable prover – a neural network that is recursively constructed from Prolog’s backward chaining algorithm. The constructed network allows us to calculate the gradient of proofs with respect to symbol representations and to learn these representations from proving facts in a knowledge base. In addition to incorporating complex first-order rules, it induces interpretable logic programs via gradient descent. Lastly, we propose recurrent neural networks with conditional encoding and a neural attention mechanism for determining the logical relationship between two natural language sentences. Impact Statement Machine learning, and representation learning in particular, is ubiquitous in many applications nowadays. Representation learning requires little or no hand-crafted features, thus avoiding the need for task-specific assumptions. At the same time, it requires a large amount of annotated training data. Many important domains lack such large training sets, for instance, because annotation is too costly or domain expert knowledge is generally hard to obtain. The combination of neural and symbolic approaches investigated in this thesis has only recently regained significant attention due to advances of representation learning research in certain domains and, more importantly, their lack of success in other domains. The research conducted under this Ph.D. investigated ways of training representation learning models using explanations in form of function-free first-order logic rules in addition to individual training facts. This opens up the possibility of taking advantage of the strong generalization of representation learning models, while still being able to express domain expert knowledge. We hope that this work will be particularly useful for applying representation learning in domains where annotated training data is scarce, and that it will empower domain experts to train machine learning models by providing explanations. Contents 1 2 Introduction 1.1 Aims . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 1.3 Thesis Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 Background 2.1 2.2 2.3 3 23 29 Function Approximation with Neural Networks . . . . . . . . . . . 29 2.1.1 Computation Graphs . . . . . . . . . . . . . . . . . . . . . 30 2.1.2 From Symbols to Subsymbolic Representations . . . . . . . 30 2.1.3 Backpropagation . . . . . . . . . . . . . . . . . . . . . . . 32 Function-free First-order Logic . . . . . . . . . . . . . . . . . . . . 34 2.2.1 Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 2.2.2 Deduction with Backward Chaining . . . . . . . . . . . . . 35 2.2.3 Inductive Logic Programming . . . . . . . . . . . . . . . . 38 Automated Knowledge Base Completion . . . . . . . . . . . . . . . 38 2.3.1 Matrix Factorization . . . . . . . . . . . . . . . . . . . . . 39 2.3.2 Other Neural Link Prediction Models . . . . . . . . . . . . 43 2.3.3 Path-based Models . . . . . . . . . . . . . . . . . . . . . . 44 Regularizing Representations by First-order Logic Rules 47 3.1 Matrix Factorization Embeds Ground Literals . . . . . . . . . . . . 47 3.2 Embedding Propositional Logic . . . . . . . . . . . . . . . . . . . 49 3.3 Embedding First-order Logic via Grounding . . . . . . . . . . . . . 52 3.4 3.3.1 Stochastic Grounding . . . . . . . . . . . . . . . . . . . . . 53 3.3.2 Baseline . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 3.4.1 3.5 Training Details . . . . . . . . . . . . . . . . . . . . . . . . 56 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . 57 14 4 Contents 3.5.1 Zero-shot Relation Learning . . . . . . . . . . . . . . . . . 57 3.5.2 Relations with Few Distant Labels . . . . . . . . . . . . . . 60 3.5.3 Comparison on Complete Data . . . . . . . . . . . . . . . . 60 3.6 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 Lifted Regularization of Predicate Representations by Implications 65 4.1 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 4.2 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 4.2.1 4.3 5 61 Training Details . . . . . . . . . . . . . . . . . . . . . . . . 69 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . 70 4.3.1 Restricted Embedding Space for Constants . . . . . . . . . 70 4.3.2 Zero-shot Relation Learning . . . . . . . . . . . . . . . . . 71 4.3.3 Relations with Few Distant Labels . . . . . . . . . . . . . . 71 4.3.4 Incorporating Background Knowledge from WordNet . . . 72 4.3.5 Computational Efficiency of Lifted Rule Injection . . . . . . 73 4.3.6 Asymmetry . . . . . . . . . . . . . . . . . . . . . . . . . . 73 4.4 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 End-to-End Differentiable Proving 5.1 5.2 5.3 77 Differentiable Prover . . . . . . . . . . . . . . . . . . . . . . . . . 79 5.1.1 Unification Module . . . . . . . . . . . . . . . . . . . . . . 80 5.1.2 OR Module . . . . . . . . . . . . . . . . . . . . . . . . . . 82 5.1.3 AND Module . . . . . . . . . . . . . . . . . . . . . . . . . 82 5.1.4 Proof Aggregation . . . . . . . . . . . . . . . . . . . . . . 83 5.1.5 Neural Inductive Logic Programming . . . . . . . . . . . . 84 Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 5.2.1 Training Objective . . . . . . . . . . . . . . . . . . . . . . 85 5.2.2 Neural Link Prediction as Auxiliary Loss . . . . . . . . . . 86 5.2.3 Computational Optimizations . . . . . . . . . . . . . . . . 87 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 5.3.1 Training Details . . . . . . . . . . . . . . . . . . . . . . . . 90 5.4 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . 91 5.5 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 Contents 6 7 Recognizing Textual Entailment with Recurrent Neural Networks 6.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1.1 Recurrent Neural Networks . . . . . . . . . . . . . . . 6.1.2 Independent Sentence Encoding . . . . . . . . . . . . . 6.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2.1 Conditional Encoding . . . . . . . . . . . . . . . . . . 6.2.2 Attention . . . . . . . . . . . . . . . . . . . . . . . . . 6.2.3 Word-by-word Attention . . . . . . . . . . . . . . . . . 6.2.4 Two-way Attention . . . . . . . . . . . . . . . . . . . . 6.3 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.1 Training Details . . . . . . . . . . . . . . . . . . . . . . 6.4 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . 6.4.1 Qualitative Analysis . . . . . . . . . . . . . . . . . . . 6.5 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.5.1 Bidirectional Conditional Encoding . . . . . . . . . . . 6.5.2 Generating Entailing Sentences . . . . . . . . . . . . . 6.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 97 97 100 101 101 102 104 105 106 106 106 108 110 110 111 111 Conclusions 115 7.1 Summary of Contributions . . . . . . . . . . . . . . . . . . . . . . 115 7.2 Limitations and Future Work . . . . . . . . . . . . . . . . . . . . . 116 Appendices 119 A Annotated Rules 119 B Annotated WordNet Rules 123 Bibliography 126 List of Figures 2.1 2.2 2.3 Two simple computation graphs with inputs (gray), parameters (blue), intermediate variables ui , and outputs (dashed). Operations are shown next to the nodes. . . . . . . . . . . . . . . . . . . . . . Computation graph for z = kC1i + P 1s − C1j k1 . . . . . . . . . . 31 31 Backward pass for the computation graph shown in Fig. 2.1a. . . . . 33 2.4 Simplified pseudocode for symbolic backward chaining (cycle detection omitted for brevity, see Russell and Norvig [2010], Gelder [1987], Gallaire and Minker [1978] for details). . . . . . . . . . . . 36 2.5 Full proof tree for a small knowledge base. . . . . . . . . . . . . . . 37 2.6 Knowledge base inference as matrix completion with true training facts (green), unobserved facts (question mark), relation representations (red), and entity pair representations (orange). . . . . . . . . . 41 2.7 A complete computation graph of a single training example for Bayesian personalized ranking with `2 -regularization. . . . . . . . . 42 3.1 Computation graph for rule in Eq. 3.7 where · denotes a placeholder for the output of the connected node. . . . . . . . . . . . . . . . . . 51 3.2 Given a set of training ground atoms, matrix factorization learns k-dimensional predicate and constant pair representations. Here, we also consider additional first-order logic rules (red) and seek to learn symbol representations such that the predictions (completed matrix) comply with these given rules. . . . . . . . . . . . . . . . . . . . . 53 3.3 Weighted MAP scores for zero-shot relation learning. . . . . . . . . 58 3.4 Weighted MAP of the various models as the fraction of Freebase training facts is varied. For 0% Freebase training facts we get the zero-shot relation learning results presented in Table 3.2. . . . . . . 59 3.5 Precision-recall curve demonstrating that the Joint method, which incorporates annotated rules derived from the data, outperforms existing factorization approaches (MF and Riedel13-F). . . . . . . 61 18 List of Figures 4.1 Example for implications that are directly captured in a vector space. 67 4.2 Weighted MAP for injecting rules as a function of the fraction of Freebase facts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 5.1 A module is mapping an upstream proof state (left) to a list of new proof states (right), thereby extending the substitution set SΨ and adding nodes to the computation graph of the neural network Sτ representing the proof success. . . . . . . . . . . . . . . . . . . . . 79 5.2 Exemplary construction of an NTP computation graph for a toy knowledge base. Indices on arrows correspond to application of the respective KB rule. Proof states (blue) are subscripted with the sequence of indices of the rules that were applied. Underlined proof states are aggregated to obtain the final proof success. Boxes visualize instantiations of modules (omitted for unify). The proofs S33 , S313 and S323 fail due to cycle-detection (the same rule cannot be applied twice). . . . . . . . . . . . . . . . . . . . . . . . . . . . 84 5.3 Overview of different tasks in the Contries dataset as visualized by Nickel et al. [2016]. The left part (a) shows which atoms are removed for each task (dotted lines), and the right part (b) illustrates the rules that can be used to infer the location of test countries. For task S1, only facts corresponding to the blue dotted line are removed from the training set. For task S2, additionally facts corresponding to the green dashed line are removed. Finally, for task S3 also facts for the red dash-dotted line are removed. . . . . . . . . . . . . . . . 89 6.1 Computation graph for the fully-connected RNN cell. . . . . . . . . 97 6.2 Computation graph for the LSTM cell. . . . . . . . . . . . . . . . . 99 6.3 Independent encoding of the premise and hypothesis using the same LSTM. Note that both sentences are compressed as dense vectors as indicated by the red mapping. . . . . . . . . . . . . . . . . . . . . . 100 6.4 Conditional encoding with two LSTMs. The first LSTM encodes the premise, and then the second LSTM processes the hypothesis conditioned on the representation of the premise (s5 ). . . . . . . . . 102 List of Figures 6.5 6.6 6.7 6.8 6.9 Attention model for RTE. Compared to Fig. 6.4, this model does not have to represent the entire premise in its cell state, but can instead output context representations (informally visualized by the red mapping) that are later queried by the attention mechanism (blue). Also note that now h1 to h5 are used. . . . . . . . . . . . . . . . . Word-by-word attention model for RTE. Compared to Fig. 6.5, querying the memory Y multiple times allows the model to store more fine-grained information in its output vectors when processing the premise. Also note that now also h6 to h8 are used. . . . . . . . . . Attention visualizations. . . . . . . . . . . . . . . . . . . . . . . . Word-by-word attention visualizations. . . . . . . . . . . . . . . . . Bidirectional encoding of a tweet conditioned on bidirectional en← coding of a target ([s→ 3 s1 ]). The stance is predicted using the last ← forward and reversed output representations ([h→ 9 h4 ]). . . . . . . . 19 103 105 109 113 114 List of Tables 2.1 2.2 3.1 3.2 4.1 4.2 5.1 6.1 Example knowledge base using Prolog syntax (left) and as list representation as used in the backward chaining algorithm (right). . . . . 34 Example proof using backward chaining. . . . . . . . . . . . . . . 37 Top rules for five different Freebase target relations. These implications were extracted from the matrix factorization model and manually annotated. The premises of these implications are shortest paths between entity arguments in dependency tree, but we present a simplified version to make these patterns more readable. See Appendix A for the list of all annotated rules. . . . . . . . . . . . . 56 (Weighted) MAP with relations that do not appear in any of the annotated rules omitted from the evaluation. The difference between Pre and Joint is significant according to the sign-test (p < 0.05). . . 58 Weighted MAP for our reimplementation of the matrix factorization model (F), compared to restricting the constant embedding space (FS) and to injecting WordNet rules (FSL). The orginial matrix factorization model by Riedel et al. [2013] is denoted as Riedel13-F. 70 Average score of facts with constants that appear in the body of facts (ei , ej ) or in the head (êi , êj ) of a rule. . . . . . . . . . . . . . . . . 73 AUC-PR results on Countries and MRR and HITS@m on Kinship, Nations, and UMLS. . . . . . . . . . . . . . . . . . . . . . . . . . 91 Results on the SNLI corpus. . . . . . . . . . . . . . . . . . . . . . 107 Chapter 1 Introduction “We are attempting to replace symbols by vectors so we can replace logic by algebra.” — Yann LeCun The vast majority of knowledge produced by mankind is nowadays available in a digital but unstructured form such as images or text. It is hard for algorithms to extract meaningful information from such data resources, let alone to reason with it. This issue is becoming more severe as the amount of unstructured data is growing very rapidly. In recent years, remarkable successes in processing unstructured data have been achieved by representation learning methods which automatically learn abstractions from large collections of training data. This is achieved by processing input data using artificial neural networks whose weights are adapted during training. Representation learning lead to breakthroughs in applications such as automated Knowledge Base (KB) completion [Nickel et al., 2012, Riedel et al., 2013, Socher et al., 2013, Chang et al., 2014, Yang et al., 2015, Neelakantan et al., 2015, Toutanova et al., 2015, Trouillon et al., 2016], as well as Natural Language Processing (NLP) applications like paraphrase detection [Socher et al., 2011, Hu et al., 2014, Yin and Schütze, 2015], machine translation [Bahdanau et al., 2014], image captioning [Xu et al., 2015], speech recognition [Chorowski et al., 2015] and sentence summarization [Rush et al., 2015], to name just a few. Representation learning methods achieve remarkable results, but they usually rely on a large amount of annotated training data. Moreover, since representation learning operates on a subsymbolic level (for instance by replacing words with lists of real numbers – so-called vector representations or embeddings), it is hard to determine why we obtain a certain prediction, let alone how to correct systematic errors or how to incorporate domain and commonsense knowledge. In fact, a recent General Data Protection Regulation by the European Union introduces the “right 24 Chapter 1. Introduction to explanation” of decisions by algorithms and machine learning models that affect users [Council of the European Union, 2016], to be enacted in 2018. This has profound implications for the future development and research of machine learning algorithms [Goodman and Flaxman, 2016], especially for nowadays commonly used representation learning methods. Moreover, for many domains of interests there is not enough annotated training data, which renders applying recent representation learning methods difficult. Many of these issues do not exist with purely symbolic approaches. For instance, given a KB of facts and first-order logic rules, we can use Prolog to obtain an answer as well as a proof for a query to this KB. Furthermore, we can easily incorporate domain knowledge by adding more rules. However, rule-based system do not generalize to new questions. For instance, given that an apple is a fruit and apples are similar to oranges, we would like to infer that oranges are likely also fruits. To summarize, symbolic rule-based systems are interpretable and easy to modify. They do not need large amounts of training data and we can easily incorporate domain knowledge. On the other hand, learning subsymbolic representations requires a lot of training data. The trained models are generally opaque and it is hard to incorporate domain knowledge. Consequently, we would like to develop methods that take the best of both worlds. 1.1 Aims In this thesis, we are investigating the combination of representation learning with first-order logic rules and reasoning. Representation learning methods achieve strong generalization by learning subsymbolic vector representations that can capture similarity and even logical relationships directly in a vector space [Mikolov et al., 2013]. Symbolic representations, on the other hand, allow us to formalize domain and commonsense knowledge using rules. For instance, we can state that every human is mortal, or that every grandfather is a father of a parent. Such rules are often worth many training facts. Furthermore, by using symbolic representations we can take advantage of algorithms for multi-hop reasoning like Prolog’s backward chaining algorithm [Colmerauer, 1990]. Backward chaining is not only widely used for multi-hop question answering in KBs, but it also provides us with proofs in addition to the answer for a question. However, such symbolic reasoning is relying on a complete specification of background and commonsense knowledge in logical form. As an example, let us assume we are asking for the grandpa of a person, but only know the grandfather of that person. If there is no explicit rule connecting grandpa to grandfather, we will not find an answer. However, given a large 1.2. Contributions 25 KB, we can use representation learning to learn that grandpa and grandfather mean the same thing, or that a lecturer is similar to a professor [Nickel et al., 2012, Riedel et al., 2013, Socher et al., 2013, Chang et al., 2014, Yang et al., 2015, Neelakantan et al., 2015, Toutanova et al., 2015, Trouillon et al., 2016]. This becomes more relevant once we do not only want to reason with structured relations but also use textual patterns as relations [Riedel et al., 2013]. The problem that we seek to address in this thesis is how symbolic logical knowledge can be combined with representation learning to make use of the best of both worlds. Specifically, we investigate the following research questions. • Can we efficiently incorporate domain and commonsense knowledge in form of rules into representation learning methods? • Can we use rules to alleviate the need for large amounts of training data while still generalizing beyond what is explicitly stated in these rules? • Can we synthesize representation learning with symbolic multi-hop reasoning as used for automated theorem proving? • Can we learn rules directly from data using representation learning? • Can we determine the logical relationship between natural language sentences using representation learning? 1.2 Contributions This thesis makes the following core contributions. Regularizing Representations by First-order Logic Rules We introduce a method for incorporating function-free first-order logic rules directly into vector representations of symbols, which avoids the need for symbolic inference. Instead of symbolic inference, we regularize symbol representations by given rules such that logical relationships hold implicitly in the vector space (Chapter 3). This is achieved by mapping propositional logical rules to differentiable loss terms so that we can calculate the gradient of a given rule with respect to symbol representations. Given a first-order logic rule, we stochastically ground free variables using constants in the domain, and add the resulting loss term for the propositional rule to the training objective of a neural link prediction model for automated KB completion. This allows us to infer relations with little or no training facts in a KB. While mapping logical rules to soft rules using algebraic operations has a long tradition (e.g. in Fuzzy logic), our contribution is the connection to representation learning, i.e., using such rules to 26 Chapter 1. Introduction directly learning better vector representations of symbols that can be used to improve performance on a downstream task such as automated KB completion. Content in this chapter first appeared in the following two publications: Tim Rocktäschel, Matko Bosnjak, Sameer Singh and Sebastian Riedel. 2014. Low-Dimensional Embeddings of Logic. In Proceedings of Association for Computational Linguistics Workshop on Semantic Parsing (SP’14). Tim Rocktäschel, Sameer Singh and Sebastian Riedel. 2015. Injecting Logical Background Knowledge into Embeddings for Relation Extraction. In Proceedings of North American Chapter of the Association for Computational Linguistics – Human Language Technologies (NAACL HLT 2015). Lifted Regularization of Predicate Representations by Implications For the subclass of first-order logic implication rules, we present a scalable method that is independent of the size of the domain of constants, that generalizes to unseen constants, and that can be used with a broader class of training objectives (Chapter 4). Instead of relying on stochastic grounding, we use implication rules directly as regularizers for predicate representations. Compared to the method in Chapter 3, this method is independent of the number of constants and ensures that a given implication between two predicates holds for any possible pair of constants at test time. Our method is based on Order Embeddings [Vendrov et al., 2016] and our contribution is the extension to the task of automated KB completion which requires constraining entity representations to be non-negative. This chapter is based on the following two publications: Thomas Demeester, Tim Rocktäschel and Sebastian Riedel. 2016. Regularizing Relation Representations by First-order Implications. In Proceedings of North American Chapter of the Association for Computational Linguistics (NAACL) Workshop on Automated Knowledge Base Construction (AKBC). Thomas Demeester, Tim Rocktäschel and Sebastian Riedel. 2016. Lifted Rule Injection for Relation Embeddings. In Proceedings of Empirical Methods in Natural Language Processing (EMNLP). My contribution to this work is the conceptualization of the model, the design of experiments, and the extraction of commonsense rules from WordNet. End-to-end Differentiable Proving Current representation learning and neural link prediction models have deficits when it comes to complex multi-hop inferences such as transitive reasoning [Bouchard et al., 2015, Nickel et al., 2016]. Automated theorem provers, on the other hand, have a long tradition in computer science and 1.2. Contributions 27 provide elegant ways to reason with symbolic knowledge. In Chapter 5, we propose Neural Theorem Provers (NTPs): end-to-end differentiable theorem provers for automated KB completion based on neural networks that are recursively constructed and inspired by Prolog’s backward chaining algorithm. By doing so, we can calculate the gradient of a proof success with respect to symbol representations in a KB. This allows us to learn symbol representations directly from facts in a KB, and to make use of the similarities of symbol representations and provided rules in proofs. In addition, we demonstrate that we can induce interpretable rules of predefined structure. On three out of four benchmark KBs, our method outperforms ComplEx [Trouillon et al., 2016], a state-of-the-art neural link prediction model. Work in this chapter appeared in: Tim Rocktäschel and Sebastian Riedel. 2016. Learning Knowledge Base Inference with Neural Theorem Provers. In Proceedings of North American Chapter of the Association for Computational Linguistics (NAACL) Workshop on Automated Knowledge Base Construction (AKBC). Tim Rocktäschel and Sebastian Riedel. 2017. End-to-End Differentiable Proving. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems (NIPS). Recognizing Textual Entailment with Recurrent Neural Networks Representation learning models such as Recurrent Neural Networks (RNNs) can be used to map natural language sentences to fixed-length vector representations, which has been successfully applied for various downstream NLP tasks including Recognizing Textual Entailment (RTE). In RTE, the task is to determine the logical relationship between two natural language sentences. This has so far been either approached by NLP pipelines with hand-crafted features, or neural network architectures that independently map the two sentences to fixed-length vector representations. Instead of encoding the two sentences independently, we propose a model that encodes the second sentence conditioned on an encoding of the first sentence. Furthermore, we apply a neural attention mechanism to bridge the hidden state bottleneck of the RNN (Chapter 6). Work in this chapter first appeared in: Tim Rocktäschel, Edward Grefenstette, Karl Moritz Hermann, Tomas Kocisky and Phil Blunsom. 2016. Reasoning about Entailment with Neural Attention. In Proceedings of International Conference on Learning Representations (ICLR). 28 1.3 Chapter 1. Introduction Thesis Structure In Chapter 2, we provide background on representation learning, computation graphs, first-order logic, and the notation used throughout the thesis. Furthermore, we explain the task of automated KB completion and describe neural link prediction and path-based approaches that have been proposed for this task. Chapter 3 introduces a method for regularizing symbol representations by first-order logic rules. In Chapter 4, we subsequently focus on direct implications between predicates, a subset of first-order logic rules. For this class of rules, we provide an efficient method by directly regularizing predicate representations. In Chapter 5, we introduce a recursive construction of a neural network for automated KB completion based on Prolog’s backward chaining algorithm. Chapter 6 presents a RNN for RTE based on conditional encoding and a neural attention mechanism. Finally, Chapter 7 concludes the thesis with a discussion of limitations, open issues, and future research avenues. Chapter 2 Background This chapter introduces core methods used in the thesis. Section 2.1 explains function approximation with neural networks and backpropagation. Subsequently, Section 2.2 introduces function-free first-order logic, the backward chaining algorithm, and inductive logic programming. Finally, Section 2.3 discusses prior work on automated knowledge base completion, linking the first two sections together. 2.1 Function Approximation with Neural Networks In this thesis, we consider models that can be formulated as differentiable functions fθ : X → Y parameterized by θ ∈ Θ. Our task is to find such functions, i.e., to learn parameters θ from a set of training examples T = {(xi , yi )} where xi ∈ X is the input and yi ∈ Y some desired output of the ith training example. Both, xi and yi , can be structured objects. For instance, xi could be a fact about the world, like directedBy(I NTERSTELLAR, N OLAN), and yi a corresponding target truth score (e.g. 1.0 for T RUE). We define a loss function L : Y × Y × Θ → R that measures the discrepancy between a provided output y and a predicted output fθ (x) on an input x, given a current setting of parameters θ. We seek to find those parameters θ ∗ that minimize this discrepancy on a training set. We denote the global loss over the entire training data as L. Our learning problem can thus be written as θ ∗ = arg min L = arg min θ θ 1 X L(fθ (x), y, θ). |T | (2.1) (x,y) ∈ T Note that L is also a function of θ, since we might not only want to measure the discrepancy between given and predicted outputs, but also use a regularizer on the parameters to improve generalization. Sometimes, we omit θ in L for brevity. As L and fθ are differentiable functions, we can use gradient-based optimization methods, 30 Chapter 2. Background such as Stochastic Gradient Descent (SGD) [Nemirovski and Yudin, 1978], for iteratively updating θ based on mini-batches B ⊆ T of the training data1 θt+1 = θt − η ∇θt 1 X L(fθ (x), y, θt ) |B| (2.2) (x,y) ∈ B where η denotes a learning rate, and ∇θ denotes the differentiation operation of the loss with respect to parameters, given the current batch at time step t. 2.1.1 Computation Graphs A useful abstraction for defining models as differentiable functions are computation graphs which illustrate the computations carried out by a model more precisely [Goodfellow et al., 2016]. In such a directed acyclic graph, nodes represent variables and directed edges from one or multiple nodes to another node correspond to a differentiable operation. As variables we consider scalars, vectors, matrices, and, more generally, higher-order tensors.2 We denote scalars by lower case letters x, vectors by bold lower case letters x, matrices by bold capital letters X, and higherorder tensors by Euler script letters X . Variables can either be inputs, outputs, or parameters of a model. Figure 2.1a shows a simple computation graph that calculates z = σ(x> y). Here, σ and sigm refer to the element-wise (or scalar) sigmoid operation σ(x) = 1 1 + e−x (2.3) and dot and x> y denote the dot product between two vectors. Furthermore, we name the ith intermediate expression as ui . Figure 2.1b shows a slightly more complex computation graph with two parameters w and b. This computation graph in fact represents logistic regression f (x) = σ(w> x + b). 2.1.2 From Symbols to Subsymbolic Representations In this thesis, we will use neural networks to learn representations of symbols. For instance, such symbols can be words, constants, or predicate names. When we say we learn a subsymbolic representation for symbols, we mean that we map symbols to fixed-dimensional vectors or more generally tensors. This can be done by first 1 Note that there are many alternative methods for minimizing the loss in Eq. 2.1, but all models in this thesis are optimized by variants of SGD. 2 For implementation purpose we will also consider structured objects over such tensors, like tuples and lists. Current deep learning libraries such as Theano [Al-Rfou et al., 2016], Torch [Collobert et al., 2011] and TensorFlow [Abadi et al., 2016] come with support for tuples and lists. However, for brevity we leave them out of the description here. 2.1. Function Approximation with Neural Networks 31 z sigm z u2 b sigm + u1 u1 w dot dot y x x (a) z = σ(x> y) (b) z = σ(w> x + b) Figure 2.1: Two simple computation graphs with inputs (gray), parameters (blue), intermediate variables ui , and outputs (dashed). Operations are shown next to the nodes. z k • k1 u2 u1 − + vs P matmul vi vj matmul 1s C matmul 1i 1j Figure 2.2: Computation graph for z = kC1i + P 1s − C1j k1 . enumerating all symbols and then assigning the number i to the ith symbol. Let S be the set of all symbols. We denote the one-hot vector for the ith symbol as 1i ∈ {0, 1}|S| , which is 1 at index i and 0 everywhere else. Figure 2.2 shows a computation graph whose inputs are one-hot vectors of some symbols with indices s, i, and j. In the first layer, these one-hot vectors are mapped to dense vector 32 Chapter 2. Background representations via a matrix multiplication with so-called embedding lookup matrices (P and C in this case). This computation graph corresponds to a neural link prediction model that we explain in more detail in Section 2.3.2. Here it serves only as an illustration for how symbols can be mapped to vector representations. In the remainder of this thesis, we will often omit the embedding layer for clarity. The goal is to learn symbol representations such as vs , vi and vj automatically from data. To this end, we need to be able to calculate the gradient of the output of the computation graph with respect to its parameters (the embedding lookup matrices in this case). 2.1.3 Backpropagation For learning from data, we need to be able to calculate the gradient of a loss with respect to all model parameters. As we assume all operations in the computation graph are differentiable, we can recursively apply the chain rule of calculus. Chain Rule of Calculus Assume we are given a composite function z = f (y) = f (g(x)) with f : Rn → Rm and g : Rl → Rn . The chain rule allows us to decompose the calculation of ∇x z, i.e., the gradient of the entire computation z with respect to x, as follows [Goodfellow et al., 2016]  ∇x z = ∂y ∂x > ∇y z. (2.4) ∂y is the Jacobian matrix of g, i.e., the matrix of partial derivatives, and ∇y z Here, ∂x is the gradient of z with respect to y. Note that this approach generalizes to matrices and higher-order tensors by reshaping them to vectors before the gradient calculation (vectorization) and back to their original shape afterwards. Backpropagation uses the chain rule to recursively define the efficient calculation of gradients of parameters (and inputs) in the computation graph by avoiding recalculation of previously calculated expressions. This is achieved via dynamic programming, i.e., storing previously calculated gradient expressions and reusing them for later gradient calculations. We refer the reader to Goodfellow et al. [2016] for details. In order to run backpropagation with a differentiable operation f that we want to use in a computation graph, all we need to ensure is that this function is differentiable with respect to each one of its inputs. Example Let us take the computation graph depicted in Fig. 2.1a as an example. Assume we are given an upstream gradient ∇z and want to compute the gradient of z with respect to the inputs x and y. The computations carried out by backpropagation 2.1. Function Approximation with Neural Networks 33 ∇z z sigm ∂z ∂u1 u1 ∂u1 ∂x x dot ∂u1 ∂y y Figure 2.3: Backward pass for the computation graph shown in Fig. 2.1a. are depicted in Fig. 2.3. For instance, by recursively applying the chain rule we can calculate ∇x z as follows: ∇x z = ∂z ∂u1 ∂z = = σ(u1 )(1 − σ(u1 ))y. ∂x ∂u1 ∂x (2.5) ∂z Note that the computation of ∂u can be reused for calculating ∇y z. We get the 1 gradient of the entire computation graph (including upstream nodes) with respect to x via ∇z ∇x z. Later we will use computation graphs where nodes are used in multiple downstream computations. Such nodes receive multiple gradients from downstream nodes during backpropagation, which are summed up to calculate the gradient of the computation graph with respect to the variable represented by the node. In Chapter 3, we will use backpropagation for computing the gradient of differentiable propositional logic rules with respect to vector representations of symbols to develop models that combine representation learning with first-order logic. In Chapter 5, we take this further and construct a computation graph for all possible proofs in a Knowledge Base (KB) using the backward chaining algorithm. This will allow us to calculate the gradient of proofs with respect to symbol representations and to induce rules using gradient descent. Finally, in Chapter 6, we will use Recurrent Neural Networks (RNNs), i.e., computation graphs that are dynamically constructed for input varying-length input sequences of word representations. 34 Chapter 2. Background 1 2 3 4 5 6 7 P C V Prolog/Datalog Syntax List Representation fatherOf(ABE, HOMER). parentOf(HOMER, LISA). parentOf(HOMER, BART). grandpaOf(ABE, LISA). grandfatherOf(ABE, MAGGIE). grandfatherOf(X1 , Y1 ) :– fatherOf(X1 , Z1 ), parentOf(Z1 , Y1 ). grandparentOf(X2 , Y2 ) :– grandfatherOf(X2 , Y2 ). [[fatherOf, ABE, HOMER]] [[parentOf, HOMER, LISA]] [[parentOf, HOMER, BART]] [[grandpaOf, ABE, LISA]] [[grandfatherOf, ABE, MAGGIE]] [[grandfatherOf, X1 , Y1 ], [fatherOf, X1 , Z1 ], [parentOf, Z1 , Y1 ]] [[grandparentOf, X2 , Y2 ], [grandfatherOf, X2 , Y2 ]] {fatherOf, parentOf, grandpaOf, grandfatherOf, grandparentOf} {ABE, HOMER, LISA, BART, MAGGIE} {X1 , Y1 , Z1 , X2 , Y2 } Table 2.1: Example knowledge base using Prolog syntax (left) and as list representation as used in the backward chaining algorithm (right). 2.2 Function-free First-order Logic We now turn to a brief introduction of function-free first-order logic to the extent it is used in subsequent chapters. This section follows the syntax of Prolog and Datalog [Gallaire and Minker, 1978], and is based on Lloyd [1987], Nilsson and Maluszynski [1990], and Dzeroski [2007]. 2.2.1 Syntax We start by defining an atom as a predicate3 symbol and a list of terms. We will use lowercase names to refer to predicate and constant symbols (e.g. fatherOf and BART), and uppercase names for variables (e.g. X, Y, Z). In Prolog, one also considers function terms and defines constants as function terms with zero arity. However, in this thesis we will work only with function-free first-order logic rules, the subset of logic that Datalog supports. Hence, for us a term can be a constant or a variable. For instance, grandfatherOf(Q, BART) is an atom with the predicate grandfatherOf, and two terms, the variable Q and the constant BART, respectively. We define the arity of a predicate to be the number of terms it takes as arguments. Thus, grandfatherOf is a binary predicate. A literal is defined as a negated or non-negated atom. A ground literal is a literal with no variables (see rules 1 to 5 in Table 2.1). Furthermore, we consider first-order logic rules4 of the form H :– B, where the body B (also called condition or premise) is a possibly empty conjunction of atoms represented as a list, and the head H (also 3 4 We will use predicate and relation synonymously throughout this thesis. We will use rule, clause and formula synonymously. 2.2. Function-free First-order Logic 35 called conclusion, consequent or hypothesis) is an atom. Examples are rules 6 and 7 in Table 2.1. Such rules with only one atom as the head are called definite rules. In this thesis we only consider definite rules. Variables are universally quantified (e.g. ∀X1 , Y1 , Z1 in rule 6). A rule is a ground rule if all its literals are ground. We call a ground rule with an empty body a fact, hence the rules 1 to 5 in Table 2.1 are facts.5 We define S = C ∪ P ∪ V to be the set of symbols, containing constant symbols C, predicate symbols P, and variable symbols V. We call a set of definite rules like the one in Table 2.1 a knowledge base or logic program. A substitution Ψ = {X1 /t1 , . . . , XN /tN } is an assignment of variable symbols Xi to terms ti , and applying a substitution to an atom replaces all occurrences of variables Xi by their respective term ti . What we have defined so far is the syntax of logic used in this thesis. To assign meaning to this language (semantics), we need to be able to derive the truth value for facts. We focus on proof theory, i.e., deriving the truth of a fact from other facts and rules in a KB.6 In the next subsection we explain backward chaining, an algorithm for deductive reasoning. It is used to derive atoms from other atoms by applying rules. 2.2.2 Deduction with Backward Chaining Representing knowledge (facts and rules) in symbolic form has the appeal that one can use automated deduction systems to infer new facts. For instance, given the logic program in Table 2.1, we can automatically deduce that grandfatherOf(ABE, LISA) is a true fact by applying rule 6 using facts 1 and 2. Backward chaining is a common method for automated theorem proving, and we refer the reader to Russell and Norvig [2010], Gelder [1987], Gallaire and Minker [1978] for details and to Fig. 2.4 for an excerpt of the pseudocode in style of a functional programming language. Particularly, we are making use of pattern matching to check for properties of arguments passed to a module. Note that ” ” matches every argument and that the order matters, i.e., if arguments match a line, subsequent lines are not evaluated. We denote sets by Euler script letters (e.g. E), lists by small capital letters (e.g. E), lists of lists by blackboard bold letters (e.g. E) and we use : to refer to prepending an element to a list (e.g. e : E or E : E). While an atom is a list of a predicate symbol and terms, a rule can be seen as a list of atoms and thus a list of lists where the head of the list is the rule head.7 5 We sometimes only call a rule a rule if it has a non-empty body. This will be clear from the context. 6 See Dzeroski [2007] for other methods for semantics. 7 For example, [[grandfatherOf, X, Y], [fatherOf, X, Z], [parentOf, Z, Y]]. 36 Chapter 2. Background 1. or(G, S) = [S 0 | S 0 ∈ and(B, unify(H, G, S)) for H :– B ∈ K] 2. and( , FAIL) = FAIL 3. and([ ], S) = S 4. and(G : G, S) = [S 00 | S 00 ∈ and(G, S 0 ) for S 0 ∈ or(substitute(G, S), S)] 5. 6. 7. 8. unify( , , FAIL) = FAIL unify([ ], [ ], S) = S unify([ ], , ) = FAIL unify( , [ ], ) = FAIL  S ∪ {h/g}     S ∪ {g/h} 9. unify(h : H, g : G, S) = unify  H, G,  S   FAIL  if h ∈ V if g ∈ V, h 6∈ V if g = h otherwise          10. substitute([ ], ) = [ ]  11. substitute(g : G, S) = x if g/x ∈ S g otherwise  : substitute(G, S) Figure 2.4: Simplified pseudocode for symbolic backward chaining (cycle detection omitted for brevity, see Russell and Norvig [2010], Gelder [1987], Gallaire and Minker [1978] for details). Given a goal such as grandparentOf(Q1 , Q2 ), backward chaining finds substitutions of free variables with constants in facts in a KB (e.g. {Q1 /ABRAHAM, Q2 /BART}). This is achieved by recursively iterating through rules that translate a goal into subgoals which we attempt to prove, thereby exploring possible proofs. For example, the KB could contain the following rule that can be applied to find answers for the above goal: grandfatherOf(X, Y) :– fatherOf(X, Z), parentOf(Z, Y). The proof exploration in backward-chaining is divided into two functions called or and and that perform a depth-first search through the space of possible proofs. The function or (line 1) attempts to prove a goal by unifying it with the head of every rule in a KB, yielding intermediate substitutions. Unification (lines 5-9) iterates through pairs of symbols in the two lists corresponding to the atoms that we want to unify and updates the substitution set if one of the two symbols is a variable. It 2.2. Function-free First-order Logic 37 Rule Remaining Goals S 7 6 1 2 [[grandparentOf, Q1 , Q2 ]] [[grandfatherOf, Q1 , Q2 ]] [[fatherOf, Q1 , Z1 ], [parentOf, Z1 , Q2 ]] [[parentOf, HOMER, Q2 ]] [] {} {X2 /Q1 , Y2 /Q2 } {X2 /Q1 , Y2 /Q2 , X1 /Q1 , Y1 /Q2 } {X2 /Q1 , Y2 /Q2 , X1 /Q1 , Y1 /Q2 , Q1 /ABE, Z1 /HOMER} {X2 /Q1 , Y2 /Q2 , X1 /Q1 , Y1 /Q2 , Q1 /ABE, Z1 /HOMER, Q2 /LISA} Table 2.2: Example proof using backward chaining. Example Knowledge Base: 1. fatherOf(abe, homer). 2. parentOf(homer, bart). 3. grandfatherOf(X, Y) :– fatherOf(X, Z), parentOf(Z, Y). Query ?– grandfatherOf(abe, bart). 1 2 failure failure success {X/abe, Y/bart} 3.1 ?– fatherOf(abe, Z). 1 2 success failure {X/abe, Y/bart, Z/homer} 3.2 ?– parentOf(bart, homer). 1 2 or 1 3 3 and 1 or 2 failure 3 or 2 failure failure success {X/abe, Y/bart, Z/homer} Figure 2.5: Full proof tree for a small knowledge base. returns a failure if two non-variable symbols are not identical or the two atoms have different arity. For rules where unification with the head of the rule succeeds, the body and substitution are passed to and (lines 2-4), which then attempts to prove every atom in the body sequentially by first applying substitutions and subsequently calling or. This is repeated recursively until unification fails, atoms are proven by unification via grounding with facts in the KB, or a certain proof-depth is exceeded. Table 2.2 shows a proof for the query grandparentOf(Q1 , Q2 ) given the KB in Table 2.1 using Fig. 2.4. The method substitute (lines 10-11) replaces variables in an atom by the corresponding symbol if there exists a substitution for that variable in the substitution list. Figure 2.5 shows the full proof tree for a small knowledge base and the query ?– grandfatherOf(ABE, BART).8 The numbers on the arrows correspond to the application of the respective rules. We visualize the recursive calls to or and and together with the proof depth on the right side. Note how most proofs can be aborted early due to unification failure. Though first-order logic can be used for complex multi-hop reasoning, a drawback is that for such symbolic inference there is no generalization beyond what we explicitly specify in the facts and rules. For instance, given a large KB we would 8 We denote queries by the prefix “?–”. 38 Chapter 2. Background like to learn automatically that in many cases where we observe fatherOf we also observe parentOf. This can be approached by statistical relational learning, inductive logic programming, and particularly neural networks for KB completion which we discuss in the remainder of this chapter. 2.2.3 Inductive Logic Programming While backward chaining is used for deduction, i.e., for inferring facts given rules and other facts in a KB, Inductive Logic Programming (ILP) [Muggleton, 1991] combines logic programming with inductive learning to learn logic programs from training data. This training data can include facts, but also a provided logic program that the ILP system is supposed to extend. Specifically, given facts and rules, the task of an ILP system is to find further regularities and form hypotheses for unseen facts [Dzeroski, 2007]. Crucially, these hypothesis are again formulated using first-order logic. There are different variants of ILP learning tasks [Raedt, 1997] and we focus on learning from entailment [Muggleton, 1991]. That is, given examples of positive facts and negative facts, the ILP system is supposed to find rules such that positive facts can be deduced, but negative facts cannot. There are many variants of ILP systems, and we refer the reader to Muggleton and Raedt [1994] and Dzeroski [2007] for an overview. One of the most prominent systems is the First Order Inductive Learner (FOIL) [Quinlan, 1990], which is a greedy algorithm that induces one rule at a time by constructing a body that satisfies the maximum number of positive facts and the minimum number of negative facts. In Chapter 5, we will construct neural networks for proving facts in a KB and introduce a method for inducing logic programs using gradient descent while learning vector representations of symbols. 2.3 Automated Knowledge Base Completion Automated KB completion is the task of inferring facts from information contained in a KB and other resources such as text. This is an important task as real-world KBs are usually incomplete. For instance, the placeOfBirth predicate is missing for 71% of people in Freebase [Dong et al., 2014]. Prominent recent approaches to automated KB completion learn vector representations of symbols via neural link prediction models. The appeal of learning such subsymbolic representations lies in their ability to capture similarity and even implicature directly in a vector space. Compared to ILP systems, neural link prediction models do not rely on a combinatorial search over the space of logic programs, but instead learn a local scoring function based on 2.3. Automated Knowledge Base Completion 39 subsymbolic representations using continuous optimization. However, this comes at the cost of uninterpretable models and no straightforward ways of incorporating logical background knowledge – drawbacks that we seek to address in this thesis. Another benefit of neural link prediction models over ILP is that inferring whether a fact is true or not often amounts to efficient algebraic operations (feed-forward passes in shallow neural networks), which makes test-time inference very scalable. In addition, representations of symbols can be compositional. For instance, we can compose a representation of a natural language predicate from a sequence of word representations [Verga et al., 2016a]. In recent years, many models for automated KB completion have been proposed. In the next sections, we discuss prominent approaches. On a high level, these methods can be categorized into (i) neural link prediction models which define a local scoring function for the truth of a fact based on estimated symbol representations, and (ii) models that use paths between two entities in a KB for predicting new relations between them. 2.3.1 Matrix Factorization In this section, we describe the matrix factorization relation extraction model by Riedel et al. [2013], which is an instance of a simple neural link prediction model. We discuss this model in detail, as it the basis on which we develop rule injection methods in Chapter 3 and 4. Assume a set of observed entity pair symbols C and a set of predicate symbols P, which can either represent structured binary relations from Freebase, a large collaborative knowledge base [Bollacker et al., 2008], or unstructured Open Information Extraction (OpenIE) [Etzioni et al., 2008] textual surface patterns collected from news articles. Examples for structured and unstructured relations are company/founders and #2-co-founder-of-#1, respectively. Here, #2-co-founder-of-#1 is a textual pattern where #1 and #2 are placeholders for entities. For instance, the relationship between Elon Musk and Tesla in the sentence “Elon Musk, the co-founder of Tesla and the CEO of SpaceX, cites The Foundation Trilogy by Isaac Asimov as a major influence on his thinking”9 could be expressed by the ground atom #2-co-founder-of-#1(TESLA, ELON MUSK ). In this example, ELON MUSK appeared first in the textual pattern, but #2 indicates that this constant will be used as the second argument in the predicate corresponding to the pattern. That way we can later introduce a rule 9 http://www.dailymail.co.uk/news/article-4045816 40 Chapter 2. Background ∀X, Y : company/founders(X, Y) :– #2-co-founder-of-#1(X, Y) without changing the order of the variables in the body of the rule. Let O = {rs (ei , ej ))} be the set of observed ground atoms. Model F by Riedel et al. [2013] maps all symbols in a knowledge base to subsymbolic representations, i.e., it learns a dense k-dimensional vector representation for every relation and entity pair. Thus, a training fact rs (ei , ej ) is represented by two vectors, vs ∈ Rk and vij ∈ Rk , respectively. We will refer to these as embeddings, subsymbolic representations, vector representations, neural representations, or simply (symbol) representations when it is clear from the context. The truth estimate of a fact is modeled via the sigmoid of the dot product of the two symbol representations: psij = σ(vs> vij ). (2.6) In fact, this expression corresponds to the computation graph shown in Fig. 2.1a with x = vs and y = vij . The score psij is measuring the compatibility between the relation and entity pair representation and can be interpreted as the probability of the fact being true conditioned on parameters of the model. We would like to train symbol representations such that true ground atoms get a score close to one and false ground atoms a score close to zero. This results in a low-rank matrix factorization corresponding to Generalized Principal Component Analysis (GPCA) [Collins et al., 2001] K ≈ σ(P C > ) ∈ R|P|×|C| (2.7) where P ∈ R|P|×k and C ∈ R|C|×k are matrices of all relation and entity pair representations, respectively. This factorization is depicted in Fig. 2.6 where known facts are shown in green and the task is to complete this matrix for cells with a question mark. Equation 2.7 leads to generalization with respect to unseen facts as every relation and entity pair is represented in a low-dimensional space, and this information bottleneck will lead to similar entity pairs being represented close in distance in the vector space (likewise for similar predicates). The distributional hypothesis states that “one shall know a word by the company it keeps” [Firth, 1957]. It has been used for learning word meaning from large collections of text [Lowe and McDonald, 2000, Padó and Lapata, 2007]. Applied to automated KB completion, one could say that the meaning of a relation or entity pair can be estimated from the entity pairs and relations that respectively appear together 2.3. Automated Knowledge Base Completion 41 #2 2 t2 -a -# at-# y t r o -a orst an #2 ri fess t- s-hi t o a t ro - he eA um is -h is-p se teac loye s u i m p - #1 - #1 em #1 #1 petrie, ucl ? T T ? T vc1 ferguson, harvard T T ? ? ? vc2 andrew, cambridge ? T ? T ? vc3 trevelyan, cambridge T ? ? ? ? vc4 Textual Patterns v p1 vp2 vp3 v p4 Freebase vp5 ∈ Rk Figure 2.6: Knowledge base inference as matrix completion with true training facts (green), unobserved facts (question mark), relation representations (red), and entity pair representations (orange). in facts in a KB. That is, if for many entity pairs we observe that both fatherOf and dadOf is true, we can assume that both relations are similar. 2.3.1.1 Bayesian Personalized Ranking A common problem in automated KB completion is that we do not observe any negative facts. In the recommendation literature, this problem is called implicit feedback [Rendle et al., 2009]. Applied to KB completion, we would like to infer (or recommend) for a target relation some (unobserved) facts that we do not know from facts that we do know. Facts that we do not know can be unknown either because they are not true or because they are true but missing in the KB. One method to address this issue is to formulate the problem in terms of a ranking loss, and sample unobserved facts as negative facts during training. Given a known fact rs (ei , ej ) ∈ O, Bayesian Personalized Ranking (BPR) [Rendle et al., 2009] samples another entity pair (em , en ) ∈ C such that rs (em , en ) 6∈ O and adds the soft-constraint vs> vij ≥ vs> vmn . (2.8) This follows a relaxation of the local Closed World Assumption (CWA) [Galárraga et al., 2013, Dong et al., 2014]. In the CWA, the knowledge about true facts 42 Chapter 2. Background loss sum u9 λp ∗ u10 u11 λc ∗ −log u8 sigm u5 k• u6 − u1 u2 u3 u4 dot dot k • k22 k • k22 vij vs P u7 + k22 matmul vmn matmul 1s C matmul 1ij 1mn Figure 2.7: A complete computation graph of a single training example for Bayesian personalized ranking with `2 -regularization. for relation rs is assumed to be complete and any sampled unobserved fact can consequently be assumed to be negative. In BPR, the assumption is that unseen facts are not necessarily false but their probability of being true should be less than for known facts. Thus, sampled unobserved facts should have a lower score than known facts. Instead of working with a fixed set of samples, we resample negative facts for every known fact in every epoch, where an epoch is a full iteration through all known facts in a KB. We denote a sample from the set of constant pairs as (em , en ) ∼ C.10 This leads to the overall approximate loss L= X rs (ei ,ej ) ∈ O, (em ,en ) ∼ C, rs (em ,en ) 6∈ O −ws log σ(vs> vij − vs> vmn ) + λp kvs k22 + λc (kvij k22 + kvmn k22 ) (2.9) 10 Note that C is the set of constant-pairs. 2.3. Automated Knowledge Base Completion 43 where λp and λc are the strength of the `2 regularization of the relation and entity pair representations, respectively. Furthermore, ws is a relation-dependent implicit weight accounting for how often (em , en ) is sampled during training time. As we resample an unobserved fact every time we visit an observed fact during training, unobserved facts for relations with many observed facts are more likely to be sampled as negative facts than for relations with few observed facts. A complete computation graph of a single training example for matrix factorization using BPR with `2 regularization is shown in Fig. 2.7. 2.3.2 Other Neural Link Prediction Models An example for a simple neural link prediction model is the matrix factorization approach from the previous section. Alternative methods define the score psij in Eq. 2.6 in different ways, use different loss functions, and parameterize relation and entity (or entity pair) representations differently. For instance, Bordes et al. [2011] train two projection matrices per relation, one for the left and one for the right-hand argument position, respectively. Subsequently, the score of a fact is defined as the `1 norm of the difference of the two projected entity argument embeddings psij = kMsleft vi − Msright vj k1 . (2.10) Note that compared to the matrix factorization model by Riedel et al. [2013] which embeds entity-pairs, here we learn individual entity embeddings. Similarly, TransE [Bordes et al., 2013] models the score as the `1 or `2 norm of the difference between the right entity embedding and a translation of the left entity embedding via the relation embedding psij = kvi + vs − vj k1 (2.11) where vi and vj are constrained to be unit-length. The computation graph for this model is shown in Fig. 2.2 in Section 2.1.2. RESCAL [Nickel et al., 2012] represents relations as matrices and defines the score of a fact as psij = vi> Ms vj . (2.12) In contrast to the other models mentioned in this section, RESCAL is not optimized with SGD but using alternating least squares [Nickel et al., 2011]. TRESCAL [Chang et al., 2014] extends RESCAL with entity type constraints for Freebase relations. 44 Chapter 2. Background Neural Tensor Networks [Socher et al., 2013] add a relation-dependent compatibility score to RESCAL psij = vi> Ms vj + vi> vsleft + vj> vsright (2.13) and are optimized with L-BFGS [Byrd et al., 1995]. DistMult by Yang et al. [2015] is modeling the score as the trilinear dot product psij = vs> (vi vj ) = X vsk vik vjk (2.14) k where is the element-wise multiplication. This model is a special case of RESCAL where Ms is constrained to be diagonal. ComplEx by Trouillon et al. [2016] uses complex vectors vs , vi , vj ∈ Ck for representing relations and entities. Let real(v) denote the real part and imag(v) the imaginary part of a complex vector v. The scoring function defined by ComplEx is psij = real(vs )> (real(vi ) real(vj )) + real(vs )> (imag(vi ) imag(vj )) + imag(vs )> (real(vi ) imag(vj )) − imag(vs )> (imag(vi ) real(vj )). (2.15) The benefit of ComplEx over RESCAL and DistMult is that by using complex vectors it can capture symmetric as well as asymmetric relations. Building upon Riedel et al. [2013], Verga et al. [2016a] developed a column-less factorization approach by encoding surface form patterns using Long Short-Term Memories (LSTMs) [Hochreiter and Schmidhuber, 1997] instead of learning a noncompositional representation. Similarly, Toutanova et al. [2015] uses Convolutional Neural Networks (CNNs) to encode surface form patterns. In a follow-up study, Verga et al. [2016b] propose a row-less method where entity pair representations are not learned but instead computed from observed relations, thereby generalizing to new entity pairs at test time. 2.3.3 Path-based Models While all methods presented in the previous section model the truth of a fact as a local scoring function of the representations of the relation and entities (or entity pairs), path-based models score facts based either on random walks over the KBs (path ranking) or by encoding entire paths in a vector space (path encoding). 2.3. Automated Knowledge Base Completion 2.3.3.1 45 Path Ranking The Path Ranking Algorithm (PRA) [Lao and Cohen, 2010, Lao et al., 2011] learns to predict a relation between two entities based on logistic regression over features collected from random walks between these entities in the KB up to some predefined length. Lao et al. [2012] extend PRA inference to OpenIE surface patterns in addition to structured relations contained in KBs. A related approach to PRA is Programming with Personalized PageRank (ProPPR) [Wang et al., 2013], which is a first-order probabilistic logic programming language. It uses Prolog’s Selective Linear Definite clause resolution (SLD) [Kowalski and Kuehner, 1971], a depth-first search strategy for theorem proving, to construct a graph of proofs. Instead of returning deterministic proofs for a given query, ProPPR defines a stochastic process on the graph of proofs using PageRank [Page et al., 1999]. Furthermore, in ProPPR one can use features on the head of rules whose weights are learned from data to guide stochastic proofs. Experiments with ProPPR were conducted on comparably small KBs and contrary to neural link prediction models and the extensions to PRA below, it has not yet been scaled to large real-world KBs [Gardner et al., 2014]. A shortcoming of PRA and ProPPR is that they are operating on symbols instead of vector representations of symbols. This limits generalization as it results in an explosion in the number of paths to consider when increasing the path length. To overcome this limitation, Gardner et al. [2013] extend PRA to include vector representations of verbs. These verb representations are obtained from pre-training via PCA on a matrix of co-occurrences of verbs and subject-object tuples collected from a large dependency-parsing corpus. Subsequently, these representations are used for clustering relations, thus avoiding an explosion of path features in prior PRA work while improving generalization. Gardner et al. [2014] take this approach further by introducing vector space similarity into random walk inference, thus dealing with paths containing unseen surface forms by measuring the similarity to surface forms seen during training, and following relations proportionally to this similarity. 2.3.3.2 Path Encoding While Gardner et al. introduced vector representations into PRA, these representations are not trained end-to-end from task data but instead pretrained on an external corpus. This means that relation representations cannot be adapted during training on a KB. Neelakantan et al. [2015] propose RNNs for learning embeddings of entire paths. The input to these RNNs are trainable relation representations. Given a known 46 Chapter 2. Background relation between two entities and a path connecting the two entities in the KB, an RNN for the target relation is trained to output an encoding of the path such that the dot product of that encoding and the relation representation is maximal. Das et al. [2017] note three limitations of the work by Neelakantan et al. [2015]. First, there is no parameter sharing of RNNs that encode different paths for different target relations. Second, there is no aggregation of information from multiple path encodings. Lastly, there is no use of entity information along the path as only relation representations are fed to the RNN. Das et al. address the first issue by using a single RNN whose parameters are shared across all paths. To address the second issue, Das et al. train an aggregation function over the encodings of multiple paths connecting two entities. Finally, to obtain entity representations that are fed into the RNN alongside relation representations they sum learned vector representations of the entity’s annotated Freebase types. Chapter 3 Regularizing Representations by First-order Logic Rules In this chapter, we introduce a paradigm for combining neural link prediction models for automated Knowledge Base (KB) completion (Section 2.3.2) with background knowledge in the form of first-order logic rules. We investigate simple baselines that enforce rules through symbolic inference before and after matrix factorization. Our main contribution is a novel joint model that learns vector representations of relations and entity pairs using both distant supervision and first-order logic rules, such that these rules are captured directly in the vector space of symbol representations. To this end, we map symbolic rules to differentiable computation graphs representing real-valued losses that can be added to the training objective of existing neural link prediction models. At test time, inference is still efficient as only a local scoring function over symbol representations is used and no logical inference is needed. We present an empirical evaluation where we incorporate automatically mined rules into a matrix factorization neural link prediction model. Our experiments demonstrate the benefits of incorporating logical knowledge for Freebase relation extraction. Specifically, we find that joint factorization of distant and logic supervision is efficient, accurate, and robust to noise (Section 3.5). By incorporating logical rules, we were able to train relation extractors for which no or only few training facts are observed. 3.1 Matrix Factorization Embeds Ground Literals In Section 2.3.1, we introduced matrix factorization as a method for learning representations of predicates and constant pairs for automated KB completion. In this section, we elaborate on how matrix factorization indeed embeds ground atoms 48 Chapter 3. Regularizing Representations by First-order Logic Rules in a vector space and lay out the foundation for developing a method that embeds first-order logic rules. Let F ∈ F denote a rule in a KB. For instance, F could be a ground rule without a body (i.e. a fact) like parentOf(HOMER, BART). Furthermore, let JFK denote the probability of this rule being true conditioned on the parameters of the model. For now, we restrict F to ground atoms and discuss ground literals, propositional rules, and first-order rules later. With slight abuse of notation, let J•K also denote the mapping of predicate or constant symbols (or a pair of constant symbols) to their subsymbolic representation as assigned by the model. Note that this mapping depends on the neural link prediction model. For matrix factorization, J•K is a function S → Rk from symbols (constant pairs and predicates) to k-dimensional dense vector representations. For RESCAL (see Section 2.3.2), J•K maps constants to Rk and predicates to Rk×k . Using this notation, matrix factorization decomposes the probability of a fact rs (ei , ej ) as Jrs (ei , ej )K = σ(Jrs K> Jei , ej K) = σ(vs> vij ). (3.1) Training Objective Riedel et al. [2013] used Bayesian Personalized Ranking (BPR) [Rendle et al., 2009] as training objective, i.e., they encouraged the score of known true facts to be higher than unknown facts (Section 2.3.1.1). However, as we will later model the probability of a rule from the probability of ground atoms scored by a neural link prediction model, we need to ensure that all scores are in the interval [0, 1]. Instead of BPR, we thus use the negative log-likelihood loss to directly maximize the probability of all rules, including ground atoms, in a KB (we omit `2 regularization for brevity): X L= − log(JFK). (3.2) F ∈F Therefore, instead of learning to rank facts, we optimize representations to assign a score close to 1.0 to rules (including facts). Our model can thus be seen as generalization of a neural link prediction model to rules beyond ground atoms. For matrix factorization as neural link prediction model, we are embedding ground atoms in a vector space of predicate and constant pair representations. Next, we will extend this to ground literals and afterwards to propositional and then first-order logic rules. 3.2. Embedding Propositional Logic 49 Negation Let F = ¬ G be the negation of a ground atom G. We can model the probability of F as follows: JFK = J¬ GK = 1 − JGK . (3.3) By using ground literals in the training objective in Eq. 3.2, we can say that matrix factorization is embedding ground literals via learning predicate and constant pair representations. In other words, given known ground literals (negated and nonnegated facts), matrix factorization embeds symbols in a low-dimensional vector space such that a scoring function assigns a high probability to these ground literals. As symbols are embedded in a low-dimensional vector space, this method can generalize to unknown facts and predict a probability for these at test time by placing similar symbols close in distance in the embedding space. Note that so far we have not gained anything over matrix factorization as explained in Section 2.3.1. Equations 3.2 and 3.3 are only introducing notation that will make it easier to embed more complex rules later. Now that we can embed negated and non-negated facts, we can ask the question whether we can also embed propositional and first-order logic rules. 3.2 Embedding Propositional Logic We know from propositional logic that with the negation and conjunction operators we can model any other Boolean operator and propositional rule. In Eq. 3.3, we effectively turned a symbolic logical operation (negation) into a differentiable operation that can be used to learn subsymbolic representations for automated KB completion. If we can find such a differentiable operation for conjunction, then we could backpropagate through any propositional logical expression, and learn vector representations of symbols that encode given background knowledge in propositional logic. Conjunction In Product Fuzzy Logic, conjunction is modeled using a Product tNorm [Lukasiewicz, 1920]. Let F = A ∧ B be the conjunction of two propositional expressions A and B. The probability of H is then defined as follows: JFK = JA ∧ BK = JAK JBK . (3.4) In other words, we replaced conjunction, a symbolic logical operation, with multiplication, a differentiable operation. Note that alternatives for modeling conjunction 50 Chapter 3. Regularizing Representations by First-order Logic Rules exist. For instance, one could take the min of JAK and JBK (Gödel t-Norm [Gödel, 1932]). Given the probability of ground atoms, we can use Product Fuzzy Logic to calculate the probability of the conjunction of these atoms. However, we will go a step further and assume that we know the ground truth probability of the conjunction of two atoms. We can then use the negative log-likelihood loss to measure the discrepancy between the predicted probability of the conjunction and the ground truth. Our contribution is backpropagating this discrepancy through the propositional rule and a neural link prediction model that scores ground atoms to calculate a gradient with respect to vector representations of symbols. Subsequently, we update these representations using gradient descent, thereby encoding the ground truth of a propositional rule directly in the vector representations of symbols. At test time, predicting a score for any unobserved ground atom rs (ei , ej ) is done efficiently by calculating Jrs (ei , ej )K. Disjunction Let F = A ∨ B be the disjunction of two propositional expressions A and B. Using De Morgan’s law and Eqs. 3.3 and 3.4, we can model the probability of F as follows: JFK = JA ∨ BK = J¬ (¬ (A ∨ B))K = J¬ (¬ A ∧ ¬ B)K = 1 − (1 − JAK)(1 − JBK) = JAK + JBK − JAK JBK . (3.5) Note that Eq. 3.3 not only holds for ground atoms, but any propositional logical expression. Furthermore, any propositional logical expression can be normalized to Conjunctive Normal Form. Thus, with Eqs. 3.3 to 3.5 we now have a way to construct a differentiable computation graph, and thus a real-valued loss term, for any symbolic expression in propositional logic. Implication A particular class of logical expressions that we care about in practice are propositional implication rules of the form H :– B, where the body B is a possibly 3.2. Embedding Propositional Logic 51 Loss loss −log u11 • +1 Rule u10 ∗ u9 ∗ u7 u8 1−• u4 u5 sigm Atoms • u6 sigm u1 −1 sigm u2 u3 dot dot dot JparentOfK Jhomer, bartK JmotherOfK JfatherOfK Figure 3.1: Computation graph for rule in Eq. 3.7 where · denotes a placeholder for the output of the connected node. empty conjunction of atoms represented as a list, and the head H is an atom. Let F = H :– B . The probability of F is then modeled as: JFK = JH :– BK = J¬ B ∨ HK = J¬ BK + JHK − J¬ BK JHK = 1 − JBK + JHK − (1 − JBK) JHK = 1 − JBK + JHK − JHK + JBK JHK = 1 − JBK + JBK JHK = JBK (JHK − 1) + 1. (3.6) 52 Chapter 3. Regularizing Representations by First-order Logic Rules Say we want to ensure that fatherOf(HOMER, BART) :– parentOf(HOMER, BART) ¬ motherOf(HOMER, BART). (3.7) We now have a way to map this rule to a differentiable expression that we can use alongside facts in Eq. 3.2 and optimize the symbol representations using gradient descent as we did previously for matrix factorization. The computation graph that allows us to calculate the gradient of this rule with respect to symbol representations is shown in Fig. 3.1. While the structure of the bottom part (Atoms) of this computation graph is determined by the neural link prediction model, the middle part (Rule) is determined by the propositional rule. Note that we can use any neural link predictor (see Section 2.3.2) instead of matrix factorization for obtaining a probability of ground atoms. The only requirement is that ground atom scores need to lie in the interval [0, 1]. However, for models where this is not the case, we can always apply a transformation such as the sigmoid. Independence Assumption Equation 3.4 underlies a strong assumption, namely that the probability of the arguments of the conjunction are conditionally independent given symbol embeddings. We already get a violation of this assumption for the simple case JF ∧ FK with 0 < JFK < 1, which results in JF ∧ FK = JFK JFK < JFK. However, for dependent arguments we get an approximation to the probability of the conjunction that can still be used for gradient updates of the symbol representations, and we demonstrate empirically that conjunction as modeled in Eq. 3.4 is useful for improving automated KB completion. In Chapter 4, we will present a way to avoid this independence assumption for implications. 3.3 Embedding First-order Logic via Grounding Now that we can backpropagate through propositional Boolean expressions, we turn to embedding first-order logic rules in the vector space of symbol representations. Note that in this process, we do not chain rules, i.e., we do not perform logical inference at training or test time. Instead, a provided rule is used to construct a loss term for optimizing vector representations of symbols. When training the model using gradient descent, we attempt to find a minimum of the global loss Eq. 3.2 such that the probability of all rules (including facts) is high. This minimum might be attained where the prediction of the neural link prediction model for scoring ground atoms not only agrees with a single but all rules, thereby predicting ground atoms as 3.3. Embedding First-order Logic via Grounding Evidence Sparse Training Matrix |P| Regularized Embeddings Completed Matrix |P| k Facts |C| |C| Rules 53 |C| |P| k KB company/founders(X, Y) :– #2-co-founder-of-#1(X, Y). author/works_written(X, Y) :– #2-review-by-#1(X, Y). person/parents(X, Y) :– #2-daughter-of-#1(X, Y). Figure 3.2: Given a set of training ground atoms, matrix factorization learns k-dimensional predicate and constant pair representations. Here, we also consider additional first-order logic rules (red) and seek to learn symbol representations such that the predictions (completed matrix) comply with these given rules. if we chained rules. A high-level overview of this process for matrix factorization as neural link prediction model of ground atoms is shown in Fig. 3.2. Assuming a finite set of constants, we can ground any first-order rule by replacing free variables with constants in the domain. Let F(X, Y) denote a universally quantified rule with two free variables X and Y, then u J∀X, Y : F(X, Y)K = v ^ (ei ,ej )∈C } F (ei , ej )~ (3.8) where C is the set of all entity pairs. With Eqs. 3.2 and 3.4 we obtain the following ground loss for F: u −log v ^ (ei ,ej )∈C 3.3.1 } F (ei , ej )~  = − log   Y (ei ,ej )∈C JF(ei , ej )K = − X (ei ,ej )∈C log JF(ei , ej )K. (3.9) Stochastic Grounding For large domains of constants, Eq. 3.9 becomes very costly to optimize as it would result in many expressions that are added to the training objective of the underlying neural link prediction model. For rules over pairs of constants, we can reduce the number of terms drastically by only considering pairs of constants C train that appeared together in training facts. Still, C train might be a prohibitively large set of constant pairs. Thus, we resort to a heuristic similar to BPR for sampling constant pairs. Given a rule, we obtain a ground propositional rule for every constant pair for which at least one atom in the rule is a known training fact when substituting free variables with these constants. In addition, we sample as many constant pairs that appeared 54 Chapter 3. Regularizing Representations by First-order Logic Rules together in training facts but for which all atoms in the rule are unknown when substituting free variables with the sampled constant pairs. 3.3.2 Baseline Background knowledge in form of first-order rules can be seen as hints that can be used to generate additional training data [Abu-Mostafa, 1990]. For pre-factorization inference we first perform symbolic logical inference on the training data using provided rules and add inferred facts as additional training data. For example, for a rule rt (X, Y) :– rs (X, Y), we add an additional observed training facts rt (ei , ej ) for any pair of constants (ei , ej ) for which rs (ei , ej ) is a true fact in the training data. This is repeated until no further facts can be inferred. Subsequently, we run matrix factorization on the extended set of observed facts. The intuition is that the additional training data generated by rules provide evidence of the logical dependencies between relations to the matrix factorization model, while at the same time allowing the factorization to generalize to unobserved facts and to deal with ambiguity and noise in the data. No further logical inference is performed during or after training of the factorization model as we expect that the learned embeddings encode the provided rules. One drawback of pre-factorization inference is that the rules are enforced only on observed atoms, i.e., first-order dependencies on predicted facts are ignored. In contrast, with the loss in Eq. 3.9 we add terms for the rule directly to the matrix factorization objective, thus jointly optimizing embeddings to reconstruct known facts, as well as to obey to provided first-order logical background knowledge. However, as we stochastically ground first-order rules, we have no guarantee that a given rule will indeed hold for all possible entity pairs at test time. While we next demonstrate that this approach is still useful for KB completion despite this limitation, in Chapter 4 we will introduce a method that overcomes this limitation for simple first-order implication rules. 3.4 Experiments There are two orthogonal questions when evaluating the method above. First, does regularizing symbol embeddings by first-order logic rules indeed capture such logical knowledge in a vector space and can it be used to improve KB completion? Second, where can we obtain background knowledge in form of rules that is useful for a particular KB completion task? The latter is a well-studied problem [Hipp et al., 2000, Schoenmackers et al., 2010a, Völker and Niepert, 2011]. Thus, we focus the 3.4. Experiments 55 evaluation on the ability of various approaches to benefit from rules that we directly extract from the training data using a simple method. Distant Supervision Evaluation We follow the procedure of Riedel et al. [2013] for evaluating knowledge base completion of Freebase [Bollacker et al., 2008] with textual data from the NYT corpus [Sandhaus, 2008]. The training matrix consists of 4 111 columns, representing 151 Freebase relations and 3 960 textual patterns, 41 913 rows (constant pairs) and 118 781 training facts of which 7 293 belong to Freebase relations. The constant pairs are divided into train and test, and we remove all Freebase facts for these test pairs from the training data. Our primary evaluation measure is (weighted) Mean Average Precision (MAP) [Manning et al., 2008, Riedel et al., 2013]. Let R be the set of test relations and let {f1j , . . . , fmj } be the set of test facts for relation rj ∈ R. Furthermore, let Rkj be the ranked list of facts for relation rj scored by a model up until fact fkj is reached. MAP is then defined as mj |R| 1 X 1 X precision(Rkj ) MAP(R) = |R| j=1 mj k=1 (3.10) where precision calculates the fraction of correctly predicted test facts of all predicted facts. For weighted MAP, the average precision for every relation is weighted by the number of true facts for the respective relation [Riedel et al., 2013]. Note that the MAP metric operates only on the ranking of facts as predicted by the model and does not take the absolute predicted score into account. Rule Extraction and Annotation We use a simple technique for extracting rules from a matrix factorization model based on Sanchez et al. [2015]. We first run matrix factorization over the complete training data to learn symbol representations. After training, we iterate over all pairs of relations (rs , rt ) where rt is a Freebase relation. For every relation pair we iterate over all training atoms rs (ei , ej ), evaluate the score Jrt (ei , ej ) :– rs (ei , ej )K using Eq. 3.6, and calculate the average to arrive at a score as the proxy for the coverage of the rule. Finally, we rank all rules by their score and manually inspect and filter the top 100, which resulted in 36 annotated high-quality rules (see Table 3.1 for the top rules for five different Freebase target relations and Appendix A for the list of all annotated rules). Note that our rule extraction approach does not observe the relations for test constant pairs and that all extracted rules are simple first-order logic expressions. All models in our experiments have access to these rules, except for the matrix factorization baseline. Methods Our proposed methods for injecting logic into symbol embeddings are prefactorization inference (Pre; Section 3.3.2) which is a baseline method that performs 56 Chapter 3. Regularizing Representations by First-order Logic Rules Rule Score org/parent/child(X, Y) :– #2-unit-of-#1(X, Y). location/containedby(X, Y) :– #2-city-of-#1(X, Y). person/nationality(X, Y) :– #2-minister-#1(X, Y). person/company(X, Y) :– #2-executive-#1(X, Y). company/founders(X, Y) :– #2-co-founder-of-#1(X, Y). 0.97 0.97 0.97 0.96 0.96 Table 3.1: Top rules for five different Freebase target relations. These implications were extracted from the matrix factorization model and manually annotated. The premises of these implications are shortest paths between entity arguments in dependency tree, but we present a simplified version to make these patterns more readable. See Appendix A for the list of all annotated rules. regular matrix factorization after propagating the logic rules in a deterministic manner, and joint optimization (Joint; Section 3.3) which maximizes an objective that combines terms from facts and first-order logic rules. Additionally, we evaluate three baselines. The matrix factorization (MF; Section 2.3.1) model uses only ground atoms to learn relation and constant representations (i.e. it has no access to any rules). Furthermore, we consider pure symbolic logical inference (Inf). Since we restrict ourselves to a set of consistent, simple rules, this inference can be performed efficiently. Our final approach, post-factorization inference (Post), first runs matrix factorization and then performs logical inference on the known and predicted facts. Post-inference is computationally expensive since for all premises of rules we have to iterate over all rows (constant pairs) in the matrix to assess whether the premise is predicted to be true or not. 3.4.1 Training Details Since we have no negative training facts, we follow Riedel et al. [2013] by sampling unobserved ground atoms that we assume to be false. For rules, we use stochastic grounding as described in Section 3.3.1. Thus, in addition to a loss over the score of training facts, we have a loss over sampled unobserved ground atoms that we assume to be negative, as well as loss terms for ground rules. In other words, we learn symbol embeddings by minimizing Eq. 3.2 where F includes known and unobserved atoms, as well as ground propositional rules sampled using stochastic grounding. In addition, we use `2 -regularization on all symbol representations. For minimizing the training loss we use AdaGrad [Duchi et al., 2011]. Remember that at test time, predicting a score for any unobserved ground atom rs (ei , ej ) is done efficiently by calculating Jrs (ei , ej )K. Note that this does not 3.5. Results and Discussion 57 involve any explicit logical inference. Instead, we expect the vector space of symbol embeddings to incorporate all given rules. Hyperparameters Based on Riedel et al. [2013], we use k = 100 as the dimension for symbol representations in all models, λ = 0.01 as parameter for `2 -regularization, and α = 0.1 as the initial learning rate for AdaGrad, which we run for 200 epochs. Runtime Each AdaGrad update is defined over a single cell of the matrix, and thus training data can be provided one ground atom at a time. For matrix factorization, each AdaGrad epoch touches all the observed ground atoms once per epoch, and as many sampled negative ground atoms. With provided rules, it additionally revisits all the observed ground atoms that appear as an atom in the rules (and as many sampled negative ground atoms), and thus more general rules will be more expensive. However, the updates on ground atoms are performed independently and thus not all the data needs to be stored in memory. All presented models take less than 15 minutes to train on a 2.8 GHz Intel Core i7 machine. 3.5 Results and Discussion To asses the utility of injecting logic rules into symbol representations, we present a comparison on a variety of benchmarks. First, we study the scenario of learning extractors for relations for which we do not have any Freebase alignments, i.e., the number of entity pairs that appear both in textual patterns and structured Freebase relations is zero. This measures how well the different models are able to generalize only from logic rules and textual patterns (Section 3.5.1). In Section 3.5.2, we then describe an experiment where the number of Freebase alignments is varied in order to assess the effect of combining distant supervision and background knowledge on the accuracy of predictions. Although the methods presented here target relations with insufficient alignments, we also provide a comparison on the complete distant supervision dataset in Section 3.5.3. 3.5.1 Zero-shot Relation Learning We start with the scenario of learning extractors for relations that do not appear in the KB, i.e., those that do not have any textual alignments. Such a scenario occurs in practice when a new relation is added to a KB for which there are no facts that connect the new relation to existing relations or textual surface forms. For accurate extraction of such relations, we can only rely on background domain knowledge, e.g., in form of rules, to identify relevant textual alignments. However, at the same time, there are correlations between textual patterns that can be utilized for improved generalization. To simulate this setup, we remove all alignments between all entity 58 Chapter 3. Regularizing Representations by First-order Logic Rules Figure 3.3: Weighted MAP scores for zero-shot relation learning. Relation # MF Inf Post Pre Joint person/company 102 location/containedby 72 author/works written 27 person/nationality 25 parent/child 19 person/place of birth 18 18 person/place of death neighborhood/neighborhood of 11 person/parents 6 company/founders 4 2 film/directed by 1 film/produced by 0.07 0.03 0.02 0.01 0.01 0.01 0.01 0.00 0.00 0.00 0.00 0.00 0.03 0.06 0.05 0.19 0.01 0.43 0.24 0.00 0.17 0.25 0.50 1.00 0.15 0.14 0.18 0.09 0.48 0.40 0.23 0.60 0.19 0.13 0.50 1.00 0.31 0.22 0.31 0.15 0.66 0.56 0.27 0.63 0.37 0.37 0.36 1.00 0.35 0.31 0.27 0.19 0.75 0.59 0.23 0.65 0.65 0.77 0.51 1.00 MAP Weighted MAP 0.01 0.03 0.23 0.10 0.34 0.21 0.43 0.33 0.52 0.38 Table 3.2: (Weighted) MAP with relations that do not appear in any of the annotated rules omitted from the evaluation. The difference between Pre and Joint is significant according to the sign-test (p < 0.05). pairs and Freebase relations from the distant supervision data, use the extracted logic rules (Section 3.4) as background knowledge, and assess the ability of the different methods to recover the lost alignments. Figure 3.3 provides detailed results. Unsurprisingly, matrix factorization (MF) performs poorly since predicate representations cannot be learned for the Freebase 3.5. Results and Discussion 59 0.7 weighted Mean Average Precision 0.6 0.5 Inf F Post Pre Joint 0.4 0.3 0.2 0.1 0.0 0.0 0.1 0.2 0.3 0.4 0.5 Fraction of Freebase training facts Figure 3.4: Weighted MAP of the various models as the fraction of Freebase training facts is varied. For 0% Freebase training facts we get the zero-shot relation learning results presented in Table 3.2. relations without any observed facts. The fact that we see a non-zero MAP score for matrix factorization is due to random predictions. Symbolic logical inference (Inf) is limited by the number of known facts that appear as a premise in one of the implications, and thus performs poorly too. Although post-factorization inference (Post) is able to achieve a large improvement over logical inference, explicitly injecting logic rules into symbol representations using pre-factorization inference (Pre) or joint optimization (Joint) gives superior results. Finally, we observe that jointly optimizing the probability of facts and rules is able to best combine logic rules and textual patterns for accurate, zero-shot learning of relation extractors. Table 3.2 shows detailed results for each of the Freebase test relations. Except for author/works written and person/place of death, jointly optimizing the probability of facts and rules yields superior results. 60 Chapter 3. Regularizing Representations by First-order Logic Rules 3.5.2 Relations with Few Distant Labels In this section, we study the scenario of learning relations that have only a few distant supervision alignments, i.e., structured Freebase relations where only few textual patterns are observed for entity-pairs. In particular, we observe the behavior of the various methods as the amount of distant supervision is varied. We run all methods on training data that contains different fractions of Freebase training facts (and therefore different degrees of relation/text pattern alignment), but keep all textual patterns in addition to the set of annotated rules. Figure 3.4 summarizes the results. The performance of symbolic logical inference does not depend on the amount of distant supervision data since it does not take advantage of the correlations in this data. Matrix factorization does not make use of logical rules, and thus is the baseline performance when only using distant supervision. For the factorization based methods, only a small fraction (15%) of the training data is needed to achieve around 0.50 weighted MAP performance, thus demonstrating that they are efficiently exploiting correlations between predicates, and generalizing to unobserved facts. Pre-factorization inference, however, does not outperform post-factorization inference and is on par with matrix factorization for most of the curve. This suggests that it is not an effective way of injecting logic into symbol representations when ground atoms are also available. In contrast, the joint model learns symbol representations that outperform all other methods in the 0 to 30% Freebase training data interval. Beyond 30%, there seem to be sufficient Freebase facts for matrix factorization to encode these rules, thus yielding diminishing returns. 3.5.3 Comparison on Complete Data Although the focus of our work is on injecting logical rules for relations without sufficient alignments to the knowledge base, we also present an evaluation on the complete distant supervision data by Riedel et al. [2013]. Compared to Riedel et al.’s matrix factorization model Riedel13-F, our reimplementation (MF) achieves a lower wMAP (64% vs 68%) and a higher MAP (66% vs 64%). We attribute this difference to the different loss function (BPR vs. negative log-likelihood). We show the precision-recall curve in Fig. 3.5, demonstrating that joint optimization provides benefits over the existing factorization and distant supervision techniques even on the complete dataset, and obtains 66% weighted MAP and 69% MAP, respectively. This improvement over the matrix factorization model can be explained by noting that the joint model reinforces high-quality annotated rules. 3.6. Related Work 61 1 Joint MF Mintz09 Yao11 Surdeanu12 Riedel13-F Precision 0.8 0.6 0.4 0.2 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Recall Figure 3.5: Precision-recall curve demonstrating that the Joint method, which incorporates annotated rules derived from the data, outperforms existing factorization approaches (MF and Riedel13-F). 3.6 Related Work Embeddings for Knowledge Base Completion Many methods for embedding predicates and constants (or pairs of constants) based on training facts for knowledge base completion have been proposed in the past (see Section 2.3.2). Our work goes further in that we learn embeddings that follow not only factual but also first-order logic knowledge. Note that the method of regularizing symbol embeddings by rules described in this chapter are generally compatible with any existing neural link prediction model that provides per-atom scores between 0.0 and 1.0. In our experiments we only worked with matrix factorization as neural link prediction model but based on our work Guo et al. [2016] were able to incorporate transitivity rules into TransE [Bordes et al., 2013] which models entities separately instead of learning a representation for every entity pair. Logical Inference A common alternative, where adding first-order logic knowledge is trivial, is to perform symbolic logical inference [Bos and Markert, 2005, Baader et al., 2007, Bos, 2008]. However, such purely symbolic approaches cannot deal with the uncertainty inherent to natural language and generalize poorly. 62 Chapter 3. Regularizing Representations by First-order Logic Rules Probabilistic Inference To ameliorate some of the drawbacks of symbolic logical inference, probabilistic logic based approaches have been proposed [Schoenmackers et al., 2008, Garrette et al., 2011, Beltagy et al., 2013, 2014]. Since logical connections between relations are modeled explicitly, such approaches are generally hard to scale to large KBs. Specifically, approaches based on Markov Logic Networks (MLNs) [Richardson and Domingos, 2006] encode logical knowledge in dense, loopy graphical models, making structure learning, parameter estimation, and inference hard for the scale of our data. In contrast, in our model the logical knowledge is captured directly in symbol representations, leading to efficient inference at test time as we only have to calculate the forward pass of a neural link prediction model. Furthermore, as symbols are embedded in a low-dimensional vector space, we have a natural way of dealing with linguistic ambiguities and label errors that appear once OpenIE textual patterns are included as predicates for automated KB completion [Riedel et al., 2013]. Stochastic grounding is related to locally grounding a query in Programming with Personalized PageRank (ProPPR) [Wang et al., 2013]. One difference is that we use stochastically grounded rules as differentiable terms in a representation learning training objective, whereas in ProPPR such grounded rules are used for stochastic inference without learning symbol representations. Weakly Supervised Learning Our work is also inspired by weakly supervised approaches [Ganchev et al., 2010] that use structural constraints as a source of indirect supervision. These methods have been used for several NLP tasks [Chang et al., 2007, Mann and McCallum, 2008, Druck et al., 2009, Singh et al., 2010]. The semi-supervised information extraction work by Carlson et al. [2010] is in spirit similar to our goal as they are using commonsense constraints to jointly train multiple information extractors. A main difference is that we are learning symbol representations and allow for arbitrarily complex logical rules to be used as regularizers for these representations. Combining Symbolic and Distributed Representations There have been a number of recent approaches that combine trainable subsymbolic representations with symbolic knowledge. Grefenstette [2013] describes an isomorphism between firstorder logic and tensor calculus, using full-rank matrices to exactly memorize facts. Based on this isomorphism, Rocktäschel et al. [2014] combine logic with matrix factorization for learning low-dimensional symbol embeddings that approximately satisfy given rules and generalize to unobserved facts on toy data. Our work extends this workshop paper by proposing a simpler formalism without tensor-based logical 3.6. Related Work 63 connectives, presenting results on a large real-world task, and demonstrating the utility of this approach for learning relations with no or few textual alignments. Chang et al. [2014] use Freebase entity types as hard constraints in a tensor factorization objective for universal schema relation extraction. In contrast, our approach is imposing soft constraints that are formulated as universally quantified first-order rules. de Lacalle and Lapata [2013] combine first-order logic knowledge with a topic model to improve surface pattern clustering for relation extraction. Since these rules only specify which relations can be clustered and which cannot, they do not capture the variety of dependencies embeddings can model, such as asymmetry. Lewis and Steedman [2013] use distributed representations to cluster predicates before logical inference. Again, this approach is not as expressive as learning subsymbolic representations for predicates, as clustering does not deal with asymmetric logical relationships between predicates. Several studies have investigated the use of symbolic representations (such as dependency trees) to guide the composition of symbol representations [Clark and Pulman, 2007, Mitchell and Lapata, 2008, Coecke et al., 2010, Hermann and Blunsom, 2013]. Instead of guiding composition, we are using first-order logic rules as prior domain knowledge in form of regularizers to directly learn better symbol representations. Combining symbolic information with neural networks has a long tradition. Towell and Shavlik [1994] introduce Knowledge-Based Artificial Neural Networks whose topology is isomorphic to a KB of facts and inference rules. There, facts are input units, intermediate conclusions hidden units, and final conclusions (inferred facts) output units. Unlike in our work, there are no learned symbol representations. Hölldobler et al. [1999] and Hitzler et al. [2004] prove that for every logic program there exists a recurrent neural network that approximates the semantics of that program. This is a theoretical insight that unfortunately does not provide a way of constructing such a neural network. Recently, Bowman [2013] demonstrated that Neural Tensor Networks (NTNs) [Socher et al., 2013] can accurately learn natural logic reasoning. The method presented in this chapter is also related to the recently introduced Neural Equivalence Networks (EqNets) [Allamanis et al., 2017]. EqNets recursively construct neural representations of symbolic expressions to learn about equivalence classes. In our approach, we recursively construct neural networks for evaluating Boolean expressions and use them as regularizers to learn better symbol representations for automated KB completion. 64 Chapter 3. Regularizing Representations by First-order Logic Rules 3.7 Summary In this chapter, we introduced a method for mapping symbolic first-order logic rules to differentiable terms that can be used to regularize symbol representations learned by neural link prediction models for automated KB completion. Specifically, we proposed a joint training objective that maximizes the probability of known training facts as well as propositional rules that we made continuous by replacing logical operators with differentiable functions. While this is inspired by Fuzzy Logic, our contribution is backpropagating a gradient from a negative log-likelihood loss through the propositional rule and a neural link prediction model that scores ground atoms to calculate a gradient with respect to vector representations of symbols. Subsequently, we update these representations using gradient descent, thereby encoding the ground truth of a propositional rule directly in the vector representations of symbols. This leads to efficient predictions at test time as we only have to calculate the forward pass of the neural link prediction model. We described a stochastic grounding process for incorporating first-order logic rules. Our experiments for automated KB completion show that the proposed method can be used to learn extractors for relations with little to no observed textual alignments, while at the same time benefiting from correlations between textual surface form patterns. Chapter 4 Lifted Regularization of Predicate Representations by Implications The method for incorporating first-order logic rules into symbol representations introduced in the previous chapter relies on stochastic grounding. Moreover, not only vector representations of predicates but also representations of pairs of constants are optimized to maximize the probability of provided rules. This is problematic for the following reasons. Scalability Even with stochastic grounding, incorporating first-order logic rules with the method described so far is dependent on the size of the domain of constants. As an example take the simple rule isMortal(X) :– isHuman(X) and assume we observe isHuman for seven billion constants. Only for this single rule, we would already add seven billion loss terms to the training objective in Eq. 3.2. Generalizability Since we backpropagate upstream gradients of a rule not only into predicate representations but also into representations of pairs of constants, there is no theoretical guarantee that the rule will indeed hold for constant pairs not observed during training. Flexibility of Training Loss The previous method is not compatible with rank-based losses such as Bayesian Personalized Ranking (BPR). Instead, we had to use the negative log-likelihood loss, which results in lower performance compared to BPR for automated Knowledge Base (KB) completion. Independence Assumption As explained in Section 3.2, we had to assume that the probability of ground atoms is conditionally independent given symbol representations. This assumption is already violated for the simple Boolean expression JF ∧ FK with 0 < JFK < 1, which results in JF ∧ FK = JFK JFK < JFK. Ideally, we would like to have a way to incorporate first-order logic rules into symbol representations in a way that (i) is independent of the size of the domain of 66 Chapter 4. Lifted Regularization of Predicate Representations by Implications constants, (ii) generalizes to unseen constant pairs, (iii) can be used with a broader class of training objectives, and (iv) does not assume that the probability of ground atoms is conditionally independent given symbol representations. In this chapter, we present a method that satisfies these desiderata, but only for simple implication rules instead of general function-free first-order logic rules, and for matrix factorization as neural link prediction model. However, we note that such simple implications are commonly used and can improve automated KB completion. The method we propose incorporates implications into vector representations of predicates while maintaining the computational efficiency of only modeling training facts. This is achieved by enforcing a partial order of predicate representations in the vector space, which is entirely independent of the number of constants in the domain. It only involves representations of the predicates that are mentioned in rules, as well as a general rule-independent constraint on the embedding space of constant pairs. In the example given above, if we require that every component of the vector representation JisHumanK is smaller or equal to the corresponding component of the predicate representation JisMortalK, then we can show that the implication holds for any non-negative representation of a constant. Hence, our method avoids the need for separate loss terms for every ground atom resulting from grounding rules. In statistical relational learning, this type of approach is often referred to as lifted inference or learning [Poole, 2003, Braz, 2007] because it deals with groups of random variables at a first-order level. In this sense, our approach is a lifted form of rule injection. This allows us to impose a large number of rules while learning distributed representations of predicates and constant pairs. Furthermore, once these constraints are satisfied, the injected rules always hold, even for unseen but inferred ground atoms. In addition, it does not rely on the assumption of conditional independence of ground atoms. 4.1 Method We want to incorporate implications of the form ∀X, Y : h(X, Y) :– b(X, Y) and we consider matrix factorization as the neural link prediction model for scoring atoms, i.e., Jr(ei , ej )K = σ(JrK> Jei , ej K) (Section 2.3.1). A necessary condition for the implication to hold is that for every possible assignment of constants ei to X and ej to Y, the score of Jh(ei , ej )K needs to be as least as large as Jb(ei , ej )K. In the discrete case this means that if Jb(ei , ej )K is 1.0 (True) within some small epsilon, Jh(ei , ej )K needs to be 1.0 too, but not vice versa. As sigmoid is a monotonic 4.1. Method 67 Jhomer, bartK JfatherOfK JparentOfK Jmarge, bartK JmotherOfK 0 Figure 4.1: Example for implications that are directly captured in a vector space. function, we can rewrite ∀X, Y : h(X, Y) :– b(X, Y) in terms of grounded rules as the following condition: ∀(ei , ej ) ∈ C : JhK> Jei , ej K ≥ JbK> Jei , ej K . (4.1) Ordering symbol representations in a vector space such that implications are directly captured is inspired by Order Embeddings [Vendrov et al., 2016]. An example where Eq. 4.1 holds is illustrated in Fig. 4.1. Here, fatherOf and motherOf both imply parentOf, since every component of JparentOfK is larger than the corresponding component in JfatherOfK or JmotherOfK. Predicate representations in the blue, red and purple area are implied by fatherOf, motherOf and parentOf, respectively. Both constant pair representations, JHOMER, BARTK and JMARGE, BARTK, are non-negative. Thus, for any score of JfatherOf(HOMER, BART)K, the score of JparentOf(HOMER, BART)K is larger, but not vice versa. Note that this holds for any representation of constant pairs Jei , ej K as long as it is non-negative (i.e. placed in the upper-right quadrant). The main insight is that we can make this condition independent of ei , ej if we make sure that Jei , ej K ∈ Rk+ , i.e., all constant representations are non-negative. Thus, Eq. 4.1 becomes JhK ≥ JbK , ∀(ei , ej ) ∈ C : Jei , ej K ∈ Rk+ (4.2) where ≥ is the component-wise comparison. In other words, for ensuring that b implies h for any pair of constants, we only need one relation-specific loss term that 68 Chapter 4. Lifted Regularization of Predicate Representations by Implications makes sure all components in JhK are at least as large as in JbK, and one general restriction on the representations of constant pairs. Non-negative Representation of Constant Pairs There are many choices for ensuring all constant pair representations are positive. One option is to initialize constant pair representations to non-negative vectors and projecting gradient updates to make sure they stay non-negative. Another option is to apply a transformation f : Rk → Rk+ before constant pair representations are used in a neural link prediction model for scoring atoms. For instance, we could use exp(x) = ex or ReLU(x) = max(0, x) for f . However, we choose to restrict constant representations even more than required, and decided to use a transformation to approximately Boolean embeddings [Kruszewski et al., 2015]. For every constant pair representation Jei , ej K ∈ Rk , we obtain a non-negative representation Jei , ej K+ ∈ [0, 1]k by applying the element-wise sigmoid function. Thus, the matrix factorization objective with BPR for facts in Eq. 2.9 becomes the following approximate loss L= X rs (ei ,ej ) ∈ O, (em ,en ) ∼ C, rs (em ,en ) 6∈ O −wr log σ(vr> σ(vij )−vr> σ(vmn ))+λp kvs k22 +λc (kvij k22 +kvmn k22 ) (4.3) where we sample constant pairs as in Section 2.3.1.1. We denote the extension with sigmoidal constant pair representations of the matrix factorization model (F) by [Riedel et al., 2013] as FS (S = sigmoidal). Implication Loss There are various ways for modeling JhK ≥ JbK to incorporate the implication h :– b. Here, we propose to use a hinge-like loss L(Jh :– bK) = k X i max(0, JbKi − JhKi + ) (4.4) X (4.5) where  is a small positive margin to ensure that the gradient does not disappear before the inequality is actually satisfied. A nice property of this loss compared to the method presented in the previous chapter is that once the implication holds (i.e. JhK is larger than JbK), the gradient is zero and both predicate representations are not further updated for this rule. For every given implication rule, we add the corresponding loss term to the fact loss in Eq. 4.3. The global approximate loss over facts and rules in a set R is thus L0 = L + h :– b ∈ R L(Jh :– bK) 4.2. Experiments 69 and we denote the resulting model as FSL (L = logic). Furthermore, we use a margin of  = 0.01 in all experiments. As in Chapter 3, we predict the probability of an atom rs (ei , ej ) at test time via Jrs (ei , ej )K, which is efficient as there is no logical inference. 4.2 Experiments We follow the experimental setup from the previous chapter (Section 3.4) and evaluate on the NYT corpus [Riedel et al., 2013]. Again, we test how well the presented models can incorporate rules when there is no alignment between textual surface forms and Freebase relations (Zero-shot Relation Learning), and when the number of Freebase training facts is increased (Relations with Few Distant Labels). In addition, we experiment with rules automatically extracted from WordNet [Miller, 1995] to improve automated KB completion on the full dataset. Incorporating Background Knowledge from WordNet We use WordNet hypernyms to generate rules for the NYT dataset. To this end, we iterate over all surface form patterns in the dataset and attempt to replace words in the pattern by their hypernyms. If the resulting surface form is contained in the dataset, we generate the corresponding rule. For instance, we generate a rule #1-official-#2(X, Y) :– #1-diplomat-#2(X, Y) since both patterns are contained in the dataset and we know from WordNet that official is a hypernym of diplomat. This resulted in 427 generated rules that we subsequently annotated manually, yielding 36 high-quality rules listed in Appendix B. Note that all of these rules are between surface form patterns. Thus, none of these rules has a Freebase relation as the head predicate. Although the test relations all originate from Freebase, we still hope to see improvements by transitive effects, such as better surface form representations that in turn help to predict Freebase facts. 4.2.1 Training Details All models were implemented in TensorFlow [Abadi et al., 2016]. We use the hyperparameters of Riedel et al. [2013], with k = 100 as the size of symbol representations and a weight of 0.01 for the `2 regularization (Eq. 4.3). We use ADAM [Kingma and Ba, 2015] for optimization with an initial learning rate of 0.005 and a batch size of 8192. The embeddings are initialized by sampling k values uniformly from [−0.1, 0.1]. 70 Chapter 4. Lifted Regularization of Predicate Representations by Implications Test relation # person/company location/containedby person/nationality author/works written person/place of birth parent/child person/place of death neighborhood/neighborhood of person/parents company/founders sports team/league team owner/teams owned team/arena stadium film/directed by broadcast/area served structure/architect composer/compositions person/religion film/produced by 106 73 28 27 21 19 19 11 6 4 4 2 2 2 2 2 2 1 1 Weighted MAP Riedel13-F F FS FSL 0.75 0.69 0.19 0.65 0.72 0.76 0.83 0.70 0.61 0.77 0.59 0.38 0.13 0.50 0.58 1.00 0.67 1.00 0.50 0.73 0.62 0.20 0.71 0.69 0.77 0.85 0.67 0.53 0.73 0.44 0.64 0.13 0.18 0.83 1.00 0.64 1.00 1.00 0.74 0.70 0.20 0.69 0.72 0.81 0.83 0.63 0.66 0.64 0.43 0.64 0.13 0.17 0.83 1.00 0.51 1.00 1.00 0.77 0.71 0.21 0.65 0.70 0.85 0.85 0.62 0.66 0.67 0.56 0.61 0.12 0.13 1.00 1.00 0.50 1.00 0.33 0.67 0.65 0.67 0.69 Table 4.1: Weighted MAP for our reimplementation of the matrix factorization model (F), compared to restricting the constant embedding space (FS) and to injecting WordNet rules (FSL). The orginial matrix factorization model by Riedel et al. [2013] is denoted as Riedel13-F. 4.3 Results and Discussion Before turning to the injection of rules, we compare model F with model FS, and show that restricting the constant embedding space has a regularization effect rather than limiting the expressiveness of the model (Section 4.3.1). We then demonstrate that model FSL is capable of zero-shot learning (Section 4.3.2), that it can take advantage of alignments between textual surface forms and Freebase relations alongside rules (Section 4.3.3), and we show that injecting high-quality WordNet rules leads to improved predictions on the full dataset (Section 4.3.4). Lastly, we provide details on the computational efficiency of the lifted rule injection method (Section 4.3.5) and demonstrate that it correctly captures the asymmetry of implication rules (Section 4.3.6). 4.3.1 Restricted Embedding Space for Constants Before incorporating external commonsense knowledge into relation representations, we were curious about how much we lose by restricting the embedding space of 4.3. Results and Discussion 71 constant symbols to approximately Boolean embeddings. Surprisingly, we find that the expressiveness of the model does not suffer from this strong restriction. From Table 4.1 we see that restricting the constant embedding space (FS) yields a 2 percentage points higher weighted Mean Average Precision (MAP) compared to a real-valued constant embedding space (F). This result suggests that the restriction has a regularization effect that improves generalization. We also provide the original results for the matrix factorization model by Riedel et al. [2013] denoted as Riedel13F for comparison. Due to a different implementation and optimization procedure, the results for our model F compared to Riedel13-F are slightly worse (65% vs 67% wMAP). 4.3.2 Zero-shot Relation Learning In Section 3.5.1, we observed that by injecting implications where the head is a Freebase relation for which no training facts are available, we can infer Freebase facts based on rules and correlations between textual surface patterns. Here, we repeat this experiment. The lifted rule injection model (FSL) reaches a weighted MAP of 35%, comparable to the 38% of the method presented in the last chapter. For this experiment, we initialized the predicate representations of Freebase relations implied by the rules with negative random vectors sampled uniformly from [−7.9, −8.1]. The reason is that without any negative training facts for these relations, their components can only go up due to the lifted implication loss. Consequently, starting with high values before optimization would impede the freedom with which these representations can be ordered in the embedding space. This demonstrates that while FSL performs a bit worse than the Joint model in Chapter 3, it can still be used for zero-shot relation learning. 4.3.3 Relations with Few Distant Labels Figure 4.2 shows how the relation extraction performance improves when more Freebase facts are added. As in the last chapter, it measures how well the proposed models, matrix factorization (F), propositionalized rule injection (Joint), and our lifted rule injection model (FSL), can make use of provided implication rules, as well as correlations between textual surface form patterns and increasing numbers of Freebase facts. Although FSL starts at a lower performance than Joint when no Freebase training facts are present, it outperforms Joint and a plain matrix factorization model by a substantial margin when provided with more than 7.5% of Freebase facts. This indicates that, in addition to being much faster than Joint, it can make better use of provided rules and few training facts. We attribute this 72 Chapter 4. Lifted Regularization of Predicate Representations by Implications 0.7 weighted Mean Average Precision 0.6 0.5 0.4 F Joint FSL 0.3 0.2 0.1 0.0 0.0 0.1 0.2 0.3 0.4 0.5 Fraction of Freebase training facts Figure 4.2: Weighted MAP for injecting rules as a function of the fraction of Freebase facts. to being able to use BPR as loss for ground atoms and the regularization effect of restricting the embedding space of constants pairs. The former is compatible with our rule-injection method, but not with the approach of maximizing the expectation of propositional rules presented in the previous chapter. 4.3.4 Incorporating Background Knowledge from WordNet In column FSL in Table 4.1, we show results obtained by injecting WordNet rules. Compared to FS, we obtain an increase of weighted MAP by 2%, as well as 4% compared to our reimplementation of the matrix factorization model F. This demonstrates that imposing a partial order based on implication rules can be used to incorporate logical commonsense knowledge and increase the quality of information extraction and automated KB completion systems. Note that our evaluation setting guarantees that only indirect effects of the rules are measured, i.e., we do not use any rules directly implying Freebase test relations. Consequently, our increase of prediction performance is due to an improved predicate embedding space beyond those predicates that are explicitly stated in provided rules. For example, injecting 4.3. Results and Discussion head Rule :– body #1-organization-#2 #1-parent-#2 #1-lawyer-#2 #1-newspaper-#2 #1-diplomat-#2 :– :– :– :– :– #1-party-#2 #1-father-#2 #1-prosecutor-#2 #1-daily-#2 #1-ambassador-#2 Average over all rules FS Jhead(ei , ej )K Jbody(êi , êj )K 0.70 0.72 0.87 0.90 0.93 0.86 0.89 0.80 0.86 0.84 0.74 0.70 73 FSL Jhead(ei , ej )K Jbody(êi , êj )K 0.99 0.96 0.99 0.98 0.31 0.22 0.00 0.01 0.79 0.05 0.95 0.28 Table 4.2: Average score of facts with constants that appear in the body of facts (ei , ej ) or in the head (êi , êj ) of a rule. the rule #1-parent-#2(X, Y) :– #1-father-#2(X, Y) can contribute to improved predictions for the Freebase test relation parent/child via co-occurring entity pairs between #1-parent-#2(X, Y) and parent/child. 4.3.5 Computational Efficiency of Lifted Rule Injection In order to assess the computational efficiency of the proposed method, we measure the time needed per training epoch when using a single 2.4GHz CPU core. We measure on average 6.33s per epoch when not using rules (model FS), against 6.76s and 6.97s when using 36 filtered and 427 unfiltered rules (model FSL), respectively. Increasing the number of rules from 36 to 427 leads to an increase of only 3% in computation time. Furthermore, using 427 rules only adds an overhead of 10% to the computation time needed for learning ground atoms. This demonstrates that lifted rule injection scales well with the number of rules. 4.3.6 Asymmetry One concern with incorporating implications into a vector space is that the vector representation of the predicates in the head and body are simply moving closer together. This would violate the asymmetry of implications. In the experiments above we might not observe that this is a problem as we are only testing how well the model predicts facts for Freebase relations and not how well we can predict textual surface form patterns. Thus, we perform the following experiment. After incorporating WordNet rules of the form head :– body, we select all constant pairs (ei , ej ) for which we observe body(ei , ej ) in the training set. If the implication holds, Jhead(ei , ej )K should yield a high score. If we conversely select (êi , êj ) based on known facts head(êi , êj ) and assume that head and body are not equivalent, then we expect a lower score for Jbody(êi , êj )K than for Jhead(êi , êj )K as body(êi , êj ) might not be true. Table 4.2 lists these scores for five sampled WordNet rules, and the average over all WordNet rules when injecting these rules (model FSL) or not (model FS). From this table, we can see that the score of the head atom is on average much higher 74 Chapter 4. Lifted Regularization of Predicate Representations by Implications than the score for the body atom once we incorporate the rule in the vector space (FSL). In contrast, if we only run matrix factorization with the restricted constant embedding space, we often see high predictions for both, the body and head atom. This suggests that matrix factorization merely captures similarity between predicates. In contrast, by injecting implications we trained predicate representations that indeed yield asymmetric predictions. Given a high score for a body ground atom, the model predicts a high score for the head, but not vice versa. The reason that we also get a high score for body ground atoms in the fourth rule is that newspaper and daily are synonymously used in training texts. 4.4 Related Work Recent research on combining rules with learned vector representations has been important for new developments in the field of automated KB completion. Wang et al. [2015] demonstrated how different types of rules can be incorporated using an Integer Linear Programming approach. Wang and Cohen [2016] learned embeddings for facts and first-order logic rules using matrix factorization. Yet, all of these approaches, and the method presented in the previous chapter, ground first-order rules with constants in the domain. This limits their scalability towards large rule sets and KBs with large domains of constants. It formed an important motivation for our lifted rule injection model, which by construction does not suffer from that limitation. Wei et al. [2015] proposed an alternative strategy to tackle the scalability problem by reasoning on a filtered subset of ground atoms. Wu et al. [2015] proposed to use the Path Ranking Algorithm (PRA) (Section 2.3.3) for capturing long-range interactions between entities in conjunction with modeling pairwise relations. Our model differs substantially from their approach, in that we consider pairs of constants instead of separate constants, and that we inject a provided set of rules. Yet, by creating a partial order in the relation embeddings as a result of injecting implication rules, model FSL can also capture interactions beyond the predicates directly mentioned in these rules, which we demonstrated in Section 4.3.4 by injecting rules between surface patterns and measuring an improvement on predictions for structured Freebase relations. Combining logic and distributed representations is also an active field of research outside of automated KB completion. Recent advances include the work by Faruqui et al. [2015], who injected ontological knowledge from WordNet into word embeddings to improve performance on downstream NLP tasks. Furthermore, Vendrov et al. [2016] proposed to enforce a partial order in an embeddings space of images and phrases. Our method is related to such Order Embeddings since we 4.5. Summary 75 define a partial order on relation embeddings. We extend this work to automated KB completion where we ensure that implications hold for all pairs of constants by introducing a restriction on the embedding space of constant pairs. Another important contribution is the recent work by Hu et al. [2016], who proposed a framework for injecting rules into general neural network architectures by jointly training on target outputs and on rule-regularized predictions provided by a so-called teacher network. Although quite different at first sight, their work could offer a way to use our model in various neural network architectures by integrating the proposed lifted loss into the teacher network. 4.5 Summary We presented a fast approach for incorporating first-order implication rules into distributed representations of predicates for automated KB completion. We termed our approach lifted rule injection, as the main contribution over the previous chapter is the fact that it avoids the costly grounding of first-order implication rules and is thus independent of the size of the domain of constants. By construction, these rules are satisfied for any observed or unobserved ground atom. The presented approach requires a restriction on the embedding space of constant pairs. However, experiments on a large-scale real-world KB show that this does not impair the expressiveness of the learned representations. On the contrary, it appears to have a beneficial regularization effect. By incorporating rules generated from WordNet hypernyms, our model improved over a matrix factorization baseline for KB completion. Especially for domains where annotation is costly and only small amounts of training facts are available, our approach provides a way to leverage external knowledge sources efficiently for inferring facts. On the downside, the lifted rule injection method presented here is only applicable for implication rules and when using matrix factorization as the neural link prediction model. Furthermore, it is unclear how far regularizing predicate representations can be pushed without constraining the embedding space too much. Specifically, it is unclear how more complex rules such as transitivity can be incorporated in a lifted way. Hence, we are exploring a more direct synthesis of representation learning and first-order logic inference in the next chapter. Chapter 5 End-to-End Differentiable Proving Current state-of-the-art methods for automated Knowledge Base (KB) completion use neural link prediction models to learn distributed vector representations of symbols (i.e. subsymbolic representations) for scoring atoms [Nickel et al., 2012, Riedel et al., 2013, Socher et al., 2013, Chang et al., 2014, Yang et al., 2015, Toutanova et al., 2015, Trouillon et al., 2016]. Such subsymbolic representations enable these models to generalize to unseen facts by encoding similarities: If the vector of the predicate symbol grandfatherOf is similar to the vector of the symbol grandpaOf, both predicates likely express a similar relation. Likewise, if the vector of the constant symbol LISA is similar to MAGGIE, similar relations likely hold for both constants (e.g. they live in the same city, have the same parents etc.). This simple form of reasoning based on similarities is remarkably effective for automatically completing large KBs. However, in practice it is often important to capture more complex reasoning patterns that involve several inference steps. For example, if ABE is the father of HOMER and HOMER is a parent of BART, we would like to infer that ABE is a grandfather of BART. Such transitive reasoning is inherently hard for neural link prediction models as they only learn to score facts locally. In contrast, symbolic theorem provers like Prolog [Gallaire and Minker, 1978] enable exactly this type of multi-hop reasoning. Furthermore, Inductive Logic Programming (ILP) [Muggleton, 1991] builds upon such provers to learn interpretable rules from data and to exploit them for reasoning in KBs. However, symbolic provers lack the ability to learn subsymbolic representations and similarities between them from large KBs, which limits their ability to generalize to queries with similar but not identical symbols. While the connection between logic and machine learning has been addressed by statistical relational learning approaches, these models traditionally do not support reasoning with subsymbolic representations (e.g. Kok and Domingos [2007]), and 78 Chapter 5. End-to-End Differentiable Proving when using subsymbolic representations they are not trained end-to-end from training data (e.g. Gardner et al. [2013, 2014], Beltagy et al. [2017]). Neural multi-hop reasoning models [Neelakantan et al., 2015, Peng et al., 2015, Das et al., 2017, Weissenborn, 2016, Shen et al., 2016] address the aforementioned limitations to some extent by encoding reasoning chains in a vector space or by iteratively refining subsymbolic representations of a question before comparison with answers. In many ways, these models operate like basic theorem provers, but they lack two of their most crucial ingredients: interpretability and straightforward ways of incorporating domain-specific knowledge in form of rules. Our approach to this problem is inspired by recent neural network architectures like Neural Turing Machines [Graves et al., 2014], Memory Networks [Weston et al., 2014], Neural Stacks/Queues [Grefenstette et al., 2015, Joulin and Mikolov, 2015], Neural Programmer [Neelakantan et al., 2016], Neural Programmer-Interpreters [Reed and de Freitas, 2016], Hierarchical Attentive Memory [Andrychowicz et al., 2016] and the Differentiable Forth Interpreter [Bosnjak et al., 2017]. These architectures replace discrete algorithms and data structures by end-to-end differentiable counterparts that operate on real-valued vectors. At the heart of our approach is the idea to translate this concept to basic symbolic theorem provers, and hence combine their advantages (multi-hop reasoning, interpretability, easy integration of domain knowledge) with the ability to reason with vector representations of predicates and constants. Specifically, we keep variable binding symbolic but compare symbols using their subsymbolic vector representations. In this chapter we introduce Neural Theorem Provers (NTPs): End-to-end differentiable provers for basic theorems formulated as queries to a KB. We use Prolog’s backward chaining algorithm as a recipe for recursively constructing neural networks that are capable of proving queries to a KB using subsymbolic representations. The success score of such proofs is differentiable with respect to vector representations of symbols, which enables us to learn such representations for predicates and constants in ground atoms, as well as parameters of function-free first-order logic rules of predefined structure. By doing so, NTPs learn to place representations of similar symbols in close proximity in a vector space and to induce rules given prior assumptions about the structure of logical relationships in a KB such as transitivity. Furthermore, NTPs can seamlessly reason with provided domain-specific rules. As NTPs operate on distributed representations of symbols, a single hand-crafted rule can be leveraged for many proofs of queries with symbols that have a similar representation. Finally, NTPs demonstrate a high degree of interpretability as they induce latent rules that we can decode to human-readable symbolic rules. 5.1. Differentiable Prover X/Q Y/BART Z/HOMER X/Q Y/BART SΨ 79 Sτ SΨ0 Sτ0 Figure 5.1: A module is mapping an upstream proof state (left) to a list of new proof states (right), thereby extending the substitution set SΨ and adding nodes to the computation graph of the neural network Sτ representing the proof success. Our contributions are threefold: (i) We present the construction of NTPs inspired by Prolog’s backward chaining algorithm and a differentiable unification operation using subsymbolic representations, (ii) we propose optimizations to this architecture by joint training with a neural link prediction model, batch proving, and approximate gradient calculation, and (iii) we experimentally show that NTPs can learn representations of symbols and function-free first-order rules of predefined structure, enabling them to learn to perform multi-hop reasoning on benchmark KBs and to outperform ComplEx [Trouillon et al., 2016], a state-of-the-art neural link prediction model, on three out of four KBs. 5.1 Differentiable Prover In the following, we describe the recursive construction of NTPs – neural networks for end-to-end differentiable proving that allow us to calculate the gradient of proof successes with respect to vector representations of symbols. We define the construction of NTPs in terms of modules similar to dynamic neural module networks [Andreas et al., 2016]. Each module takes as inputs discrete objects (atoms and rules) and a proof state, and returns a list of new proof states (see Figure 5.1 for a graphical representation). A proof state S = (Ψ, τ ) is a tuple consisting of the substitution set Ψ constructed in the proof so far and a neural network τ that outputs a real-valued success score of a (partial) proof. While discrete objects and the substitution set are only used during construction of the neural network, once the network is constructed a continuous proof success score can be calculated for many different goals at training and test time. To summarize, modules are instantiated by discrete objects and the substitution set. They construct a neural network representing the (partial) proof success score and recursively instantiate submodules to continue the proof. The shared signature of modules is D × S → S N where D is a domain that controls the construction of the network, S is the domain of proof states, and N is the number of output proof states. Furthermore, let SΨ denote the substitution set of the 80 Chapter 5. End-to-End Differentiable Proving proof state S and let Sτ denote the neural network for calculating the proof success. Akin to the pseudocode of backward chaining in Chapter 2, we use pseudocode in style of a functional programming language to define the behavior of modules and auxiliary functions. 5.1.1 Unification Module Unification of two atoms, e.g., a goal that we want to prove and a rule head, is a central operation in backward chaining. Two non-variable symbols (predicates or constants) are checked for equality and the proof can be aborted if this check fails. However, we want to be able to apply rules even if symbols in the goal and head are not equal but similar in meaning (e.g. grandfatherOf and grandpaOf) and thus replace symbolic comparison with a computation that measures the similarity of both symbols in a vector space. The module unify updates a substitution set and creates a neural network for comparing the vector representations of non-variable symbols in two sequences of terms. The signature of this module is L × L × S → S where L is the domain of lists of terms. unify takes two atoms represented as lists of terms and an upstream proof state, and maps these to a new proof state (substitution set and proof success). To this end, unify iterates through the list of terms of two atoms and compares their symbols. If one of the symbols is a variable, a substitution is added to the substitution set. Otherwise, the vector representations of the two non-variable symbols are compared using a Radial Basis Function (RBF) kernel [Broomhead and Lowe, 1988] where µ is a hyperparameter that we set to √12 in our experiments. The following pseudocode implements unify. Note that ” ” matches every argument and that the order matters, i.e., if arguments match a line, subsequent lines are not evaluated. 1. unifyθ ([ ], [ ], S) = S 2. unifyθ ([ ], , ) = FAIL 3. unifyθ ( , [ ], ) = FAIL 4. unifyθ (h : H, g : G, S) = unifyθ (H, G, S 0 ) = (SΨ0 , Sτ0 ) where     S ∪ {h/g} if h ∈ V Ψ   0 SΨ = SΨ ∪ {g/h} if g ∈ V, h 6∈ V     SΨ otherwise   ( )! −kθh: −θg: k2 exp if h, g ∈ 6 V 2 2µ Sτ0 = min Sτ , 1 otherwise 5.1. Differentiable Prover 81 Here, S 0 refers to the new proof state, V refers to the set of variable symbols, h/g is a substitution from the variable symbol h to the symbol g, and θg: denotes the embedding lookup of the non-variable symbol with index g. unify is parameterized by an embedding matrix θ ∈ R|Z|×k where Z is the set of non-variables symbols and k is the dimension of vector representations of symbols. Furthermore, FAIL represents a unification failure due to mismatching arity of two atoms. Once a failure is reached, we abort the creation of the neural network for this branch of proving. In addition, we constrain proofs to be cycle-free by checking whether a variable is already bound. Note that this is a simple heuristic that prohibits applying the same non-ground rule twice. There are more sophisticated ways for finding and avoiding cycles in a proof graph such that the same rule can still be applied multiple times (e.g. Gelder [1987]), but we leave this for future work. Example Assume that we are unifying two atoms [grandpaOf, ABE, BART] and [s, Q, i] given an upstream proof state S = (∅, τ ) where the latter input atom has placeholders for a predicate s and a constant i, and the neural network τ would output 0.7 when evaluated. Furthermore, assume grandpaOf, ABE and BART represent the indices of the respective symbols in a global symbol vocabulary. Then, the new proof state constructed by unify is: unifyθ ([grandpaOf, ABE, BART], [s, Q, i], (∅, τ )) = (SΨ0 , Sτ0 ) = ({Q/ABE}, min (τ, exp(−kθgrandpaOf: − θs: k2 ), exp(−kθBART: − θi: k2 ))) Thus, the output score of the neural network Sτ0 will be high if the subsymbolic representation of the input s is close to grandpaOf and the input i is close to BART. However, the score cannot be higher than 0.7 due to the upstream proof success score in the forward pass of the neural network τ . Note that in addition to extending the neural network τ to Sτ0 , this module also outputs a substitution set {Q/ABE} at graph creation time that will be used to instantiate submodules. Furthermore, note that unification is applied multiple times for a proof that involves more than one step, resulting in chained application of the RBF kernel and min operations. The choice of min stems from the property that for a successful proof, all unifications should be successful (conjunction). This could also be realized with a multiplication of unification scores along the proof, but it would likely result in unstable optimization for longer proofs due to exploding gradients. 82 Chapter 5. End-to-End Differentiable Proving 5.1.2 OR Module Based on unify, we now define the or module which attempts to apply rules in a KB. The signature of or is L × N × S → S N where L is the domain of goal atoms and N is the domain of integers used for specifying the maximum proof depth of the neural network. Furthermore, N is the number of possible output proof states for a goal of a given structure and a provided KB.1 We implement or as 1. orKθ (G, d, S) = [S 0 | S 0 ∈ andKθ (B, d, unifyθ (H, G, S)) for H :– B ∈ K] where H :– B denotes a rule in a given KB K with a head atom H and a list of body atoms B. In contrast to the symbolic O R method, the or module is able to use the grandfatherOf rule above for a query involving grandpaOf provided that the subsymbolic representations of both predicates are similar as measured by the RBF kernel in the unify module. Example For a goal [s, Q, i], or would instantiate an and submodule based on the rule [grandfatherOf, X, Y] :– [[fatherOf, X, Z], [parentOf, Z, Y]] as follows orKθ ([s, Q, i], d, S) = [S 0 | S 0 ∈ andKθ ([[fatherOf, X, Z], [parentOf, Z, Y]], d, ({X/Q, Y/i}, Ŝτ )), . . .] | {z } result of unify 5.1.3 AND Module For implementing and we first define an auxiliary function called substitute which applies substitutions to variables in an atom if possible. This is realized via 1. substitute([ ], ) = [ ] ( 2. substitute(g : G, Ψ) = x if g/x ∈ Ψ g otherwise ) : substitute(G, Ψ) For example, substitute([fatherOf, X, Z], {X/Q, Y/i}) = [fatherOf, Q, Z]. 1 The creation of the neural network is dependent on the KB but also the structure of the goal. For instance, the goal s(Q, i) would result in a different neural network, and hence a different number of output proof states, than s(i, j). 5.1. Differentiable Prover 83 The signature of and is L × N × S → S N where L is the domain of lists of atoms and N is the number of possible output proof states for a list of atoms with a known structure and a provided KB. This module is implemented as 1. andKθ ( , , FAIL) = FAIL 2. andKθ ( , 0, ) = FAIL 3. andKθ ([ ], , S) = S 4. andKθ (G : G, d, S) = [S 00 | S 00 ∈ andKθ (G, d, S 0 ) for S 0 ∈ orKθ (substitute(G, SΨ ), d − 1, S)] where the first two lines define the failure of a proof, either because of an upstream unification failure that has been passed from the or module (line 1), or because the maximum proof depth has been reached (line 2). Line 3 specifies a proof success, i.e., the list of subgoals is empty before the maximum proof depth has been reached. Lastly, line 4 defines the recursion: The first subgoal G is proven by instantiating an or module after substitutions are applied, and every resulting proof state S 0 is used for proving the remaining subgoals G by again instantiating and modules. Example Continuing the example from Section 5.1.2, the and module would instantiate submodules as follows: andKθ ([[fatherOf, X, Z], [parentOf, Z, Y]], d, ({X/Q, Y/i}, Ŝτ ) ) = {z } | result of unify in or [S 00 | S 00 ∈ andKθ ([[parentOf, Z, Y]], d, S 0 ) for S 0 ∈ orKθ ([fatherOf, Q, Z], d − 1, ({X/Q, Y/i}, Ŝτ ) )] | {z } | {z } result of substitute result of unify in or 5.1.4 Proof Aggregation Finally, we define the overall success score of proving a goal G using a KB K with parameters θ as ntpKθ (G, d) = S ∈ arg max Sτ orKθ (G,d,(∅,1)) S6=FAIL where d is a predefined maximum proof depth and the initial proof state is set to an empty substitution set and a proof success score of 1. Hence, the success of proving 84 Chapter 5. End-to-End Differentiable Proving 1. orK θ ([s, i, j], 2, (∅, 1)) unifyθ ([fatherOf, abe, homer], [s, i, j], (∅, 1)) 2. ... S1 = (∅, τ1 ) S2 = (∅, τ2 ) 3. unifyθ ([grandfatherOf, X, Y], [s, i, j], (∅, 1)) S3 = ({X/i, Y/j}, τ3 ) Example Knowledge Base: 1. fatherOf(abe, homer). 2. parentOf(homer, bart). 3. grandfatherOf(X, Y) :– fatherOf(X, Z), parentOf(Z, Y). andK θ ([[fatherOf, X, Z], [parentOf, Z, Y]], 2, S3 ) substitute orK θ ([fatherOf, i, Z], 1, S3 ) 1. 2. 3. ... unifyθ ([fatherOf, abe, homer], [fatherOf, i, Z], S3 ) unifyθ ([parentOf, homer, bart], [fatherOf, i, Z], S3 ) S31 = ({X/i, Y/j, Z/homer}, τ31 ) andK θ ([parentOf, Z, Y], 2, S31 ) substitute orK θ ([parentOf, homer, j], 1, S31 ) . . . 1. S311 = ({X/i, Y/j, Z/homer}, τ311 ) 2. ... S33 = FAIL S32 = ({X/i, Y/j, Z/bart}, τ32 ) andK θ ([parentOf, Z, Y], 2, S32 ) substitute orK θ ([parentOf, bart, j], 1, S32 ) 3. . . . ... S313 = FAIL S312 = ({X/i, Y/j, Z/homer}, τ312 ) 3. S323 = FAIL 2. 1. . . . ... S321 = ({X/i, Y/j, Z/bart}, τ321 ) S322 = ({X/i, Y/j, Z/bart}, τ322 ) Figure 5.2: Exemplary construction of an NTP computation graph for a toy knowledge base. Indices on arrows correspond to application of the respective KB rule. Proof states (blue) are subscripted with the sequence of indices of the rules that were applied. Underlined proof states are aggregated to obtain the final proof success. Boxes visualize instantiations of modules (omitted for unify). The proofs S33 , S313 and S323 fail due to cycle-detection (the same rule cannot be applied twice). a goal is a max-pooling operation over the output of neural networks representing all possible proofs up to some depth. Example Figure 5.2 illustrates an examplary NTP computation graph constructed for a toy KB. Note that such an NTP is constructed once before training, and can then be used for proving goals of the structure [s, i, j] at training and test time where s is the index of an input predicate, and i and j are indices of input constants. Final proof states which are used in proof aggregation are underlined. 5.1.5 Neural Inductive Logic Programming We can use NTPs for ILP by gradient descent instead of a combinatorial search over the space of rules as, for example, done by the First Order Inductive Learner (FOIL) [Quinlan, 1990]. Specifically, we are using the concept of learning from entailment [Muggleton, 1991] to induce rules that let us prove known ground atoms, but that do not give high proof success scores to sampled unknown ground atoms. Let θr: , θs: , θt: ∈ Rk be representations of some unknown predicates with indices r, s and t respectively. The prior knowledge of a transitivity between three unknown predicates can be specified via r(X, Y) :– s(X, Z), t(Z, Y). We call this a parameterized rule as the corresponding predicates are unknown and their representations are learned from data. Such a rule can be used for proofs at training and 5.2. Optimization 85 test time in the same way as any other given rule. During training, the predicate representations of parameterized rules are optimized jointly with all other subsymbolic representations. Thus, the model can adapt parameterized rules such that proofs for known facts succeed while proofs for sampled unknown ground atoms fail, thereby inducing rules of predefined structures like the one above. Inspired by Wang and Cohen [2015], we use rule templates for conveniently defining the structure of multiple parameterized rules by specifying the number of parameterized rules that should be instantiated for a given rule structure (see Section 5.3.1 for examples). 5.1.5.1 Rule Decoding and Implicit Rule Confidence For inspection after training, we decode a parameterized rule by searching for the closest representations of known predicates. Given an induced rule such as θ1: (X, Y) :– θ2: (X, Z), θ2: (Z, Y) where θ1: and θ2: have been trained, we find the closest representation of a known predicate for every parameterized predicate representation in the rule (e.g., θ1: → θgrandparentOf: , θ2: → θparentOf: ). Formally, we decode θi: to a predicate symbol from the set of all predicates P using decode(θi: ) = arg max exp(−kθi: − θrs : k2 ). rs ∈P (5.1) In addition, we provide users with a rule confidence by taking the minimum similarity between unknown and decoded predicate representations using the RBF kernel in unify. Let Θ = [θi ] be the list of predicate representations of a parameterized rule. The confidence of that rule is then calculated as γ = min max exp(−kθi: − θrs : k2 ). θi: ∈Θ rs ∈P (5.2) This confidence score is an upper bound on the proof success score that can be achieved when the induced rule is used in proofs. 5.2 Optimization In this section, we present the basic training loss that we use for NTPs, a training loss where a neural link prediction models is used as auxiliary task, as well as various computational optimizations. 5.2.1 Training Objective Let K be the set of known facts in a given KB. Usually, we do not observe negative facts and thus resort to sampling corrupted ground atoms as done in previous work [Bordes et al., 2013]. Specifically, for every [s, i, j] ∈ K we obtain corrupted ground atoms [s, î, j], [s, i, ĵ], [s, ĩ, j̃] 6∈ K by sampling î, ĵ, ĩ and j̃ from the set of 86 Chapter 5. End-to-End Differentiable Proving constants. These corrupted ground atoms are resampled in every iteration of training, and we denote the set of known and corrupted ground atoms together with their target score (1.0 for known ground atoms and 0.0 for corrupted ones) as T . We use the negative log-likelihood of the proof success score as loss function for an NTP with parameters θ and a given KB K LntpK = θ X ([s,i,j],y) ∈ T −y log(ntpKθ ([s, i, j], d)τ ) − (1 − y) log(1 − ntpKθ ([s, i, j], d)τ ) where [s, i, j] is a training ground atom and y its target proof success score. Note that since in our application all training facts are ground atoms, we only make use of the proof success score τ and not the substitution list of the resulting proof state. We can prove known facts trivially by a unification with themselves, resulting in no parameter updates during training and hence no generalization. Therefore, during training we are masking the calculation of the unification success of a known ground atom that we want to prove. Specifically, we set the unification score to 0 to temporarily hide that training fact and assume it can be proven from other facts and rules in the KB. 5.2.2 Neural Link Prediction as Auxiliary Loss At the beginning of training all subsymbolic representations are initialized randomly. When unifying a goal with all facts in a KB we consequently get very noisy success scores in early stages of training. Moreover, as only the maximum success score will result in gradient updates for the respective subsymbolic representations along the maximum proof path, it can take a long time until NTPs learn to place similar symbols close to each other in the vector space and to make effective use of rules. To speed up learning subsymbolic representations, we train NTPs jointly with ComplEx [Trouillon et al., 2016] (Eq. 2.15 in Chapter 2). ComplEx and the NTP share the same subsymbolic representations, which is feasible as the RBF kernel in unify is also defined for complex vectors. While the NTP is responsible for multi-hop reasoning, the neural link prediction model learns to score ground atoms locally. At test time, only the NTP is used for predictions. Thus, the training loss for ComplEx can be seen as an auxiliary loss for the subsymbolic representations learned 5.2. Optimization 87 by the NTP. We term the resulting model NTPλ. Based on the loss in Section 5.2.1, the joint training loss is defined as LntpλKθ = LntpK + θ X − y log(complexθ (s, i, j)) ([s,i,j],y) ∈ T  − (1 − y) log(1 − complexθ (s, i, j)) where [s, i, j] is a training atom, y is its ground truth target, and complex(s, i, j) = psij as defined in Eq. 2.15. 5.2.3 Computational Optimizations NTPs as described above suffer from severe computational limitations since the neural network is representing all possible proofs up to some predefined depth. In contrast to symbolic backward chaining where a proof can be aborted as soon as unification fails, in differentiable proving we only get a unification failure for atoms whose arity does not match or when we detect cyclic rule application. We propose two optimizations to speed up NTPs. First, we make use of modern GPUs by batch processing many proofs in parallel (Section 5.2.3.1). Second, we exploit the sparseness of gradients caused by the min and max operations used in the unification and proof aggregation respectively to derive a heuristic for a truncated forward and backward pass that drastically reduces the number of proofs that have to be considered for calculating gradients (Section 5.2.3.2). 5.2.3.1 Batch Proving Let A ∈ RN ×k be a matrix of N subsymbolic representations that are to be unified with M other representations B ∈ RM ×k . We can adapt the unification module to calculate the unification success in a batched way using  v u P u k 2  u i=1 A1i  u .. u exp  . −u  t Pk 2 i=1 AN i     P k      1>  +  1N   M    2 i=1 B1i .. . Pk 2 i=1 BM i  >      >    ∈ RN ×M − 2AB      where 1N and 1M are vectors of N and M ones respectively, and the square root is taken element-wise. In practice, we partition the KB into rules that have the same structure and batch-unify goals with all rule heads per partition at the same time on a Graphics Processing Unit (GPU). Furthermore, substitution sets bind variables to vectors of symbol indices instead of single symbol indices, and min and max operations are taken per goal. 88 Chapter 5. End-to-End Differentiable Proving 5.2.3.2 K max Gradient Approximation NTPs allow us to calculate the gradient of proof success scores with respect to subsymbolic representations and rule parameters. While backpropagating through this large computation graph will give us the exact gradient, it is computationally infeasible for any reasonably-sized KB. Consider the parameterized rule θ1: (X, Y) :– θ2: (X, Z), θ3: (Z, Y) and let us assume the given KB contains 1 000 facts with binary predicates. While X and Y will be bound to the respective representations in the goal, Z we will be substituted with every possible second argument of the 1 000 facts in the KB when proving the first atom in the body. Moreover, for each of these 1 000 substitutions, we will again need to compare with all facts in the KB when proving the second atom in the body of the rule, resulting in 1 000 000 proof success scores. However, note that since we use the max operator for aggregating the success of different proofs, only subsymbolic representations in one out of 1 000 000 proofs will receive gradients. To overcome this computational limitation, we propose the following heuristic. We assume that when unifying the first atom with facts in the KB, it is unlikely for any unification successes below the top K successes to attain the maximum proof success when unifying the remaining atoms in the body of a rule with facts in the KB. That is, after the unification of the first atom, we only keep the top K substitutions and their success scores, and continue proving only with these. This means that all other partial proofs will not contribute to the forward pass at this stage, and consequently not receive any gradients on the backward pass of backpropagation. We term this the K max heuristic. Note that we cannot guarantee anymore that the gradient of the proof success is the exact gradient, but for a large enough K we get a close approximation to the true gradient. 5.3 Experiments Consistent with previous work, we carry out experiments on four benchmark KBs and compare ComplEx with the NTP and NTPλ in terms of area under the PrecisionRecall-curve (AUC-PR) on the Countries KB, and Mean Reciprocal Rank (MRR) and HITS@m [Bordes et al., 2013] on the other KBs described below. Training details, including hyperparameters and rule templates, can be found in Section 5.3.1. Countries The Countries KB is a dataset introduced by Bouchard et al. [2015] for testing reasoning capabilities of neural link prediction models. It consists of 244 countries, 5 regions (e.g. EUROPE), 23 subregions (e.g. WESTERN EUROPE, NORTHERN AMERICA ), and 1158 facts about the neighborhood of countries, and 5.3. Experiments Test Country Region 89 Test Country Region locatedIn locatedIn neighborOf locatedIn locatedIn neighborOf locatedIn locatedIn locatedIn Train Country locatedIn Subregion (a) locatedIn Train Country Subregion (b) Figure 5.3: Overview of different tasks in the Contries dataset as visualized by Nickel et al. [2016]. The left part (a) shows which atoms are removed for each task (dotted lines), and the right part (b) illustrates the rules that can be used to infer the location of test countries. For task S1, only facts corresponding to the blue dotted line are removed from the training set. For task S2, additionally facts corresponding to the green dashed line are removed. Finally, for task S3 also facts for the red dash-dotted line are removed. the location of countries and subregions. We follow Nickel et al. [2016] and split countries randomly into a training set of 204 countries (train), a development set of 20 countries (dev), and a test set of 20 countries (test), such that every dev and test country has at least one neighbor in the training set. Subsequently, three different task datasets are created. For all tasks, the goal is to predict locatedIn(c, r) for every test country c and all five regions r, but the access to training atoms in the KB varies (see Fig. 5.3). S1: All ground atoms locatedIn(c, r) where c is a test country and r is a region are removed from the KB. Since information about the subregion of test countries is still contained in the KB, this task can be solved by using the transitivity rule locatedIn(X, Y) :– locatedIn(X, Z), locatedIn(Z, Y). S2: In addition to S1, all ground atoms locatedIn(c, s) are removed where c is a test country and s is a subregion. The location of test countries needs to be inferred from the location of its neighboring countries: locatedIn(X, Y) :– neighborOf(X, Z), locatedIn(Z, Y). This task is more difficult than S1, as neighboring countries might not be in the same region, so the rule above will not always hold. S3: In addition to S2, all ground atoms locatedIn(c, r) where r is a region and c 90 Chapter 5. End-to-End Differentiable Proving is a training country that has a test or dev country as a neighbor are also removed. The location of test countries can for instance be inferred using the three-hop rule locatedIn(X, Y) :– neighborOf(X, Z), neighborOf(Z, W), locatedIn(W, Y). Kinship, Nations & UMLS We use the Nations, Alyawarra kinship (Kinship) and Unified Medical Language System (UMLS) KBs from Kok and Domingos [2007]. We left out the Animals dataset as it only contains unary predicates and can thus not be used for evaluating multi-hop reasoning. Nations contains 56 binary predicates, 111 unary predicates, 14 constants and 2565 true facts, Kinship contains 26 predicates, 104 constants and 10686 true facts, and UMLS contains 49 predicates, 135 constants and 6529 true facts. Since our baseline ComplEx cannot deal with unary predicates, we remove unary atoms from Nations. We split every KB into 80% training facts, 10% development facts and 10% test facts. For evaluation, we take a test fact and corrupt its first and second argument in all possible ways such that the corrupted fact is not in the original KB. Subsequently, we predict a ranking of every test fact and its corruptions to calculate MRR and HITS@m. 5.3.1 Training Details We use ADAM [Kingma and Ba, 2015] with an initial learning rate of 0.001 and a mini-batch size of 50 (10 known and 40 corrupted atoms) for optimization. We apply an `2 regularization of 0.01 to all model parameters, and clip gradient values at [−1.0, 1.0]. All subsymbolic representations and rule parameters are initialized using Xavier initialization [Glorot and Bengio, 2010]. We train all models for 100 epochs and repeat every experiment on the Countries corpus ten times. Statistical significance is tested using the independent t-test. All models are implemented in TensorFlow [Abadi et al., 2016]. We use a maximum proof depth of d = 2 and add the following rule templates where the number in front of the rule template indicates how often a parameterized rule of the given structure will be instantiated. Note that a rule template such as #1(X, Y) :– #2(X, Z), #2(Z, Y) specifies that the two predicate representations in the body are shared. Countries S1 3 #1(X, Y) :– #1(Y, X). 3 #1(X, Y) :– #2(X, Z), #2(Z, Y). Countries S2 3 #1(X, Y) :– #1(Y, X). 3 #1(X, Y) :– #2(X, Z), #2(Z, Y). 5.4. Results and Discussion 91 Table 5.1: AUC-PR results on Countries and MRR and HITS@m on Kinship, Nations, and UMLS. Corpus Metric Model ComplEx Countries S1 AUC-PR S2 AUC-PR S3 AUC-PR 99.37 ± 0.4 87.95 ± 2.8 48.44 ± 6.3 NTP Examples of induced rules and their confidence NTPλ 90.83 ± 15.4 100.00 ± 0.0 0.90 locatedIn(X,Y) :– locatedIn(X,Z), locatedIn(Z,Y). 87.40 ± 11.7 93.04 ± 0.4 0.63 locatedIn(X,Y) :– neighborOf(X,Z), locatedIn(Z,Y). 56.68 ± 17.6 77.26 ± 17.0 0.32 locatedIn(X,Y) :– neighborOf(X,Z), neighborOf(Z,W), locatedIn(W,Y). Kinship MRR HITS@1 HITS@3 HITS@10 0.81 0.70 0.89 0.98 0.60 0.48 0.70 0.78 0.80 0.76 0.82 0.89 0.98 term15(X,Y) :– term5(Y,X) 0.97 term18(X,Y) :– term18(Y,X) 0.86 term4(X,Y) :– term4(Y,X) 0.73 term12(X,Y) :– term10(X, Z), term12(Z, Y). Nations MRR HITS@1 HITS@3 HITS@10 0.75 0.62 0.84 0.99 0.75 0.62 0.86 0.99 0.74 0.59 0.89 0.99 0.68 blockpositionindex(X,Y) :– blockpositionindex(Y,X). 0.46 expeldiplomats(X,Y) :– negativebehavior(X,Y). 0.38 negativecomm(X,Y) :– commonbloc0(X,Y). 0.38 intergovorgs3(X,Y) :– intergovorgs(Y,X). UMLS MRR HITS@1 HITS@3 HITS@10 0.89 0.82 0.96 1.00 0.88 0.82 0.92 0.97 0.93 0.87 0.98 1.00 0.88 interacts with(X,Y) :– interacts with(X,Z), interacts with(Z,Y). 0.77 isa(X,Y) :– isa(X,Z), isa(Z,Y). 0.71 derivative of(X,Y) :– derivative of(X,Z), derivative of(Z,Y). 3 #1(X, Y) :– #2(X, Z), #3(Z, Y). Countries S3 3 #1(X, Y) :– #1(Y, X). 3 #1(X, Y) :– #2(X, Z), #2(Z, Y). 3 #1(X, Y) :– #2(X, Z), #3(Z, Y). 3 #1(X, Y) :– #2(X, Z), #3(Z, W), #4(W, Y). Kinship, Nations & UMLS 20 #1(X, Y) :– #2(X, Y). 20 #1(X, Y) :– #2(Y, X). 20 #1(X, Y) :– #2(X, Z), #3(Z, Y). 5.4 Results and Discussion Results for the different model variants on the benchmark KBs are shown in Table 6.1. Another method for inducing rules in a differentiable way for automated KB completion has been introduced recently by Yang et al. [2017] and our evaluation setup is equivalent to their Protocol II. However, our neural link prediction baseline, ComplEx, already achieves much higher HITS@10 results (1.00 vs. 0.70 on UMLS and 0.98 vs. 0.73 on Kinship). We thus focus on the comparison of NTPs with ComplEx. First, we note that vanilla NTPs alone do not work particularly well compared to ComplEx. They only outperform ComplEx on Countries S3 and Nations, but not on Kinship or UMLS. This demonstrates the difficulty of learning subsymbolic 92 Chapter 5. End-to-End Differentiable Proving representations in a differentiable prover from unification alone, and the need for auxiliary losses. The NTPλ with ComplEx as auxiliary loss outperforms the other models in the majority of tasks. The difference in AUC-PR between ComplEx and NTPλ is significant for all Countries tasks (p < 0.0001). A major advantage of NTPs is that we can inspect induced rules which provide us with an interpretable representation of what the model has learned. The right column in Table 6.1 shows examples of induced rules by NTPλ (note that predicates on Kinship are anonymized). For Countries, the NTP recovered those rules that are needed for solving the three different tasks. On UMLS, the NTP induced transitivity rules. Those relationships are particularly hard to encode by neural link prediction models like ComplEx, as they are optimized to locally predict the score of a fact. 5.5 Related Work Combining neural and symbolic approaches to relational learning and reasoning has a long tradition and let to various proposed architectures over the past decades (see d’Avila Garcez et al. [2012] for a review). Early proposals for neural-symbolic networks are limited to propositional rules (e.g., EBL-ANN [Shavlik and Towell, 1989], KBANN [Towell and Shavlik, 1994] and C-IL2 P [d’Avila Garcez and Zaverucha, 1999]). Other neural-symbolic approaches focus on first-order inference, but do not learn subsymbolic vector representations from training facts in a KB (e.g., SHRUTI [Shastri, 1992], Neural Prolog [Ding, 1995], CLIP++ [França et al., 2014], Lifted Relational Neural Networks [Sourek et al., 2015], and TensorLog [Cohen, 2016]). Logic Tensor Networks [Serafini and d’Avila Garcez, 2016] are in spirit similar to NTPs, but need to fully ground first-order logic rules. However, they support function terms, whereas NTPs currently only support function-free terms. Recent question-answering architectures such as [Peng et al., 2015, Weissenborn, 2016, Shen et al., 2016] translate query representations implicitly in a vector space without explicit rule representations and can thus not easily incorporate domain-specific knowledge. In addition, NTPs are related to random walk [Lao et al., 2011, 2012, Gardner et al., 2013, 2014] and path encoding models [Neelakantan et al., 2015, Das et al., 2017]. However, instead of aggregating paths from random walks or encoding paths to predict a target predicate, reasoning steps in NTPs are explicit and only unification uses subsymbolic representations. This allows us to induce interpretable rules, as well as to incorporate prior knowledge either in the form of rules or in the form of rule templates which define the structure of logical relationships that we expect to hold in a KB. Another line of work [Rocktäschel et al., 2014, Rocktäschel et al., 2015, Vendrov et al., 2016, Hu et al., 2016, Demeester et al., 5.6. Summary 93 2016] regularizes distributed representations via domain-specific rules, but these approaches do not learn such rules from data and only support a restricted subset of first-order logic. NTPs are constructed from Prolog’s backward chaining and are thus related to Unification Neural Networks [Komendantskaya, 2011, Hölldobler, 1990]. However, NTPs operate on vector representations of symbols instead of scalar values, which are more expressive. As NTPs can learn rules from data, they are related to ILP systems such as FOIL [Quinlan, 1990], Sherlock [Schoenmackers et al., 2010b] and meta-interpretive learning of higher-order dyadic Datalog (Metagol) [Muggleton et al., 2015]. While these ILP systems operate on symbols and search over the discrete space of logical rules, NTPs work with subsymbolic representations and induce rules using gradient descent. Recently, Yang et al. [2017] introduced a differentiable rule learning system based on TensorLog and a neural network controller similar to LSTMs [Hochreiter and Schmidhuber, 1997]. Their method is more scalable than the NTPs introduced here. However, on UMLS and Kinship our baseline already achieved stronger generalization by learning subsymbolic representations. Still, scaling NTPs to larger KBs for competing with more scalable relational learning methods is an open problem that we seek to address in future work. 5.6 Summary We proposed an end-to-end differentiable prover for automated KB completion that operates on subsymbolic representations. To this end, we used Prolog’s backward chaining algorithm as a recipe for recursively constructing neural networks that can be used to prove queries to a KB. Specifically, our contribution is the use of a differentiable unification operation between vector representations of symbols to construct such neural networks. This allowed us to compute the gradient of proof successes with respect to vector representations of symbols, and thus enabled us to train subsymbolic representations end-to-end from facts in a KB, and to induce function-free first-order logic rules using gradient descent. On benchmark KBs, our model outperformed ComplEx, a state-of-the-art neural link prediction model, on three out of four KBs while at the same time inducing interpretable rules. Chapter 6 Recognizing Textual Entailment with Recurrent Neural Networks “You can’t cram the meaning of a whole %&!$# sentence into a single $&!#* vector!” — Raymond J. Mooney The ability to determine the logical relationship between two natural language sentences is an integral part for machines that are supposed to understand and reason with language. In previous chapters, we have discussed ways of combining symbolic logical knowledge with subsymbolic representations trained via neural networks. As first steps towards models that reason with natural language, we have used textual surface form patterns as predicates for automated Knowledge Base (KB) completion. However, for automated KB completion, we assumed surface patterns to be atomic, which does not generalize to unseen patterns as there is no compositional representation. In this chapter, we are using Recurrent Neural Networks (RNNs) for learning compositional representations of natural language sentences. Specifically, we are tackling the task of Recognizing Textual Entailment (RTE), i.e., determining (i) whether two natural language sentences are contradicting each other, (ii) whether they are unrelated, or (iii) whether the first sentence (called the premise) entails the second sentence (called the hypothesis). For instance, the sentence “Two girls and a guy are involved in a pie eating contest” entails “Three people are stuffing their faces”, but contradicts “Three people are drinking beer on a boat”. This task is important since many Natural Language Processing (NLP) tasks, such as information extraction, relation extraction, text summarization or machine translation, rely on it explicitly or implicitly and could benefit from more accurate RTE systems [Dagan et al., 2005]. State-of-the-art systems for RTE so far relied heavily on engineered NLP pipelines, extensive manual creation of features, as well as various external resources 96 Chapter 6. Recognizing Textual Entailment with Recurrent Neural Networks and specialized subcomponents such as negation detection [e.g. Lai and Hockenmaier, 2014, Jiménez et al., 2014, Zhao et al., 2014, Beltagy et al., 2015]. Despite the success of neural networks for paraphrase detection [e.g. Socher et al., 2011, Hu et al., 2014, Yin and Schütze, 2015], end-to-end differentiable neural architectures so far failed to reach acceptable performance for RTE due to the lack of large high-quality datasets. An end-to-end differentiable solution to RTE is desirable since it avoids specific assumptions about the underlying language. In particular, there is no need for language features like part-of-speech tags or dependency parses. Furthermore, a generic sequence-to-sequence solution allows us to extend the concept of capturing entailment across any sequential data, not only natural language. Recently, Bowman et al. [2015] published the Stanford Natural Language Inference (SNLI) corpus accompanied by a long short-term memory (LSTM) baseline [Hochreiter and Schmidhuber, 1997] which achieves an accuracy of 77.6% for RTE on this dataset. It is the first instance of a generic neural model without hand-crafted features that got close to the accuracy of a simple lexicalized classifier with engineered features for RTE. This can be explained by the high quality and size of SNLI compared to the two orders of magnitude smaller and partly synthetic datasets used so far to evaluate RTE systems. Bowman et al.’s LSTM encodes the premise and hypothesis independently as dense fixed-length vectors whose concatenation is subsequently used in a Multi-layer Perceptron (MLP) for classification (Section 6.1.2). In contrast, we are proposing a neural network that is capable of fine-grained comparison of pairs of words and phrases by processing the hypothesis conditioned on the premise and using a neural attention mechanism. Our contributions are threefold: (i) we present a neural model based on two LSTMs that read two sentences in one go to determine entailment as opposed to mapping each sentence independently into a vector space (Section 6.2.1), (ii) we extend this model with a neural word-by-word attention mechanism to encourage more fine-grained comparison of pairs of words and phrases (Section 6.2.3), and (iii) we provide a detailed qualitative analysis of neural attention for RTE (Section 6.4.1). Our benchmark LSTM achieves an accuracy of 80.9% on SNLI, outperforming a simple lexicalized classifier tailored to RTE by 2.7 percentage points. An extension with word-by-word neural attention surpasses this strong benchmark LSTM result by 2.6 percentage points, achieving an accuracy of 83.5% for recognizing entailment on SNLI. 6.1. Background 97 ht RNN cell u3 tanh u2 b + u1 W matmul st−1 zt st concat xt Figure 6.1: Computation graph for the fully-connected RNN cell. 6.1 Background In this section, we describe RNNs and LSTMs for sequence modeling, before explaining how LSTMs are used in the independent sentence encoding model for RTE by Bowman et al. [2015]. 6.1.1 Recurrent Neural Networks An RNN is parameterized by a differentiable cell function fθ : Rk × Rs → Ro × Rs that maps an input vector xt ∈ Rk and previous state st−1 ∈ Rs to an output vector ht ∈ Ro and next state st ∈ Rs . For simplicity, we assume that the input size and output size are the same, i.e., xt , ht ∈ Rk . By applying the cell function at time-step t, we obtain an output vector ht and the next state st : ht , st = fθ (xt , st−1 ). (6.1) Given a sequence of input representations X = [x1 , . . . , xT ] and a start state s0 , the output of the RNN over the entire input sequence is then obtained by recursively applying the cell function: RNN(fθ , X, s0 ) = [fθ (x1 , s0 ), . . . , fθ (xT , sT −1 )]. (6.2) 98 Chapter 6. Recognizing Textual Entailment with Recurrent Neural Networks Usually, this output is separated into a list of output vectors H = [h1 , . . . , hT ] and states S = [s1 , . . . , sT ]. Note that RNN can be applied to input sequences of varying length, such as sequences of word representations. 6.1.1.1 Fully-connected Recurrent Neural Network The most basic RNN is parameterized by the following cell function " zt = xt st−1 # (6.3) ht = tanh(W zt + b) (6.4) st = ht (6.5) fθRNN (xt , st−1 ) = (ht , st ) (6.6) where W ∈ R2k×k is a trainable transition matrix, b ∈ Rk a trainable bias, and tanh the element-wise application of the hyperbolic function. We call this a fullyconnected RNN, as the cell function is modeled by a dense layer. We illustrate the computation graph for a single application of the fully-connected RNN cell in Fig. 6.1. Note that for recurrent application of this cell function to a sequence of inputs, all parameters (transition matrix and bias) are shared between all time steps. 6.1.1.2 Long Short-Term Memory RNNs with LSTM units [Hochreiter and Schmidhuber, 1997] have been successfully applied to a wide range of NLP tasks, such as machine translation [Sutskever et al., 2014], constituency parsing [Vinyals et al., 2015b], language modeling [Zaremba et al., 2014] and recently RTE [Bowman et al., 2015]. LSTMs encompass memory cells that can store information for a long period of time, as well as three types of gates that control the flow of information into and out of these cells: input gates (Eq. 6.8), forget gates (Eq. 6.9) and output gates (Eq. 6.10). Given an input vector 6.1. Background 99 ht LSTM cell ht−1 u11 u12 matmul Wo st−1 ot + ht ht u10 st ct ct sigm Output Gates bo tanh u7 ct−1 u8 matmul zt u9 ft + sigm Forget Gates + bf Wf concat u4 u5 matmul Wi bi u1 u2 matmul Wc u6 it + sigm Input Gates u3 tanh + bc xt Figure 6.2: Computation graph for the LSTM cell. xt at time step t, the previous output ht−1 and cell state ct−1 , an LSTM with hidden size k computes the next output ht and cell state ct as " zt = xt ht−1 # (6.7) it = σ(W i zt + bi ) (6.8) ft = σ(W f zt + bf ) (6.9) ot = σ(W o zt + bo ) (6.10) ct = ft ct−1 + it ht = ot " tanh(ct ) # st = ht ct fθLSTM (xt , st−1 ) = (ht , st ) tanh(W c zt + bc ) (6.11) (6.12) (6.13) (6.14) 100 Chapter 6. Recognizing Textual Entailment with Recurrent Neural Networks h1 h2 h3 h4 h5 h7 h8 h9 s1 s2 s3 s4 s5 s7 s8 s9 x1 x2 x3 x4 x5 x7 x8 x9 A wedding party taking pictures Someone got Premise married Hypothesis Figure 6.3: Independent encoding of the premise and hypothesis using the same LSTM. Note that both sentences are compressed as dense vectors as indicated by the red mapping. where W i , W f , W o , W c ∈ R2k×k are trained matrices and bi , bf , bo , bc ∈ Rk trained biases that parameterize the gates and transformations of the input. As in previous chapters, σ denotes the element-wise application of the sigmoid function and the element-wise multiplication of two vectors. The corresponding computation graph is illustrated in Fig. 6.2. 6.1.2 Independent Sentence Encoding LSTMs can readily be used for RTE by independently encoding the premise and hypothesis as dense vectors and taking their concatenation as input to an MLP classifier [Bowman et al., 2015]. Let XP = [xP1 , . . . , xPN ] be a sequence of words representing the premise H and let XH = [xH 1 , . . . , xM ] represent the hypothesis. Both, the premise and the hypothesis, can be encoded as fixed-dimensional vectors by taking the last output vector when applying the RNN function (Eq. 6.2) with an LSTM cell function fθLSTM . Subsequently, the prediction for the three RTE classes is obtained by an MLP (Eqs. 6.17 and 6.18) followed by a softmax (Eq. 6.19): , SP = RNN(fθLSTM , XP , s0 ) H H LSTM , XH , s ) H , S = RNN(f H P θ (6.15) (6.16) 0 " h = tanh W2 tanh W1 hPN hH M h∗ = tanh (W3 h + b3 ) # ! + b1 ! + b2 (6.17) (6.18) e yi = softmax(h∗ )i = P h∗i ∗ hj je (6.19) 6.2. Methods 101 where W1 , W2 ∈ R2k×2k , W3 ∈ R2k×3 , b1 , b2 ∈ R2k and b3 ∈ R3 . Furthermore, s0 is a trainable start state and hPN denotes the last element from the list of premise output vectors (similarly for hH M ). The independent sentence encoding (Eqs. 6.15 and 6.16) is visualized in Fig. 6.3. Given the predicted output distribution over the three RTE classes y (E NTAILMENT, N EUTRAL or C ONTRADICTION) and a target one-hot vector ŷ encoding the correct class, a cross-entropy loss is commonly used: L(y, ŷ) = − X ŷi log(yi ). (6.20) i Independent sentence encoding is a straightforward model for RTE. However, it is questionable how efficiently an entire sentence can be represented in a single fixed-dimensional vector. Hence, in the next section, we investigate various neural architectures that are tailored towards more fine-grained comparison of the premise and hypothesis and thus do not require to represent entire sentences as fixed-sized vectors in an embedding space. 6.2 Methods First, we propose to encode the hypothesis conditioned on a representation of the premise (Section 6.2.1). Subsequently, we introduce an extension of an LSTM for RTE with neural attention (Section 6.2.2) and word-by-word attention (Section 6.2.3). Finally, we show how such attentive models can easily be used for attending both ways: over the premise conditioned on the hypothesis and over the hypothesis conditioned on the premise (Section 6.2.4). All these models are trained using the loss in Eq. 6.20, and predict the probability of the RTE class using Eq. 6.19, but they differ in the way h∗ (Eq. 6.18) is calculated. 6.2.1 Conditional Encoding In contrast to learning sentence representations, we are interested in neural models that read both sentences to determine entailment, thereby comparing pairs of words and phrases. Figure 6.4 shows the high-level structure of this model. The premise (left) is read by an LSTM. A second LSTM with different parameters is reading a delimiter and the hypothesis (right), but its memory state is initialized with the last state of the previous LSTM (s5 in the example). That is, it processes the hypothesis 102 Chapter 6. Recognizing Textual Entailment with Recurrent Neural Networks h1 h2 h3 h4 h5 h6 h7 h8 h9 s1 s2 s3 s4 s5 s6 s7 s8 s9 x1 x2 x3 x4 x5 x6 x7 x8 x9 A wedding party taking pictures :: Someone got Premise married Hypothesis Figure 6.4: Conditional encoding with two LSTMs. The first LSTM encodes the premise, and then the second LSTM processes the hypothesis conditioned on the representation of the premise (s5 ). conditioned on the representation that the first LSTM built for the premise. Formally, we replace Eqs. 6.15 to 6.17 with , SP = RNN(fθLSTM , X P , s0 ) P H H LSTM , XH , sP ) H , S = RNN(f H P θH h = hH M N (6.21) (6.22) (6.23) where sPN denotes the last state of the LSTM that encoded the premise and hPM denotes the last element from the list of hypothesis output vectors. This model is illustrated for an RTE example in Fig. 6.4. Note that while the premise still has to be encoded in a fixed-dimensional vector, the LSTM that processes the hypothesis has to only keep track of whether incoming words contradict the premise, whether they are entailed by it, or whether they are unrelated. This is inspired by finite-state automata proposed for natural logic inference [Angeli and Manning, 2014]. 6.2.2 Attention Attentive neural networks have recently demonstrated success in a wide range of tasks ranging from handwriting synthesis [Graves, 2013], digit classification [Mnih et al., 2014], machine translation [Bahdanau et al., 2014], image captioning [Xu et al., 2015], speech recognition [Chorowski et al., 2015], sentence summarization [Rush et al., 2015], and code summarization [Allamanis et al., 2016] to geometric reasoning [Vinyals et al., 2015a]. The idea is to allow the model to attend over past output vectors. For LSTMs this mitigates the cell state bottleneck, i.e., the fact that 6.2. Methods Y 103 α1 α2 α3 α4 α5 h1 h2 h3 h4 h5 h6 h7 h8 h9 s1 s2 s3 s4 s5 s6 s7 s8 s9 x1 x2 x3 x4 x5 x6 x7 x8 x9 A wedding party taking pictures Premise :: Someone got married Hypothesis Figure 6.5: Attention model for RTE. Compared to Fig. 6.4, this model does not have to represent the entire premise in its cell state, but can instead output context representations (informally visualized by the red mapping) that are later queried by the attention mechanism (blue). Also note that now h1 to h5 are used. a standard LSTM has to store all relevant information for future time steps in its internal memory ct (see ct in Fig. 6.2 and compare Fig. 6.4 with Fig. 6.5). An LSTM with attention for RTE does not have to capture the entire content of the premise in its cell state. Instead, it is sufficient to output vectors while reading the premise (i.e. populating a differentiable memory of the premise) and accumulating a representation in the cell state that informs the second LSTM which of the output vectors of the premise it needs to attend over to determine the RTE class. Formally, let Y ∈ Rk×N be a matrix consisting of output vectors [h1 · · · hN ] that the first LSTM produced when reading the N words of the premise. Furthermore, let 1 ∈ RN be a vector of ones and hH M be the last output vector after the premise and hypothesis were processed by the two LSTMs. The attention mechanism will produce a vector α ∈ RN of attention weights and a weighted representation r of the premise via  T M = tanh W y Y + W h hH M 1 α = softmax(wT M ) r = Y αT M ∈ Rk×N α ∈ R1×N r ∈ Rk (6.24) (6.25) (6.26) where W y , W h ∈ Rk×k are trainable projection matrices, and w ∈ Rk is a trainable > parameter vector. Note that the outer product (W h hH M )1 is repeating the linearly 104 Chapter 6. Recognizing Textual Entailment with Recurrent Neural Networks transformed hH M as many times as there are words in the premise (i.e. N times). Hence, the intermediate attention representation mi (ith column vector in M ) of the ith word in the premise is obtained from a non-linear combination of the premise’s output vector hi (ith column vector in Y ) and the transformed hH M . The attention weight for the ith word in the premise is the result of a weighted combination (parameterized by w) of values in mi . The final sentence pair representation is obtained from a non-linear combination of the attention-weighted representation r of the premise and the last output vector hH M , thus replacing Eq. 6.23 by h = tanh(W p r + W x hH M) (6.27) where W p , W x ∈ Rk×k are trainable projection matrices. The attention model is illustrated in Fig. 6.5. Note that this model does not have to represent the entire premise in its cell state, but can instead output context representations that are later queried by the attention mechanism. This is informally illustrated by the red mapping of input phrases to output context representations.1 6.2.3 Word-by-word Attention For determining whether one sentence entails another it is desirable to check for entailment or contradiction of individual word and phrase pairs. To encourage such behavior we employ neural word-by-word attention similar to Bahdanau et al. [2014], Hermann et al. [2015] and Rush et al. [2015]. The difference is that we do not use attention to generate words, but to obtain a sentence pair encoding from fine-grained comparison via soft-alignment of word and phrase pairs in the premise and hypothesis. In our case, this amounts to attending over the first LSTM’s output vectors for the premise while the second LSTM processes the hypothesis one word at a time. Consequently, we obtain attention weight-vectors αt over premise output vectors for every word in the hypothesis. This can be modeled as follows:  T r Mt = tanh W y Y + W h hH t + W rt−1 1  αt = softmax wT Mt  rt = Y αTt + tanh W t rt−1 1 Mt ∈ Rk×N αt ∈ R1×N rt ∈ Rk (6.28) (6.29) (6.30) Note that the number of word representations that are used to output a context representation is not known. For illustration purposes we depicted the case that information of three words contribute to an output context representation. 6.2. Methods Y 105 h1 h2 h3 h4 h5 h6 h7 h8 h9 s1 s2 s3 s4 s5 s6 s7 s8 s9 x1 x2 x3 x4 x5 x6 x7 x8 x9 A wedding party taking pictures :: Someone got Premise married Hypothesis Figure 6.6: Word-by-word attention model for RTE. Compared to Fig. 6.5, querying the memory Y multiple times allows the model to store more fine-grained information in its output vectors when processing the premise. Also note that now also h6 to h8 are used. where W r , W t ∈ Rk×k are trainable projection matrices. Note that rt is dependent on the previous attention representation rt−1 to inform the model about what was attended over in previous steps (see Eqs. 6.28 and 6.30). As in the previous section, the final sentence pair representation is obtained from a non-linear combination of the last attention-weighted representation of the premise (here conditioned on the last word of the hypothesis) rN and the last output vector using h = tanh(W p rM + W x hH M ). (6.31) The word-by-word attention model is illustrated in Fig. 6.6. Compared to the attention model introduced earlier, querying the memory Y multiple times allows the model to store more fine-grained information in its output vectors when processing the premise. We informally illustrate this by fewer words contributing to the output context representations (red mapping). 6.2.4 Two-way Attention Inspired by bidirectional LSTMs that read a sequence and its reverse for improved encoding [Graves and Schmidhuber, 2005], we experiment with two-way attention for RTE. The idea is to use the same model (i.e. same structure and weights) to 106 Chapter 6. Recognizing Textual Entailment with Recurrent Neural Networks attend over the premise conditioned on the hypothesis, as well as to attend over the hypothesis conditioned on the premise, by simply swapping the two sequences. This produces two sentence pair representations that we concatenate for classification. 6.3 Experiments We conduct experiments on the Stanford Natural Language Inference corpus [SNLI, Bowman et al., 2015]. This corpus is two orders of magnitude larger than other existing RTE corpora such as Sentences Involving Compositional Knowledge [SICK, Marelli et al., 2014]. Furthermore, a large part of training examples in SICK were generated heuristically from other examples. In contrast, all sentence pairs in SNLI stem from human annotators. The size and quality of SNLI make it a suitable resource for training neural architectures such as the ones proposed in this chapter. 6.3.1 Training Details We use pretrained word2vec vectors [Mikolov et al., 2013] as word representations, which we keep fixed during training. Out-of-vocabulary words in the training set are randomly initialized by uniformly sampling values from [−0.05, 0.05] and are optimized during training.2 Out-of-vocabulary words encountered at inference time on the validation and test corpus are set to fixed random vectors. By not tuning representations of words for which we have word2vec vectors, we ensure that at test time their representation stays close to unseen similar words that are contained in word2vec. We use ADAM [Kingma and Ba, 2015] for optimization with a first momentum coefficient of 0.9 and a second momentum coefficient of 0.999.3 For every model we perform a grid search over combinations of the initial learning rate [1E-4, 3E-4, 1E-3], dropout4 [0.0, 0.1, 0.2] and `2 -regularization strength [0.0, 1E-4, 3E-4, 1E-3]. Subsequently, we take the best configuration based on performance on the validation set, and evaluate only that configuration on the test set. 6.4 Results and Discussion Results on the SNLI corpus are summarized in Table 6.1. The total number of model parameters, including tunable word representations, is denoted by |θ|W+M (without word representations |θ|M ). To ensure a comparable number of parameters to Bowman et al.’s model that encodes the premise and hypothesis independently 2 We found 12.1k words in SNLI for which we could not obtain word2vec embeddings, resulting in 3.65M tunable parameters. 3 Standard configuration recommended by Kingma and Ba. 4 As in Zaremba et al. [2014], we apply dropout only on the inputs and outputs of the network. 6.4. Results and Discussion 107 Table 6.1: Results on the SNLI corpus. Model k |θ|W+M |θ|M Train Dev Test Lexicalized classifier [Bowman et al., 2015] 99.7 LSTM [Bowman et al., 2015] 100 ≈ 10M 221k 84.4 - 78.2 77.6 Conditional encoding, shared Conditional encoding, shared Conditional encoding 100 3.8M 159 3.9M 116 3.9M 111k 83.7 252k 84.4 252k 83.5 81.9 80.9 83.0 81.4 82.1 80.9 Attention Attention, two-way 100 3.9M 100 3.9M 242k 85.4 242k 86.5 83.2 82.3 83.0 82.4 Word-by-word attention Word-by-word attention, two-way 100 3.9M 100 3.9M 252k 85.3 252k 86.6 83.7 83.5 83.6 83.2 using one LSTM, we also run experiments for conditional encoding where the parameters between both LSTMs are shared (“Conditional encoding, shared” with k = 100), as opposed to using two independent LSTMs. In addition, we compare our attentive models to two benchmark LSTMs whose hidden sizes were chosen so that they have at least as many parameters as the attentive models (k set to 159 and 116 respectively). Since we are not tuning word vectors for which we have word2vec embeddings, the total number of parameters |θ|W+M of our models is considerably smaller. We also compare our models against the benchmark lexicalized classifier used by Bowman et al., which uses features based on the BLEU score between the premise and hypothesis, length difference, word overlap, uni- and bigrams, part-of-speech tags, as well as cross uni- and bigrams. Conditional Encoding We found that processing the hypothesis conditioned on the premise instead of encoding both sentences independently gives an improvement of 3.3 percentage points in accuracy over Bowman et al.’s LSTM. We argue this is due to information being able to flow from the part of the model that processes the premise to the part that processes the hypothesis. Specifically, the model does not waste capacity on encoding the hypothesis (in fact it does not need to encode the hypothesis at all), but can read the hypothesis in a more focused way by checking words and phrases for contradiction or entailment based on the semantic representation of the premise (see Fig. 6.4). One interpretation is that the LSTM is approximating a finite-state automaton for RTE [c.f. Angeli and Manning, 2014]. Another difference to Bowman et al.’s model is that we are using word2vec instead of GloVe for word representations and, more importantly, do not fine-tune these word embeddings. The drop in accuracy from the train to the test set is less severe for our models, which suggest that fine-tuning word embeddings could be a cause of overfitting. 108 Chapter 6. Recognizing Textual Entailment with Recurrent Neural Networks Our LSTM outperforms a simple lexicalized classifier by 2.7 percentage points. To the best of our knowledge, at the time of publication this was the first instance of a neural end-to-end differentiable model outperforming a hand-crafted NLP pipeline on a textual entailment dataset. Attention By incorporating an attention mechanism we observe a 0.9 percentage point improvement over a single LSTM with a hidden size of 159 and a 1.4 percentage point increase over a benchmark model that uses two LSTMs for conditional encoding (one for the premise and one for the hypothesis conditioned on the representation of the premise). The attention model produces output vectors summarizing contextual information of the premise that is useful to attend over later when reading the hypothesis. Therefore, when reading the premise, the model does not have to build up a semantic representation of the whole premise, but instead a representation that helps attending over the premise’s output vectors when processing the hypothesis (see Fig. 6.5). In contrast, the output vectors of the premise are not used by the baseline conditional model. Thus, these models have to build up a representation of the entire premise and carry it over through the cell state to the part that processes the hypothesis—a bottleneck that can be overcome to some degree by using attention. Word-by-word Attention Enabling the model to attend over output vectors of the premise for every word in the hypothesis yields another 1.2 percentage point improvement compared to attending only once. We argue that this can be explained by the model being able to check for entailment or contradiction of individual word and phrase pairs, and we demonstrate this effect in the qualitative analysis below. Two-way Attention Allowing the model to also attend over the hypothesis based on the premise did not improve performance for RTE in our experiments. We suspect that this is due to entailment being an asymmetric relation. Hence, using the same LSTM to encode the hypothesis (in one direction) and the premise (in the other direction) might lead to noise in the training signal. This could be addressed by training different LSTMs at the cost of doubling the number of model parameters. 6.4.1 Qualitative Analysis It is instructive to analyze which output representations the model is attending over when deciding the class of an RTE example. Note that interpretations based on attention weights have to be taken with care since the model is not forced to solely rely on representations obtained from attention (see hH M in Eqs. 6.27 and 6.31). In the following, we visualize and discuss attention patterns of the presented attentive models. For each attentive model, we hand-picked examples from ten sentence pairs that were randomly drawn from the development set. 6.4. Results and Discussion (a) 109 (b) (c) (d) Figure 6.7: Attention visualizations. Attention Figure 6.7 shows to what extent the attentive model focuses on contextual representations of the premise after both LSTMs processed the premise and hypothesis, respectively. Note how the model pays attention to output vectors of words that are semantically coherent with the premise (“riding” and “rides”, “animal” and “camel”, 6.7a) or in contradiction, as caused by a single word (“blue” vs. “pink”, 6.7b) or multiple words (“swim” and “lake” vs. “frolicking” and “grass”, 6.7c). Interestingly, the model shows sensitivity to context by not attending over “yellow”, the color of the toy, but “pink”, the color of the coat. However, for more involved examples with longer premises, we found that attention is more uniformly distributed (6.7d). This suggests that conditioning attention only on the last output representation has limitations when multiple words need to be considered for deciding the RTE class. Word-by-word Attention Visualizations of word-by-word attention are depicted in Fig. 6.8. We found that word-by-word attention can easily detect if the hypothesis is simply a reordering of words in the premise (6.8a). Furthermore, it is able to resolve synonyms (“airplane” and “aircraft”, 6.8c) and capable of matching multi-word expressions to single words (“garbage can” to “trashcan”, 6.8b). It is also noteworthy that irrelevant parts of the premise, such as words capturing little meaning or whole uninformative relative clauses, are correctly neglected for determining entailment (“which also has a rope leading out of it”, 6.8b). Word-by-word attention seems to also work well when words in the premise and hypothesis are connected via deeper semantics (“snow” can be found “outside” 110 Chapter 6. Recognizing Textual Entailment with Recurrent Neural Networks and a “mother” is an “adult”, 6.8e and 6.8g). Furthermore, the model is able to resolve one-to-many relationships (“kids” to “boy” and “girl”, 6.8d). Attention can fail, for example when the two sentences and their words are entirely unrelated (6.8f). In such cases, the model seems to back off to attending over function words, and the sentence pair representation is likely dominated by the last output vector (hH M ) instead of the attention-weighted representation (see Eq. 6.31). 6.5 Related Work The methods in this chapter were published in Rocktäschel et al. [2016] and since then many new models have been proposed. They can be roughly classified into sentence encoding models which extend the independent encoding LSTM by Bowman et al. [2015] (Section 6.1.2), and models that are related to the conditional encoding architecture presented in Section 6.2.1. Results for these follow-up works are collected in a leaderboard at http://nlp.stanford.edu/projects/snli/. The current best result5 is held by a bidirectional LSTM with matching and aggregation layers introduced by Wang et al. [2017]. It achieves a test accuracy of 88.8% and outperforms the best independent encoding model by 4.2 percentage points. In fact, most independent encoding models [e.g. Vendrov et al., 2016, Mou et al., 2016, Bowman et al., 2016, Munkhdalai and Yu, 2016a] do not reach the performance of our conditional model with word-by-word attention. Exceptions are the recently introduced two-stage bidirectional LSTM model by Liu et al. [2016] and the Neural Semantic Encoder by Munkhdalai and Yu [2016b]. 6.5.1 Bidirectional Conditional Encoding As the models presented in this chapter make little assumptions about the input data, they can be applied in other domains too. Augenstein et al. [2016] introduced a conditional encoding model for determining the stance of a tweet (e.g. “A foetus has rights too!”) with respect to a target (e.g. “Legalization of Abortion”). They 5 Checked last on 26th of April, 2017. 6.6. Summary 111 further extended the conditional encoding model with bidirectional LSTMs [Graves and Schmidhuber, 2005], thus replacing Eqs. 6.21 to 6.23 with   − → − → − P → , S P = RNN fθLSTM , X , s 0 → −  P − → − → − → − → H H LSTM , X H , h P H , S = RNN fθ→ − N  H  ← − ← − ← − − H H LSTM , X H , ← H , S = RNN fθ← s 0 −  H ← − ← − ← − ← − P P LSTM , X P , h H H , S = RNN fθ← − M P − − → − → H ← − ← P h = tanh W hM + W hN H − → P (6.32) (6.33) (6.34) (6.35) (6.36) − → ← − → − → − ← − ← − where P and H denotes the forward, and P and H the reversed sequence, W , W ∈ − − Rk×k are trainable projection matrices, and → s 0 and ← s 0 are the trainable forward and reverse start state, respectively. This architecture is illustrated in Fig. 6.9. It achieved the second best result on the SemEval 2016 Task 6 Twitter Stance Detection corpus [Mohammad et al., 2016]. 6.5.2 Generating Entailing Sentences Instead of predicting the logical relationship between sentences, Kolesnyk et al. [2016] used entailment pairs of the SNLI corpus to learn to generate an entailed sentence given a premise. Their model is an encoder-decoder with attention, as used in neural machine translation [Bahdanau et al., 2014]. On a manually annotated test corpus of 100 generated sentences, their model generated correct entailed sentences in 82% of the cases. By recursively applying this encoder-decoder to the produced outputs, their model is able to generate natural language inference chains such as “A wedding party looks happy on the picture ⇒ A bride and groom smiles ⇒ Couple smiling ⇒ Some people smiling.” 6.6 Summary In this chapter, we demonstrated that LSTMs that read pairs of sequences to produce a final representation from which a simple classifier predicts entailment, outperform an LSTM baseline encoding the two sequences independently, as well as a classifier with hand-engineered features. Besides contributing this conditional model for RTE, our main contribution is to extend this model with attention over the premise which provides further improvements to the predictive abilities of the system. In a qualitative analysis, we showed that a word-by-word attention model is able to compare word and phrase pairs for deciding the logical relationship sentences. With 112 Chapter 6. Recognizing Textual Entailment with Recurrent Neural Networks an accuracy of 84.5%, it held the state-of-the-art on a large RTE corpus at the time of publication. 6.6. Summary (a) 113 (b) (c) (d) (e) (f) (g) Figure 6.8: Word-by-word attention visualizations. 114 Chapter 6. Recognizing Textual Entailment with Recurrent Neural Networks h← 1 h← 2 h← 3 h← 4 h← 5 h← 6 h← 7 h← 8 h← 9 h→ 1 h→ 2 h→ 3 h→ 4 h→ 5 h→ 6 h→ 7 h→ 8 h→ 9 s← 1 s← 2 s← 3 s← 4 s← 5 s← 6 s← 7 s← 8 s← 9 s→ 1 s→ 2 s→ 3 s→ 4 s→ 5 s→ 6 s→ 7 s→ 8 s→ 9 x1 x2 x3 x4 x5 x6 x7 x8 x9 Abortion A foetus has rights too ! Legalization of Target Tweet Figure 6.9: Bidirectional encoding of a tweet conditioned on bidirectional encoding of a ← target ([s→ 3 s1 ]). The stance is predicted using the last forward and reversed ← output representations ([h→ 9 h4 ]). Chapter 7 Conclusions 7.1 Summary of Contributions In this thesis, we have presented various combinations of representation learning models with logic. First, we proposed a way to calculate the gradient of propositional logic rules with respect to parameters of a neural link prediction model (Chapter 3). By stochastically grounding first-order logic rules, we were able to use these rules as regularizers in a matrix factorization neural link prediction model for automated Knowledge Base (KB) completion. This allowed us to embed background knowledge in form of logical rules in the vector space of predicate and entity pair representations. Using this method, we were able to train relation extractors for predicates with provided rules but little or no known training facts. In Chapter 4, we identified various shortcomings of stochastic grounding and proposed a model in which implication rules are only used to regularize predicate representations. This has the two advantages that the method becomes independent of the size of the domain of entity pairs, and that we can guarantee that the provided rules will hold for any test entity pair. By restricting the entity pair embedding space to be non-negative, we were able to impose implications as a partial order on the predicate representation space similar to Order Embeddings [Vendrov et al., 2016]. We showed empirically that by restricting the entity pair embedding space, the model generalizes better to predicting facts in the test set, which we attribute to a regularization effect. Furthermore, we showed that incorporating implication rules with this method scales well with the number of rules. After investigating two ways of regularizing symbol representations based on rules, in Chapter 5 we proposed a differentiable prover that performs KB inference with symbol representations in a more explicit way. To this end, we used Prolog’s backward chaining algorithm as a recipe for recursively constructing neural networks 116 Chapter 7. Conclusions that can be used to prove facts in a KB. Specifically, we proposed a differentiable unification operation between symbol representations. The constructed neural network allows us to compute the gradient of a proof success with respect to symbol representations, and thus train symbol representation end-to-end from the proof outcome. Furthermore, given templates for unknown rules of predefined structure, we can induce first-order logic rules using gradient descent. We proposed three optimizations for this model: (i) we implemented unification of multiple symbol representations as batch-operation which allows us to make use of modern Graphics Processing Units (GPUs) for efficient proving, (ii) we proposed an approximation of the gradient by only following K max proofs, and (iii) we used neural link prediction models as regularizers for the prover to learn better symbol representations more quickly. On three out of four benchmark knowledge bases, our method outperforms ComplEx, a state-of-the-art neural link prediction model, while at the same time inducing interpretable rules. Lastly, we developed neural models for Recognizing Textual Entailment (RTE), i.e., for determining the logical relationship between two natural language sentences (Chapter 6). We used one long short-term memory (LSTM) to encode the first sentence, and then conditioned on that representation encoded the second sentence using a second LSTM for deciding the label of the sentence pair. Furthermore, we extended this model with a neural word-by-word attention mechanism that enables more fine-grained comparison of word and phrase pairs. On a large RTE corpus, these models outperform a classifier with hand-engineered features and a strong LSTM baseline. In addition, we qualitatively analyzed the attention the model pays to words in the first sentence, and we were able to confirm the presence of fine-grained reasoning patterns. 7.2 Limitations and Future Work The integration of neural representations with symbolic logic and reasoning remains an exciting and open research area, and we except to see much more systems improving representation learning models by taking inspiration from formal logic in the future. While we demonstrated the benefit of regularizing symbol representations by logical rules for automated KB completion, we were only able to do this efficiently for simple implication rules. For future work, it would be interesting to use more general first-order logic rules as regularizers on predicate representations in a lifted way, for instance, by a more informed grounding of first-order rules. However, it is likely that the approach of regularizing predicate representations using rules has theoretical limitations that need to be investigated further. Thus, we believe 7.2. Limitations and Future Work 117 an interesting alternative direction is synthesizing symbolic reasoning and neural representations in more explicit ways. The end-to-end differentiable Neural Theorem Prover (NTP) introduced in this thesis is only a first proposal towards a tight integration of symbolic reasoning systems with trainable rules and symbol representations. The major obstacle that we encountered has to do with the computational complexity of making the proof success differentiable so that we can calculate the gradient with respect to symbol representations. While it is possible to approximate the gradient by only maintaining the K max proofs for a given a query, at some point a unification of a query with all facts in a KB is necessary. As real-world KBs can contain millions of facts, this grounding becomes impossible to do efficiently without applying further heuristics even when using modern GPUs. A possible future direction could be the use of hierarchical attention [Andrychowicz et al., 2016], or recent methods for reinforcement learning such as Monte Carlo tree search [Coulom, 2006, Kocsis and Szepesvári, 2006] as used, for instance, for learning to play Go [Silver et al., 2016] or chemical synthesis planning [Segler et al., 2017]. Specifically, the idea would be to train a model that learns to select promising rules instead of trying all rules for proving a goal. Orthogonal to that, more flexible individual components of end-to-end differentiable provers are conceivable. For instance, unification, rule selection, and rule application could be modeled as parameterized functions, and thus could be used to learn a more optimal behavior from data in a KB than the behavior that we specified by closely following the backward chaining algorithm. Furthermore, while the NTP is constructed from Prolog’s backward chaining, we currently only support Datalog logic programs, i.e., function-free first-order logic. An open question is how we can enable support for function terms in end-to-end differentiable provers. Another open research direction is the extension of automated provers to handle natural language sentences, questions, and perform multi-hop reasoning with natural language sentences. A starting point could be the combination of models that we proposed for determining the logical relationship between two natural language sentences, and the differentiable prover. As the end-to-end differentiable prover introduced in this thesis can be used to calculate the gradient of proof success with respect to symbol representations, these symbol representations can itself be composed by an RNN encoder that is trained jointly with the prover. The vision is a prover that directly operates on natural language statements and explanations, avoiding the need for semantic parsing [Zettlemoyer and Collins, 2005], i.e., parsing text into logical form. As NTPs decompose inference in a more explicit way, it would 118 Chapter 7. Conclusions be worthwhile to investigate whether we can obtain interpretable natural language proofs. Furthermore, it would be interesting to scale the methods presented here to larger units of text such as entire documents. Again, this needs model extensions such as hierarchical attention to ensure computational efficiency. In addition, it would be worthwhile exploring how other, more structured forms of attention [e.g. Graves et al., 2014, Sukhbaatar et al., 2015], or other forms of differentiable memory [e.g. Grefenstette et al., 2015, Joulin and Mikolov, 2015] could help improve performance of neural networks for RTE and differentiable proving. Lastly, we are interested in applying NTPs to automated proving of mathematical theorems, either in logical or natural language form, similar to the recent work by Kaliszyk et al. [2017] and Loos et al. [2017]. Appendix A Annotated Rules Manually filtered rules and their score for experiments in Chapter 3. 0.97 organization/parent/child(X,Y) :– #2-nn<-unit->prep->of->pobj-#1(X,Y). 0.97 organization/parent/child(X,Y) :– #2->appos->subsidiary->prep->of->pobj-#1(X,Y). 0.97 organization/parent/child(X,Y) :– #2->rcmod->own->prep->by->pobj-#1(X,Y). 0.97 location/location/containedby(X,Y) :– #2-nn<-city->prep->of->pobj-#1(X,Y). 0.97 organization/parent/child(X,Y) :– #2->appos->subsidiary->nn-#1(X,Y). 0.97 people/person/nationality(X,Y) :– #2-poss<-minister->appos-#1(X,Y). 0.97 organization/parent/child(X,Y) :– #2->appos->unit->prep->of->pobj-#1(X,Y). 0.96 organization/parent/child(X,Y) :– #2->appos->division->prep->of->pobj-#1(X,Y). 0.96 business/person/company(X,Y) :– #2-poss<-executive->appos-#1(X,Y). 0.96 business/company/founders(X,Y) :– #2->appos->co-founder->prep->of->pobj-#1(X,Y). 0.96 book/author/works written(X,Y) :– #2-dobj<-review->prep->by->pobj-#1(X,Y). 0.95 business/company/founders(X,Y) :– #2->appos->founder->prep->of->pobj-#1(X,Y). 0.95 location/location/containedby(X,Y) :– 120 Appendix A. Annotated Rules #2-nn<-town->prep->of->pobj-#1(X,Y). 0.95 location/neighborhood/neighborhood of(X,Y) :– #2-nn<-neighborhood->prep->of->pobj-#1(X,Y). 0.95 film/film/directed by(X,Y) :– #2->appos->director->dep-#1(X,Y). 0.95 location/location/containedby(X,Y) :– #2-poss<-region->nn-#1(X,Y). 0.94 film/film/produced by(X,Y) :– #2->appos->producer->dep-#1(X,Y). 0.94 film/film/directed by(X,Y) :– #2-poss<-film->dep-#1(X,Y). 0.94 location/location/containedby(X,Y) :– #2-nsubj<-professor->prep->at->pobj-#1(X,Y). 0.94 film/film/directed by(X,Y) :– #2-poss<-movie->dep-#1(X,Y). 0.93 people/person/nationality(X,Y) :– #2-poss<-leader->appos-#1(X,Y). 0.93 film/film/directed by(X,Y) :– #2-nn<-film->dep-#1(X,Y). 0.93 location/location/containedby(X,Y) :– #2-nn<-suburb->prep->of->pobj-#1(X,Y). 0.93 people/person/parents(X,Y) :– #1->appos->daughter->prep->of->pobj-#2(X,Y). 0.93 business/person/company(X,Y) :– #2-poss<-chairman->appos-#1(X,Y). 0.93 location/location/containedby(X,Y) :– #2-nn<-side->prep->of->pobj-#1(X,Y). 0.93 people/deceased person/place of death(X,Y) :– #1-nsubj<-die->prep->in->pobj-#2(X,Y). 0.93 location/neighborhood/neighborhood of(X,Y) :– #2-poss<-neighborhood->nn-#1(X,Y). 0.91 location/location/containedby(X,Y) :– #2-nsubj<-professor->appos-#1(X,Y). 0.91 people/deceased person/place of death(X,Y) :– #1-nsubj<-die->prep->at->pobj->hospital->prep->in->pobj-#2(X,Y). 0.91 book/author/works written(X,Y) :– #1-poss<-book->dep-#2(X,Y). 121 0.90 business/person/company(X,Y) :– #2-nsubj<-name->dobj-#1(X,Y). 0.90 people/person/place of birth(X,Y) :– #1-nsubjpass<-bear->prep->in->pobj-#2(X,Y). 0.90 people/person/nationality(X,Y) :– #1->appos->minister->poss-#2(X,Y). 0.88 location/location/containedby(X,Y) :– #1->appos->capital->prep->of->pobj-#2(X,Y). 0.87 location/location/containedby(X,Y) :– #1-nsubj<-city->prep->in->pobj-#2(X,Y). Appendix B Annotated WordNet Rules Manually filtered rules derived from WordNet for experiments in Chapter 4. #1->appos->organization->amod-#2(X,Y) :– #1->appos->party->amod-#2(X,Y). #1-nsubj<-push->dobj-#2(X,Y) :– #1-nsubj<-press->dobj-#2(X,Y). #1->appos->artist->amod-#2(X,Y) :– #1->appos->painter->amod-#2(X,Y). #1->appos->artist->nn-#2(X,Y) :– #1->appos->painter->nn-#2(X,Y). #1->appos->writer->amod-#2(X,Y) :– #1->appos->journalist->amod-#2(X,Y). #1->appos->writer->amod-#2(X,Y) :– #1->appos->poet->amod-#2(X,Y). #1-poss<-parent->appos-#2(X,Y) :– #1-poss<-father->appos-#2(X,Y). #1->appos->lawyer->nn-#2(X,Y) :– #1->appos->prosecutor->nn-#2(X,Y). #1->appos->expert->nn-#2(X,Y) :– #1->appos->specialist->nn-#2(X,Y). #1->appos->newspaper->amod-#2(X,Y) :– #1->appos->daily->amod-#2(X,Y). #1->appos->leader->nn-#2(X,Y) :– #1->appos->boss->nn-#2(X,Y). #1->appos->firm->nn-#2(X,Y) :– #1->appos->publisher->nn-#2(X,Y). #1->appos->journalist->nn-#2(X,Y) :– 124 Appendix B. Annotated WordNet Rules #1->appos->correspondent->nn-#2(X,Y). #1->appos->company->amod-#2(X,Y) :– #1->appos->subsidiary->amod-#2(X,Y). #1-nsubj<-purchase->dobj-#2(X,Y) :– #1-nsubj<-buy->dobj-#2(X,Y). #1->appos->leader->nn-#2(X,Y) :– #1->appos->chief->nn-#2(X,Y). #1->appos->player->poss-#2(X,Y) :– #1->appos->scorer->poss-#2(X,Y). #1->appos->organization->nn-#2(X,Y) :– #1->appos->institution->nn-#2(X,Y). #1->appos->center->amod-#2(X,Y) :– #1->appos->capital->amod-#2(X,Y). #1->appos->center->poss-#2(X,Y) :– #1->appos->capital->poss-#2(X,Y). #1->appos->representative->poss-#2(X,Y) :– #1->appos->envoy->poss-#2(X,Y). #1->appos->expert->amod-#2(X,Y) :– #1->appos->specialist->amod-#2(X,Y). #1->appos->center->nn-#2(X,Y) :– #1->appos->capital->nn-#2(X,Y). #1->appos->writer->amod-#2(X,Y) :– #1->appos->novelist->amod-#2(X,Y). #1->appos->diplomat->amod-#2(X,Y) :– #1->appos->ambassador->amod-#2(X,Y). #1->appos->expert->amod-#2(X,Y) :– #1->appos->analyst->amod-#2(X,Y). #1->appos->scholar->nn-#2(X,Y) :– #1->appos->historian->nn-#2(X,Y). #1->appos->maker->amod-#2(X,Y) :– #1->appos->producer->amod-#2(X,Y). #1->appos->maker->amod-#2(X,Y) :– #1->appos->manufacturer->amod-#2(X,Y). #1->appos->official->amod-#2(X,Y) :– #1->appos->diplomat->amod-#2(X,Y). #1->appos->trainer->poss-#2(X,Y) :– #1->appos->coach->poss-#2(X,Y). 125 #1->appos->member->amod-#2(X,Y) :– #1->appos->commissioner->amod-#2(X,Y). #1->appos->institution->nn-#2(X,Y) :– #1->appos->company->nn-#2(X,Y). #1->appos->representative->amod-#2(X,Y) :– #1->appos->envoy->amod-#2(X,Y). #1->appos->scientist->nn-#2(X,Y) :– #1->appos->physicist->nn-#2(X,Y). #1->appos->representative->nn-#2(X,Y) :– #1->appos->envoy->nn-#2(X,Y). Bibliography Martı́n Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Gregory S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian J. Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Józefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Gordon Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul A. Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda B. Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. Tensorflow: Large-scale machine learning on heterogeneous distributed systems. CoRR, abs/1603.04467, 2016. URL http://arxiv.org/abs/1603.04467. Yaser S. Abu-Mostafa. Learning from hints in neural networks. J. Complexity, 6(2): 192–198, 1990. doi: 10.1016/0885-064X(90)90006-Y. URL http://dx.doi. org/10.1016/0885-064X(90)90006-Y. Rami Al-Rfou, Guillaume Alain, Amjad Almahairi, Christof Angermüller, Dzmitry Bahdanau, Nicolas Ballas, Frédéric Bastien, Justin Bayer, Anatoly Belikov, Alexander Belopolsky, Yoshua Bengio, Arnaud Bergeron, James Bergstra, Valentin Bisson, Josh Bleecher Snyder, Nicolas Bouchard, Nicolas BoulangerLewandowski, Xavier Bouthillier, Alexandre de Brébisson, Olivier Breuleux, Pierre Luc Carrier, Kyunghyun Cho, Jan Chorowski, Paul Christiano, Tim Cooijmans, Marc-Alexandre Côté, Myriam Côté, Aaron C. Courville, Yann N. Dauphin, Olivier Delalleau, Julien Demouth, Guillaume Desjardins, Sander Dieleman, Laurent Dinh, Melanie Ducoffe, Vincent Dumoulin, Samira Ebrahimi Kahou, Dumitru Erhan, Ziye Fan, Orhan Firat, Mathieu Germain, Xavier Glorot, Ian J. Goodfellow, Matthew Graham, Çaglar Gülçehre, Philippe Hamel, Iban Harlouchet, Jean-Philippe Heng, Balázs Hidasi, Sina Honari, Arjun Jain, Sébastien Jean, Kai Jia, Mikhail Korobov, Vivek Kulkarni, Alex Lamb, Pascal Lamblin, Eric Larsen, César Laurent, Sean Lee, Simon Lefrançois, Simon Lemieux, Nicholas Léonard, 128 BIBLIOGRAPHY Zhouhan Lin, Jesse A. Livezey, Cory Lorenz, Jeremiah Lowin, Qianli Ma, PierreAntoine Manzagol, Olivier Mastropietro, Robert McGibbon, Roland Memisevic, Bart van Merriënboer, Vincent Michalski, Mehdi Mirza, Alberto Orlandi, Christopher Joseph Pal, Razvan Pascanu, Mohammad Pezeshki, Colin Raffel, Daniel Renshaw, Matthew Rocklin, Adriana Romero, Markus Roth, Peter Sadowski, John Salvatier, François Savard, Jan Schlüter, John Schulman, Gabriel Schwartz, Iulian Vlad Serban, Dmitriy Serdyuk, Samira Shabanian, Étienne Simon, Sigurd Spieckermann, S. Ramana Subramanyam, Jakub Sygnowski, Jérémie Tanguay, Gijs van Tulder, Joseph P. Turian, Sebastian Urban, Pascal Vincent, Francesco Visin, Harm de Vries, David Warde-Farley, Dustin J. Webb, Matthew Willson, Kelvin Xu, Lijun Xue, Li Yao, Saizheng Zhang, and Ying Zhang. Theano: A python framework for fast computation of mathematical expressions. CoRR, abs/1605.02688, 2016. URL http://arxiv.org/abs/1605.02688. Miltiadis Allamanis, Hao Peng, and Charles A. Sutton. A convolutional attention network for extreme summarization of source code. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, pages 2091–2100, 2016. URL http://jmlr.org/ proceedings/papers/v48/allamanis16.html. Miltiadis Allamanis, Pankajan Chanthirasegaran, Pushmeet Kohli, and Charles A. Sutton. Learning continuous semantic representations of symbolic expressions. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, pages 80–88, 2017. URL http://proceedings.mlr.press/v70/allamanis17a.html. Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Learning to compose neural networks for question answering. In NAACL HLT 2016, The 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, San Diego California, USA, June 12-17, 2016, pages 1545–1554, 2016. URL http://aclweb.org/ anthology/N/N16/N16-1181.pdf. Marcin Andrychowicz, Misha Denil, Sergio Gomez Colmenarejo, Matthew W. Hoffman, David Pfau, Tom Schaul, and Nando de Freitas. Learning to learn by gradient descent by gradient descent. In Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pages 3981–3989, BIBLIOGRAPHY 129 2016. URL http://papers.nips.cc/paper/6461-learning-tolearn-by-gradient-descent-by-gradient-descent. Gabor Angeli and Christopher D. Manning. Naturalli: Natural logic inference for common sense reasoning. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP 2014, October 25-29, 2014, Doha, Qatar, A meeting of SIGDAT, a Special Interest Group of the ACL, pages 534–545, 2014. URL http://aclweb.org/anthology/D/D14/D141059.pdf. Isabelle Augenstein, Tim Rocktäschel, Andreas Vlachos, and Kalina Bontcheva. Stance detection with bidirectional conditional encoding. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, EMNLP 2016, Austin, Texas, USA, November 1-4, 2016, pages 876–885, 2016. URL http://aclweb.org/anthology/D/D16/D16-1084.pdf. Franz Baader, Bernhard Ganter, Baris Sertkaya, and Ulrike Sattler. Completing description logic knowledge bases using formal concept analysis. In IJCAI 2007, Proceedings of the 20th International Joint Conference on Artificial Intelligence, Hyderabad, India, January 6-12, 2007, pages 230–235, 2007. URL http: //ijcai.org/Proceedings/07/Papers/035.pdf. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. CoRR, abs/1409.0473, 2014. URL http://arxiv.org/abs/1409.0473. Islam Beltagy, Cuong Chau, Gemma Boleda, Dan Garrette, Katrin Erk, and Raymond J. Mooney. Montague Meets Markov: Deep Semantics with Probabilistic Logical Form. In Proceedings of the Second Joint Conference on Lexical and Computational Semantics, *SEM 2013, June 13-14, 2013, Atlanta, Georgia, USA., pages 11–21, 2013. URL http://aclweb.org/anthology/S/ S13/S13-1002.pdf. Islam Beltagy, Katrin Erk, and Raymond J. Mooney. Probabilistic soft logic for semantic textual similarity. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, ACL 2014, June 22-27, 2014, Baltimore, MD, USA, Volume 1: Long Papers, pages 1210–1219, 2014. URL http://aclweb.org/anthology/P/P14/P14-1114.pdf. 130 BIBLIOGRAPHY Islam Beltagy, Stephen Roller, Pengxiang Cheng, Katrin Erk, and Raymond J. Mooney. Representing meaning with a combination of logical form and vectors. CoRR, abs/1505.06816, 2015. URL http://arxiv.org/abs/1505. 06816. Islam Beltagy, Stephen Roller, Pengxiang Cheng, Katrin Erk, and Raymond J Mooney. Representing meaning with a combination of logical and distributional models. Computational Linguistics, 2017. Kurt D. Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. Freebase: a collaboratively created graph database for structuring human knowledge. In Proceedings of the ACM SIGMOD International Conference on Management of Data, SIGMOD 2008, Vancouver, BC, Canada, June 1012, 2008, pages 1247–1250, 2008. doi: 10.1145/1376616.1376746. URL http://doi.acm.org/10.1145/1376616.1376746. Antoine Bordes, Jason Weston, Ronan Collobert, and Yoshua Bengio. Learning structured embeddings of knowledge bases. In Proceedings of the Twenty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2011, San Francisco, California, USA, August 7-11, 2011, 2011. URL http://www.aaai.org/ocs/index. php/AAAI/AAAI11/paper/view/3659. Antoine Bordes, Nicolas Usunier, Alberto Garcı́a-Durán, Jason Weston, and Oksana Yakhnenko. Translating embeddings for modeling multi-relational data. In Advances in Neural Information Processing Systems 26: 27th Annual Conference on Neural Information Processing Systems 2013. Proceedings of a meeting held December 5-8, 2013, Lake Tahoe, Nevada, United States., pages 2787–2795, 2013. URL http://papers.nips.cc/paper/5071-translatingembeddings-for-modeling-multi-relational-data. Johan Bos. Wide-coverage semantic analysis with boxer. In Semantics in Text Processing. STEP 2008 Conference Proceedings, volume 1 of Research in Computational Semantics, pages 277–286. College Publications, 2008. Johan Bos and Katja Markert. Recognising textual entailment with logical inference. In HLT/EMNLP 2005, Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing, Proceedings of the Conference, 6-8 October 2005, Vancouver, British Columbia, Canada, pages 628–635, 2005. URL http://aclweb.org/anthology/H/H05/H051079.pdf. BIBLIOGRAPHY 131 Matko Bosnjak, Tim Rocktäschel, Jason Naradowsky, and Sebastian Riedel. Programming with a differentiable forth interpreter. In International Conference on Machine Learning (ICML), 2017. URL http://arxiv.org/abs/1605. 06640. Guillaume Bouchard, Sameer Singh, and Theo Trouillon. On approximate reasoning capabilities of low-rank vector spaces. In Proceedings of the 2015 AAAI Spring Symposium on Knowledge Representation and Reasoning (KRR): Integrating Symbolic and Neural Approaches, 2015. Samuel R. Bowman. Can recursive neural tensor networks learn logical reasoning? volume abs/1312.6192, 2013. URL http://arxiv.org/abs/1312.6192. Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, EMNLP 2015, Lisbon, Portugal, September 17-21, 2015, pages 632–642, 2015. URL http://aclweb.org/anthology/D/D15/D15-1075.pdf. Samuel R. Bowman, Jon Gauthier, Abhinav Rastogi, Raghav Gupta, Christopher D. Manning, and Christopher Potts. A fast unified model for parsing and sentence understanding. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers, 2016. URL http://aclweb.org/anthology/P/ P16/P16-1139.pdf. Rodrigo De Salvo Braz. Lifted First-order Probabilistic Inference. PhD thesis, Champaign, IL, USA, 2007. AAI3290183. David S Broomhead and David Lowe. Radial basis functions, multi-variable functional interpolation and adaptive networks. Technical report, DTIC Document, 1988. Richard H. Byrd, Peihuang Lu, Jorge Nocedal, and Ciyou Zhu. A limited memory algorithm for bound constrained optimization. SIAM J. Scientific Computing, 16 (5):1190–1208, 1995. doi: 10.1137/0916069. URL http://dx.doi.org/ 10.1137/0916069. Andrew Carlson, Justin Betteridge, Richard C. Wang, Estevam R. Hruschka Jr., and Tom M. Mitchell. Coupled semi-supervised learning for information extraction. In 132 BIBLIOGRAPHY Proceedings of the Third International Conference on Web Search and Web Data Mining, WSDM 2010, New York, NY, USA, February 4-6, 2010, pages 101–110, 2010. doi: 10.1145/1718487.1718501. URL http://doi.acm.org/10. 1145/1718487.1718501. Kai-Wei Chang, Wen-tau Yih, Bishan Yang, and Christopher Meek. Typed tensor decomposition of knowledge bases for relation extraction. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP 2014, October 25-29, 2014, Doha, Qatar, A meeting of SIGDAT, a Special Interest Group of the ACL, pages 1568–1579, 2014. URL http://aclweb.org/ anthology/D/D14/D14-1165.pdf. Ming-Wei Chang, Lev-Arie Ratinov, and Dan Roth. Guiding semi-supervision with constraint-driven learning. In ACL 2007, Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics, June 23-30, 2007, Prague, Czech Republic, 2007. URL http://aclweb.org/anthologynew/P/P07/P07-1036.pdf. Jan Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk, Kyunghyun Cho, and Yoshua Bengio. Attention-based models for speech recognition. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 577–585, 2015. URL http://papers.nips.cc/paper/5847attention-based-models-for-speech-recognition. Stephen Clark and Stephen Pulman. Combining symbolic and distributional models of meaning. In Quantum Interaction, Papers from the 2007 AAAI Spring Symposium, Technical Report SS-07-08, Stanford, California, USA, March 26-28, 2007, pages 52–55, 2007. URL http://www.aaai.org/Library/Symposia/ Spring/2007/ss07-08-008.php. Bob Coecke, Mehrnoosh Sadrzadeh, and Stephen Clark. Mathematical foundations for a compositional distributional model of meaning. CoRR, abs/1003.4394, 2010. URL http://arxiv.org/abs/1003.4394. William W. Cohen. Tensorlog: A differentiable deductive database. CoRR, abs/1605.06523, 2016. URL http://arxiv.org/abs/1605.06523. Michael Collins, Sanjoy Dasgupta, and Robert E. Schapire. A generalization of principal components analysis to the exponential family. In BIBLIOGRAPHY 133 Advances in Neural Information Processing Systems 14 [Neural Information Processing Systems: Natural and Synthetic, NIPS 2001, December 3-8, 2001, Vancouver, British Columbia, Canada], pages 617–624, 2001. URL http://papers.nips.cc/paper/2078-a-generalizationof-principal-components-analysis-to-the-exponentialfamily. Ronan Collobert, Koray Kavukcuoglu, and Clément Farabet. Torch7: A matlablike environment for machine learning. In BigLearn, NIPS Workshop, number EPFL-CONF-192376, 2011. Alain Colmerauer. An introduction to prolog III. Commun. ACM, 33(7):69–90, 1990. doi: 10.1145/79204.79210. URL http://doi.acm.org/10.1145/ 79204.79210. Rémi Coulom. Efficient selectivity and backup operators in monte-carlo tree search. In Computers and Games, 5th International Conference, CG 2006, Turin, Italy, May 29-31, 2006. Revised Papers, pages 72–83, 2006. doi: 10.1007/978-3-54075538-8 7. URL http://dx.doi.org/10.1007/978-3-540-755388_7. Council of the European Union. Position of the council on general data protection regulation. http://www.europarl.europa.eu/sed/doc/news/ document/CONS_CONS(2016)05418(REV1)_EN.docx, 2016. 8 April 2016. Ido Dagan, Oren Glickman, and Bernardo Magnini. The PASCAL recognising textual entailment challenge. In Machine Learning Challenges, Evaluating Predictive Uncertainty, Visual Object Classification and Recognizing Textual Entailment, First PASCAL Machine Learning Challenges Workshop, MLCW 2005, Southampton, UK, April 11-13, 2005, Revised Selected Papers, pages 177–190, 2005. doi: 10.1007/11736790 9. URL http://dx.doi.org/10.1007/11736790_ 9. Rajarshi Das, Arvind Neelakantan, David Belanger, and Andrew McCallum. Chains of reasoning over entities, relations, and text using recurrent neural networks. In Conference of the European Chapter of the Association for Computational Linguistics (EACL), 2017. URL http://arxiv.org/abs/1607.01426. 134 BIBLIOGRAPHY Artur S. d’Avila Garcez and Gerson Zaverucha. The connectionist inductive learning and logic programming system. Appl. Intell., 11(1):59–77, 1999. doi: 10.1023/A:1008328630915. URL http://dx.doi.org/10.1023/A: 1008328630915. Artur S. d’Avila Garcez, Krysia Broda, and Dov M. Gabbay. Neural-symbolic learning systems: foundations and applications. Springer Science & Business Media, 2012. Oier Lopez de Lacalle and Mirella Lapata. Unsupervised relation extraction with general domain knowledge. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, EMNLP 2013, 18-21 October 2013, Grand Hyatt Seattle, Seattle, Washington, USA, A meeting of SIGDAT, a Special Interest Group of the ACL, pages 415–425, 2013. URL http://aclweb.org/ anthology/D/D13/D13-1040.pdf. Thomas Demeester, Tim Rocktäschel, and Sebastian Riedel. Lifted rule injection for relation embeddings. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, EMNLP 2016, Austin, Texas, USA, November 1-4, 2016, pages 1389–1399, 2016. URL http://aclweb.org/ anthology/D/D16/D16-1146.pdf. Liya Ding. Neural prolog-the concepts, construction and mechanism. In Systems, Man and Cybernetics, 1995. Intelligent Systems for the 21st Century., IEEE International Conference on, volume 4, pages 3603–3608. IEEE, 1995. Xin Dong, Evgeniy Gabrilovich, Geremy Heitz, Wilko Horn, Ni Lao, Kevin Murphy, Thomas Strohmann, Shaohua Sun, and Wei Zhang. Knowledge vault: a web-scale approach to probabilistic knowledge fusion. In The 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’14, New York, NY, USA - August 24 - 27, 2014, pages 601–610, 2014. doi: 10.1145/2623330. 2623623. URL http://doi.acm.org/10.1145/2623330.2623623. Gregory Druck, Gideon S. Mann, and Andrew McCallum. Semi-supervised learning of dependency parsers using generalized expectation criteria. In ACL 2009, Proceedings of the 47th Annual Meeting of the Association for Computational Linguistics and the 4th International Joint Conference on Natural Language Processing of the AFNLP, 2-7 August 2009, Singapore, pages 360–368, 2009. URL http://www.aclweb.org/anthology/P09-1041. BIBLIOGRAPHY 135 John C. Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12:2121–2159, 2011. URL http://dl.acm.org/citation. cfm?id=2021068. Saso Dzeroski. Inductive logic programming in a nutshell. In Lise Getoor and Ben Taskar, editors, Introduction to Statistical Relational Learning, chapter 3, pages 57–92. 2007. Oren Etzioni, Michele Banko, Stephen Soderland, and Daniel S. Weld. Open information extraction from the web. Commun. ACM, 51(12):68–74, 2008. doi: 10.1145/1409360.1409378. URL http://doi.acm.org/10.1145/ 1409360.1409378. Manaal Faruqui, Jesse Dodge, Sujay Kumar Jauhar, Chris Dyer, Eduard H. Hovy, and Noah A. Smith. Retrofitting word vectors to semantic lexicons. In NAACL HLT 2015, The 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Denver, Colorado, USA, May 31 - June 5, 2015, pages 1606–1615, 2015. URL http://aclweb. org/anthology/N/N15/N15-1184.pdf. John R Firth. A synopsis of linguistic theory 1930–55, 1957. Manoel V. M. França, Gerson Zaverucha, and Artur S. d’Avila Garcez. Fast relational learning using bottom clause propositionalization with artificial neural networks. Machine Learning, 94(1):81–104, 2014. doi: 10.1007/s10994-013-5392-1. URL http://dx.doi.org/10.1007/s10994-013-5392-1. Luis Antonio Galárraga, Nicoleta Preda, and Fabian M. Suchanek. Mining rules to align knowledge bases. In Proceedings of the 2013 workshop on Automated knowledge base construction, AKBC@CIKM 13, San Francisco, California, USA, October 27-28, 2013, pages 43–48, 2013. doi: 10.1145/2509558.2509566. URL http://doi.acm.org/10.1145/2509558.2509566. Hervé Gallaire and Jack Minker, editors. Logic and Data Bases, Symposium on Logic and Data Bases, Centre d’études et de recherches de Toulouse, 1977, Advances in Data Base Theory, New York, 1978. Plemum Press. ISBN 0-306-40060-X. Kuzman Ganchev, João Graça, Jennifer Gillenwater, and Ben Taskar. Posterior regularization for structured latent variable models. Journal of Machine Learning Re- 136 BIBLIOGRAPHY search, 11:2001–2049, 2010. URL http://portal.acm.org/citation. cfm?id=1859918. Matt Gardner, Partha Pratim Talukdar, Bryan Kisiel, and Tom M. Mitchell. Improving learning and inference in a large knowledge-base using latent syntactic cues. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, EMNLP 2013, 18-21 October 2013, Grand Hyatt Seattle, Seattle, Washington, USA, A meeting of SIGDAT, a Special Interest Group of the ACL, pages 833–838, 2013. URL http://aclweb.org/anthology/D/D13/ D13-1080.pdf. Matt Gardner, Partha Pratim Talukdar, Jayant Krishnamurthy, and Tom M. Mitchell. Incorporating vector space similarity in random walk inference over knowledge bases. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP 2014, October 25-29, 2014, Doha, Qatar, A meeting of SIGDAT, a Special Interest Group of the ACL, pages 397–406, 2014. URL http://aclweb.org/anthology/D/D14/D14-1044.pdf. Dan Garrette, Katrin Erk, and Raymond J. Mooney. Integrating logical representations with probabilistic information using markov logic. In Proceedings of the Ninth International Conference on Computational Semantics, IWCS 2011, January 12-14, 2011, Oxford, UK, 2011. URL http://aclweb.org/anthology/ W/W11/W11-0112.pdf. Allen Van Gelder. Efficient loop detection in prolog using the tortoise-and-hare technique. J. Log. Program., 4(1):23–31, 1987. doi: 10.1016/0743-1066(87) 90020-3. URL https://doi.org/10.1016/0743-1066(87)900203. Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, AISTATS 2010, Chia Laguna Resort, Sardinia, Italy, May 13-15, 2010, pages 249–256, 2010. URL http: //www.jmlr.org/proceedings/papers/v9/glorot10a.html. Kurt Gödel. Zum intuitionistischen aussagenkalkül. Anzeiger der Akademie der Wissenschaften in Wien, 69:65–66, 1932. BIBLIOGRAPHY 137 Ian J. Goodfellow, Yoshua Bengio, and Aaron C. Courville. Deep Learning. Adaptive computation and machine learning. MIT Press, 2016. ISBN 978-0-262-03561-3. URL http://www.deeplearningbook.org/. Bryce Goodman and Seth Flaxman. EU regulations on algorithmic decision-making and a ”right to explanation”. CoRR, abs/1606.08813, 2016. URL http:// arxiv.org/abs/1606.08813. Alex Graves. Generating sequences with recurrent neural networks. CoRR, abs/1308.0850, 2013. URL http://arxiv.org/abs/1308.0850. Alex Graves and Jürgen Schmidhuber. Framewise phoneme classification with bidirectional LSTM and other neural network architectures. Neural Networks, 18(5-6):602–610, 2005. doi: 10.1016/j.neunet.2005.06.042. URL http://dx. doi.org/10.1016/j.neunet.2005.06.042. Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. CoRR, abs/1410.5401, 2014. URL http://arxiv.org/abs/1410.5401. Edward Grefenstette. Towards a formal distributional semantics: Simulating logical calculi with tensors. In Proceedings of the Second Joint Conference on Lexical and Computational Semantics, *SEM 2013, June 13-14, 2013, Atlanta, Georgia, USA., pages 1–10, 2013. URL http://aclweb.org/anthology/S/S13/S131001.pdf. Edward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman, and Phil Blunsom. Learning to transduce with unbounded memory. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 1828–1836, 2015. URL http://papers.nips.cc/paper/5648learning-to-transduce-with-unbounded-memory. Shu Guo, Quan Wang, Lihong Wang, Bin Wang, and Li Guo. Jointly embedding knowledge graphs and logical rules. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, EMNLP 2016, Austin, Texas, USA, November 1-4, 2016, pages 192–202, 2016. URL http://aclweb. org/anthology/D/D16/D16-1019.pdf. Karl Moritz Hermann and Phil Blunsom. The role of syntax in vector space models of compositional semantics. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, ACL 2013, 4-9 August 138 BIBLIOGRAPHY 2013, Sofia, Bulgaria, Volume 1: Long Papers, pages 894–904, 2013. URL http://aclweb.org/anthology/P/P13/P13-1088.pdf. Karl Moritz Hermann, Tomás Kociský, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. Teaching machines to read and comprehend. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 1693– 1701, 2015. URL http://papers.nips.cc/paper/5945-teachingmachines-to-read-and-comprehend. Jochen Hipp, Ulrich Güntzer, and Gholamreza Nakhaeizadeh. Algorithms for association rule mining - A general survey and comparison. SIGKDD Explorations, 2(1):58–64, 2000. doi: 10.1145/360402.360421. URL http://doi.acm. org/10.1145/360402.360421. Pascal Hitzler, Steffen Hölldobler, and Anthony Karel Seda. Logic programs and connectionist networks. J. Applied Logic, 2(3):245–272, 2004. doi: 10.1016/ j.jal.2004.03.002. URL http://dx.doi.org/10.1016/j.jal.2004. 03.002. Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735–1780, 1997. doi: 10.1162/neco.1997.9.8.1735. URL http://dx.doi.org/10.1162/neco.1997.9.8.1735. Steffen Hölldobler. A structured connectionist unification algorithm. In Proceedings of the 8th National Conference on Artificial Intelligence. Boston, Massachusetts, July 29 - August 3, 1990, 2 Volumes., pages 587–593, 1990. URL http://www. aaai.org/Library/AAAI/1990/aaai90-088.php. Steffen Hölldobler, Yvonne Kalinke, and Hans-Peter Störr. Approximating the semantics of logic programs by recurrent neural networks. Appl. Intell., 11(1): 45–58, 1999. doi: 10.1023/A:1008376514077. URL http://dx.doi.org/ 10.1023/A:1008376514077. Baotian Hu, Zhengdong Lu, Hang Li, and Qingcai Chen. Convolutional neural network architectures for matching natural language sentences. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pages 2042–2050, 2014. BIBLIOGRAPHY 139 URL http://papers.nips.cc/paper/5550-convolutionalneural-network-architectures-for-matching-naturallanguage-sentences. Zhiting Hu, Xuezhe Ma, Zhengzhong Liu, Eduard H. Hovy, and Eric P. Xing. Harnessing deep neural networks with logic rules. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers, 2016. URL http: //aclweb.org/anthology/P/P16/P16-1228.pdf. Sergio Jiménez, George Dueñas, Julia Baquero, and Alexander F. Gelbukh. UNALNLP: combining soft cardinality features for semantic textual similarity, relatedness and entailment. In Proceedings of the 8th International Workshop on Semantic Evaluation, SemEval@COLING 2014, Dublin, Ireland, August 23-24, 2014., pages 732–742, 2014. URL http://aclweb.org/anthology/S/S14/S142131.pdf. Armand Joulin and Tomas Mikolov. Inferring algorithmic patterns with stackaugmented recurrent nets. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 190–198, 2015. URL http://papers.nips.cc/paper/5857-inferringalgorithmic-patterns-with-stack-augmented-recurrentnets. Cezary Kaliszyk, François Chollet, and Christian Szegedy. Holstep: A machine learning dataset for higher-order logic theorem proving. In International Conference on Learning Representations (ICLR), 2017. URL http: //arxiv.org/abs/1703.00426. Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. URL http://arxiv.org/abs/1412.6980. Levente Kocsis and Csaba Szepesvári. Bandit based monte-carlo planning. In Machine Learning: ECML 2006, 17th European Conference on Machine Learning, Berlin, Germany, September 18-22, 2006, Proceedings, pages 282–293, 2006. doi: 10.1007/11871842 29. URL http://dx.doi.org/10.1007/ 11871842_29. 140 BIBLIOGRAPHY Stanley Kok and Pedro M. Domingos. Statistical predicate invention. In Machine Learning, Proceedings of the Twenty-Fourth International Conference (ICML 2007), Corvallis, Oregon, USA, June 20-24, 2007, pages 433–440, 2007. doi: 10.1145/1273496.1273551. URL http://doi.acm.org/10.1145/ 1273496.1273551. Vladyslav Kolesnyk, Tim Rocktäschel, and Sebastian Riedel. Generating natural language inference chains. CoRR, abs/1606.01404, 2016. URL http://arxiv. org/abs/1606.01404. Ekaterina Komendantskaya. Unification neural networks: unification by errorcorrection learning. Logic Journal of the IGPL, 19(6):821–847, 2011. doi: 10.1093/jigpal/jzq012. URL http://dx.doi.org/10.1093/jigpal/ jzq012. Robert A. Kowalski and Donald Kuehner. Linear resolution with selection function. Artif. Intell., 2(3/4):227–260, 1971. doi: 10.1016/0004-3702(71)90012-9. URL http://dx.doi.org/10.1016/0004-3702(71)90012-9. Germán Kruszewski, Denis Paperno, and Marco Baroni. Deriving boolean structures from distributional vectors. TACL, 3:375–388, 2015. URL https://tacl2013.cs.columbia.edu/ojs/index.php/tacl/ article/view/616. Alice Lai and Julia Hockenmaier. Illinois-lh: A denotational and distributional approach to semantics. In Proceedings of the 8th International Workshop on Semantic Evaluation, SemEval@COLING 2014, Dublin, Ireland, August 23-24, 2014., pages 329–334, 2014. URL http://aclweb.org/anthology/S/ S14/S14-2055.pdf. Ni Lao and William W. Cohen. Relational retrieval using a combination of pathconstrained random walks. Machine Learning, 81(1):53–67, 2010. doi: 10. 1007/s10994-010-5205-8. URL http://dx.doi.org/10.1007/s10994010-5205-8. Ni Lao, Tom M. Mitchell, and William W. Cohen. Random walk inference and learning in A large scale knowledge base. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, EMNLP 2011, 27-31 July 2011, John McIntyre Conference Centre, Edinburgh, UK, A meeting of BIBLIOGRAPHY 141 SIGDAT, a Special Interest Group of the ACL, pages 529–539, 2011. URL http://www.aclweb.org/anthology/D11-1049. Ni Lao, Amarnag Subramanya, Fernando C. N. Pereira, and William W. Cohen. Reading the web with learned syntactic-semantic inference rules. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, EMNLP-CoNLL 2012, July 12-14, 2012, Jeju Island, Korea, pages 1017–1026, 2012. URL http://www. aclweb.org/anthology/D12-1093. Mike Lewis and Mark Steedman. Combined distributional and logical semantics. TACL, 1:179–192, 2013. URL https://tacl2013.cs.columbia.edu/ ojs/index.php/tacl/article/view/93. Yang Liu, Chengjie Sun, Lei Lin, and Xiaolong Wang. Learning natural language inference using bidirectional LSTM model and inner-attention. CoRR, abs/1605.09090, 2016. URL http://arxiv.org/abs/1605.09090. John W. Lloyd. Foundations of Logic Programming, 2nd Edition. Springer, 1987. ISBN 3-540-18199-7. Sarah M. Loos, Geoffrey Irving, Christian Szegedy, and Cezary Kaliszyk. Deep network guided proof search. In LPAR-21, 21st International Conference on Logic for Programming, Artificial Intelligence and Reasoning, Maun, Botswana, May 7-12, 2017, pages 85–105, 2017. URL http://www.easychair.org/ publications/paper/340345. Will Lowe and Scott McDonald. The direct route: Mediated priming in semantic space. Technical report, The University of Edinburgh, 2000. Jan Lukasiewicz. O logice trojwartosciowej. Ruch filozoficzny, 5(169-171):32, 1920. Gideon S. Mann and Andrew McCallum. Generalized expectation criteria for semisupervised learning of conditional random fields. In ACL 2008, Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics, June 15-20, 2008, Columbus, Ohio, USA, pages 870–878, 2008. URL http: //www.aclweb.org/anthology/P08-1099. Christopher D. Manning, Prabhakar Raghavan, and Hinrich Schütze. Introduction to information retrieval. Cambridge University Press, 2008. ISBN 978-0-52186571-5. 142 BIBLIOGRAPHY Marco Marelli, Luisa Bentivogli, Marco Baroni, Raffaella Bernardi, Stefano Menini, and Roberto Zamparelli. Semeval-2014 task 1: Evaluation of compositional distributional semantic models on full sentences through semantic relatedness and textual entailment. In Proceedings of the 8th International Workshop on Semantic Evaluation, SemEval@COLING 2014, Dublin, Ireland, August 23-24, 2014., pages 1–8, 2014. URL http://aclweb.org/anthology/S/S14/S142001.pdf. Tomas Mikolov, Ilya Sutskever, Kai Chen, Gregory S. Corrado, and Jeffrey Dean. Distributed representations of words and phrases and their compositionality. In Advances in Neural Information Processing Systems 26: 27th Annual Conference on Neural Information Processing Systems 2013. Proceedings of a meeting held December 5-8, 2013, Lake Tahoe, Nevada, United States., pages 3111–3119, 2013. URL http://papers.nips.cc/paper/5021distributed-representations-of-words-and-phrases-andtheir-compositionality. George A. Miller. Wordnet: A lexical database for english. Commun. ACM, 38(11): 39–41, 1995. doi: 10.1145/219717.219748. URL http://doi.acm.org/ 10.1145/219717.219748. Jeff Mitchell and Mirella Lapata. Vector-based models of semantic composition. In ACL 2008, Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics, June 15-20, 2008, Columbus, Ohio, USA, pages 236– 244, 2008. URL http://www.aclweb.org/anthology/P08-1028. Volodymyr Mnih, Nicolas Heess, Alex Graves, and Koray Kavukcuoglu. Recurrent models of visual attention. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pages 2204–2212, 2014. URL http://papers.nips.cc/paper/5542-recurrent-modelsof-visual-attention. Saif Mohammad, Svetlana Kiritchenko, Parinaz Sobhani, Xiao-Dan Zhu, and Colin Cherry. Semeval-2016 task 6: Detecting stance in tweets. In Proceedings of the 10th International Workshop on Semantic Evaluation, SemEval@NAACLHLT 2016, San Diego, CA, USA, June 16-17, 2016, pages 31–41, 2016. URL http://aclweb.org/anthology/S/S16/S16-1003.pdf. BIBLIOGRAPHY 143 Lili Mou, Rui Men, Ge Li, Yan Xu, Lu Zhang, Rui Yan, and Zhi Jin. Natural language inference by tree-based convolution and heuristic matching. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 2: Short Papers, 2016. URL http://aclweb.org/anthology/P/P16/P16-2022.pdf. Stephen Muggleton. Inductive logic programming. New Generation Comput., 8(4): 295–318, 1991. doi: 10.1007/BF03037089. URL http://dx.doi.org/10. 1007/BF03037089. Stephen Muggleton and Luc De Raedt. Inductive logic programming: Theory and methods. J. Log. Program., 19/20:629–679, 1994. doi: 10.1016/ 0743-1066(94)90035-3. URL http://dx.doi.org/10.1016/07431066(94)90035-3. Stephen H Muggleton, Dianhuan Lin, and Alireza Tamaddoni-Nezhad. Metainterpretive learning of higher-order dyadic datalog: Predicate invention revisited. Machine Learning, 100(1):49–73, 2015. Tsendsuren Munkhdalai and Hong Yu. Neural tree indexers for text understanding. CoRR, abs/1607.04492, 2016a. URL http://arxiv.org/abs/1607. 04492. Tsendsuren Munkhdalai and Hong Yu. Neural semantic encoders. CoRR, abs/1607.04315, 2016b. URL http://arxiv.org/abs/1607.04315. Arvind Neelakantan, Benjamin Roth, and Andrew McCallum. Compositional vector space models for knowledge base completion. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing of the Asian Federation of Natural Language Processing, ACL 2015, July 26-31, 2015, Beijing, China, Volume 1: Long Papers, pages 156–166, 2015. URL http://aclweb.org/anthology/P/P15/P15-1016.pdf. Arvind Neelakantan, Quoc V. Le, and Ilya Sutskever. Neural programmer: Inducing latent programs with gradient descent. In International Conference on Learning Representations (ICLR), 2016. URL http://arxiv.org/abs/1511. 04834. 144 BIBLIOGRAPHY A Nemirovski and D Yudin. On Cezari’s convergence of the steepest descent method for approximating saddle point of convex-concave functions. In Soviet Mathematics Doklady, volume 19, 1978. Maximilian Nickel, Volker Tresp, and Hans-Peter Kriegel. A three-way model for collective learning on multi-relational data. In Proceedings of the 28th International Conference on Machine Learning, ICML 2011, Bellevue, Washington, USA, June 28 - July 2, 2011, pages 809–816, 2011. Maximilian Nickel, Volker Tresp, and Hans-Peter Kriegel. Factorizing YAGO: scalable machine learning for linked data. In Proceedings of the 21st World Wide Web Conference 2012, WWW 2012, Lyon, France, April 16-20, 2012, pages 271– 280, 2012. doi: 10.1145/2187836.2187874. URL http://doi.acm.org/ 10.1145/2187836.2187874. Maximilian Nickel, Lorenzo Rosasco, and Tomaso A. Poggio. Holographic embeddings of knowledge graphs. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, February 12-17, 2016, Phoenix, Arizona, USA., pages 1955–1961, 2016. URL http://www.aaai.org/ocs/index.php/ AAAI/AAAI16/paper/view/12484. Ulf Nilsson and Jan Maluszynski. Logic, programming and Prolog. Wiley, 1990. ISBN 978-0-471-92625-2. Sebastian Padó and Mirella Lapata. Dependency-based construction of semantic space models. Computational Linguistics, 33(2):161–199, 2007. doi: 10.1162/ coli.2007.33.2.161. URL http://dx.doi.org/10.1162/coli.2007. 33.2.161. Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. The pagerank citation ranking: Bringing order to the web. Technical report, Stanford InfoLab, 1999. Baolin Peng, Zhengdong Lu, Hang Li, and Kam-Fai Wong. Towards neural networkbased reasoning. CoRR, abs/1508.05508, 2015. URL http://arxiv.org/ abs/1508.05508. David Poole. First-order probabilistic inference. In IJCAI-03, Proceedings of the Eighteenth International Joint Conference on Artificial Intelligence, Acapulco, Mexico, August 9-15, 2003, pages 985–991, 2003. URL http://ijcai.org/ Proceedings/03/Papers/142.pdf. BIBLIOGRAPHY 145 J. Ross Quinlan. Learning logical definitions from relations. Machine Learning, 5: 239–266, 1990. doi: 10.1007/BF00117105. URL http://dx.doi.org/10. 1007/BF00117105. Luc De Raedt. Logical settings for concept-learning. Artif. Intell., 95(1):187–201, 1997. doi: 10.1016/S0004-3702(97)00041-6. URL http://dx.doi.org/ 10.1016/S0004-3702(97)00041-6. Scott E. Reed and Nando de Freitas. Neural programmer-interpreters. In International Conference on Learning Representations (ICLR), 2016. URL http: //arxiv.org/abs/1511.06279. Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme. BPR: bayesian personalized ranking from implicit feedback. In UAI 2009, Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence, Montreal, QC, Canada, June 18-21, 2009, pages 452–461, 2009. URL https://dslpitt.org/uai/displayArticleDetails.jsp? mmnu=1&smnu=2&article_id=1630&proceeding_id=25. Matthew Richardson and Pedro M. Domingos. Markov logic networks. Machine Learning, 62(1-2):107–136, 2006. doi: 10.1007/s10994-006-5833-1. URL http: //dx.doi.org/10.1007/s10994-006-5833-1. Sebastian Riedel, Limin Yao, Andrew McCallum, and Benjamin M. Marlin. Relation extraction with matrix factorization and universal schemas. In Human Language Technologies: Conference of the North American Chapter of the Association of Computational Linguistics, Proceedings, June 9-14, 2013, Westin Peachtree Plaza Hotel, Atlanta, Georgia, USA, pages 74–84, 2013. URL http://aclweb. org/anthology/N/N13/N13-1008.pdf. Tim Rocktäschel and Sebastian Riedel. Learning knowledge base inference with neural theorem provers. In Proceedings of the 5th Workshop on Automated Knowledge Base Construction, AKBC@NAACL-HLT 2016, San Diego, CA, USA, June 17, 2016, pages 45–50, 2016. URL http://aclweb.org/anthology/W/ W16/W16-1309.pdf. Tim Rocktäschel and Sebastian Riedel. End-to-end differentiable proving. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long 146 BIBLIOGRAPHY Beach, California, United States, volume abs/1705.11040, 2017. URL http: //arxiv.org/abs/1705.11040. Tim Rocktäschel, Matko Bosnjak, Sameer Singh, and Sebastian Riedel. LowDimensional Embeddings of Logic. In ACL Workshop on Semantic Parsing (SP’14), 2014. Tim Rocktäschel, Sameer Singh, and Sebastian Riedel. Injecting logical background knowledge into embeddings for relation extraction. In NAACL HLT 2015, The 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Denver, Colorado, USA, May 31 - June 5, 2015, pages 1119–1129, 2015. URL http://aclweb.org/anthology/N/N15/N15-1118.pdf. Tim Rocktäschel, Edward Grefenstette, Karl Moritz Hermann, Tomas Kocisky, and Phil Blunsom. Reasoning about entailment with neural attention. In International Conference on Learning Representations (ICLR), 2016. Alexander M. Rush, Sumit Chopra, and Jason Weston. A neural attention model for abstractive sentence summarization. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, EMNLP 2015, Lisbon, Portugal, September 17-21, 2015, pages 379–389, 2015. URL http://aclweb. org/anthology/D/D15/D15-1044.pdf. Stuart J. Russell and Peter Norvig. Artificial Intelligence - A Modern Approach (3. internat. ed.). Pearson Education, 2010. ISBN 978-0-13-207148-2. Ivan Sanchez, Tim Rocktäschel, Sebastian Riedel, and Sameer Singh. Towards Extracting Faithful and Descriptive Representations of Latent Variable Models. In AAAI Spring Symposium on Knowledge Representation and Reasoning (KRR), 2015. Evan Sandhaus. The New York Times annotated corpus. Linguistic Data Consortium, 2008. Stefan Schoenmackers, Oren Etzioni, and Daniel S. Weld. Scaling textual inference to the web. In 2008 Conference on Empirical Methods in Natural Language Processing, EMNLP 2008, Proceedings of the Conference, 25-27 October 2008, Honolulu, Hawaii, USA, A meeting of SIGDAT, a Special Interest Group of the ACL, pages 79–88, 2008. URL http://www.aclweb.org/anthology/D081009. BIBLIOGRAPHY 147 Stefan Schoenmackers, Jesse Davis, Oren Etzioni, and Daniel Weld. Learning firstorder horn clauses from web text. In Empirical Methods in Natural Language Processing (EMNLP), 2010a. URL http://www.aclweb.org/anthology/ D10-1106. Stefan Schoenmackers, Jesse Davis, Oren Etzioni, and Daniel S. Weld. Learning first-order horn clauses from web text. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, EMNLP 2010, 9-11 October 2010, MIT Stata Center, Massachusetts, USA, A meeting of SIGDAT, a Special Interest Group of the ACL, pages 1088–1098, 2010b. URL http: //www.aclweb.org/anthology/D10-1106. Marwin H. S. Segler, Mike Preuß, and Mark P. Waller. Towards ”alphachem”: Chemical synthesis planning with tree search and deep neural network policies. CoRR, abs/1702.00020, 2017. URL http://arxiv.org/abs/1702.00020. Luciano Serafini and Artur S. d’Avila Garcez. Logic tensor networks: Deep learning and logical reasoning from data and knowledge. In Proceedings of the 11th International Workshop on Neural-Symbolic Learning and Reasoning (NeSy’16) co-located with the Joint Multi-Conference on Human-Level Artificial Intelligence (HLAI 2016), New York City, NY, USA, July 16-17, 2016., 2016. URL http: //ceur-ws.org/Vol-1768/NESY16_paper3.pdf. Lokendra Shastri. Neurally motivated constraints on the working memory capacity of a production system for parallel processing: Implications of a connectionist model based on temporal synchrony. In Proceedings of the Fourteenth Annual Conference of the Cognitive Science Society: July 29 to August 1, 1992, Cognitive Science Program, Indiana University, Bloomington, volume 14, page 159. Psychology Press, 1992. Jude W Shavlik and Geoffrey G Towell. An approach to combining explanationbased and neural learning algorithms. Connection Science, 1(3):231–253, 1989. Yelong Shen, Po-Sen Huang, Jianfeng Gao, and Weizhu Chen. Reasonet: Learning to stop reading in machine comprehension. In Proceedings of the Workshop on Cognitive Computation: Integrating neural and symbolic approaches 2016 co-located with the 30th Annual Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain, December 9, 2016., 2016. URL http: //ceur-ws.org/Vol-1773/CoCoNIPS_2016_paper2.pdf. 148 BIBLIOGRAPHY David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Vedavyas Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy P. Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016. doi: 10. 1038/nature16961. URL http://dx.doi.org/10.1038/nature16961. Sameer Singh, Dustin Hillard, and Chris Leggetter. Minimally-supervised extraction of entities from text advertisements. In Human Language Technologies: Conference of the North American Chapter of the Association of Computational Linguistics, Proceedings, June 2-4, 2010, Los Angeles, California, USA, pages 73–81, 2010. URL http://www.aclweb.org/anthology/N10-1009. Richard Socher, Eric H. Huang, Jeffrey Pennington, Andrew Y. Ng, and Christopher D. Manning. Dynamic pooling and unfolding recursive autoencoders for paraphrase detection. In Advances in Neural Information Processing Systems 24: 25th Annual Conference on Neural Information Processing Systems 2011. Proceedings of a meeting held 12-14 December 2011, Granada, Spain., pages 801–809, 2011. URL http://papers.nips.cc/paper/4204dynamic-pooling-and-unfolding-recursive-autoencodersfor-paraphrase-detection. Richard Socher, Danqi Chen, Christopher D. Manning, and Andrew Y. Ng. Reasoning with neural tensor networks for knowledge base completion. In Advances in Neural Information Processing Systems 26: 27th Annual Conference on Neural Information Processing Systems 2013. Proceedings of a meeting held December 5-8, 2013, Lake Tahoe, Nevada, United States., pages 926–934, 2013. URL http://papers.nips.cc/paper/5028-reasoning-withneural-tensor-networks-for-knowledge-base-completion. Gustav Sourek, Vojtech Aschenbrenner, Filip Zelezný, and Ondrej Kuzelka. Lifted relational neural networks. In Proceedings of the NIPS Workshop on Cognitive Computation: Integrating Neural and Symbolic Approaches co-located with the 29th Annual Conference on Neural Information Processing Systems (NIPS 2015), Montreal, Canada, December 11-12, 2015., 2015. URL http://ceur-ws. org/Vol-1583/CoCoNIPS_2015_paper_7.pdf. BIBLIOGRAPHY 149 Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory networks. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 2440–2448, 2015. URL http:// papers.nips.cc/paper/5846-end-to-end-memory-networks. Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pages 3104–3112, 2014. URL http://papers.nips.cc/paper/5346-sequence-tosequence-learning-with-neural-networks. Kristina Toutanova, Danqi Chen, Patrick Pantel, Hoifung Poon, Pallavi Choudhury, and Michael Gamon. Representing text for joint embedding of text and knowledge bases. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, EMNLP 2015, Lisbon, Portugal, September 17-21, 2015, pages 1499–1509, 2015. URL http://aclweb.org/anthology/D/D15/ D15-1174.pdf. Geoffrey G. Towell and Jude W. Shavlik. Knowledge-based artificial neural networks. Artif. Intell., 70(1-2):119–165, 1994. doi: 10.1016/0004-3702(94)90105-8. URL http://dx.doi.org/10.1016/0004-3702(94)90105-8. Théo Trouillon, Johannes Welbl, Sebastian Riedel, Éric Gaussier, and Guillaume Bouchard. Complex embeddings for simple link prediction. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, pages 2071–2080, 2016. URL http: //jmlr.org/proceedings/papers/v48/trouillon16.html. Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel Urtasun. Order-embeddings of images and language. In International Conference on Learning Representations (ICLR), 2016. URL http://arxiv.org/abs/1511.06361. Patrick Verga, David Belanger, Emma Strubell, Benjamin Roth, and Andrew McCallum. Multilingual relation extraction using compositional universal schema. In NAACL HLT 2016, The 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, San Diego California, USA, June 12-17, 2016, pages 886–896, 2016a. URL http://aclweb.org/anthology/N/N16/N16-1103.pdf. 150 BIBLIOGRAPHY Patrick Verga, Arvind Neelakantan, and Andrew McCallum. Generalizing to unseen entities and entity pairs with row-less universal schema. CoRR, abs/1606.05804, 2016b. URL http://arxiv.org/abs/1606.05804. Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 2692–2700, 2015a. URL http://papers.nips.cc/paper/5866pointer-networks. Oriol Vinyals, Lukasz Kaiser, Terry Koo, Slav Petrov, Ilya Sutskever, and Geoffrey E. Hinton. Grammar as a foreign language. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 2773– 2781, 2015b. URL http://papers.nips.cc/paper/5635-grammaras-a-foreign-language. Johanna Völker and Mathias Niepert. Statistical schema induction. In The Semantic Web: Research and Applications - 8th Extended Semantic Web Conference, ESWC 2011, Heraklion, Crete, Greece, May 29-June 2, 2011, Proceedings, Part I, pages 124–138, 2011. doi: 10.1007/978-3-642-21034-1 9. URL http://dx.doi. org/10.1007/978-3-642-21034-1_9. Quan Wang, Bin Wang, and Li Guo. Knowledge base completion using embeddings and rules. In Proceedings of the Twenty-Fourth International Joint Conference on Artificial Intelligence, IJCAI 2015, Buenos Aires, Argentina, July 25-31, 2015, pages 1859–1866, 2015. URL http://ijcai.org/Abstract/15/264. William Yang Wang and William W. Cohen. Joint information extraction and reasoning: A scalable statistical relational learning approach. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing of the Asian Federation of Natural Language Processing, ACL 2015, July 26-31, 2015, Beijing, China, Volume 1: Long Papers, pages 355–364, 2015. URL http://aclweb.org/anthology/P/P15/P15-1035.pdf. William Yang Wang and William W. Cohen. Learning first-order logic embeddings via matrix factorization. In Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence, IJCAI 2016, New York, NY, USA, 9-15 July BIBLIOGRAPHY 151 2016, pages 2132–2138, 2016. URL http://www.ijcai.org/Abstract/ 16/304. William Yang Wang, Kathryn Mazaitis, and William W. Cohen. Programming with personalized pagerank: a locally groundable first-order probabilistic logic. In 22nd ACM International Conference on Information and Knowledge Management, CIKM’13, San Francisco, CA, USA, October 27 - November 1, 2013, pages 2129– 2138, 2013. doi: 10.1145/2505515.2505573. URL http://doi.acm.org/ 10.1145/2505515.2505573. Zhiguo Wang, Wael Hamza, and Radu Florian. Bilateral multi-perspective matching for natural language sentences. CoRR, abs/1702.03814, 2017. URL http: //arxiv.org/abs/1702.03814. Zhuoyu Wei, Jun Zhao, Kang Liu, Zhenyu Qi, Zhengya Sun, and Guanhua Tian. Large-scale knowledge base completion: Inferring via grounding network sampling over selected instances. In Proceedings of the 24th ACM International Conference on Information and Knowledge Management, CIKM 2015, Melbourne, VIC, Australia, October 19 - 23, 2015, pages 1331–1340, 2015. doi: 10.1145/2806416.2806513. URL http://doi.acm.org/10.1145/ 2806416.2806513. Dirk Weissenborn. Separating answers from queries for neural reading comprehension. CoRR, abs/1607.03316, 2016. URL http://arxiv.org/abs/1607. 03316. Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. CoRR, abs/1410.3916, 2014. URL http://arxiv.org/abs/1410.3916. Fei Wu, Jun Song, Yi Yang, Xi Li, Zhongfei (Mark) Zhang, and Yueting Zhuang. Structured embedding via pairwise relations and long-range interactions in knowledge base. In Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence, January 25-30, 2015, Austin, Texas, USA., pages 1663–1670, 2015. URL http://www.aaai.org/ocs/index.php/AAAI/AAAI15/ paper/view/9342. Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron C. Courville, Ruslan Salakhutdinov, Richard S. Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, 152 BIBLIOGRAPHY France, 6-11 July 2015, pages 2048–2057, 2015. URL http://jmlr.org/ proceedings/papers/v37/xuc15.html. Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. Embedding entities and relations for learning and inference in knowledge bases. In International Conference on Learning Representations (ICLR), 2015. URL http://arxiv.org/abs/1412.6575. Fan Yang, Zhilin Yang, and William W. Cohen. Differentiable learning of logical rules for knowledge base completion. CoRR, abs/1702.08367, 2017. URL http: //arxiv.org/abs/1702.08367. Wenpeng Yin and Hinrich Schütze. Convolutional neural network for paraphrase identification. In NAACL HLT 2015, The 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Denver, Colorado, USA, May 31 - June 5, 2015, pages 901–911, 2015. URL http://aclweb.org/anthology/N/N15/N15-1091.pdf. Wojciech Zaremba, Ilya Sutskever, and Oriol Vinyals. Recurrent neural network regularization. CoRR, abs/1409.2329, 2014. URL http://arxiv.org/abs/ 1409.2329. Luke S. Zettlemoyer and Michael Collins. Learning to map sentences to logical form: Structured classification with probabilistic categorial grammars. In UAI ’05, Proceedings of the 21st Conference in Uncertainty in Artificial Intelligence, Edinburgh, Scotland, July 26-29, 2005, pages 658–666, 2005. URL https://dslpitt.org/uai/displayArticleDetails.jsp? mmnu=1&smnu=2&article_id=1209&proceeding_id=21. Jiang Zhao, Tiantian Zhu, and Man Lan. ECNU: one stone two birds: Ensemble of heterogenous measures for semantic relatedness and textual entailment. In Proceedings of the 8th International Workshop on Semantic Evaluation, SemEval@COLING 2014, Dublin, Ireland, August 23-24, 2014., pages 271–277, 2014. URL http://aclweb.org/anthology/S/S14/S142044.pdf.
9cs.NE
Road Damage Detection Using Deep Neural Networks with Images Captured Through a Smartphone arXiv:1801.09454v2 [cs.CV] 2 Feb 2018 Hiroya Maeda∗, Yoshihide Sekimoto, Toshikazu Seto, Takehiro Kashiyama, Hiroshi Omata University of Tokyo, 4-6-1 Komaba, Tokyo, Japan 1 Abstract: Research on damage detection of road surfaces using image processing techniques has been actively conducted, achieving considerably high detection accuracies. Many studies only focus on the detection of the presence or absence of damage. However, in a real-world scenario, when the road managers from a governing body need to repair such damage, they need to clearly understand the type of damage in order to take effective action. In addition, in many of these previous studies, the researchers acquire their own data using different methods. Hence, there is no uniform road damage dataset available openly, leading to the absence of a benchmark for road damage detection. This study makes three contributions to address these issues. First, to the best of our knowledge, for the first time, a large-scale road damage dataset is prepared. This dataset is composed of 9,053 road damage images captured with a smartphone installed on a car, with 15,435 instances of road surface damage included in these road images. In order to generate this dataset, we cooperated with 7 municipalities in Japan and acquired road images for more than 40 hours. These images were captured in a wide variety of weather and illuminance conditions. In each image, we annotated the bounding box representing the location and type of damage. Next, we used a state-of-the-art object detection method using convolutional neural networks to train the damage detection model with our dataset, and compared the accuracy and runtime speed on both, using a GPU server and a smartphone. Finally, we demonstrate that the type of damage can be classified into eight types with high accuracy by applying the proposed object detection method. The road damage dataset, our experimental results, and the developed smartphone application used in this study are publicly available (https://github.com/sekilab/RoadDamageDetector/). Introduction During the period of high economic growth in Japan from 1954 to 1973, infrastructure such as roads, bridges, and tunnels were constructed extensively; however, because many of these were constructed more than 50 years ago (MLIT, 2016), they are now aged, and the number of structures that are to be inspected is expected to increase rapidly in the next few decades. In addition, the discovery of the aged and affected parts of infrastructure has thus far depended solely on the expertise of veteran field engineers. However, owing to the increasing demand for inspections, a shortage of field technicians (experts) and financial resources has resulted in many areas. In particular, the number of municipalities that have neglected conducting appropriate inspections owing to the lack of resources or experts has been increasing (Kazuya et al., 2013). The United States also has similar infrastructure aging problems (AAoSHaT, 2008). The prevailing problems in infrastructure maintenance and management are likely to be experienced by countries all over the world. Considering this negative trend in infrastructure maintenance and management, it is evident that efficient and sophisticated infrastructure maintenance methods are urgently required. In response to abovementioned problem, many methods to efficiently inspect infrastructure, especially road conditions, have been studied, such as methods using laser technology or image processing. Moreover, there are quite a few studies using neural networks for civil engineering problems in the 11 years from 1989 to 2000 (Adeli, 2001). Furthermore, recently, computer vision and machine learning techniques have been successfully applied to automate road surface inspection (Chun et al., 2015; Zalama et al., 2014; Jo and Ryu, 2015). However, thus far, with respect to methods of inspections using image processing, we believe these methods ∗ [email protected] 1 suffer from three major disadvantages: 1. There is no common dataset for a comparison of results; in each research, the proposed method is evaluated using its own dataset of road damage images. Motivated by the field of general object recognition, wherein large common datasets such as ImageNet (Deng et al., 2009) and PASCAL VOC (Everingham et al., 2010, 2015) exist, we believe there is a need for a common dataset on road scratches. 2. Although current state-of-the-art object detection methods use end-to-end deep learning techniques, no such method exists for road damage detection. 3. Though road surface damage is distinguished into several categories (in Japan, eight categories according to the Road Maintenance and Repair Guidebook 2013 (JRA, 2013)), many studies have been limited to the detection or classification of damage in only the longitudinal and lateral directions (Chun et al., 2015; Zalama et al., 2014; Zhang et al., 2016; Akarsu et al., 2016; Maeda et al., 2016). Therefore, it is difficult for road managers to apply these research results directly in practical scenarios. Considering the abovementioned disadvantages, in this study, we develop a new, large-scale road damage dataset, and then train and evaluate a damage detection model that is based on the state-of-the-art convolutional neural network (CNN) method. The contributions of this study are as follows. 1.We created and released 9,053 road damage images containing 15,435 damages. The dataset contains the bounding box of each class for the eight types of road damage. Each image is extracted from an image set created by capturing pictures of a large number of roads obtained using a vehicle-mounted smartphone. The 9,053 images of the dataset contain a wide variety of weather and illuminance conditions. In addition, in assessing the type of damage, the expertise of a professional road administrator was employed, rendering the dataset considerably reliable. 2. Using our developed dataset, we have evaluated the state-of-the art object detection method based on deep learning and made benchmark results. All the trained models are also publicly available on our website1 . 3. Furthermore, we showed that the type of damage from among the eight types can be identified with high accuracy. The rest of the paper is organized as follows. In Sec- tion 2, we discuss the related works. Details of our new dataset are presented in Section 3. The experimental settings are explained in Section 4. Then, the results are provided in Section 5. Finally, Section 6 concludes the paper. 2 2.1 Related Works Road Damage Detection Road surface inspection is primarily based on visual observations by humans and quantitative analysis using expensive machines. Among these, the visual inspection approach not only requires experienced road managers, but also is timeconsuming and expensive. Furthermore, visual inspection tends to be inconsistent and unsustainable, which increases the risk associated with aging road infrastructure. Considering these issues, municipalities lacking the required resources do not conduct infrastructure inspections appropriately and frequently, increasing the risk posed by deteriorating structures. In contrast, quantitative determination based on large-scale inspection, such as using a mobile measurement system (MMS) (KOKUSAI KOGYO CO., 2016) or laser-scanning method (Yu and Salari, 2011) is also widely conducted. An MMS obtains highly accurate geospatial information using a moving vehicle; this system comprises a global positioning system (GPS) unit, an internal measurement unit, digital measurable images, a digital camera, a laser scanner, and an omnidirectional video recorder. Though quantitative inspection is highly accurate, it is considerably expensive to conduct such comprehensive inspections especially for small municipalities that lack the required financial resources. Therefore, considering the abovementioned issues, several attempts have been made to develop a method for analyzing road properties by using a combination of recordings by in-vehicle cameras and image processing technology to more efficiently inspect a road surface. For example, a previous study proposed an automated asphalt pavement crack detection method using image processing techniques and a naive Bayes-based machine-learning approach (Chun et al., 2015). In addition, a pothole-detection system using a commercial black-box camera has been previously proposed (Jo and Ryu, 2015). In recent times, it has become possible to quite accurately analyze the damage to road surfaces using deep neural networks (Zhang et al., 2016; 1 https://github.com/sekilab/RoadDamageDetector/ 2 Maeda et al., 2016; Zhang et al., 2017). For instance, Zhang et al. (Zhang et al., 2017) introduced CrackNet, which predicts class scores for all pixels. However, such road damage detection methods focus only on the determination of the existence of damage. Though some studies do classify the damage based on types— for example, Zalama et al. (Zalama et al., 2014) classified damage types vertically and horizontally, and Akarsu et al. (Akarsu et al., 2016) categorized damage into three types, namely, vertical, horizontal, and crocodile—most studies primarily focus on classifying damages between a few types. Therefore, for a practical damage detection model for use by municipalities, it is necessary to clearly distinguish and detect different types of road damage; this is because, depending on the type of damage, the road administrator needs to follow different approaches to rectify the damage. Furthermore, the application of deep learning for road surface damage identification has been proposed by few studies, for example, studies by Maeda et al. (Maeda et al., 2016) and Zhang et al. (Zhang et al., 2016). However, the method proposed by Maeda et al. (Maeda et al., 2016), which uses 256 × 256 pixel images, identifies the damaged road surfaces, but does not classify them into different types. In addition, the method of Zhang et al. (Zhang et al., 2016) identifies whether damage occurred exclusively using a 99 × 99 patch obtained from a 3264 × 2448 pixel image. Further, a 256 × 256 pixel damage classifier is applied using a sliding window approach (Felzenszwalb et al., 2010) for 5,888 × 3,584 pixel images in order to detect cracks on the concrete surface (Cha et al., 2017). In these studies, classification methods are applied to input images and damage is detected. Recently, it has been reported that object detection using end-to-end deep learning is more accurate and has a faster processing speed than using a combination of classification methods; this will be discussed in detail in 2.3. As an example of a method using end-to-end deep learning performing better than tradition methods, white line detection based on end-to-end deep learning using OverFeat (Sermanet et al., 2013) outperformed a previously proposed empirical method (Huval et al., 2015). However, to the best of our knowledge, no example of the application of end-to-end deep learning method for road damage detection exists. It is important to note that classification refers to labeling an image rather than an object, whereas detection means assigning an image a label and identifying the objects coordinates as exemplified by the ImageNet competition (Deng et al., 2009). Therefore, considering this, we apply the end-to-end object detection method based on deep learning to the road surface damage detection problem, and verify its detection accuracy and processing speed. In particular, we examine whether we can detect eight classes of road damage by applying state-of-the-art object detection methods (discussed later in 2.3) with the newly created road damage dataset (explained in Section 3). Although many excellent methods have been proposed, such as segmentation of cracks on the concrete surface (O’Byrne et al., 2014; Nishikawa et al., 2012), our research uses an object detection method. 2.2 Image Dataset of Road Surface Damage Though an image dataset of the road surface exists, called the kitti dataset (Geiger et al., 2013), it is primarily used for applications related to automatic driving. However, to the best of our knowledge, no dataset tagged for road damage exists in the field. In all the studies focusing on road damage detection described in 2.1, in each study, the researchers independently propose unique methods using acquired road images. Therefore, a comparison between the methods presented in these studies is difficult. Furthermore, according to Mohan et al. (Mohan and Poobal, 2017), there are few studies that construct damage detection models using real data, and 20 of these studies use road images taken directly from above the road. In fact, it is difficult to reproduce the road images taken directly from above the roads, because doing so involves installing a camera outside the car body, which, in many countries, is a violation of the law; in addition, it is costly to maintain a dedicated car solely for road images. Therefore, we have developed a dataset of road damage images using the road images captured using a smartphone on the dashboard of a general passenger car; in addition, we made this dataset publicly available. Moreover, we show that road surface damage can be detected with considerably high accuracy even with images acquired by employing such a simple method. 2.3 Object Detection System In general, for object detection, methods that apply an image classifier to an object detection task have become mainstream; these methods entail varying the size and position of the object in the test image, and then using 3 the classifier to identify the object. The sliding window approach is a well-known example (Felzenszwalb et al., 2010). In the past few years, an approach involving the extraction of multiple candidate regions of objects using region proposals as typified by R-CNN, then making a classification decision with candidate regions using classifiers has also been reported (Girshick et al., 2014). However, the R-CNN approach can be time consuming because it requires more crops, leading to significant duplicate computation from overlapping crops. This calculation redundancy was solved using a Fast R-CNN (Girshick, 2015), which inputs the entire image once through a feature extractor so that crops share the computation load of feature extraction. As described above, image processing methods have historically developed at a considerable pace. In our study, we primarily focus on four recent object detection systems: the Faster R-CNN (Ren et al., 2015), the You Look Only Once (YOLO) (Redmon et al., 2016; Redmon and Farhadi, 2016) system, the Region-based Fully Convolutional Networks (R-FCN) system (Dai et al., 2016), and the Single Shot Multibox Detector (SSD) system (Liu et al., 2016). 2.3.1 erably fast because it solves the problem as a mere regression, detecting objects by considering background information. The YOLO algorithm outputs the coordinates of the bounding box of the object candidate and the confidence of the inference after receiving an image as input. 2.3.3 R-FCN is another object detection framework, which was proposed by Dai et al. (Dai et al., 2016). Its architecture is that of a region-based, fully convolutional network for accurate and efficient object detection. Although Faster R-CNN is several times faster than Fast R-CNN, the region-specific component must be applied several hundred times per image. Instead of cropping features from the same layer where the region proposals are predicted like in the case of the Faster R-CNN method, in the R-FCN method, crops are taken from the last layer of the features prior to prediction. This approach of pushing cropping to the last layer minimizes the amount of per-region computation that must be performed. Dai et al. (Dai et al., 2016) showed that the R-FCN model (using Resnet 101) could achieve accuracy comparable to Faster R-CNN often at faster running speeds. Faster R-CNN The Faster R-CNN (Ren et al., 2015) has two stages for detection. In the first stage, images are processed using a feature extractor (e.g., VGG, MobileNet) called the Region Proposal Network (RPN) and simultaneously, some intermediate level layers (e.g., ”conv5”) are used to predict class bounding box proposals. In the second stage, these box proposals are used to crop features from the same intermediate feature map, which are subsequently input to the remainder of the feature extractor in order to predict a class label and its bounding box refinement for each proposal. It is important to note that Faster R-CNN does not crop proposals directly from the image and re-runs crops through the feature extractor, which would lead to duplicated computations. 2.3.2 R-FCN 2.3.4 SSD SSD (Liu et al., 2016) is an object detection framework that uses a single feed-forward convolutional network to directly predict classes and anchor offsets without requiring a second stage per-proposal classification operation. The key feature of this framework is the use of multi-scale convolutional bounding box outputs attached to multiple feature maps at the top of the network. 2.4 Base Network In all these object detection systems, a convolutional feature extractor as a base network is applied to the input image in order to obtain high-level features. The selection of the feature extractor is considerably important because the number of parameters and layers, the type of layers, and other properties directly affect the performance of the detector. We have selected seven representative base networks, which are explained in 2.4, and three base networks to evaluate the results in Section 5. The six feature extractors we have selected are YOLO YOLO is an object detection framework that can achieve high mean average precision (mAP) and speed (Redmon et al., 2016; Redmon and Farhadi, 2016). In addition, YOLO can predict the region and class of objects with a single CNN. An advantageous feature of YOLO is that its processing speed is consid4 widely used in the field of computer vision. 2.4.1 blocks are depthwise separable convolutions that factorize a standard convolution into a depthwise convolution and a 1 × 1 convolution, effectively reducing both the computational cost and number of parameters. darknet-19 Darknet-19 (Redmon and Farhadi, 2016) is a base model of the YOLO framework. The model has 19 convolutional layers and 5 maxpooling layers. 2.4.2 3 In this section, we describe our proposed new dataset, including how the data was obtained, how it was annotated, its contents, and issues related to privacy. VGG-16 VGG 16 (Simonyan and Zisserman, 2014) is a CNN with a total of 16 layers consisting of 13 convolution layers and 3 fully connected layers proposed in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) in 2014. This model achieved good results in ILSVRC and COCO 2015 (classification, detection, and segmentation) considering the depth of the layers. 2.4.3 3.1 Resnet Inception V2 Inception V2 (Ioffe and Szegedy, 2015) and Inception V3 (Szegedy et al., 2016) enable one to increase the depth and breadth of the network without increasing the number of parameters or the computational complexity by introducing so-called inception units. 2.4.5 Inception Resnet Inception Resnet V2 (Szegedy et al., 2017) improves recognition accuracy by combining both residual connections and Inception units effectively. 2.4.6 Data Collection Thus far, in the study of damage detection on the road surface, images are either captured from above the road surface or using on-board cameras on vehicles. When models are trained with images captured from above, the situations that can be applied in practice are limited, considering the difficulty of capturing such images. In contrast, when a model is constructed with images captured from an on-board vehicle camera, it is easy to apply these images to train the model for practical situations. For example, using a readily available camera like on smartphones and general passenger cars, any individual can easily detect road damages by running the model on the smartphone or by transferring the images to an external server and processing it on the server. We selected seven local governments in Japan2 and cooperated with the road administrators of each local government to collect 163,664 road images3 . Seven municipalities have snowy areas and urban areas that are very diverse in terms of regional characteristics such as the weather and fiscal constraints. We installed a smartphone (LG Nexus 5X) on the dashboard of a car, as shown in Figure 1, and photographed images of 600 × 600 pixels once per second. The reason we select a photographing interval of 1 s is because it is possible to photograph images while traveling on the road without leakage or duplication when the average speed of the car is approximately 40 km/h (or approximately 10 m/s). For this purpose, we created a smartphone application that can capture images of the roads and record the location information once per second; this application is also publicly available on our website. Resnet, which refers to Deep Residual Learning, (He et al., 2016), is a structure for deep learning, particularly for CNNs, that enables high-precision learning in a very deep network; it was released by Microsoft Research in 2015. Accuracy beyond human ability is obtained by learning images with 154 layers. Resnet achieved an error rate of 3.57% with the ImageNet test set and won the first place in ILSVRC 2015 classification task. 2.4.4 Proposed Dataset MobileNet MobileNet (Howard et al., 2017) has been shown to achieve an accuracy comparable to VGG-16 on ImageNet with only 1/30th of the computational cost and model size. MobileNet is designed for efficient inference in various mobile vision applications. Its building 2 Ichihara city, Chiba city, Sumida ward, Nagakute city, Adachi city, Muroran city, and Numazu city. 3 We traveled through every municipality covering approximately 1,500 km in total 5 3.3 Data Annotation The collected images were then annotated manually. We illustrate our annotation pipeline in Figure 3. Because our dataset format is designed in a manner similar to the PASCAL VOC (Everingham et al., 2010, 2015), it is easy to apply it to many existing methods used in the field of image processing. 3.4 Figure 1: Installation setup of the smartphone on the car. It is mounted on the dashboard of a general passenger car. Our developed application can capture a photograph of the road surface approximately 10 m ahead, which indicates that this application can photograph images while traveling on the road without leakage or duplication when the care moves at an average speed of about 40 km/h (about 10 m/s) if photographing every second. In addition, it can detect road damages in 1.5 s with high accuracy (see Section 5). 3.2 Data Statistics Our dataset is composed of 9,053 labeled road damage images. Of these 9,053 images, 15,435 bounding boxes of damage are annotated. Figure 4 shows the number of instances per label that were collected in each municipality. We photographed a number of road images in various regions of Japan, but could not avoid biasing some of the data. For example, damages such as D40 pose a more significant danger, and therefore, road managers repair these damages as soon as they occur; thus, there are not many instances of D40 in reality. In many studies, the blurring of white lines is not considered to be damage; however, in this study, white line blur is also considered as damage. In summary, our new dataset includes 9,053 damage images and 15,435 damage bounding boxes. The resolution of the images is 600 × 600 pixels. The area and the weather in the area are diverse, and thus, the dataset closely resembles the real world. We used this dataset to evaluate the damage detection model. Data Category Table 1 list the different damage types and their definition. In this paper, each damage type is represented with a Class Name such as D00. Each type of damage is illustrated in the examples in Figure 2. As can be seen from the table, the damage types are divided into eight categories. First, the damage is classified into cracks or other corruptions. Then, the cracks are divided into linear cracks and alligator cracks (crocodile cracks). Other corruptions, include not only pot holes and rutting, but also other road damage such as blurring of white lines. 3.5 Privacy Matters Our dataset is openly accessible by the public. Therefore, considering issues with privacy, based on visual inspection, when a person’s face or a car license plate are clearly reflected in the image, they are blurred out. To the best of our knowledge, no previous research covers such a wide variety of road damages, especially in the case of image processing. For example, the method proposed by Jo et al. (Jo and Ryu, 2015) detects only potholes in D40, and that of Zalama et al. (Zalama et al., 2014) classifies damage types exclusively as longitudinal and lateral, whereas the method proposed by Akarsu et al. (Akarsu et al., 2016) categorizes damage types into longitudinal, lateral, and alligator cracks. Further, the previous study using deep learning (Zhang et al., 2016; Maeda et al., 2016) only detects the presence or absence of damage. 4 Experimental Setup Based on a previous study in which many neural networks and object detection methods were compared in detail (Huang et al., 2016), among the state-of-theart object detection methods, the SSD using Inception V2 and SSD using MobileNet are those with relatively small CPU loads and low memory consumption, even while maintaining high accuracy. However, it is important to note that the results of the abovementioned research were obtained using the COCO dataset (Lin 6 (a) D00 (b) D01 (c) D10 (d) D11 (e) D20 (f) D40 (g) D43 (h) D44 (i) Class name D00 : Liner crack, longitudinal, wheel mark part D01 : Liner crack, longitudinal, construction joint part D10 : Liner crack, lateral, equal interval D11 : Liner crack, lateral, construction joint part D20 : Alligator crack D40 : Rutting, bump, pothole, separation D43 : White line blur D44 : Cross walk blur Figure 2: Sample images of our dataset: (a) to (h) correspond to each one of the eight categories, and (i) shows the legend. Our benchmark contains 163,664 road images and of these, 9,053 images include cracks. A total of 9,053 images were annotated with class labels and bounding boxes. The images were captured using a smartphone camera in realistic scenarios. 7 Table 1: Road damage types in our dataset and their definitions. Damage Type Longitudinal Linear Crack Lateral Detail Class Name Wheel mark part D00 Crack Construction joint part D01 Equal interval D10 Construction joint part D11 Alligator Crack Partial pavement, overall pavement D20 Rutting, bump, pothole, separation D40 Other Corruption White line blur D43 Cross walk blur D44 Source: Road Maintenance and Repair Guidebook 2013 JRA (2013) in Japan. Note: In reality, rutting, bump, pothole, and separation are different types of road damage, but it was difficult to distinguish these four types using images. Therefore, they were classified as one class, viz., D40. 4.1.2 As in the previous case, we followed the methodology mentioned in the original paper (Liu et al., 2016) as well. Similar to Huang et al. (Huang et al., 2016), we initialize the weights with a truncated normal distribution with a standard deviation of 0.03. The initial learning rate is 0.003 with a learning rate decay of 0.95 every 10,000 iterations. The input image size in this case is 300 × 300 pixels as well. D00 D01 Figure 3: Annotation pipeline. First, we draw the bounding box. Then, the class label is attached. et al., 2014). Because we believe that an object detection method that can be executed on a smartphone (or a small computational resource) is desirable, in this study, we trained the model using the SSD Inception V2 and SSD MobileNet frameworks. We analyze the cases of applying the SSD using Inception and SSD using MobileNet to our dataset in detail. 4.1 SSD using MobileNet 4.2 Training and Evaluation We conducted training and evaluation using our dataset. For our experiment, the dataset was randomly divided in a ratio of 8:2; the former part was set as training data, and the latter as evaluation data. Thus, the training data included 7,240 images, and the evaluation data had 1,813 images. Parameter Settings 5 In the object detection algorithm using deep learning, the parameters learned from the data are enormous; in addition, the number of hyper parameters set by humans is large. The parameter setting in the case of each algorithm is described below. Results In our experiment, training was performed on an PC running the Ubuntu 16.04 operating system with an NVIDIA GRID K520 GPU and 15 GB RAM memory. In the evaluation, the Intersection Over Union (IOU) threshold was set to 0.5. The detected samples are illustrated in Figures 6 and 7. We compared the results provided by the SSD Incep4.1.1 SSD using Inception V2 tion V2 and SSD MobileNet. These results are listed We followed the methodology mentioned in the orig- in Table 2. Although D01 and D44 were detected with inal paper (Liu et al., 2016). The initial learning rate relatively high recall and precision, the value of recall is is 0.002, which is reduced by a learning rate decay of low in the case of D11 and D40; This can be attributed 0.95 per 10,000 iterations. The input image size is 300 to the number of training data (see Figure 4). On the × 300 pixels, which indicates that the original images contrary, D43 was detected with high recall and preciare resized from 600 × 600 to 300 × 300. sion even though the number of training data is small; 8 4000 # OF DAMAGES 3500 3000 2500 2000 D00 : Liner crack, longitudinal, wheel mark part D01 : Liner crack, longitudinal, construction joint part D10 : Liner crack, lateral, equal interval D11 : Liner crack, lateral, construction joint part D20 : Alligator crack D40 : Rutting, bump, pothole, separation D43 : White line blur D44 : Cross walk blur 1500 1000 500 0 D00 D01 D10 D11 D20 D40 D43 D44 TOTAL Ichihara city 175 71 18 9 43 8 20 138 482 Chiba city 183 187 13 12 27 3 104 267 796 Sumida ward 168 660 20 61 21 19 201 482 1632 Nagakute city 482 477 169 58 351 14 90 659 2300 Adachi ward 529 1013 153 279 172 11 191 567 2915 Muroran city 671 574 124 88 1192 189 50 712 3600 Numazu city 560 807 245 129 735 165 161 908 3710 TOTAL 2768 3789 742 636 2541 409 817 3733 15435 Figure 4: Number of damage instances in each class in each municipality. We can see that the distribution of damage type differs for each local government. For example, in Muroran city, there are many D20 damages (1,192 damages) compared to other municipalities. This is because Muroran city is a snowy zone, therefore, alligator cracks tend to occur during the thaw of snow. this is because D43 (blur of the pedestrian crossing) occupies a large proportion in the image and the feature is clear (i.e. stripped pattern). Overall, the SSD MobileNet yields better results. Next, the inference speed of each model is described in Table 3. The speed was tested on a PC with the same specifications as in the previous case and a Nexus 5X smartphone with an MSM8992 CPU and 2 RAM GB memory. In this case, the SSD Inception V2 is two times slower than the SSD MobileNet, which is consistent with the result of Huang et al. (Huang et al., 2016). In addition, because the smartphone processed data in 1.5 s, when it is installed in a moving car, damage to the road surface can be detected in real time and with the same accuracy as in Table 2. Our smartphone application, which we used to detect road damage using the trained model with our dataset (SSD with MobileNet. See Figure 5) is publicly available on our website. 6 with road damage were visually confirmed and classified into eight classes; out of these, 9,053 images were annotated and released as a training dataset. To the best of our knowledge, this dataset is the first one for road damage detection. We strongly believe this dataset provides a new avenue for road damage detection. In addition, we trained and evaluated the damage detection model using our dataset. Based on the results, in the best-detectable category, we achieved recalls and precisions greater than 75% with an inference time of 1.5 s on a smartphone. We believe that a simple road inspection method using only a smartphone will be useful in regions where experts and financial resources are lacking. To support research in this field, we have made the dataset, trained models, source code, and smartphone application publicly available. In the future, we plan to develop methods that can detect rare types of damage that are uncommon in our dataset. Acknowledgement Conclusions In this study, we developed a new large-scale dataset for road damage detection and classification. In collaboration with seven local governments in Japan, we collected 163,664 road images. Then, these images This research was supported by the National Institute of Information and Communication Technology (NICT) under contract research: “Social Big Data Utilization R and D of Basic Technology” (Issue D: Knowledge 9 Table 2: Detection and classification results for each class using the SSD Inception and SSD MobileNet. SIR: SSD Inception V2 Recall, SIP: SSD Inception V2 Precision, SIA: SSD Inception V2 Accuracy, SMR: SSD Recall, SMP: SSD Precision, SMA: SSD Accuracy. Class D00 D01 D10 D11 D20 D40 D43 D44 SIR 0.22 0.60 0.10 0.05 0.68 0.03 0.81 0.62 SIP 0.73 0.84 0.99 0.95 0.73 0.67 0.77 0.81 SIA 0.78 0.80 0.94 0.92 0.85 0.95 0.95 0.83 SMR 0.40 0.89 0.20 0.05 0.68 0.02 0.71 0.85 SMP 0.73 0.64 0.99 0.95 0.68 0.99 0.85 0.66 SMA 0.81 0.77 0.92 0.94 0.83 0.95 0.95 0.81 Table 3: Inference speed (ms) for each model for image resolution of a 300 × 300-pixel image Model Details Inference speed (ms) SSD using Inception V2 (GPU) 63.1 SSD using MobileNet (GPU) 30.6 SSD using MobileNet (smartphone) 1500 of the Site, Citizen’ s Knowledge Organically, Development of Next-generation Citizen Collaborative Platform). Additionally, we would like to express our gratitude to Ichihara city, Chiba city, Sumida ward, Nagakute city, Adachi city, Muroran city, and Numazu city for their cooperation with the experiment. of Civil Engineers, Ser. E1 (Pavement Engineering), 70(3). Dai, J., Li, Y., He, K., and Sun, J. (2016). R-fcn: Object detection via region-based fully convolutional networks. In Advances in neural information processing systems, pages 379–387. Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. (2009). Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages 248–255. IEEE. References AAoSHaT, O. (2008). Bridging the gap–restoring and rebuilding the nations bridges. Washington (DC): American Association of State Highway and Transportation Officials. Adeli, H. (2001). Neural networks in civil engineering: 1989–2000. Computer-Aided Civil and Infrastructure Engineering, 16(2):126–142. Akarsu, B., KARAKÖSE, M., PARLAK, K., Erhan, A., and SARIMADEN, A. (2016). A fast and adaptive road defect detection approach using computer vision with real time implementation. Everingham, M., Eslami, S. A., Van Gool, L., Williams, C. K., Winn, J., and Zisserman, A. (2015). The pascal visual object classes challenge: A retrospective. International journal of computer vision, 111(1):98– 136. Everingham, M., Van Gool, L., Williams, C. K., Winn, J., and Zisserman, A. (2010). The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303–338. Cha, Y.-J., Choi, W., and Büyüköztürk, O. (2017). Deep learning-based crack damage detection using convolutional neural networks. Computer-Aided Civil and Infrastructure Engineering, 32(5):361–378. Chun, P.-j., Hashimoto, K., Kataoka, N., Kuramoto, N., and Ohga, M. (2015). Asphalt pavement crack detection using image processing and naı̈ve bayes based machine learning approach. Journal of Japan Society 10 Felzenszwalb, P. F., Girshick, R. B., McAllester, D., and Ramanan, D. (2010). Object detection with discriminatively trained part-based models. IEEE transactions on pattern analysis and machine intelligence, 32(9):1627–1645. Geiger, A., Lenz, P., Stiller, C., and Urtasun, R. (2013). Vision meets robotics: The kitti dataset. The Inter- Huval, B., Wang, T., Tandon, S., Kiske, J., Song, W., Pazhayampallil, J., Andriluka, M., Rajpurkar, P., Migimatsu, T., Cheng-Yue, R., et al. (2015). An empirical evaluation of deep learning on highway driving. arXiv preprint arXiv:1504.01716. D20 START DETECTION Ioffe, S. and Szegedy, C. (2015). Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning, pages 448–456. STOP DETECTION Jo, Y. and Ryu, S. (2015). Pothole detection system using a black-box camera. Sensors, 15(11):29316– 29331. Figure 5: Operating screen of our smartphone application. It is supposed to be mounted on the dashboard of a general passenger car (See Figure 1). Detection of road surface damage is initiated by pressing the ”START DETECTION” button. An image of the damaged part and the position information are transmitted to the external server only when damage is found. Using the SSD with MobileNet, this application can detect road damages (of eight types) in 1.5 s with the same accuracy as shown in Table 2. JRA (2013). Maintenance and repair guide book of the pavement 2013. Japan Road Association, 1st. edition. Kazuya, T., Akira, K., Shun, F., and Takeki, I. (2013). An effective surface inspection method of urban roads according to the pavement management situation of local governments. Japan Scienc and Technology Information Aggregator. national Journal of Robotics Research, 32(11):1231– KOKUSAI KOGYO CO., L. (2016). Mms(mobile mea1237. surement system). Girshick, R. (2015). Fast r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., and Zitnick, C. L. (2014). pages 1440–1448. Microsoft coco: Common objects in context. In EuGirshick, R., Donahue, J., Darrell, T., and Malik, J. ropean conference on computer vision, pages 740– (2014). Rich feature hierarchies for accurate object 755. Springer. detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pat- Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, tern recognition, pages 580–587. S., Fu, C.-Y., and Berg, A. C. (2016). Ssd: Single shot multibox detector. In European conference on He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep computer vision, pages 21–37. Springer. residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and Maeda, H., Sekimoto, Y., and Seto, T. (2016). pattern recognition, pages 770–778. Lightweight road manager: smartphone-based autoHoward, A. G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., and Adam, H. (2017). Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861. matic determination of road damage status by deep neural network. In Proceedings of the 5th ACM SIGSPATIAL International Workshop on Mobile Geographic Information Systems, pages 37–45. ACM. MLIT (2016). Present state and future of social capital Huang, J., Rathod, V., Sun, C., Zhu, M., Korattikara, A., aging. infrastructure maintenance information. Fathi, A., Fischer, I., Wojna, Z., Song, Y., Guadarrama, S., et al. (2016). Speed/accuracy trade-offs Mohan, A. and Poobal, S. (2017). Crack detection usfor modern convolutional object detectors. arXiv ing image processing: A critical review and analysis. preprint arXiv:1611.10012. Alexandria Engineering Journal. 11 Nishikawa, T., Yoshida, J., Sugiyama, T., and Fujino, Y. (2012). Concrete crack detection by multiple sequential image filtering. Computer-Aided Civil and Infrastructure Engineering, 27(1):29–47. Computer-Aided Civil and Infrastructure Engineering, 29(5):342–358. Zhang, A., Wang, K. C., Li, B., Yang, E., Dai, X., Peng, Y., Fei, Y., Liu, Y., Li, J. Q., and Chen, C. (2017). Automated pixel-level pavement crack detection on 3d asphalt surfaces using a deep-learning network. Computer-Aided Civil and Infrastructure Engineering, 32(10):805–819. O’Byrne, M., Ghosh, B., Schoefs, F., and Pakrashi, V. (2014). Regionally enhanced multiphase segmentation technique for damaged surfaces. Computer-Aided Civil and Infrastructure Engineering, 29(9):644–658. Redmon, J., Divvala, S., Girshick, R., and Farhadi, A. (2016). You only look once: Unified, real-time object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 779–788. Redmon, J. and Farhadi, A. (2016). Yolo9000: better, faster, stronger. arXiv preprint arXiv:1612.08242. Ren, S., He, K., Girshick, R., and Sun, J. (2015). Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pages 91–99. Sermanet, P., Eigen, D., Zhang, X., Mathieu, M., Fergus, R., and LeCun, Y. (2013). Overfeat: Integrated recognition, localization and detection using convolutional networks. arXiv preprint arXiv:1312.6229. Simonyan, K. and Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556. Szegedy, C., Ioffe, S., Vanhoucke, V., and Alemi, A. A. (2017). Inception-v4, inception-resnet and the impact of residual connections on learning. In AAAI, pages 4278–4284. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., and Wojna, Z. (2016). Rethinking the inception architecture for computer vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2818–2826. Yu, X. and Salari, E. (2011). Pavement pothole detection and severity measurement using laser imaging. In Electro/Information Technology (EIT), 2011 IEEE International Conference on, pages 1–5. IEEE. Zalama, E., Gómez-Garcı́a-Bermejo, J., Medina, R., and Llamas, J. (2014). Road crack detection using visual features extracted by gabor filters. 12 Zhang, L., Yang, F., Zhang, Y. D., and Zhu, Y. J. (2016). Road crack detection using deep convolutional neural network. In Image Processing (ICIP), 2016 IEEE International Conference on, pages 3708–3712. IEEE. (a) D20, D44 (b) D43 (c) D01 (d) D00, D01 (e) D01, D44 (f) D00, D44 (g) D20 (h) D43, D44 (i) D00, D44 (j) D01 (k) D00 (i) Class name D00 : Liner crack, longitudinal, wheel mark part D01 : Liner crack, longitudinal, construction joint part D10 : Liner crack, lateral, equal interval D11 : Liner crack, lateral, construction joint part D20 : Alligator crack D40 : Rutting, bump, pothole, separation D43 : White line blur D44 : Cross walk blur Figure 6: Detected samples using the SSD MobileNet. 13 (a) D44 (b) D01 (c) D01, D44 (d) D00, D44 (e) D00 (f) D00, D44 (g) D43 (h) D43 (i) D20 (j) D20 (k) D00 (i) Class name D00 : Liner crack, longitudinal, wheel mark part D01 : Liner crack, longitudinal, construction joint part D10 : Liner crack, lateral, equal interval D11 : Liner crack, lateral, construction joint part D20 : Alligator crack D40 : Rutting, bump, pothole, separation D43 : White line blur D44 : Cross walk blur Figure 7: Detected samples using the SSD Inception V2. 14
1cs.CV
arXiv:1603.02521v3 [math.RT] 20 Dec 2017 TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS JIARUI FEI Abstract. For each valued quiver Q of Dynkin type, we construct a valued ice quiver ∆2Q . Let G be a simply connected Lie group with Dynkin diagram the underlying valued graph of Q. The upper cluster algebra of ∆2Q is graded by the triple dominant weights (µ, ν, λ) of G. We prove that when G is simply-laced, the dimension of each graded component counts the tensor multiplicity cλ µ,ν . We conjecture that this is also true if G is not simply-laced, and sketch a possible approach. Using this construction, we improve BerensteinZelevinsky’s model, or in some sense generalize Knutson-Tao’s hive model in type A. Contents Introduction 2 Part 1. Construction of iARt QPs 1. Graded Upper Cluster Algebras 2. AR-theory of Presentations 3. iARt Quivers 4. Cluster Character from Quivers with Potentials 5. iARt QPs Appendices A. A List of some iARt Quivers 6 6 10 15 19 21 26 26 Part 2. Isomorphism to k[Conf 2,1 ] 6. Cluster Structure of Maximal Unipotent Groups 7. Maps Relating Unipotent Groups 8. Cluster Structure on k[Conf 2,1 ] 9. Epilogue Appendices B. The Twisted Cyclic Shift via Mutations Acknowledgement References 28 28 32 37 38 40 40 46 46 2010 Mathematics Subject Classification. Primary 13F60, 16G20; Secondary 13A50, 52B20. Key words and phrases. Graded Upper Cluster Algebra, iARt Quiver, Quiver Representation, Quiver with Potential, Mutation, Cluster Character, Configuration Space, Tensor Multiplicity, Littlewood-Richardson Coefficient, g-vector Cone, Lattice Point. 1 2 JIARUI FEI Introduction Finding the polyhedral model for the tensor multiplicities in Lie theory is a longstanding problem. By tensor multiplicities we mean the multiplicities of irreducible summands in the tensor product of any two finite-dimensional irreducible representations of a simply connected Lie group G. The problem asks to express the multiplicity as the number of lattice points in some convex polytope. Accumulating from the works of Gelfand, Berenstein and Zelevinsky since 1970’s, a first quite satisfying model for G of type A was invented in [5]. Finally around 1999, building upon their work, Knutson and Tao invented their hive model, which led to the solution of the saturation conjecture [34]. In fact, the reduction of Horn’s problem to the Saturation conjecture is an important driving force for the evolution of the models. Outside type A, up to now Berenstein and Zelevinsky’s models [6] are still the only known polyhedral models. Those models lose a few nice features of KnutsonTao’s hive model. We will have a short discussion on this in Section 0.1. Despite of a lot of effort to improve the Berenstein-Zelevinsky model, to author’s best knowledge there is no very satisfying further results in this direction. Recently an interesting link between the hive model and the cluster algebra theory was established in [14] through the Derksen-Weyman-Zelevinsky’s quiver with potential model [12, 13] for cluster algebras. A similar but different link between the polyhedral models and tropical geometry was established by Goncharov and Shen in [28]. In fact, from the work of Berenstein, Fomin and Zelevinsky [6, 7], those links may not be a big surprise. There are two goals in the current paper. First we want to generalize the work [14] to other types. More specifically, we hope to prove that the algebras of regular functions on certain configuration spaces are all upper cluster algebras. Second we want to improve the Berenstein-Zelevinsky’s model in the spirit of Knutson-Tao. In fact, as we shall see, we accomplish these two goals almost simultaneously. Namely, we use our conjectural models to establish the cluster algebra structures. Once the cluster structures are established, the conjectural models are proved as well. The key to making conjectural models is the construction of the iARt quivers. Let Q be a valued quiver of Dynkin type. Let C 2 Q be the category of projective presentations of Q. We can associate this category an Auslander-Reiten quiver ∆(C 2 Q) with translation (ARt quiver in short). The ice ARt quiver (iARt quiver in short) ∆2Q is obtained from ∆(C 2 Q) by freezing three sets of vertices, which correspond to the negative, positive, and neutral presentations in C 2 Q. We can put a (quite canonical) potential WQ2 on the iARt quiver ∆2Q . A quiver with potential (or QP in short) (∆, W ) is related to Berenstein-FominZelevinsky’s upper cluster algebras [7] through cluster characters evaluating on µsupported g-vectors introduced in [14]. The cluster character CW considered in this paper is the generic one [41], but it can be replaced by fancier ones. As we have seen in many different situations [14, 15, 16] that the set G(∆, W ) of µ-supported g-vectors is given by lattice points in some rational polyhedral cone. This is also the case for the iARt QPs (∆2Q , WQ2 ). The whole Part 1 is devoted to the construction of the iARt QP (∆2Q , WQ2 ) and the polyhedral cone G∆2Q . It turns out that the cone G∆2Q has a very neat 2 hyperplane presentation {x ∈ R(∆Q )0 | xH ≥ 0}, where the columns of the matrix TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 3 H are given by the dimension vectors of subrepresentations of 3|Q0 | representations of ∆2Q . These 3|Q0 | representations are in bijection with the frozen vertices of ∆2Q . They also have a very simple and nice description (see Theorem 5.3). The main result of Part 1 is the following. 2 Theorem 0.1 (Theorem 5.8). The set G∆2Q ∩ Z(∆Q )0 is exactly G(∆2Q , WQ2 ). The upper cluster algebra C(∆2Q ) has a natural grading by the weight vectors of presentations. This grading can be extended to a triple-weight grading σ 2Q : 2 3|Q | Z(∆Q )0 → Z>0 0 . This grading slices the cone G∆2Q into polytopes n o G∆2Q (µ, ν, λ) := g ∈ G∆2Q | σ 2Q (g) = (µ, ν, λ) . Let G := GQ be the simply connected simple Lie group with Dynkin diagram the underlying valued graph of Q. Our conjectural model is that the lattice points in G∆2Q (µ, ν, λ) counts the tensor multiplicity cλµ ν for G. Here, cλµ ν is the multiplicity of the irreducible representation L(λ) of highest weight λ in the tensor product L(µ) ⊗ L(ν). More often than not we identify a dominant weight by a non-negative integral vector. To prove this model, we follow a similar line as [14]. However, we do not have a quiver setting to work with in general. We replace the semi-invariant rings of triple-flag quiver representations by the ring of regular functions on certain configuration space introduced in [18]. Fix an opposite pair of maximal unipotent subgroups (U − , U ) of G. The quotient space A := U − \G is called base affine space, and the quotient space A∨ := G/U is called its dual. The configuration space Conf 2,1 is by definition (A × A × A∨ )/G, where G acts multi-diagonally. The ring of regular functions k[Conf 2,1 ] is just the − − invariant ring (k[G]U ⊗ k[G]U ⊗ k[G]U )G . The ring k[Conf 2,1 ] is multigraded by λ a triple of weights (µ, ν, λ). Each graded component Cµ,ν := k[Conf 2,1 ]µ,ν,λ is given ∨ G λ by the G-invariant space (L(µ) ⊗ L(ν) ⊗ L(λ) ) . So the dimension of Cµ,ν counts λ the tensor multiplicity cµ ν . Here is the main result of Part 2. Theorem 0.2 (Theorem 8.1). Suppose that Q is trivially valued. Then the ring 2 of regular functions on Conf 2,1 is the graded upper cluster algebra C(∆2Q , SQ ; σ 2Q ). Moreover, the generic character maps the lattice points in G∆2Q onto a basis of this algebra. In particular, cλµ ν is counted by lattice points in G∆2Q (µ, ν, λ). We will show by an example that the upper cluster algebra strictly contains the corresponding cluster algebra in general. We conjecture that the trivially valued assumption can be dropped in the above theorem and the theorem below. It is pointed in the end that the only missing ingredient for proving the conjecture is the analogue of [13, Lemma 5.2] for species with potentials [37]. Fock and Goncharov studied in [18] the similar spaces Conf 3 1 as cluster varieties. However, to author’s best knowledge it is not clear from their discussion what an initial seed is if G is not of type A. Moreover the equality established in the theorem does not seem to fellow from any result there. In fact, Fock and Goncharov later conjectured in [19] that the tropical points in their cluster X -varieties parametrize bases in the corresponding (upper) cluster algebras. Our result can be viewed as 1They considered the generic part of the quotient stack [(A∨ )3 /G]. We will work with the categorical quotient as its partial compactification. 4 JIARUI FEI an algebraic analog of their conjecture for the space Conf 2,1 . Instead of working with the tropical points, we work with the g-vectors. To sketch our ideas, we first observe that if we forget the frozen vertices corresponding to the positive and neutral presentations, then we get a valued ice quiver denoted by ∆Q whose cluster algebra is isomorphic to the coordinate ring k[U ]. Roughly speaking, this procedure corresponds to an open embedding i : H × H × U ֒→ Conf 2,1 , or more precisely Corollary 7.9. We will define the cluster 2 SQ in Theorem 0.2 through the pullback map i∗ . It is then not hard to show that 2 k[Conf 2,1 ] contains the upper cluster algebra C(∆2Q , SQ ; σ 2Q ) as a graded subalgebra. The detail will be given in Section 7.1. So far we have the graded inclusions   2 ; σ 2Q ) ⊆ k[Conf 2,1 ]. Span CW (G∆2Q ) ⊆ C(∆2Q , SQ   To finish the proof, it suffices to show the containment k[Conf 2,1 ] ⊆ Span CW (G∆2Q ) . For this, we come back to the cluster structure of k[U ]. It turns out that the analog of Theorem 0.2 for U is rather easy to prove. The set G(∆Q , WQ ) contains exactly lattice points in the polytope G∆Q , which is defined by one of the three sets of relations of G∆2Q . On the other hand, we have two other embeddings il , ir : U ֒→ Conf 2,1 . They are the map iu := i|U followed by the twisted cyclic shift of Conf 2,1 . Another crucial ingredient in this paper is an interpretation of the twisted cyclic shift in terms of a sequence of mutations µl . Applying µl and µ−1 to l the QP (∆Q , WQ ), we get two other QPs (∆lQ , WQl ) and (∆rQ , WQr ). The analogous polytopes G∆lQ and G∆rQ for them are defined by the other two sets of relations of G∆2Q . Finally, after showing the good behavior of g-vectors under the pullback of the three embeddings, the required inclusion will follow from the fact that  2 (0.1) k[Conf 2,1 ] ⊆ s ∈ L(SQ ) | i∗# (s) ∈ k[U ] for # = u, l, r , 2 2 where L(SQ ) is the Laurent polynomial ring in the cluster SQ . The detail will be given in Section 7.2. Except for these two main results, we have a side result for the base affine spaces. The author would like to thank B. Leclerc and M. Yakimov for confirming that the following theorem was an open problem. It turns out that the cluster structure of A lies between that of U and Conf 2,1 . Let ∆♯Q be the valued ice quiver obtained from ∆2Q by deleting frozen vertices corresponding to neutral presentations. Theorem 0.3 (Theorem 9.2). Suppose that Q is trivially valued. Then the ring of regular functions on A is the graded upper cluster algebra C(∆♯Q , M♯Q ; ̟(σ ♯Q )). Moreover, the generic character maps the lattice points in G∆♯ onto a basis of Q C(∆♯Q ). In particular, the weight multiplicity dim L(µ)λ is counted by lattice points in G∆♯ (µ, λ). Q 0.1. The Models. In [34] Knutson and Tao invented a remarkable polyhedral model called hives or honeycomb. The author personally thinks that it has at least three advantages over Berenstein-Zelevinsky’s model [6]. First, the hive polytopes  have a nice presentation x ∈ R3n | xH ≥ 0, xσ = (µ, ν, λ) . Second, the cyclic symmetry of the type-A tensor multiplicity is lucid from the hive model. Actually TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 5 other symmetries can also follow from the hive model. Last and most importantly, there is an operation called overlaying for honeycombs [34]. In appropriate sense, our models share these nice properties. The first one is clear from our result. Our H-matrices even have all non-negative entries. However, if readers prefer the rhombus-type inequalities of the hives, one can transform our model through a totally unimodular map as in [14]. However, the rhombus-type inequalities are not always as neat as the ones in type A. We will discuss the transformation and the analogous overlaying elsewhere. Although the cyclic symmetry is not immediately clear from H itself, we understand from our construction and proofs that it is just hidden there. We believe that this is probably the best we can do outside type A. In a more general context of Kac-Moody algebras, the tensor multiplicity problem can be solved by P. Littelmann’s path model [38]. As pointed out in [6], his model can be transformed into polyhedral ones (with some non-trivial work). However, in general it involves a union of several convex polytopes. 0.2. Relation to the work of Berenstein-Zelevinsky and Goncharov-Shen. In a groundbreaking work [6] Berenstein and Zelevinsky invented their polyhedral model for all Dynkin types. Their main tools are Lusztig’s canonical basis and tropical relations in double Bruhat cells. The polytopes are defined explicitly in terms of their i-trails. But the author feels that i-trails are hard to compute especially in type E. By contrast, the subrepresentations defining our H are rather easy to list in most cases. In few difficult cases, such as type E7 and E8 , we provide an algorithm suitable for computers. Recently Goncharov and Shen made some further progress in [28]. Using tropical geometry and geometric Satake, they proved a more symmetric polyhedral model (see [28, Theorem 2.6 and (214)]). However, there is no further explicit description on the polytopes. The equality of (0.1) as an intermediate byproduct of our proof is similar to this result. Loosely speaking, our work is independent of their results, though the author did benefit a lot from reading their papers. The construction of iARt quivers ∆2Q is new. We believe that the construction and results, especially the ideas behind, are beyond just solving the tensor multiplicity problem for simple Lie groups. The proofs in Part 1 are similar to those in [14]. In Part 2 what we heavily rely on is the cluster structure of k[U ] and a mutation interpretation of the twisted cyclic shift. Throughout the quiver with potential model for cluster algebras is most important. Outline of the Paper. In Section 1.1 we recall the basics on valued quivers and their representations. We define the graded upper cluster algebra attached to a valued quiver in Section 1.2 and 1.3. In Section 2.1 we recall the Auslander-Reiten theory from a functorial point of view. We specialize the theory to the category of presentations mostly for hereditary algebras in Section 2.2. In Section 3.1 we define the iARt quivers in general. We then consider the hereditary cases in more detail in Section 3.2. Proposition 3.6 compares the ARt quivers of presentations with the more familiar ARt quivers of representations. In Section 4 we review the generic cluster character in the setting of quivers with potentials. In Section 5 we study the iARt QPs and their µ-supported g-vectors. We prove the two main results of Part 1 – Theorem 5.3 and 5.8. In Appendix A, we provide more examples of iARt quivers. 6 JIARUI FEI In Section 6 we review the rings of regular functions on base affine spaces and maximal unipotent groups, especially the cluster structure of the latter (Theorem 6.5 and Proposition 6.6). In Section 7 we study maps relating the configuration spaces to the corresponding unipotent groups. These are almost all the technical work required for proving the main result. In Section 8 we prove our main result – Theorem 8.1. In Section 9 we prove the side result – Theorem 9.2. In the end we make some remark on the possible generalization to the non-simply laced cases. In Appendix B we prove the mutation interpretation of the twisted cyclic shift in Theorem B.14. As a consequence, we produce an algorithm for computing the (µ-supported) g-vector cones. Notations and Conventions. Our vectors are exclusively row vectors. All modules are right modules. Arrows are composed from left to right, i.e., ab is the path a b ·− →·− → ·. Unless otherwise stated, unadorned Hom and ⊗ are all over the base field k, and the superscript ∗ is the trivial dual for vector spaces. For direct sum of n copies of M , we write nM instead of the traditional M ⊕n . Part 1. Construction of iARt QPs 1. Graded Upper Cluster Algebras 1.1. Valued Quivers and their Representations. If you are familiar with the usual quiver representations and only care about our results on the simply laced cases, you can skip this subsection. Readers may find more detailed treatment on valued quivers in [9]. Definition 1.1. A valued quiver is a triple Q = (Q0 , Q1 , C) where (1) Q0 is a set of vertices, usually labelled by natural numbers 1, 2, . . . , n; (2) Q1 is a set of arrows, which is a subset of Q0 × Q0 ; (3) C = {(ci,j , cj,i ) ∈ N × N | (i, j) ∈ Q1 } is called the valuation of Q. It is called symmetrizable if there is d = {di ∈ N | i ∈ Q0 } such that di ci,j = cj,i dj for every (i, j) ∈ Q1 . For such a valued quiver, the pair (Q0 , Q1 ) is called its ordinary quiver. Throughout this paper, all valued quivers are assumed to have no loops or oriented 2-cycles in their ordinary quivers. If ci,j = cj,i for every (ci,j , cj,i ) ∈ C, then Q is called equally valued. To draw a valued quiver (Q0 , Q1 , C), we first draw its ordinary quiver, then (ci,j ,cj,i ) put valuations above its arrows, eg., i −−−−−−→ j. We will omit the valuation if (i, j) is trivially valued, i.e., ci,j = cj,i = 1. All valued quivers in this paper will be symmetrizable. We always fix a choice of d, so readers may view d as a part of the defining data for Q. We let di,j = gcd(di , dj ). Let F be a finite field. We write F for an algebraic closure of F. For each positive integer k denote by Fk the degree k extension of F in F. Note that the largest subfield of F contained in both Fk and Fl is Fgcd(k,l) = Fk ∩ Fl . If k | l we can fix a basis of Fl over Fk and thus freely identify Fl as a vector space over Fk . A representation M of Q is an assignment for each i ∈ Q0 a Fdi -vector space M (i), and for each arrow (i, j) ∈ Q1 an Fdi,j -linear map M (i, j).2 The dimension 2This definition is slightly different from the classical reference [9], but it is more adapted to the cluster algebra theory. TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 7 vector dimM is the integer vector (dimFdi M (i))i∈Q0 . Similar to the usual quiver representations, we can define a morphism φ : M → N as the set n o such that φj M (i, j) = N (i, j)φi for all (i, j) ∈ Q1 . φi ∈ HomFdi (M (i), N (i)) i∈Q0 The category Rep(Q) of all (finite-dimensional) representations of Q is an abelian category, in which the kernels and cokernels are taken vertex-wise. The category Rep(Q) is also Krull-Schmidt, that is every representation can be written uniquely as a direct sum of indecomposable representations. Just as with usual quivers it is useful to consider an equivalent category of modules over the path algebra. Such an analog Lfor valued quivers is the notion of Q F-species. Define Γ0 = i∈Q0 Fdi and Γ1 = (i,j)∈Q1 Fdi ci,j . Notice that Fdi ci,j contains both Fdi and Fdj and thus we have a Γ0 -Γ0 -bimodule structure on Γ1 . Now we define the F-species ΓQ to be the tensor algebra TΓ0 (Γ1 ) of Γ1 over Γ0 . If ΓQ is finite-dimensional, then it is clear that the indecomposable projective (resp. injective) modules are precisely Pi = ei ΓQ (resp. Ii = (ΓQ ei )∗ ) for i ∈ Q0 , where ei is the identity element in Fdi . The category Rep(Q) has enough projective and injective objects. The top of Pi is the simple representation Si supported on the vertex i, which is also the socle of Ii . The algebra ΓQ is hereditary, that is, it has global dimension 1. So for M, N ∈ Rep(Q), hM, N i = dimF HomQ (M, N ) − dimF Ext1Q (M, N ) is a bilinear form only depending on the dimension vectors of M and N . This is called “Ringel-Euler” form, and we denote the matrix of this form by E(Q). We also define the matrix El (Q) := (elj,i ) and Er (Q) := (eri,j ) by ( ( 1 i = j; 1 i = j; eli,j = eri,j = −cj,i (i, j) ∈ Q1 , −ci,j (i, j) ∈ Q1 . These matrices are related by E(Q) = El (Q)D = DEr (Q), where D is the diagonal matrix with diagonal entries di,i = di . 3,1 Example 1.2 (G2 ). Consider the valued quiver 1 −−→ 2 of type G2 with d = (1, 3). Its module category has six indecomposable objects • • • • The The The The simple injective S1 : F → 0, and its projective cover P1 : F ֒→ F3 ; simple projective S2 : 0 → F3 , and its injective hull I2 : F3 ֒→ F3 ; module M1 : F2 ֒→ F3 , which is presented by P2 ֒→ 2P1 ; module M2 : F3 ֒→ F23 , which is presented by P2 ֒→ 3P1 . In this paper, we will encounter two kinds of valued quivers. One is valued quivers Q of Dynkin type, and the other is bigger valued quivers ∆Q and ∆2Q constructed from Q (see Section 3). We will define upper cluster algebras attached to the latter. 1.2. Upper Cluster Algebras. We mostly follow [7, 23, 20]. To define the upper cluster algebra, we need to introduce the notion of the quiver mutation. The mutation of valued quivers is defined through Fomin-Zelevinsky’s mutation of the associated skew-symmetrizable matrix. 8 JIARUI FEI Every symmetrizable valued quiver ∆ corresponds to a skew symmetrizable integer matrix B(∆) := −El (∆) + Er (∆)T . So the entries (bu,v )u,v∈∆0 are given by   if (u, v) ∈ ∆1 cu,v , bu,v = −cu,v , if (v, u) ∈ ∆1   0 otherwise. The matrix B(∆) is skew symmetrizable because DB is skew-symmetric for the diagonal matrix D. Conversely, given a skew symmetrizable matrix B, a unique valued quiver ∆ can be easily defined such that B(∆) = B. Definition 1.3. The mutation of a skew symmetrizable matrix B on the direction u ∈ ∆0 is given by µu (B) = (b′v,w ), where ( −bv,w , if u ∈ {v, w}, b′v,w = bv,w + sign(bv,u ) max(0, bv,u bu,w ), otherwise. We denote the induced operation on its valued quiver also by µu . The cluster algebras that we will consider in this paper are skew-symmetrizable cluster algebras with coefficients. The combinatorial data defining such a cluster algebra is encoded in a symmetrizable valued quiver ∆ with frozen vertices. Frozen vertices are forbidden to be mutated, and the remaining vertices are mutable. Such a valued quiver is called valued ice quiver (or VIQ in short). The mutable part ∆µ is the full subquiver of ∆ consisting of mutable vertices. In general, to define an (upper) cluster algebra only ∆µ is required to be symmetrizable. However, in this paper all VIQs happen to be “globally” symmetrizable. We usually label the mutable vertices as the first p out of q vertices of ∆. The restricted B-matrix B∆ of ∆ is the first p rows of B(∆). Let k be a field, not necessarily related in any sense to the finite field F or the base field in the rest of Part 1. Definition 1.4. Let F be a field containing k. A seed in F is a pair (∆, x) consisting of a VIQ ∆ as above together with a collection x = {x1 , x2 , . . . , xq }, called an extended cluster, consisting of algebraically independent (over k) elements of F , one for each vertex of ∆. The elements of x associated with the mutable vertices are called cluster variables; they form a cluster. The elements associated with the frozen vertices are called frozen variables, or coefficient variables. A seed mutation µu at a (mutable) vertex u transforms (∆, x) into the seed (∆′ , x′ ) = µu (∆, x) defined as follows. The new VIQ is ∆′ = µu (∆). The new extended cluster is x′ = x ∪ {x′u } \ {xu } where the new cluster variable x′u replacing xu is determined by the exchange relation Y Y xcwu,w . xcvv,u + (1.1) xu x′u = (v,u)∈∆1 (u,w)∈∆1 We note that the mutated seed (∆′ , x′ ) contains the same coefficient variables as the original seed (∆, x). It is easy to check that one can recover (∆, x) from (∆′ , x′ ) by performing a seed mutation again at u. Two seeds (∆, x) and (∆′ , x′ ) that can be obtained from each other by a sequence of mutations are called mutation-equivalent, denoted by (∆, x) ∼ (∆′ , x′ ). TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 9 Definition 1.5. The cluster algebra C(∆, x) associated to a seed (∆, x) is defined as the subring of F generated by all elements of all extended clusters of the seeds mutation-equivalent to (∆, x). Note that the above construction of C(∆, x) depends only, up to a natural isomorphism, on the mutation equivalence class of the initial VIQ ∆. In fact, it only depends on the mutation equivalence class of the restricted B-matrix of ∆. So we may drop x and simply write C(∆) or C(B∆ ). An amazing property of cluster algebras is the Laurent Phenomenon. Theorem 1.6 ([22, 7]). Any element of a cluster algebra C(∆, x) can be expressed in terms of the extended cluster x as a Laurent polynomial, which is polynomial in coefficient variables. Since C(∆, x) is generated by cluster variables from the seeds mutation equivalent to (∆, x), Theorem 1.6 can be rephrased as \ Lx′ , C(∆, x) ⊆ (∆′ ,x′ )∼(∆,x) ±1 k[x±1 1 , . . . , xp , xp+1 , . . . xq ]. where Lx := Note that our definition of Lx is slightly different from the original one in [7], where Lx is replaced by the Laurent polynomial ±1 ±1 ±1 L(x) := k[x±1 1 , . . . , xp , xp+1 , . . . xq ]. Definition 1.7. The upper cluster algebra with seed (∆, x) is \ Lx′ . C(∆, x) := (∆′ ,x′ )∼(∆,x) Any (upper) cluster algebra, being a subring of a field, is an integral domain (and under our conventions, a k-algebra). Conversely, given such a domain R, one may be interested in identifying R as an (upper) cluster algebra. The following useful lemma is a specialization of [20, Proposition 3.6] to the case when R is a unique factorization domain. Lemma 1.8. Let R be a finitely generated UFD over k. Suppose that (∆, x) is a seed contained in R, and each adjacent cluster variable x′u is also in R. Moreover, each pair in x and each pair (xu , x′u ) are relatively prime. Then R ⊇ C(∆, x). 1.3. g-vectors and Gradings. Let x = {x1 , x2 , . . . , xq } be an (extended) cluster. g(1) g(2) g(q) For a vector g ∈ Zq , we write xg for the monomial x1 x2 · · · xq . For u = 1, 2, . . . , p, we set yu = x−bu where bu is the u-th row of the matrix B∆ , and let y = {y1 , y2 , . . . , yp }. Suppose that an element z ∈ L(x) can be written as (1.2) z = xg(z) F (y1 , y2 , . . . , yp ), where F is a rational polynomial not divisible by any yi , and g(z) ∈ Zq . If we assume that the matrix B∆ has full rank, then the elements y1 , y2 , . . . , yp are algebraically independent so that the vector g(z) is uniquely determined [23]. We call the vector g(z) the (extended) g-vector of z with respect to the pair (∆, x). Definition implies at once that for two such elements z1 , z2 we have that g(z1 z2 ) = g(z1 ) + g(z2 ). So the set of all g-vectors in any subalgebra of L(x) forms a sub-semigroup of Zq . Lemma 1.9 ([14, Lemma 5.5], cf. [41]). If the matrix B∆ has full rank, then any subset of L(x) with distinct well-defined g-vectors is linearly independent over k. 10 JIARUI FEI Definition 1.10. A weight configuration σ of a lattice L ⊆ Rm on a VIQ ∆ is an assignment for each vertex v of ∆ a weight vector σ(v) ∈ L such that for each mutable vertex u, we have that X X cu,w σ(w). cv,u σ(v) = (1.3) (v,u)∈∆1 (u,w)∈∆1 The mutation µu also transforms σ into a weight configuration σ ′ on the mutated quiver µu (∆) defined as  X  cu,w σ(w) − σ(u) if v = u,  σ ′ (v) = (u,w)∈∆1  σ(v) otherwise. By slight abuse of notation, we can view σ as a matrix whose v-th row is the weight vector σ(v). In this matrix notation, the condition (1.3) is equivalent to that B∆ σ is a zero matrix. So we call the cokernel of B∆ as the grading space of C(∆). A weight configuration σ is called full if the corank of B∆ is equal to the rank of σ. It is easy to see that for any weight configuration of ∆, the mutation can be iterated. Given a weight configuration (∆; σ), we can assign a multidegree (or weight) to the upper cluster algebra C(∆, x) by setting deg(xv ) = σ(v) for v = 1, 2, . . . , q. Then mutation preserves multihomogeneousity. We say that this upper cluster algebra is σ-graded, and denoted by C(∆, x; σ). We refer to (∆, x; σ) as a graded seed. Note that the variables in y have zero degrees. So if z has a well-defined g-vector as in (1.2), then z is homogeneous of degree gσ. 2. AR-theory of Presentations 2.1. Review of categories of functors and AR-theory. We briefly review Auslander-Reiten theory emphasizing a functorial point of view following [3, IV.6]. The theory were developed originally for module categories of Artin algebras, but without much difficulty almost all can be generalized to k-linear, Hom-finite, and Krull-Schmidt categories. Let k be a field, and A be a k-linear, Hom-finite, and Krull-Schmidt category. For any two objects M, N ∈ A, let radA (M, N ) be the space of all non-invertible morphisms from M to N . We define rad2A (M, N ) to consist of all morphisms of form gf , where f ∈ radA (M, L) and g ∈ radA (L, N ) for some L ∈ A. We denote IrrA (M, N ) := radA (M, N )/ rad2A (M, N ). If M is indecomposable, then EndA (M ) is local, and thus EM := EndA (M )/ rad EndA (M ) is a division k-algebra. Let ind(A) be the full subcategory of all indecomposable objects in A. Definition 2.1. For M, N ∈ ind(A), an irreducible morphism f : M → N is an element in IrrA (M, N ). Lt Let M = i=1 mi Mi be an object in A with Mi indecomposable and pairwise non-isomorphic. For f ∈ HomA (L, M ) with L indecomposable, we can write f as f = (f1 , . . . , ft ) where fi = (fi,1 , . . . , fi,mi ) : L → mi Mi . Then f is called left minimal almost split if the residual classes of fi,j ’s in IrrA (L, Mi ) form a EL -basis for all i, and if there is an indecomposable M ′ ∈ A such that IrrA (L, M ′ ) 6= 0, TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 11 then M ′ ∼ = Mi for some i. Similarly we can define right minimal almost split for g : M → N. An exact sequence 0 → L → M → N → 0 in A is called almost split if L → M is left minimal almost split, and M → N is right minimal almost split. In this case, L is called the translation of N , denoted by τ N . The definition for left (right) minimal almost split is not the original one (see [3, Definition IV.1.1]), but it follows from [3, Proposition IV.4.2]. We also recall a basic fact [2] that (2.1) op Irr (L, M ). dimEMi IrrA (Mi , N ) = dimEM A i i Let A be a finite dimensional k-algebra, and mod A be the category of finite dimensional (right) A-modules. Proposition 2.2 ([2, Theorem V.1.15]). The translation τ is defined for every indecomposable non-projective A-module. It is given by the trivial dual of Auslander’s transpose functor (see [2, IV.1]). We denote by Aop the opposite category of A, and by mod Aop the category of k-linear, additive, contravariant functors from A to mod k. For an object M of A, the functor HomA (−, M ) is an object of mod Aop . From the Yoneda embedding Hommod Aop (HomA (−, M ), F ) ∼ = F (M ) ∀F ∈ mod Aop , we can conclude [3, Corollary IV.6.4] that HomA (−, M ) is a projective object in mod Aop . In fact, every (finitely generated) projective object is of this form ([3, Lemma IV.6.5]). We can ask readers to formulate the dual statements for the injective object HomA (M, −)∗ . For each object M , we can associate a unique simple functor SM characterized by SM (M ) = k and SM (N ) = 0 for all N ≇ M . Let us come back to the category mod A. We consider the subcategory proj -A of all projective A-modules. Through the Yoneda embedding, mod(proj -A)op is equivalent to mod A. Lemma 2.3 ([3, Theorem IV.6.11]). Let N be an indecomposable A-module. (1) If N is non-projective, then the exact sequence 0 → L − →M − → N → 0 is almost split if and only if the induced sequence of functors 0 → HomA (−, L) → HomA (−, M ) → HomA (−, N ) → SN → 0 is a minimal projective resolution of SN in mod(mod Aop ). (2) If N is projective, then M → N is right minimal almost split if and only if the induced sequence of functors 0 → HomA (−, M ) → HomA (−, N ) → SN → 0 is a minimal projective resolution of SN in mod(mod Aop ). 2.2. Presentations. Let A be some finite dimensional k-algebra with valued quiver Q (see [2, III.1]). If you do not know what a valued quiver associated to A is, then you can just take A to be the F-species defined in Section 1.1. Let C 2 A := Ch2 (proj -A) be the category of projective presentations. To be more precise, the f → P− in proj -A (with P+ and P− in some objects in C 2 A are 2-term complexes P+ − fixed degrees). The morphisms are commutative diagrams. Note that this category is Krull-Schmidt, but not abelian. Let Pi be the indecomposable projective module L 0 β(i)P we denote corresponding to i ∈ Q0 . For any β ∈ ZQ i by P (β). If i∈Q0 >0 12 JIARUI FEI P± = P (β± ), then the weight vector (f− , f+ ) of f is (β− , β+ ). The reduced weight vector f is the difference f+ − f− . Id Presentations of forms 0 → P, P → 0, and P −→ P are called negative, positive, + and neutral. They are also denoted by O− P , OP and IdP respectively. If P = Pi , then they are called i-th negative, positive, and neutral presentation, and denoted + by O− i , Oi and Idi respectively. Lemma 2.4. [10] Any presentation f decomposes as f = f+ ⊕ fId ⊕ f ′ , where f+ is positive, fId is neutral, and f ′ is the minimal presentation of Coker(f ). Corollary 2.5. An indecomposable presentation is one of the following four kinds. They are i-th negative, positive, neutral presentations, and minimal presentations of indecomposable non-projective representations of A. The following lemma is easy to verify. f → P− ∈ C 2 A and P ∈ mod A, we have Lemma 2.6. For any P+ − ∼ (1) HomC 2 A (O− P , f ) = HomA (P, P− ); + ∼ (2) HomC 2 A (f, OP ) = HomA (P+ , P ); (3) HomC 2 A (IdP , f ) ∼ = HomA (P, P+ ); (4) HomC 2 A (f, IdP ) ∼ = HomA (P− , P ). Let f and g be two presentations of representations M and N , namely, M = Coker f and N = Coker g. For any morphism in ϕ ∈ HomC 2 A (f, g), we get an induced morphism φ ∈ HomA (M, N ). P+ f ϕ+  R+ / P− ϕ− g  / R− /M /0 φ  /N /0 Conversely, any φ ∈ HomA (M, N ) lifts to a morphism in HomC 2 A (f, g). So we obtain a surjection π : HomC 2 A (f, g) ։ HomA (Coker f, Coker g). π maps to a zero morphism if and only if the image of ϕ− is contained in the image of g. In this case, ϕ− lifts to a map in HomA (P− , R+ ) because P− is projective. Hence the kernel of π is the image of the map ι ι : HomA (P− , R+ ) → HomC 2 A (f, g), h 7→ gh − hf. Moreover, if g is injective, ι is injective as well. We record this in terms of functors. Lemma 2.7. We have the exact sequence HomC 2 A (−, IdR+ ) → HomC 2 A (−, g) − → HomA (Coker −, Coker g) → 0. Moreover, if g is injective, then the leftmost map is injective as well. From now on let us assume A is the F-species ΓQ . We denote C 2 ΓQ by C 2 Q. For a representation M of Q, we write f (M ) : P+ (M ) → P− (M ) for its minimal presentation. In the following two lemmas, we write Hom(−, −) for HomC 2 Q (−, −). Some of the results below were also proved in [8] in a slightly different setting. TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 13 Lemma 2.8. Suppose that 0 → L → M → N → 0 is an almost split sequence in Rep(Q). Then we have the projective resolution of the simple in mod(C 2 Q)op (2.2) 0 → Hom(−, f (L)) → Hom(−, f (M )) → Hom(−, f (N )) → Sf (N ) → 0 unless L is simple. If L = Si , then the resolution becomes (2.2’) 0 → Hom(−, f (Si )) → Hom(−, f (M )) ⊕ Hom(−, Idi ) → Hom(−, f (N )) → Sf (N ) → 0. Proof. Suppose that L is not simple. We can splice the minimal presentations of L and N together to form a presentation of M f P+ (L) ⊕ P+ (N ) − → P− (L) ⊕ P− (N ) → M → 0. By construction, we have the exact sequence 0 → f (L) → f → f (N ) → 0. We claim that f is minimal. This is equivalent to that dim HomQ (M, Si ) = dim HomQ (L, Si ) + dim HomQ (N, Si ) and dim Ext1Q (M, Si ) = dim Ext1Q (L, Si ) + dim Ext1Q (N, Si ) for each Si . Since 0 → L → M → N → 0 is almost split and L is non-simple, it follows that 0 → HomQ (N, Si ) → HomQ (M, Si ) → HomQ (L, Si ) → 0, 0 → Ext1Q (N, Si ) → Ext1Q (M, Si ) → Ext1Q (L, Si ) → 0 are both exact. Using Lemma 2.7, we expand (2.2) to a double complex with exact columns, in which (2.2) sits as the middle row. The bottom row is 0 → HomQ (Coker −, L) → HomQ (Coker −, M ) → HomQ (Coker −, N ) → SN → 0, which is exact by Lemma 2.3. The top row is 0 → Hom(−, IdP+ (L) ) → Hom(−, IdP+ (L)⊕P+ (N ) ) → Hom(−, IdP+ (N ) ) → 0, which is clearly exact. It is easy to see that (2.2) holds when evaluating on f (N ). When evaluating outside f (N ), the exactness of (2.2) follows from the 3 × 3 lemma. In the case when L = Si , by Auslander-Reiten formula [3, Theorem IV.2.13] HomQ (τ −1 L, Si ) = Ext1Q (Si , Si )∗ = 0, Ext1Q (τ −1 L, Si ) = HomQ (Si , Si )∗ = k. So we have the exact sequence 0 → HomQ (M, Si ) → HomQ (L, Si ) ∼ = Ext1Q (τ −1 L, Si ) → Ext1Q (M, Si ) → 0. Hence HomQ (M, Si ) = Ext1Q (M, Si ) = 0. This implies 0 → f (L) → f (M ) ⊕ Idi → f (N ) → 0 is exact. The same argument as before shows that (2.2’) holds.  14 JIARUI FEI Lemma 2.9. We have the following projective resolutions of simple objects in mod(C 2 Q)op (2.3) 0 → M − cj,i Hom(−, O− j ) → Hom(−, Oi ) → SO− → 0; i (i,j)∈Q1 (2.4) 0 → Hom(−, f (Ii )) → → Hom(−, O+ i ) M (k,i)∈Q1 ci,k Hom(−, f (Ik )) ⊕ → S O+ → 0 i M (2.4’) 0 → Hom(−, f (Ii )) → ci,k Hom(−, f (Ik )) ⊕ (k,i)∈Q1 ⊕ Hom(−, Idi ) → Hom(−, O+ i ) → S O+ → 0 M cj,i Hom(−, O+ j ) (i,j)∈Q1 if i is not a source; M cj,i Hom(−, O+ j ) (i,j)∈Q1 if i is a source; i (2.5) 0 → Hom(−, f (Si )) → Hom(−, Idi ) → SIdi → 0. ∼ Proof. By Lemma 2.7 Hom(f, O− i ) = HomQ (Coker f, Pi ), so the exactness of (2.3) follows from Lemma 2.3.(2). We prove the rest by evaluating these sequences on indecomposable presentations f : P− → P+ . For (2.5), it is clear for f = Idi . If f 6= Idi , it is equivalent to that any morphism ϕ− ∈ Hom L Q (P− , Pi ) uniquely determines a morphism in Hom(f, f (Si )). Since f (Si ) : (i,j)∈Q1 cj,i Pj ֒→ Pi is right minimal almost split ([3, Proposition IV.3.5]), we have a unique lift ϕ e∈ L e ϕ− ) is the unique morphism HomA (P− , (i,j)∈Q1 cj,i Pj ). Then the map (ϕ+ := ϕf, in Hom(f, f (Si )). The proof of resolution (2.4) and (2.4’) of SO+ is similar to that of Lemma 2.8. It i contains two cases, but we only prove theL first one because the second is similar. We have the exact sequence 0 → Si → Ii → (i,k)∈Q1 ci,k Ik → 0. Similar argument as in Lemma 2.8 shows that 0 → f (Si ) → f (Ii ) ⊕ Idi → M (i,k)∈Q1 ci,k f (Ik ) → 0 is exact if i is not a source (i.e., Ii is not simple). We assume that Ii is minimally i i presented by R+ → R− , then by construction we have an exact sequence (2.6) 0→ M (i,j)∈Q1 i cj,i Pj → R+ ⊕ Pi → M (k,i)∈Q1 k ci,k R+ → 0. L k contains Pi as a direct summand. In particular, we see that (k,i)∈Q1 ci,k R+ + It is easy to verify (2.4) for Oi , so we assume that f 6= O+ i , and moreover f : P− → P+ presents M . Using Lemma 2.7, we expand (2.4) to a double complex TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS with exact columns. (2.7) 0 15 0  M  i HomQ (P− , R+ ) ci,k HomQ (P− , Rk+ ) (k,i)∈Q1 / M ⊕ / HomQ (P+ , Pi ) cj,i HomQ (P+ , Pj ) (i,j)∈Q1  / Hom(f, f (Ii )) 0  M / ci,k Hom(f, f (Ik )) (k,i)∈Q1 ⊕ M / Hom(f, O+ ) i cj,i Hom(f, O+ j ) /0 (i,j)∈Q1  HomQ (M, Ii ) l / M  ci,k HomQ (M, Ik ) (k,i)∈Q1  0  0 Due to (2.6), we write    .   M M k i cj,i Pj . ci,k R+ Pi ⊕HomQ P− , HomQ (P− , R+ ) = HomQ P− , (i,j)∈Q1 (k,i)∈Q1 After reflecting the top row of (2.7) for a while, we find that it is (2.8) HomQ   P− , M ci,k Rk + .  Pi HomQ   P− , M ci,k Rk + . Pi  ❨0❨❨, (k,i)∈Q1 / HomQ (P+ , Pi ) ⊕ Hom (P , P ) Q − i 2 2 ❢ ❞❞❞❞❞❞❞ ❢ ❞ ❢ ❞ ❞ ❢ ❞ ❢ ❢ M M /⊕ cj,i HomQ (P− , Pj ) cj,i HomQ (P+ , Pj ) (k,i)∈Q1 ⊕ (i,j)∈Q1 (i,j)∈Q1 So the top row of (2.7) reduces to the last two rows of (2.8). It is nothing but the double complex obtained by taking HomQ to the presentations f (M ) and f (Si ). We know that this complex has cohomology 0, HomQ (M, Si ) and Ext1Q (M, Si ) from left to right. On the other hand, the map l has kernel and cokernel HomQ (M, Si ) and Ext1Q (M, Si ) as well. Finally, we can easily chase through the diagram and conclude the exactness of the middle row.  According to the above two lemmas, it makes sense (as least in hereditary cases) to extend the classical Auslander-Reiten translation τ from projective-free representations to the negative-and-neutral-free presentations as in [10]. For f = f (M ), we define τ f := f (τ M ). For each i ∈ Q0 we define τ (O+ i ) := f (Ii ). 3. iARt Quivers 3.1. iARt Quivers. We slightly upgrade the classical Auslander-Reiten quiver by adding the translation arrows. The following definition is basically taken from [2, 16 JIARUI FEI VII.1]. Let A be a category as in Section 2.1. Recall that for each M ∈ ind(A), EM := EndA (M )/ rad EndA (M ) is a division k-algebra. Definition 3.1 (ARt quiver). The ARt valued quiver ∆(A) of A is defined as follows: (1) The vertex of ∆(A) are the isomorphism classes of objects in ind A. (2) There is a morphism arrow M → N if IrrA (M, N ) is non-empty. We assign the valuation (a, b) to this arrow, where a = dimEM IrrA (M, N ) and op Irr (M, N ). b = dimEN A (3) There is a translation arrow from N to τ N with trivial valuation if τ N is defined. A vertex u in an ARt quiver is called transitive if the translation and its inverse are both defined at u. Note that the number a in the valuation (a, b) can be alternatively interpreted as the (direct sum) multiplicity of M in E for E → N minimal right almost split. Similarly b is the multiplicity of N in E ′ for M → E ′ minimal left almost split. Moreover, if A is k-elementary, i.e., EM = k for any M ∈ ind(A), then all morphism arrows are equally valued. Definition 3.2 (iARt quivers). • The iARt quiver ∆A is obtained from the ARt quiver ∆(A) by freezing all vertices whose translations are not defined. • The iARt quiver ∆2A is obtained from the ARt quiver ∆(C 2 A) by freezing all non-transitive vertices. Remark 3.3. When A is the module category of a finite-dimensional algebra, the frozen vertices of ∆A are precisely indecomposable projective modules. However, it is not clear which vertices are frozen in ∆2A . The author think it is an interesting and doable problem to determine the non-transitive vertices. We will see that if the algebra is hereditary, then they are precisely the negative, positive and neutral presentations. We use the notation L→M if there is a morphism arrow from L to M ; L 99K M if there is a translation arrow from L to M ; L֌M if there is an arrow from L to M in the ARt quiver. Lemma 3.4. Let θ be an additive function from A to some abelian group, that is, θ(L) + θ(N ) = θ(M ) for each exact sequence N → 0 in A. Then P 0 → L → M →P at each transitive vertex L, we have that M֌L cM,L θ(M ) = L֌N cL,N θ(N ). Proof. We have two almost split sequences M 0 → τL → cM,L M → L → 0, M→L 0→L→ M L→N cL,N N → τ −1 L → 0. TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 17 By the additivity of θ, we have that X X θ(L) = cM,L θ(M ) − θ(τ L) = cL,N θ(N ) − θ(τ −1 L), M→L ⇒ X L→N cM,L θ(M ) = M֌L X cL,N θ(N ). L֌N  A typical additive function in C 2 A is the weight vector. In some special cases including examples below, indecomposable presentations are uniquely determined by their weight vectors. So we can label them on an iARt quiver by their weight vectors. We will use the “exponential form” as a shorthand. For example, a vector (3, 1, 0, 0, −2, 0, −1) is written as (52 7, 13 2). Example 3.5. Let A be the Jacobian algebra of the quiver with potential (Q, W ) @3❂ ✁✁ ❂❂❂ ✁ ✁  (see Section 4.1), where Q = 2 ❂o 1 and W is the difference of two oriented ❂❂ ✁@ ✁ ❂ ✁✁ 4 triangles. The iARt quiver ∆2mod A is drawn below. We always put frozen vertices in boxes. 1,1 4,4 2,2 ? ❄❄ F ✵ ❄❄ ⑧⑧ ✍G ✶✶✶ ✌✌ ✵✵✵ ❄ ⑧⑧ ✶ ⑧ ✌ ✍✍✍ ✶ ❴ ❴ ❴ 4,1 o❴ ❴ ❴ ❴ 2,4 o❴ ❴ ❴ ✌✌❴ 0,3 ✵✵ ✍ 3,0 o❴ ❉❉ = ❈❈❈ ❉❉✶✶✶ ❇❇ ✵✵ ③< ❉❉ ✌✌③③< ✍✍✍⑤⑤= ④④ ❉ ❇ ❈ ③③ ✍⑤⑤o❴ ❴ ❴ ❴ ❉✶"  ③③o❴ ❴ ❴ ❴ ❈! ④④o❴④ ❴ ❴ ❴ ❉" ✌✌③③o❴ ❴ ❴ ❴ ❇✵!  2,34 1,2 2,1 1,2 34,1 ✵✵❇❇ ④= F ✵ ❈❈ ⑤= G ③< F ✶ ❉❉ ③< F ✶ ❉❉ ✵✵ ❇! ⑤⑤✍✍✍ ③③ ✌✌✌ ✶✶✶ ❉❉" ④④④④✍✍✍ ✵✵✵ ❈❈❈! ③③③ ✌✌✌ ✶✶✶ ❉" ✵✵ 4,0 o❴✌✌ ❴ ❴ ✶❴✶ 3,1 o❴✍✍ ❴ ❴ ✵❴✵ 2,3 o❴✌✌ ❴ ❴ ✶❴✶ 0,4 ✍✍ ✵? ✶✶ ✍✍ ✶✶ ✍❄✍❄ ✵✵ ✌✌ ✌ ✶ ✍✍ ✶ ✍✍ ❄❄ ⑧⑧⑧⑧✵✵ ✌✌✌ ✵ ✌✌ o❴ ❴ 0,2 o❴ ❴ 2,0 f 0,1 1,0 ◗ 3,3 ❴ ♠ Two vertices with the same weight label (1, 2) are identified. The translation arrow going out from (0, 2) ends in (1, 0). 3.2. Hereditary Cases. In particular, if we take A := Rep(Q) for some valued quiver Q, we get two ARt quivers ∆(Q) := ∆(Rep(Q)) and ∆(C 2 Q). We denote the corresponding iARt quivers by ∆Q and ∆2Q . Proposition 3.6. The ARt quiver ∆(C 2 Q) can be obtained from ∆(Q) as follows. (1) We add |Q0 | vertices corresponding to O+ i and another |Q0 | vertices corresponding to Idi . (a,b) (a,b) (2) For each i −−−→ j in Q, we draw morphism arrows O+ −−→ O+ j − i , and (b,a) + f (Ii ) −−−→ O+ j . We add translation arrows from Oi to f (Ii ). (3) We draw morphism arrows f (Si ) − → Idi , and Idi − → τ −1 f (Si ). In particular, the non-transitive vertices are precisely negative, positive and neutral presentations. 18 JIARUI FEI Proof. The vertices of ∆(Q) are identified with vertices of ∆(C 2 Q) via minimal presentations. By Corollary 2.5, we only need to add the vertices as in (1). Step (2) is due to (2.4) and (2.4’). Note that τ −1 f (Si ) is equal to f (τ −1 Si ) if i is not a source, otherwise it is equal to O+ i . So Step (3) is due to (2.5), (2.2’), and (2.4’). We do not need anything else because of (2.2) and (2.3).  Due to this proposition, we will freely identify ∆Q as a subquiver of ∆2Q . From now on, we let Q be a valued quiver of Dynkin type. In this case, any indecomposable presentation f is uniquely determined by its weight vector. The quiver ∆Q was already consider in [7, 25]. In [7] the authors associated an ice quiver to any reduced expression of the longest element w0 in the Weyl group of Q. The iARt quiver ∆Q only corresponds to those reduced expressions adapted to Q. Example 3.7. The iARt quiver ∆2Q for Q of type An is the ice hive quiver ∆n constructed in [14] up to some arrows between frozen vertices. Example 3.8 (The iARt quiver ∆2Q for Q a D4 -quiver). 4,4 ❄❄ ? ❄❄ ⑧⑧  ⑧⑧ ⑧ ❴ o ❴ ❴ ❴ ❴ o ❴ ❴ ❴ ❴ ❴ o ❴ ❴ ❴ ❴ 4,0 0,4 4,2 3,1 ❈ ❄❄ ❅❅ ? ? ❈❈ ⑤= ⑥> ⑧ ⑧ ❄ ⑤ ⑥ ❅ ❈ ⑧ ⑧ ❄ ⑤ ❅❅ ⑥⑥ ⑧ ❈❈ ⑤ ⑧⑧ ! ⑤ o❴ ❴ ❴ ❴  ⑧ o❴ ❴ ❴ ❴ 34,1⑥ o❴ ❴ ❴ ❴ 34,12 2,0 0,2 > G ✵❆ ?G ❄❄ ④= F ✶✶❇❇ G ✴ ❄❄❄ ❄❄ ⑦⑦ ✍ ✵✵ ❆❆❆ ④④✌✌✌ ✶✶ ❇❇❇ ⑧⑧⑧⑧✎✎ ✎✎✎ ✴✴✴ ❄ ⑦⑦ ✍✍ ❆  ④ ✵ ④ ✶ ! ⑧ ✎ ✎✎ ✴✴ 3,0 o❴✍✍✍ ❴ ❴ ❴ ✵✵❴✵ 4,1 o❴✌✌✌❴ ❴ ❴ ✶❴✶ 3,2 o❴✎✎ ❴ ❴ ❴ 0,3 ✶✶ ✎❄✎❄ ? ✴✴ ✍✍ ✵✵ ✌✌ ✎✎ ✶ ✎✎ ❄❄ ⑧⑧⑧⑧ ✴ ✍✍ ✵ ✌✌ ✎✎ o❴ ❴ ❴ ❴ 2,1 o❴ ❴ ❴ ❴ 34,2 o❴ ❴ ❴❴ ❴ ❴ 0,1 1,0 3,3 ❅❅ ? > ❄❄ ❅❅ ❄❄ ⑧⑧ ⑥⑥ ⑧ ⑥  ⑧ ⑥ 1,1 2,2 Readers can find a few other iARt quivers in Appendix A. Remark 3.9. One natural question is whether the iARt quivers ∆2Q and ∆2Q′ (or ∆Q and ∆Q′ ) are mutation-equivalent if Q and Q′ are reflection-equivalent. The answer is positive at least in trivially valued cases. We conjecture that this is also true in general. As pointed in [7, Remark 2.14], for ∆Q with Q trivially valued, by the Tits lemma every two reduced words can be obtained form each other by a sequence of elementary 2- and 3-moves (see [43, Section 2.1]); by [43, Theorem 3.5] every such move either leaves the seed unchanged, or replaces it by an adjacent seed. Finally, similar to the proof of Corollary B.9, the result can be extended from ∆Q to ∆2Q . However, reflection-equivalent quivers can not replaced with mutation-equivalent Jacobian algebras (see Section 4.1). The Jacobian algbera in Example 3.5 is obtained from the above path algebra of D4 by mutating at the vertex 2. According to [31], the iARt quiver in Example 3.8 is mutation-equivalent to a finite mutation (1,1) type quiver E6 , while the one in Example 3.5 is mutation equivalent to a wild acyclic quiver, which is of infinite mutation type. It follows from (2.1) and the fact that each Q is symmetrizable that Lemma 3.10. The B-matrix of ∆2Q is skew-symmetrizable. TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 19 The following lemma is an easy exercise in linear algebra. Lemma 3.11. The restricted B-matrices of ∆Q , and thus of ∆2Q , have full ranks. By Lemma 3.4, the assignment f 7→ (f− , f+ ) is a weight configuration of ∆2Q . However, it is not full (see Section 1.3). We want to extend it to a full one which is useful for the second half of the paper. Since Q is of finite representation type, each non-neutral f ∈ ind(C 2 Q) is translated from a unique indecomposable positive presentation, that is, f = τ t (O+ i ) for some i ∈ Q0 and t ∈ Z>0 . Now for each f ∈ ind(C 2 Q), we assign a triple-weight vector as follows. 3|Q0 | e Definition 3.12. If f is translated from O+ i , then the triple weights f ∈ Z of f is given by (e(f ), f− , f+ ) where e(f ) := ei the unit vector supported on i. If f = Idi , then we set ef := (0, ei , ei ). We also define another weight vector f ∈ Z|Q0 | attached to f by f := ei + f− − f+ . Corollary 3.13. The assignment σ 2 : f 7→ ef (resp. σ Q : f 7→ f) defines a full Q weight configuration for the iARt quiver ∆2Q (resp. ∆Q ). Proof. Due to Lemma 3.4, it suffices to show that for each mutable f X X cg,f e(g) = cf,h e(h). g֌f f ֌h We call a vertex u regular if it is transitive and τ −1 v is defined for each v → u, and τ w is defined for each u → w. It is clear from (2.1) that the equation holds at each regular vertex. From the description of Proposition 3.6, we see that all transitive vertices are regular except for f (Si ) and τ −1 f (Si ). The problem is that these vertices may have (morphism) arrows to neutral frozen vertices, whose translation is not defined. But the first component of the triple weights of Idi is a zero vector so the equality still holds at these vertices. For the case of ∆Q , it is enough to observe that the weight vector f is zero on the positive and neutral frozen vertices of ∆2Q , and ∆Q is obtained from ∆2Q by deleting these vertices.  We shall consider the graded upper cluster algebra C(∆2Q ; σ 2Q ) and the graded cluster algebra C(∆Q ; σ Q ) later. 4. Cluster Character from Quivers with Potentials 4.1. Quivers with Potentials. The mutation of quivers with potentials is invented in [12] and [13] to model the cluster algebras. In this and next section and Appendix B, we switch back to the usual quiver notation. A quiver ∆ is a quadruple (∆0 , ∆1 , h, t), where the maps h and t map an arrow a ∈ ∆1 to its head and tail h(a), t(a) ∈ ∆0 . Following [12], we define a potential W on an ice quiver ∆ as a (possibly infinite) linear combination of oriented cycles in ∆. More precisely, a c := k∆/[ c k∆, c k∆], c where k∆ c is the potential is an element of the trace space Tr(k∆) c c completion of the path algebra k∆ and [k∆, k∆] is the closure of the commutator c The pair (∆, W ) is an ice quiver with potential, or IQP for short. subspace of k∆. c is defined to be the linear For each arrow a ∈ ∆1 , the cyclic derivative ∂a on k∆ extension of d X ∂a (a1 · · · ad ) = a∗ (ak )ak+1 · · · ad a1 · · · ak−1 . k=1 20 JIARUI FEI For each potential W , its Jacobian ideal ∂W is the (closed two-sided) ideal in c generated by all ∂a W . The Jacobian algebra J(∆, W ) is the quotient algebra k∆ c k∆/∂W . 3 If W is polynomial and J(∆, W ) is finite-dimensional, then the completion is unnecessary to define J(∆, W ). This is the case throughout this paper. The key notion introduced in [12, 13] is the mutation of quivers with potentials and their decorated representations. For an ice quiver with nondegenerate potential (see [12]), the mutation in certain sense “lifts” the mutation in Definition 1.3. We have a short review in Appendix B.1. Definition 4.1. A decorated representation of a Jacobian algebra J := J(∆, W ) is a pair M = (M, M + ), where M ∈ Rep(J), and M + is a finite-dimensional k ∆0 -module. Let Rep(J) be the set of decorated representations of J(∆, W ) up to isomorphism. Let K 2 J be the homotopy category of C 2 J. There is a bijection between two additive categories Rep(J) and K 2 J mapping any representation M to its minimal presentation in Rep(J), and the simple representation Su+ of k ∆0 to Pu → 0. Suppose that M corresponds to a projective presentation P (β+ ) → P (β− ). Definition 4.2. The g-vector g(M) of a decorated representation M is the reduced weight vector β+ − β− . Definition 4.3. A potential W is called rigid on a quiver ∆ if every potential on ∆ is cyclically equivalent to an element in the Jacobian ideal ∂W . Such a QP (∆, W ) is also called rigid. A potential W is called µ-rigid on an ice quiver ∆ if its restriction to the mutable part ∆µ is rigid. It is known [12, Proposition 8.1, Corollary 6.11] that every rigid QP is 2-acyclic, and the rigidity is preserved under mutations. In particular, any rigid QP is nondegenerate. Definition 4.4. Two QPs (∆, W ) and (∆′ , W ′ ) on the same vertex set ∆0 are call right-equivalent if there is an isomorphism ϕ : k∆ → k∆′ such that ϕ |k∆0 = Id and ϕ(W ) is cyclically equivalent to W ′ . Two IQPs (∆, W ) and (∆′ , W ′ ) are call µright-equivalent if they are right-equivalent when restricted to their mutable parts. Definition 4.5. A representation is called µ-supported if its supporting vertices are all mutable. We denote by Repµ (J) the category of all µ-supported decorated representations of J. Remark 4.6. If (∆, W ) and (∆′ , W ′ ) are µ-right-equivalent, then Repµ (J) and Repµ (J ′ ) are equivalent. Indeed, we write W = Wµ +W⋄ where Wµ is the restriction of W to the mutable part of ∆. We find that any cyclic derivative ∂a W⋄ is a sum of paths passing some frozen vertices. Such a sum gives rise to a trivial relation on the µ-supported representations. 4.2. The Generic Cluster Character. Definition 4.7. To any g ∈ Z∆0 we associate the reduced presentation space PHomJ (g) := HomJ (P ([g]+ ), P ([−g]+ )), where [g]+ is the vector satisfying [g]+ (u) = max(g(u), 0). We denote by Coker(g) the cokernel of a general presentation in PHomJ (g). 3Unlike the definition in [4], we need to include ∂ W in the ideal ∂W even if a is an arrow a between frozen vertices. TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 21 Reader should be aware that Coker(g) is just a notation rather than a specific representation. If we write M = Coker(g), this simply means that we take a presentation general enough (according to context) in PHomJ (g), then let M to be its cokernel. Definition 4.8. A g-vector g is called µ-supported if Coker(g) is µ-supported. Let G(∆, W ) be the set of all µ-supported g-vectors in Z∆0 . It turns out that for a large class of IQPs the set G(∆, W ) is given by lattice points in some rational polyhedral cone. Such a class includes the IQPs introduced in [14, 15, 16], and the ones to be introduced in Section 5.1. Definition 4.9 ([41]). We define the generic character CW : G(∆, W ) → Z(x) by X  (4.1) CW (g) = xg χ Gre (Coker(g)) ye , e e where Gr (M ) is the variety parametrizing e-dimensional quotient representations of M , and χ(−) denotes the topological Euler-characteristic. Theorem 4.10 ([14, Corollary 5.14], cf. [41, Theorem 1.1]). Suppose that IQP (∆, W ) is non-degenerate and B∆ has full rank. The generic character CW maps G(∆, W ) (bijectively) to a set of linearly independent elements in C(∆) containing all cluster monomials. Definition 4.11. We say that an IQP (∆, W ) models an algebra C if the generic cluster character maps G(∆, W ) (bijectively) onto a basis of C. If C is the upper cluster algebra C(∆), then we simply say that (∆, W ) is a cluster model. Remark 4.12. Suppose that (∆, W ) and (∆′ , W ′ ) are µ-right-equivalent. By Remark 4.6, Repµ (J) and Repµ (J ′ ) are equivalent via some isomorphism ϕ : k∆µ → k(∆′ )µ . By abuse of notation we denote the equivalence also by ϕ. Since ϕ |∆0 = Id, ϕ(M ) and M are isomorphic and have the same g-vector. We see from (4.1) that if (∆, W ) is a cluster model, then so is (∆′ , W ′ ). 5. iARt QPs 5.1. The iARt QP (∆2Q , WQ2 ). For the time being, let us assume Q is a trivially valued Dynkin quiver. A translation triangle in an iARt quiver is an oriented cycle of the form =M ❆ ④④ ❆❆❆ ④ ❆❆ ④④ ❆ ④④ τ L o❴ ❴ ❴ ❴ ❴ ❴ L For each iARt quiver ∆2Q , we define the potential WQ2 as an alternating sum of all translation triangles. We make this more precise as follows. We can also label each ∗ ∗ ∗ 2 non-neutral f = τ −t O− i by the pair (i, t) = (i (f ), t (f )). The arrows of ∆Q are thus classified into three classes Type A arrows Type B arrows Type C arrows (i, t)∗ → (j, t + 1)∗ , f (Si ) → Idi ; (i, t)∗ → (j, t)∗ , Idi → τ −1 f (Si ); (i, t)∗ 99K (i, t − 1)∗ . 22 JIARUI FEI Let ȧ (resp. ḃ and ċ) denote the sum of all type A (B and C) arrows. The potential WQ2 is defined as ȧċḃ − ȧḃċ. Thus the Jacobian ideal is generated by the elements (5.1) (5.2) eu (ȧċ − ċȧ)ev , eu (cb − bc)ev , eu (ḃȧ − ȧḃ)ev . J(∆2Q , WQ2 ) Let J := be the Jacobian algebra of (∆2Q , WQ2 ). For the rest of this section, we denote a single arrow by the lowercase letter of its type with some superscript (eg. a and a′ ). We observe that for each non-neutral vertex in ∆2Q there are exactly one incoming arrow and one outgoing arrow of type C. Moreover, if non-neutral u and v are connected by an arrow of type B or A, then the relations (5.1) say that (5.1.1) eu acev = eu c′ a′ ev or eu cbev = eu b′ c′ ev for some a, a′ , b, b′ , c, c′ . In general, the relations (5.2) do not have a similar implication because there is a trivalent vertex for Q of type D or E. We have that X (eu a′ b′ ev or eu b′ a′ ev ). (5.2.1) eu baev = a′ ,b′ If u and v are not trivalent, then the right sum has only one summand. If both u and v are negative (resp. positive), then some translation arrows are undefined so the relations (5.1.1) reduce to the following (5.1.2) eu acev = 0 (resp. eu caev = 0 or eu cbev = 0). Similarly if u (resp. v) is neutral, then (5.1.3) eu bcev = 0 (resp. eu caev = 0). Lemma 5.1. The IQP (∆2Q , WQ2 ) is rigid and J(∆2Q , WQ2 ) is finite-dimensional. Proof. To show that J(∆2Q , WQ2 ) is finite-dimensional, it suffices to observe that any nonzero path from f to g in J can be uniquely identified as an element in Lt∗ (f ) t t=0 HomC 2 Q (τ f, g). Indeed, suppose that p is a path from f to g. By (5.2.1) and (5.1.3), we can make p avoid any neutral vertex. By (5.1.1) we can move all arrows of type C to the left. If we remove all arrows of type C, then the truncated path can be interpreted as a morphism from τ t f to g. Due to relations (5.1.1) and (5.2.1), any cycle in the Jacobian algebra is equivalent to a sum of composition of cycles eu (acb)eu′ with u and u′ mutable. It suffices to show that each eu acbeu′ is in fact zero in the Jacobian algebra. Applying the relation (5.1.1) twice (if u is not negative), we see that eu acbeu′ is equivalent to ew a′ c′ b′ eu where u and w are connected by an arrow of type C. If u is mutable, then there is some negative vertex v connected to u by arrows of type C. So eu acbeu′ is equivalent to ev a′′ c′′ b′′ , which is zero by (5.1.2).  We delete all translation arrows of ∆2Q , and obtain a subquiver denoted by ∧2Q . Let R be the direct sum of all presentations in ind(C 2 Q). It is well-known that the Auslander algebra A2Q := EndC 2 Q (R) is equal to k∧2Q modulo the mesh relations [2, VII.1]. The Auslander algebra A2Q is the quotient of Jacobian algebra J by the ideal generated by translation arrows. Let f : P+ − → P− be a presentation in ind(C 2 Q). We denote by P f (resp. I f ) the indecomposable projective (resp. injective) representation of J corresponding to the vertex f . TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 23 Lemma 5.2. We have the following for the module P f in J(∆2Q , WQ2 ) ∼ (1) P f (O− i ) = HomQ (Pi∗ (f ) , Pi ); + ∼ (2) P f (Oi ) = HomQ (P+ , Pi ); (3) P f (Idi ) ∼ = HomQ (P− , Pi ). Lt∗ (f ) − t Proof. (1). Recall from the proof of Lemma 5.1 that P f (O− i )⊂ t=0 HomC 2 Q (τ f, Oi ). − We know that HomC 2 (Q) (f, Oi ) = 0 unless f is negative. This implies that P f (O− i ) ⊂ HomQ (Pi∗ (f ) , Pi ). Conversely, we identify an element in HomQ (Pi∗ (f ) , Pi ) by a path consisting solely of arrows of type B. By adjoining t∗ (f ) arrows of type C we get a path from f to O− i , which is easily seen to be nonzero. (2). We observe that any path from f to O+ i containing a translation arrow must be equivalent to zero due to the relations (5.1) and (5.1.2). So P f (O+ i ) is the same as P f (O+ ) restricted to the Auslander algebra. The result follows from Lemma i 2.6.(4). (3). Similar to (2), any path from f to Idi containing a translation arrow must be equivalent to zero. The result follows from Lemma 2.6.(2).  5.2. The Cone G∆2Q . We consider the following set of representations Tv (5.3) (5.4) (5.5) 0 → Tv → I v → 0 → Tv → I v → 0 → Tv → I v → M I O− for v = O− i , I O+ for v = O+ i , I Idj for v = Idi . j i→j M j i→j M i→j The maps in (5.3) and (5.4) are canonical, that is, the map from I O± to I O± is i j ± given by the morphism arrow O± j → Oi . For the map from I Idi to I Idj , let us recall from Lemma 2.8.(3) that Hom(I Idi , I Idj ) ∼ = P Idj (Idi ) ∼ = HomQ (Pj , Pi ). We take the map to be the irreducible map in HomQ (Pj , Pi ). It will follow from the proof of Theorem 5.3 that the rightmost maps in (5.3)–(5.5) are all surjective. ∗ For j ∈ Q0 , let j ∗ = i∗ (O+ j ) ∈ Q0 . It is well-known that j 7→ j is a (possibly trivial) involution. The involution does not depend on the orientation of Q. Its formula is listed in [25, Section 2.3]. For any map between projective modules f : P (β1 ) → P (β2 ), the i-th top restriction of f is the induced map top P (β1 (i)) → top P (β2 (i)). Theorem 5.3. We have the following description for the modules Tv . (1) The module TO− is the indecomposable module supported on all vertices i translated from O+ i∗ with dimension vector (1, 1, . . . , 1); (2) The defining linear map f → g in TO+ is given by the i-th top restriction i of ϕ+ ; (3) The defining linear map f → g in TIdi is given by the i-th top restriction of ϕ− , where ϕ = (ϕ+ , ϕ− ) is the irreducible morphism from f to g. 24 JIARUI FEI In particular, the dimension vector θv of Tv is given by θv (f ) = e(f )(i∗ ) for v = O− i , θv (f ) = f+ (i) for v = O+ i , θv (f ) = f− (i) for v = Idi . Proof. By Lemma 5.2.(1), we can identify I O− (f ) with HomQ (Pi∗ (f ) , Pi )∗ . From i the definition of TO− and the exact sequence i M HomQ (Pi∗ (f ) , Pj ) → HomQ (Pi∗ (f ) , Pi ) → HomQ (Pi∗ (f ) , Si ) → 0, 0→ i→j we conclude that TO− (f ) = HomQ (Pi∗ (f ) , Si )∗ . i By Lemma 5.2.(2), we can identify I O+ (f ) with HomQ (P+ , Pi )∗ . From the i definition of TO+ and the exact sequence i M HomQ (P+ , Pj ) → HomQ (P+ , Pi ) → HomQ (P+ , Si ) → 0, 0→ i→j we conclude that TO+ (f ) = HomQ (P+ , Si )∗ . The description of maps follows from i the naturality. By Lemma 5.2.(3), we can identify I Idi (f ) with HomQ (P− , Pi )∗ . From the definition of TIdi and the exact sequence M HomQ (P− , Pj ) → HomQ (P− , Pi ) → HomQ (P− , Si ) → 0, 0→ i→j we conclude that TIdi (f ) = HomQ (P− , Si )∗ and the description of maps follows from the naturality.  Definition 5.4. A vertex v is called maximal in a representation M if dim M (v) = 1 and all strict subrepresentations of M are not supported on v. We note that every Tv contains a maximal vertex w, which is another frozen vertex. − − + For example, w = O+ i∗ (resp. Idi and Oi ) for v = Oi (resp. Oi and Idi ). For the rest of this article, when we write g(−), we view g as a linear functional via the usual dot product. Lemma 5.5. [14, Lemma 6.5] Suppose that a representation T contains a maximal vertex. Let M = Coker(g), then HomJ (M, T ) = 0 if and only if g(dimS) ≥ 0 for all subrepresentations S of T . Lemma 5.6. Let M = Coker(g), then HomJ (M, Tv ) = 0 for each frozen v if and only if HomJ (M, Iv ) = 0 for each frozen v. Proof. Since each subrepresentation of Tv is also a subrepresentation of Iv , one direction is clear. Conversely, let us assume that HomJ (M, Tv ) = 0 for each frozen vertex v. We prove that HomJ (M, Iv ) = 0 by induction. + We first notice that for v = O− i , Oi , Idi , Tv = Iv if i is a sink. In general, for an injective presentation of T : 0 → T → I1 → I0 with HomJ (M, I0 ) = 0, we have that HomJ (M, I1 ) = 0 is equivalent to HomJ (M, T ) = 0. Now we perform the induction from a sink i in an appropriate order using (5.3)–(5.5). We can conclude that HomJ (M, Iv ) = 0 for all frozen v.  TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 25 2 Definition 5.7. We define a cone G∆2Q ⊂ R(∆Q )0 by g(dimS) ≥ 0 for all strict subrepresentations S ⊆ Tv and all frozen v. 2 Theorem 5.8. The set of lattice points G∆2Q ∩ Z(∆Q )0 is exactly G(∆2Q , WQ2 ). Proof. Due to Lemma 5.5 and 5.6, it suffices to show that G∆2Q is defined by g(dimS) ≥ 0 for all subrepresentations S of Tv and all v frozen. We notice that these conditions are the union of the defining conditions of G∆2Q and g(dimTv ) ≥ 0. But the latter conditions are redundant because dimTv = ew + (dimTv − ew ) and dimTv − ew is the dimension vector of a strict subrepresentation of Tv , where w is the maximal (frozen) vertex of Tv .  Example 5.9 (Example 3.8 continued). In this case, it is almost trivial to list all strict subrepresentations for all Tv . Readers can easily find that there are 3 + + + + subrepresentations for all O− i , and 7, 6, 1, 1 subrepresentations for O1 , O2 , O3 , O4 respectively, and 1, 2, 7, 7 subrepresentations for Id1 , Id2 , Id3 , Id4 respectively. All these subrepresentations are needed to define G∆2Q , so there are 44 inequalities. Readers can find an extended version of this example in [17]. This example can be easily generalized to Q of type Dn with a similar orientation. Remark 5.10 (The Valued Cases). To deal with the general valued quiver Q, we could have worked with species analogue of QP, but this require some lengthy preparation. To avoid this Q we can define the analogous Jacobian algebra by the Ext-completion algebra i≥0 Ext2A (A∗ , A)⊗A i for the Auslander algebra A := A2Q . We can even define G∆2Q without introducing the Jacobian algebra. Without the Jacobian algebra we are unable to define the module Tv by injective presentations, but it still makes perfect sense to define Tv via Theorem 5.3. Once Tv ’s are defined, we define the cones G∆2Q by Definition 5.7. The same remark also applies for the cone G∆Q to be defined below. In general, some of the defining conditions of G∆2Q may be redundant as shown in the following example. Example 5.11. The most complicated Tv for Q of type G2 (see Example 1.2 and 1O ❄ ❄❄ ❄❄ ❴ o ❴ ❴ o ❴ ❴ o ❴ 0 1O ❄ 2O ❄ 1O ❄ ❄ ❄❄ ❄ O ❄❄ ❄❄❄ ❄❄❄ , and it strict subAppendix A) is TId1 . Its dimension vector is   o❴ ❴ 3 o❴ ❴ 3 o❴ ❴  0 0 ❄ ❄❄ O ❄❄  0 representations have 13 distinct dimension vectors. However, we only need 5 of them to define G∆2Q . Readers can find a full list of inequalities in [17]. Now we come back to the iARt quiver ∆Q . We define a potential WQ on ∆Q by the same formula defining WQ2 . The iARt QP (∆Q , WQ ) is nothing but the restriction of (∆2Q , WQ2 ) to the subquiver ∆Q . For each O− i , we define the representation TO− of J(∆Q , WQ ) by the same injective presentation (5.3). Similar to Theorem i 5.3.(1), the module TO− is the indecomposable module supported on all vertices i translated from O+ i∗ with dimension vector (1, 1, . . . , 1). 26 JIARUI FEI We define a cone G∆Q ⊂ R(∆Q )0 by g(dimS) ≥ 0 for all subrepresentations S ⊆ TO− and all i. Note that we ask all subrepresentations not just strict ones. i We observe that the defining conditions of G∆Q and those of G∆2Q are related as follows. We group the defining conditions of G∆2Q into three sets gHu ≥ 0, gHl ≥ 0, gHr ≥ 0. They arise from the subrepresentations of Tv for v negative, neutral and positive respectively. Then the defining conditions of G∆Q are exactly gHu ≥ 0. Similar to Theorem 5.8, we have the following proposition. Proposition 5.12. The set of lattice points G∆Q ∩ Z(∆Q )0 is exactly G(∆Q , WQ ). Definition 5.13. Given a weight configuration σ of a quiver ∆ and a convex polyhedral cone G ⊂ R∆0 , we define the (not necessarily bounded) convex polytope G(σ) as G cut out by the hyperplane sections gσ = σ. Our conjectural model for the tensor multiplicity is that the multiplicity cλµ,ν is counted by G∆2Q (µ, ν, λ) for the weight configuration σ 2Q . We will prove this model for G of type ADE in Part 2. A. A List of some iARt Quivers For Q of type B, C and D, we only draw the iARt quiver ∆2Q for B3 , C3 and D5 (D4 is our running example). Reader should have no difficulty to draw the general ones. The cases of E7 and E8 are not hard, but they are too large to draw. We label the vertices of Q in the same way as the software LiE [36] so that you can check things conveniently. 1,1 G2 O ❆❆❆ ❆❆ ❆❆ ❆❆ ❴ o ❴ ❴ ❴ o ❴ ❴ 0,1 ❴ o ❴ ❴ 2 1,0 1,2 1 ,2 O ❆❆ O ❇❇ O ❅❅❅❅ O ❆❆ ❇❇ 3,1 3,1 1,3❅ 1,3❆ 3,1 1,3❇ 3,1 ❅❅ ❇❇ ❆❆ ❅❅ ❇❇ ❆  ❴ o ❴ ❴ ❴ o ❴ ❴ ❴ o ❴ ❴ 2,0 0,2 13 ,2 13 ,22 O ❅❅ ❅❅ ❅❅ ❅❅ ❅ 2,2 TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 2,2 B3 1,1 O ❈❈❈ O ❊❊❊ ❊❊ ❈❈ ❈❈ ❊❊ ! ❊" ❴ o ❴ ❴ ❴ o ❴ ❴ ❴ o ❴ ❴ 0,1 1,0 1,2 ❈ 2,32 O ❈ O ❊❊ O ❊❊❊ O ❈❈ ❊❊ ❊❊ ❈❈ ❊ ❊❊ ❊❊ ❈! ❊" ❊" o❴ ❴ ❴ 12,3 2 o❴ ❴ ❴ 1,32 o❴ ❴ ❴ 0,2 2,0 O ❊❊❊ O ❈❈❈ O ❊❊❊ O 1,2 1,2 2,1❊ 2,1❊ 2,1❈ 1,2 ❈❈ 1,2 ❊❊ ❊❊ ! ❊" ❊" o❴ ❴ ❴ 2,3 o❴ ❴ ❴ 1,3 o❴ ❴ ❴ 0,3 3,0 ❊❊ O ❊❊ ❊❊ ❊" 2,2 C3 1,1 O ❉❉❉ O ❈❈❈ ❈❈ ❉❉ ❈❈ ❉❉ ! ❉! ❴ o ❴ ❴ ❴ o ❴ ❴ ❴ o ❴ ❴ 0,1 1,0 2,3 ❉ 1,2 ❈ O ❉ O ❈ O ❈❈❈ O ❈❈ ❉❉ ❈❈ ❈❈ ❉ ❉❉ ❈❈ ❈! ❈! ❉! o❴ ❴ ❴ 12,3 o❴ ❴ ❴ 1,3 o❴ ❴ ❴ 0,2 2,0 O ❉❉❉ O ❈❈❈ O ❈❈❈ O 2,1 2,1 1,2❈ 1,2❉ 1,2❈ 2,1 ❈❈ 2,1 ❈❈ ❉❉ ❈ ❉ ! o❴ ❴ ❴ 2! 2 ,3 o❴ ❴ ❴ 1! 2 ,3 o❴ ❴ ❴ 0,3 3,0 ❈❈ O ❈❈ ❈❈ ❈! 3,3 3,3 5,5 D5 ✿✿ ☎A ✿ ☎☎ ☎ ❴ o ❴ ❴ ❴ ❴ o ❴ ❴ ❴ ❴ o ❴ ❴ ❴ ❴ o ❴ ❴ 4,0 0,5 5,1 > 4,2 ❆❆ ❄❄ @ 5,3 ✿✿ ❀❀ ✂@ ✿✿ ☎☎A ❆❆ ❄❄ ⑥⑥ ☎A ✂ ❀ ⑥ ☎ ✂ ❀ ✂✂ ❄ ⑥⑥ ☎ ❆  ☎ o❴ ❴ ❴ 45,1 o❴ ❴ ❴ 45,12 o❴ ❴ ❴ ❴ 45,23 o❴ ❴ ❴ 0,3 3,0 A H ✳❂ ? G ✵❆ > F ✳❃ H ✲ ✿✿ ⑥⑥ ✍ ✳ ❃❃ ✄✄ ✑ ✳ ❂❂ ⑧⑧ ✎ ✵ ❆❆ ☎A H ✿✿✿  ✑✑✑ ✲✲✲ ✿ ✄✄✑✑✑ ✳✳✳ ❂❂ ⑧⑧⑧✎✎✎ ✵✵✵ ❆❆ ⑥⑥⑥ ✍✍✍ ✳✳✳ ❃❃ ☎☎☎✑✑✑ ✲✲ 5,0 ✑o❴✑ ❴ ❴ ❴✳✳ 4,1 o❴✎✎ ❴ ❴ ❴✵✵ 5,2 o❴✍✍ ❴ ❴ ❴✳ 4,3 o❴✑ ❴ ❴ 0,4 ✑✑✑ ✑ ✳ ✿ ✲✲ ✑✑ ✳✳ ✎✎ ✵✵ ✍✍ ✳✳ ✑✑ ✿✿ ☎A ✲ ✑✑ ✳ ✎✎ ✵ ✍✍ ✑✑  ✑ ✿ ☎☎ o❴ ❴ ❴ 3,1 o❴ ❴ ❴ ❴ 45,2 o❴ ❴ ❴ ❴ 45,13 o❴ ❴❴ ❴ 0,2 4,4 2,0 ❆❆ ❄❄ A ❀❀ @ @ A☎ ✿✿ ⑥> ❆❆ ❀❀ ✂✂✂ ❄❄ ✿✿ ☎☎☎☎ ⑥⑥ ❀ ❄ ☎☎ ⑥ ✂    ☎ ✂ ⑥ o❴ ❴ ❴ 2,1 o❴ ❴ ❴ ❴ 3,2 o❴ ❴ ❴ ❴ 45,3 o❴ ❴ ❴ ❴ 0,1 1,0 ❂❂ A ✿✿ A ? ✿✿ ✿✿ ✄✄✄ ❂❂ ⑧⑧⑧ ✿ ☎☎☎☎  ✄  ⑧ 1,1 2,2 3,3 2,2 F4 1,1 O ❉❉❉❉ O ❉❉❉❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉" ❉❉ " ❴ o ❴ ❴ ❴ ❴ o ❴ ❴ ❴ o ❴ ❴ ❴ ❴ o ❴ ❴ ❴ o ❴ ❴ ❴ o ❴ ❴ 2 2 2 2 1,0 1,2 ❉ 0,1 2,3 1,4 2,4 12,3 O ❉❉ O ❉❉ O ❉❉ O ❉❉ O ❉❉ O ❉❉❉❉ O ❉ ❉ ❉ ❉ ❉ ❉ ❉ ❉ ❉ ❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉" ❉" ❉" ❉" " " ❴ o ❴ ❴ 0,2 ❴ o ❴ ❴ ❴ o ❴ ❴ o ❴ ❴ o ❴ ❴ ❴ o ❴ 2 2 2 2 2 2 2 2 2 2 2,0 12,4 12 ,3 4 1 2,3 4 12,3 4 1,3 O ❉❉❉ O ❉❉❉ O ❉❉❉ O ❉❉❉ O ❉❉❉ O ❉❉❉❉ O ❉ ❉ ❉ ❉ ❉ 1,2 1,2 1,2 2,1❉ 1,2 1,2 2,1❉ 2,1❉ 2,1❉ 2,1❉ 2,1❉ 1,2 ❉❉ 1,2 ❉❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉" ❉❉ ❉❉ ❉❉ ❉ ❉❉ " " " o❴ ❴ ❴ ❉" " o❴ ❴ ❴ ❴ 2,4 o❴ ❴ ❴ 12,34 2 o❴ ❴ ❴ 12,32 4 o❴ ❴ ❴ 1,34 o❴ ❴ ❴ 0,3 12,34 3,0 O ❉❉❉ O ❉❉ O ❉❉ O ❉❉ O ❉❉ O ❉❉❉❉ O ❉❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉ ❉ ❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉❉ " o❴ ❴ ❴ ❴❉" o❴ ❴ ❴ ❴ ❉" o❴ ❴ ❴ ❴ ❉" o❴ ❴ ❴ ❴ ❉" o❴ ❴ ❴ ❴ " ❴ o ❴ ❴ ❴ 4,0 3,4 ❉ 2,3 0,4 1,4 2,34 1,3 ❉❉ O ❉❉ O ❉❉ ❉ ❉❉ ❉❉ ❉❉ ❉❉ ❉" ❉" 4,4 3,3 27 28 JIARUI FEI 6,6 5,5 4,4 3,3 1,1 B ✿✿ ✿✿ ✿✿ ✿✿ B B ✿✿ ☎B ☎B ✿ ✿✿ ☎☎☎ ✿✿ ☎☎☎ ✿✿ ✿✿ ☎☎☎ ☎☎ ☎☎ ☎     ☎ ☎ ☎ o❴ ❴ ❴ 5,6 o❴ ❴ ❴ 4,5 o❴ ❴ ❴ 23,4 o❴ ❴ ❴ 1,6 o❴ ❴ ❴ 2,5 o❴ ❴ ❴ 3,4 o❴ ❴ ❴ 1,3 o❴ ❴ ❴ 0,1 6,0 ✿✿ B ☎B ✿✿✿ ☎B ✿✿✿ ☎B ✿✿✿ ☎B ✿✿✿ ☎B ✿✿✿ ☎B ✿✿✿ ☎B ✿✿✿ ✿ ☎☎☎ ✿ ☎☎☎ ✿✿ ☎☎☎ ✿✿ ☎☎☎ ✿✿ ☎☎☎ ✿✿ ☎☎☎ ✿✿ ☎☎☎ ✿✿ ☎☎☎  ☎ ☎      ☎ ☎ ☎ ☎ o❴ ❴ ❴ 4,6 o❴ ❴ ❴ 23,5 o❴ ❴ ❴ 123,46 o❴ ❴ 12,56 o❴ ❴ ❴ 23,45 o❴ ❴ ❴ 1,4 o❴ ❴ ❴ 0,3 5,0 B ✿✿ ☎B ✿✿✿ ☎B ✿✿✿ ☎B ✿✿✿ ☎B ✿✿✿ ☎B ✿✿✿ ☎B ✿✿✿ ✿ ☎☎☎ ✿✿ ☎☎☎ ✿ ☎☎☎ ✿✿ ☎☎☎ ✿✿ ☎☎☎ ✿✿ ☎☎☎ ✿✿ ☎☎☎ ☎  ☎ ☎ ☎  ☎    o❴ ❴ ❴ 23,6 o❴ ❴ ❴ 123,56 o❴ ❴ 122 3,456 o❴ ❴ 123,456 o❴ ❴ 12,45 o❴ ❴ ❴ 0,4 4,0 B K ✭✿ B K ✭✿ B K ✭✿ B K ✭✿ B K ✭✿ ✭✭ ✿✿ ☎☎ ✖ ✭ ✿✿ ☎☎ ✖ ✭ ✿✿ ☎☎ ✖ ✭ ✿✿ ☎☎ ✖ ✭ ✿✿ ☎B K ✭ ☎☎ ✖ ✭ ✿✿ E6 ✭✭ ✿ ☎☎ ✖✖ ✭✭ ✿✿ ☎☎☎ ✖✖ ✭✭ ✿✿ ☎☎☎ ✖✖ ✭✭ ✿✿ ☎☎☎ ✖✖ ✭✭ ✿✿ ☎☎☎ ✖✖ ✭✭ ✿ ☎☎ ✖✖✖ ✭✭✭ ✭✭ 3,0 o❴✖✖ ❴ ✭❴✭✭ 12,6 o❴✖✖ ❴ ❴✭✭✭ 23,56 o❴✖✖ ❴✭✭ 123,45 o❴✖✖ ❴✭✭✭ 12,46 o❴✖✖ ❴ ✭✭❴✭ 0,5 ✖✖ ✭✭ ✭✭☎B ✿✿✖✖ ✿✿✖✖ ✖ ✭✭ ✭✭ ✭✭B ✭✭☎B ✿✿✖✿✖ ✭✭☎B ✿✿✖✿✖✖ ✭✭☎B ✿✿✖✿✖✖ ✿ ☎ ☎ ☎ ☎ ☎ ✖ ✖ ✖ ✖ ✿✿ ☎☎ ✭✭ ✖ ✖✖ ✿✿ ☎☎☎ ✭✭ ✖✖ ✿✿ ☎☎☎ ✭✭ ✖✖ ✿ ☎☎ ✭✭ ✭✭ ✭✭ ✖✖ ✿ ☎☎ ✭✭ ✖  ☎ ✖ ✭✭ ✖ 1,0 o❴ ✭✭ ❴ ✖✖ ❴ 2,6 o❴ ✭✭ ❴ ✖✖ ❴ 3,5 o❴ ✭✭ ❴ ✖✖❴ 12,4 o❴ ✭✭ ❴ ✖✖❴ 0,6 ✭✭ ✖✖ ✭✭ ✭✭ ✖ ✭✭ ✖✖ ✭✭ ✖✖ ✭✭ ✖✖ ✭✭ ✖✖ ✭✭ ✖✖✖ ✭✭ ✖ ✭ ✖✖ ✭ ✖✖ ✭ ✖✖ ✭ ✖✖ ✭ ✖✖ ✖  o❴ ❴ ❴ 3,6 o❴ ❴ ❴ 12,5 o❴ ❴ ❴ 23,46 o❴ ❴ ❴ 1,5 o❴ ❴ ❴ 2,4 o❴ ❴❴ ❴ 0,2 2,0 ✿✿ ✿✿ ☎☎B  ☎ 2,2 Part 2. Isomorphism to k[Conf 2,1 ] 6. Cluster Structure of Maximal Unipotent Groups 6.1. Basic Notation for a Simple Lie Group. Let k be an algebraically closed field of characteristic zero. From now on, G will always be a simply connected linear algebraic group over k with Lie algebra g. We assume that the Dynkin diagram of g is the underlying valued graph of Q. The Lie algebra g has the Cartan decomposition g = n− ⊕h⊕n. Let ei , α∨ i , and fi (i ∈ Q0 ) be the Chevalley generators of g. The simple coroots α∨ of g form a basis of a Cartan subalgebra h. The simple i roots αi (i ∈ Q0 ) form a basis in the dual space h∗ such that [h, ei ] = αi (h)ei , and [h, fi ] = −αi (h)fi for any h ∈ h and i ∈ Q0 . The structure of g is uniquely determined by the Cartan matrix C(G) = (c′i,j ) given by c′i,j = αj (α∨ i ). We have ′ ′ T that ci,i = 2 and ci,j = −cj,i so we have that C(G) = El (Q) + Er (Q) . Let U − , H and U := U + be closed subgroups of G with Lie algebras n− , h and n. Thus H is a maximal torus, and U − and U are two opposite maximal unipotent subgroups of G. Let Ui± (i ∈ Q0 ) be the simple root subgroup of U ± . By abuse of notation, we let α∨ i : Gm → H be the simple coroot corresponding to the root αi : H → Gm . For all i ∈ Q0 , there are isomorphisms xi : Ga → Ui+ and xi : Ga → Ui− such that the maps       1 a 1 0 t 0 7→ xi (a), 7→ xi (b), 7→ α∨ i (t) 0 1 b 1 0 t−1 provide homomorphisms φi : SL2 → G. We denote by g 7→ g T the transpose antiautomorphism of G defined by xi (a)T = xi (a), xi (b)T = xi (b), hT = h, (i ∈ Q0 , h ∈ H). Let si (i ∈ Q0 ) be the simple reflections generating the Weyl group of G. Set si := xi (−1)xi (1)xi (−1). TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 29 The elements si satisfy the braid relations. So we can associate to each w ∈ W its representative w in such a way that for any reduced decomposition w = si1 · · · siℓ one has w = si1 · · · siℓ . Denote by w0 be the longest element of the Weyl group. In general w0 2 is not the identity but an order two central element in G. It is well-known that w0 (αi ) = −αi∗ , where i 7→ i∗ is the same involution in Section 5.1. The weight lattice P (G) of G consists of all γ ∈ h∗ such that γ(α∨ i ) ∈ Z for all i. Thus P (G) has a Z-basis {̟i }i∈Q0 of fundamental weights given by ̟j (α∨ i ) = δi,j . Q0 . We write ̟(λ) for We can thus identify a weight by an integral vector λ ∈ Z P λ(i)̟ . In this notation, ̟ = ̟(e ). We stress that throughout the paper, this i i i i identification is widely used. A weight λ ∈ ZQ0 is dominant if it is non-negative. 6.2. The Base Affine Space. The natural G × G-action on G: (g1 , g2 ) · g = g1 gg2−1 , induces the left and right translation of G on k[G]: (g1 , g2 )ϕ(g) = ϕ(g1−1 gg2 ) for ϕ ∈ k[G]. The algebraic Peter-Weyl theorem [29, Theorem 4.2.7] says that as a G × G-module k[G] decomposes as M k[G] = L(λ)∨ ⊗ L(λ), Q λ∈Z>00 where L(λ) is the irreducible G-module of highest-weight ̟(λ), and L(λ)∨ is its dual. Quotienting out the left translation of U − , we get a G-module decomposition of the ring of regular functions on the base affine space A := U − \G: M  k[A] = ϕ ∈ k[G] | ϕ(u− g) = ϕ(g) for u− ∈ U − , g ∈ G = L(λ). Q λ∈Z>00 Each G-module L(λ) can be realized as the subspace of k[A]: n o (6.1) L(λ) ∼ = ϕ ∈ k[A] | ϕ(hg) = h̟(λ) ϕ(g) for h ∈ H, g ∈ G . Similarly for the dual base affine space A∨ := G/U , we have that M k[A∨ ] = {ϕ ∈ k[G] | ϕ(gu) = ϕ(g) for u ∈ U, g ∈ G} = L(λ)∨ , Q λ∈Z>00 (6.2) n o L(λ)∨ ∼ = ϕ ∈ k[A∨ ] | ϕ(gh) = h̟(λ) ϕ(g) for h ∈ H, g ∈ G . Keep in mind that the G-actions on k[U − \G] and k[G/U ] are via the right and left translations respectively. We fix an additive character U − → Ga , then a coset U − g (resp. gU ) determines a point in A (resp. A∨ ). We refer readers to [28, 1.1.1] for the detail. For each fundamental representation L(ei ) and its dual L(ei )∨ , we choose a U − ∨ fixed vector ui and a U -fixed vector u∨ i , normalized by hui | ui i = 1. Following [21] we define for each fundamental weight ̟i and each pair w, w′ ∈ W the generalized i minor m̟ w,w ′ ∈ k[G] (6.3) ′ g 7→ hwu∨ i | gw ui i. ` − More concretely, the Bruhat decomposition G = w∈W U HwU implies that − any g ∈ G can be written as g = u hwu. Suppose that g lies in the open set 30 JIARUI FEI i G0 := U − HU , then the regular function m̟i := m̟ e,e restricted to G0 is given by ̟i ̟i m (g) = h (For the type Ar , this is a principal minor of the matrix g). Then the definition of m̟i can be extended to whole G as in [21, Proposition 2.4]. Proposition 6.1. For g = u− hwu ∈ G, we have that m̟i (g) = h̟i if w(̟i ) = ̟i and m̟i (g) = 0 otherwise. i The function m̟ w,w ′ is then given by ̟i −1 i m̟ gw′ ). w,w ′ (g) = m (w (6.4) It follows from (6.4) or (6.3) that i ′ Lemma 6.2. The weight for the H × H-action on m̟ w,w ′ is (w(̟i ), −w (̟i )), i.e., w(̟i ) −w ′ (̟i ) ̟i h2 mw,w′ (g). −1 i m̟ w,w ′ (h1 gh2 ) = h1 6.3. Cluster Structure on k[U ]. Now we are ready to recall the cluster algebra structure of k[U ]. For this, we associate to each non-neutral indecomposable presentation f , a generalized minor mf as follows. Suppose that e(f ) = ei , i.e., ̟i f is translated from O+ i , and moreover w(̟i ) = ̟(f), then we put mf := me,w . ̟i Note that if f is positive, then f = me,e is a principal minor. We take the iARt quiver ∆Q , and let MQ := {mf }f ∈(∆Q )0 . It is known [7, 25] that k[U ] is the upper cluster algebra with this standard seed (∆Q , MQ ). Here we again identify ∆Q as the subquiver of ∆2Q . Moreover, the cluster algebra is equal to its upper cluster algebra when Q is simply-laced [26]. Recall the weight configuration σ Q in Definition 3.12. For each vertex f ∈ (∆Q )0 , we set σ Q (f ) = e(f ) − f. By Lemma 6.2 the degree of mf for the conjugation action of H is exactly ̟(σ Q (f )). Let ρQ be the matrix with rows also indexed by (∆ P Q )0 such that ρQ (f ) is the positive root of G corresponding to f , i.e., ρQ (f ) = i dim(M (i))α(i) where M = Coker(f ). Note that by the generalized Gabriel’s theorem [9], ρQ contains exactly all positive roots of G. The Kostant’s partition function pQ (γ) by definition counts the lattice points in the polytope K(γ) := {h ≥ 0 | hρQ = γ}. Now we consider a labelling dual to the one in Section 5.1. Let (i, t) = (i(f ), t(f )) + t ∗ ∗ ∗ be the number such that   f = τ (Oi ). Note that i(f ) = i (f ) and t(f ) + t (f ) = − + ∗ t Oi∗ (f ) = t Oi(f ) . Lemma 6.3. ̟ (σ Q (f )) = Pt(f )−1 t=0 ρQ (τ −t f ). Proof. Let M = Coker(f ). The equality is equivalent to that e(f ) − f =  t(f )−1 X t=0 dim(τ −t  M ) C(G). Recall that the Cartan matrix C(G) is equal to El (Q) + Er (Q)T . It follows from the definition that matrix El (Q) transforms −dimM to the reduced weight vector of f (M ), and Er (Q)T transforms dimM to the reduced weight vector f ∨ (M ) of the minimal injective presentation of M , or equivalently, f ∨ (M ) = −f(τ −1 M ). So the TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 31 righthand side is equal to t(f )−1 X t=0 −f(τ −t M ) − f ∨ (τ −t M ). After some cancellations, only two terms survive. They are e(f ) and −f = −f(M ).  For the weight configuration ̟(σ Q ) := {̟(σ Q (f ))}f ∈(∆Q )0 and the polyhedral cone G∆Q , we consider the polytope G∆Q (γ) as in Definition 5.13. Lemma 6.4. The function |G∆Q (−) ∩ Z(∆Q )0 | is the partition function pQ (−). Pt∗ (f ) t Proof. Recall that G∆Q (γ) is defined by t=0 g(τ f ) ≥ 0 for f ∈ (∆Q )0 and g̟(σ Q ) = γ. We introduce new variables h(f ) for each f satisfying that h(f ) = Pt∗ (f ) t t=0 g(τ f ). Then ∗ hρQ = (f ) X tX f t=0 g(τ t f )ρQ (f ) = )−1 X t(f X f g(f )ρQ (τ −t f ) = g̟(σ Q ). t=0 The second equality is established through an easy bijection and the last one is due to Lemma 6.3. So the defining condition for G∆Q (γ) is equivalent to that for the polytope K(γ). Finally, we observe that the transformation from g to h is totally unimodular. In particular, the transformation and its inverse preserve lattice points.  Theorem 6.5. The coordinate ring of U is equal to the graded cluster algebra C (∆Q , MQ ; ̟(σ Q )). Moreover, if Q is trivially valued, then (∆Q , WQ ) is a cluster model (Definition 4.11). Proof. We only need to prove the second statement. Recall that the universal enveloping algebra U (n) has a standard grading by deg(ei ) = αi . It is a classical fact [32] that the partition function pQ (γ) counts the dimension of the homogeneous component U (n)γ . The algebra k[U ] is graded dual to U (n) with (see [44])  k[U ]γ = ϕ ∈ k[U ] | ϕ(huh−1 ) = hγ ϕ(u) for h ∈ H, u ∈ U . We have seen that the degree of MQ (f ) = mf is ̟(σ Q (f )). So the second statement follows from Theorem 4.10, Lemma 6.4, and the fact that C(∆Q ) = C(∆Q ).  There are another two seeds of this cluster structure of k[U ]. One is called left standard, and the other is called right standard. Both are obtained from the standard seed by a sequence of mutations. The sequences of mutations µl and µr will be defined and studied in Appendix B. Let  e #, W f # , M# ; σ # = µ (∆Q , WQ , MQ ; σ Q ) for # = l, r. ∆ # Q Q Q Q For what follows in this Section, # always represents l and r. Let ∆lQ (resp. ∆rQ ) be the ice quiver obtained from ∆2Q by deleting the negative and positive (resp. e # and ∆# are equal up negative and neutral) frozen vertices. By Corollary B.11, ∆ Q Q e #, W f # ). to arrows between frozen vertices. Since the QP (∆Q , WQ ) is rigid, so is (∆ Q Q # # Let WQ# be the potential WQ2 restricted to ∆# Q . By [12, Proposition 8.9] (∆Q , WQ ) 32 JIARUI FEI e #, W f # ) and (∆# , W # ) are µ-rigid. By the is also rigid. In particular, both (∆ Q Q Q Q e #, W f # ) is equivalent definition of rigidity ([12, Definition 6.10, Remark 6.8]), (∆ Q Q # µ-right-equivalent to (∆# Q , WQ ). [14, Proposition 5.15] implies that being a cluster model is mutation-invariant. f # ) is also a cluster model. In view of Remark 4.12, we can e #, W It follows that (∆ Q Q e #, W f # ) by (∆# , W # ). replace (∆ Q Q Q Q Proposition 6.6. (∆lQ , MlQ ; σ lQ ) and (∆rQ , MrQ ; σ rQ ) are another two graded seeds for the cluster algebra k[U ]. Moreover, if Q is trivially valued, both (∆lQ , WQl ) and (∆rQ , WQr ) are cluster models. Remark 6.7. Later we will also need a concrete description of G(∆lQ , WQl ) and G(∆rQ , WQr ). Both are given by lattice points in rational polyhedral cones. Recall that the three sets of defining conditions of the polyhedral cone G∆2Q gHu ≥ 0, gHl ≥ 0, gHr ≥ 0. l l We define the polyhedral cones G∆lQ ⊂ R(∆Q )0 and G∆rQ ⊂ R(∆Q )0 by the relations gHl ≥ 0 and gHr ≥ 0 respectively. Almost the same proof as Theorem 5.8 can r l show that G(∆lQ , WQl ) = G∆lQ ∩ Z(∆Q )0 and G(∆rQ , WQr ) = G∆rQ ∩ Z(∆Q )0 . 7. Maps Relating Unipotent Groups 7.1. Standard Maps. Let Conf n,1 := (An × A∨ )/G be the categorical quotient in the category of varieties 4. Lemma 7.1. The ring of regular functions on Conf n,1 is a unique factorization domain. Proof. It is well-known [33] that k[G] is a UFD. Since the group U and G have no multiplicative characters, by [42, Theorem 3.17] k[Conf n,1 ] is also a UFD.  By the Bruhat decomposition, any pair (A1 , A∨ 0 ) ∈ Conf 1,1 has a representative (U hw, U ) for some h ∈ H, w ∈ W . A pair (A1 , A∨ 0 ) ∈ Conf 1,1 is called generic if the w ∈ W can be chosen as the identity. Let Conf ◦n,1 be the open subset of Conf n,1 where each pair (Ai , A∨ 0 ) is generic (but we do not impose any condition among ◦ Ai ’s). By definition, we have an isomorphism H ∼ = Conf 1,1 . Let ι : H ֒→ Conf 1,1 be − the open embedding h 7→ (U h, U ). For each i ∈ Q0 we define the regular function ̟ e i on Conf 1,1 by ̟ e i (U − g1 , g0 U ) = m̟i (g1 g0 ). It is clear that the definition does not depend on the representatives. It follows from Proposition 6.1 that − Lemma 7.2. k[H] is exactly the localization of k[Conf 1,1 ] at all ̟ e i. From now on, we will focus on the space Conf 2,1 . Its ring of regular functions has a triple-weight decomposition M G k[Conf 2,1 ] = (L(µ) ⊗ L(ν) ⊗ L(λ)∨ ) . Q µ,ν,λ∈Z>00 λ We write very often Cµ,ν for the graded component k[Conf 2,1 ]µ,ν,λ . It is clear that λ λ cµ ν = dim Cµ,ν . 4For our purpose, it is enough to work with the categorical quotient, which has the same ring of regular functions on the corresponding quotient stack. TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 33 We recall several rational maps defined in [18, 28]. Let i : H × H × U ֒→ Conf 2,1 be the open embedding (h1 , h2 , u) 7→ (U − h1 , U − h2 u, U ). It is an embedding because the stabilizer of the generic pair (U − h, U ) is 1G . It is clear that the image of i is exactly Conf ◦2,1 . By restriction we get an embedding i∗ : k[Conf 2,1 ] ֒→ k[H × H × U ]. We will view two H’s and U as subgroups of H × H × U in the natural way, and write i1 , i2 and iu for the restriction of i on λ first H, second H and U respectively. A function s in Cµ,ν is uniquely determined ∗ by its restriction iu (s) on U because (7.1) ∗ ̟(µ) ̟(ν) h2 s(u) s(h1 , h2 , u) = h1 for (h1 , h2 , u) ∈ H × H × U. λ into k[U ] (in fact into k[U ]̟(µ+ν−λ) by embeds Cµ,ν λ that each Cµ,ν is not disjoint under this embedding. easy calculation). We So i note We recall a classical interpretation of the multiplicity cλµ ν . Let L(µ)γ be the weight-γ subspace of the irreducible G-module L(µ). We denote o n ν(i)+1 (ϕ) = 0 for i ∈ Q0 . L(µ)νγ := ϕ ∈ L(µ)γ | ei Lemma 7.3. [40] We have that cλµ ν = dim L(µ)ν̟(λ−ν) . Lemma 7.4. For any f ∈ ind(C 2 Q), the generalized minor mf spans the space f+ f− ⊂ k[A]. In particular, ce(f L (e(f ))̟(f) ),f− = 1. Proof. Recall that the Chevalley generator ei acts on k[G] by d ϕ(gxi (t)). dt t=0 For any ϕ ∈ k[G], the coefficient of each tn in ϕ(xi (t)) is equal to eni ϕ(1G )/n!. If ϕ is homogeneous of degree γ, then ϕ(1G ) can be nonzero only if γ = 0. It follows that ϕ(xi (t)) contains only tn with nαi = γ. So for γ = ̟(e(f ) − f) ei ϕ(g) = f (i)+1 ei− ϕ = 0 if and only if (f− (i) + 1)ci 6= e(f ) − f, where ci is the i-th column of the Cartan matrix C(G). Since e(f )(i) ≤ 1 and ci (i) = 2, we can never have (e(f ) + f− − f+ )(i) = (f− (i) + 1)ci (i). Hence we proved f− f (i)+1 . mf = 0 for i ∈ Q0 so that mf ∈ L (e(f ))̟(f) that ei− ̟j . The generalized minor mf spans the space because Suppose that mf = me,w dim L(ej )w(̟j ) = dim L(ej )̟j = 1.  We recall that each mO+ is a principal minor, which evaluates to 1 on U ; and now i we set mIdi = 1. Lemma-Definition 7.5. For any f ∈ ind(C 2 Q), there is a unique function sf ∈ k[Conf 2,1 ]e(f ),f− ,f+ such that i∗u (sf ) = mf . Q µ(i) λ Moreover, s ∈ Cµ,ν satisfies i∗u (s) = 1 if and only if s = sµ+ sν0 , where sµ+ := i sO+ i Q ν(i) and sν0 := i sIdi . 34 JIARUI FEI Proof. The first statement follows from Lemma 7.4. For the last statement, it suffices to prove the “only if” part. If i∗u (s) = 1, then µ + ν = λ because i∗u (s) has degree ̟(µ+ ν − λ). It is clear that cµ+ν µ,ν = 1 (consider tensoring the highest weight vectors of L(µ) and L(ν)), but sµ+ sν0 also has degree (µ, ν, µ + ν), so s = sµ+ sν0 .  Remark 7.6. For Q of type A, the analogous map i∗ was also considered in a quiverinvariant theory setting (see [14, Example 3.10, 3.12]). In that setting, the map comes from a semi-orthogonal decomposition of the module category of a triple flag quiver. Corollary 7.7. For each f ∈ ind(C 2 Q), sf is irreducible in k[Conf 2,1 ]. Proof. It is clear that the zero-degree component of k[Conf 2,1 ] is k. If f is positive or neutral, then the degree of f is obviously indecomposable. The same argument as in [14, Lemma 1.8] shows that sf is irreducible. We remain to consider the case when f is the minimal presentation of a general representation. It is known that all generalized minors mf are irreducible in k[U ]. Suppose that sf factors as sf = s1 s2 , then i∗u (sf ) = i∗u (s1 )i∗u (s2 ) = mf . So one of them, say i∗u (s1 ), has to be a unit. According to Lemma 7.5, s1 is a polynomial in sO+ ’s and sIdi ’s. So there is a homogenous component s′2 of s2 and some µ and i f + ν such that s′2 sµ+ sν0 has the same degree as sf , i.e., s′2 sµ+ sν0 ∈ Ce(f ),f− . Since f is a minimal presentation of a general representation, we must have ν = 0. If µ is nonzero, then we must have µ = e(f ). Then the weight of s′2 is (0, f− , f+ − e(f )), f+ and f− has to be equal to f+ − e(f ). But i∗u embeds Ce(f ),f− into k[U ]e(f )−f . Since f is not positive or neutral, e(f ) 6= f, which is a contradiction.  Let p be a rational inverse of i, which is regular on Conf ◦2,1 = i(H × H × U ). Let p1 , p2 and pu be the composition of p with the natural projection to first H, second H, and U respectively. Corollary 7.8. On the open subset Conf ◦2,1 ⊂ Conf 2,1 −1  e(f ) f , p∗u (mf ) = sf s+ s0− p∗1 (̟i ) = sO+ , p∗2 (̟i ) = sIdi . i for any indecomposable presentation f . Proof. They follow from straightforward calculation (p∗u mf ) (i(h1 , h2 , u)) = mf (pu i(h1 , h2 , u)) = mf (u); ̟(e(f )) ̟(f− ) e(f ) f ̟(e(f )) ̟(f− ) h2 mf (u), (s+ s0− )(i(h1 , h2 , u)) = h1 h2 . ∗ ∗ statement for p1 , p2 is rather obvious. For example, ∗ i 0 ∗ i sO+ (i(h1 , h2 , u)) = h̟ )(u) = h̟ 1 h2 iu (sO+ 1 = p1 (̟i ) (i(h1 , h2 , u)) . i i sf (i(h1 , h2 , u)) = h1 The  Corollary 7.9. The localization of k[Conf 2,1 ] at all sf ’s for f positive and neutral is exactly k[H × H × U ]. Proof. Let pe1 and pe2 be the natural projections Conf 2,1 → Conf 1,1 given by ◦ ∨ ∨ ∨ (A1 , A2 , A∨ 0 ) 7→ (A1 , A0 ) and (A1 , A2 , A0 ) 7→ (A2 , A0 ). Then Conf 2,1 is the in◦ ◦ e∗1 (̟ e i ) = sO+ and e−1 tersection pe−1 2 (Conf 1,1 ). It suffices to show that p 1 (Conf 1,1 ) ∩ p i TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 35 pe∗2 (̟ e i ) = sIdi . The map ιpi agrees with pei on a dense subset of Conf 2,1 for i = 1, 2. So pe∗1 (̟ e i ) = p∗1 ι∗ (̟ e i ) = p∗1 (̟i ) = sO+ and pe∗2 (̟ e i ) = p∗2 ι∗ (̟ e i ) = p∗2 (̟i ) = sIdi .  i 2 Corollary 7.10. The set of functions SQ := {sf }f ∈ind(C 2 Q) is algebraically independent over the base field k. Proof. Since the map p : Conf 2,1 → H × H × U is birational, the pullbacks of (two copies of) ̟i ∈ H and mf ∈ U are algebraically independent. We have seen that 2 up to a factor of some monomial in sO+ , sIdi ∈ SQ , they are exactly the functions i 2 in SQ . Our claim follows.  Suppose that f as above is mutable. Let m′f be a function in k[U ] in the exchange relation Y Y (7.2) mf m′f = mg + mh . g֌f ′ ′ ′ Let (µ , ν , λ ) = P f ֌h (e(h), h− , h+ ) f ֌h − (e(f ), f− , f+ ). ′ ′ Lemma 7.11. The function s′f := p∗u (m′f )sµ+ sν0 is regular on Conf 2,1 satisfying i∗u (s′f ) = m′f and the exchange relation Y Y sf s′f = sg + sh . g֌f Moreover, s′f f ֌h and sf are relatively prime in k[Conf 2,1 ]. Proof. Since i is a regular section of p, we have that i∗u (s′f ) = m′f and s′f is regular ′ ′ on Conf ◦2,1 = i(H × H × U ). We pull back (7.2) through pu , and multiply sµ+ sν0 on both sides. Then we get from Corollary 7.8 that   Y Y ′ ′ sg + sh . (7.3) sf p∗u (m′f )sµ+ sν0 = g֌f f ֌h We remain to show that s′f is in fact regular on whole Conf 2,1 . By (7.3), the locus of indeterminacy of s′f is contained in the zero locus Z(sf ) of sf . Since sf |U = mf , the intersection of Z(sf ) and Conf ◦2,1 is nonempty. In other words, Z(sf ) is not contained in the complement of Conf ◦2,1 . Since Z(sf ) is irreducible (Corollary 7.7), we conclude that s′f is regular outside a codimension 2 subvariety of Conf 2,1 . Since Conf 2,1 is factorial (Lemma 7.1) and thus normal, by the algebraic Hartogs s′f is a regular function on Conf 2,1 . For the last statement, we suppose the contrary. Since sf is irreducible, sf is then a factor of s′f . But this is clearly impossible by comparing the first component of their triple-weight.  7.2. Twisted Maps. Let ∨ (written exponentially) be the morphism U − \G → G/U given by (U − g)∨ → 7 g T U . Let ˜l be the twisted cyclic shift on A × A × A∨ : ∨ (A1 , A2 , A∨ 3 ) 7→ (A2 , A3 w0 , (A1 w0 ) ). We have that 2 ∨ ˜ l2 (A1 , A2 , A∨ 3 ) = (A3 w0 , A1 w0 , (A2 w0 ) ), 36 JIARUI FEI and l̃3 is the identity. By the universal property of categorical quotients, ˜l induces an endomorphism l on Conf 2,1 , which is an automorphism on the generic part of Conf 2,1 . We define il := liu : U ֒→ Conf 2,1 , that is u 7→ (U − u, U − w0 , w0 T U ). We write r for l2 . Similarly we denote ir := riu : U ֒→ Conf 2,1 . Conf 2,1 O [✽ ✝✝ ✽✽✽ ✝ ✽✽ ✝ iu ✽✽ ✝✝ ✝ ✽✽l l ✝✝ ✝ ✽ U ❏❏ ✝ t ❏❏ ✽✽✽ ✝ tt ✝ ❏ t ✝ ❏❏ ✽✽ t ❏❏ ✽ ✝✝ tttit ❏❏ ✽✽ ✝ i r l t ❏% ✝✝ytt / Conf 2,1 Conf 2,1 l It turns out that the twisted cyclic shift is also related to the sequence µl of mutations considered in Section 6.3. We will see in Theorem B.14 that l∗ (sf ) = µl (sf ) for f mutable. Moreover, µl (∆2Q ) = π 2 (∆2Q ), where π 2 is a relabelling of frozen vertices of ∆2Q . Lemma 7.12. We have the following i∗l (sf ) = µl (mf ), i∗r (sf ) = µr (mf ) for f mutable; i∗l (sIdi ) = mO−∗ , i∗l (sO− ) = i∗l (sO+ ) = 1; i i i i∗r (sO+ ) = mO− , i∗r (sO− ) = i∗r (sIdi ) = 1. i i i Proof. First consider the case when f is mutable, then l∗ is given by a sequence of mutations µl . The pullback i∗u clearly commutes with any sequence of mutations, so by Lemma 7.5 i∗l (sf ) = i∗u l∗ (sf ) = i∗u (µl (sf )) = µl (i∗u (sf )) = µl (mf ). For f frozen, by Lemma 7.5 we have that i∗l (sIdi ) = i∗u (sO−∗ ) = mO−∗ , i∗l (sO− ) = i∗u (sO+ ) = 1, i∗l (sO+ ) = i∗u (sIdi∗ ) = 1. i i The argument for i∗r is similar. i i i  We set ∆uQ := ∆Q and MuQ := MQ . We define three linear natural projections 2 # i∗# : R(∆Q )0 → R(∆Q )0 for # ∈ {u, l, r}, where i∗u (resp. i∗l ; i∗r ) forgets the coordinates corresponding to the neutral and positive (resp. positive and negative; negative and neutral) frozen vertices. Recall the definition of the g-vector with respect to a pair (∆, x) as in Section 2 2 2 1.3. Let L(SQ ) be the Laurent polynomial ring in SQ . Note that L(SQ ) has a basis 2 2 parameterized by all possible g-vectors (with respect to (∆Q , SQ )), which can be 2 identified with Z(∆Q )0 . 2 Lemma 7.13. If s ∈ L(SQ ) ⊂ k(Conf 2,1 ) has a well-defined g-vector g with respect # 2 2 to (∆Q , SQ ), then so is the g-vectors of i∗# (s) with respect to (∆# Q , MQ ), which is ∗ equal to i# (g) for # ∈ {u, l, r}. TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 37 Proof. We will only prove the statement for il because the argument for the other two is similar. Suppose that s = xg F (y) where x(f ) = sf and y is as in Section 1.3. We have seen in Lemma 7.12 that i∗l (sf ) = µl (mf ) = MlQ (f ) if f is not negative ∗ or positive. If f is negative or positive, then i∗l (sf ) = 1 so i∗l (xg ) = (xl )il (g) where xl (f ) = µl (mf ). Recall that the quiver ∆lQ is obtained from ∆2Q by deleting the negative and positive frozen vertices. According to this description, we have that −blf i∗l (y(f )) = i∗l (x−bf ) = xl = yl (f ), where bf (resp. blf ) is the row of B∆2Q (resp. ∗ B∆lQ ) corresponding to f . Hence, we get i∗l (s) = (xl )il (g) F (yl ).  Lemma 7.14. The polytope G∆2Q (µ, ν, λ) has lattice points no less than cλµ ν . 2 Proof. By Corollary 7.9, we have that k[Conf 2,1 ] ⊂ L(SQ ). Since iu , il , ir are all regular, we trivially have that  2 ) | i∗# (s) ∈ k[U ] for # = u, l, r . (7.4) k[Conf 2,1 ] ⊆ s ∈ L(SQ 2 If s ∈ L(SQ ) has a well-defined g-vector, then by the previous lemma the g-vector ∗ of i# (s) is equal to i∗# (g). So if i∗# (s) ∈ k[U ], then i∗# (g) ∈ G∆# by Lemma 1.9, Q Proposition 5.12, 6.6 and Remark 6.7. Due to (7.4), it suffices to show that the lattice points in G∆2Q (µ, ν, λ) can be 2 identified with points g ∈ Z(∆Q )0 of weight (µ, ν, λ) such that i∗# (g) ∈ G∆# for Q # ∈ {u, l, r}. But this follows from the description of the cones G∆2Q , G∆Q , G∆lQ and G∆rQ (Theorem 5.8, Proposition 5.12, and Remark 6.7).  8. Cluster Structure on k[Conf 2,1 ] Theorem 8.1. Suppose that Q is trivially valued. Then the ring of regular func2 tions on Conf 2,1 is the graded upper cluster algebra C(∆2Q , SQ ; σ 2Q ). Moreover, (∆2Q , WQ2 ) is a cluster model. In particular, cλµ ν is counted by lattice points in G∆2Q (µ, ν, λ). 2 Proof. By Lemma 7.10 and Corollary 3.13, (∆2Q , SQ ; σ 2Q ) form a graded seed. Due to Lemma 7.11 and Corollary 7.7, we can apply Lemma 1.8 to conclude that 2 C(∆2Q , SQ ; σ 2Q ) is a graded subalgebra of k[Conf 2,1 ].   2 By Theorem 4.10, CW G∆2Q (µ, ν, λ) ∩ Z(∆Q )0 is a linearly independent set λ in Cµ,ν for any triple weights. (∆2Q )0 G∆2Q (µ, ν, λ) ∩ Z is at least But Lemma 7.14 says that the cardinality of cλµ ν . λ So they actually span the vector space Cµ,ν . 2 We conclude that C(∆2Q , SQ ; σ 2Q ) is equal to k[Conf 2,1 ], and (∆2Q , WQ2 ) is a cluster model.  2 ; σ 2Q ) is It follows that (7.4) is in fact an equality. The proof shows that C(∆2Q , SQ a graded subalgebra of k[Conf 2,1 ] no matter if Q is trivially valued. Conjecture 8.2. The first and last statement of Theorem 8.1 is true even if Q is not trivially valued. We illustrate by example that in general the upper cluster algebra C(∆2Q ) strictly contains the cluster algebra C(∆2Q ). 38 JIARUI FEI Example 8.3. Let Q be of type D4 as in Example 3.8. It can be checked by the 44 inequalities in Example 5.9 that g = e34,12 − e34,1 + e2,0 (in weight labelling) is µ-supported. Its triple weights is (e2 , e2 , e2 ), and moreover cee22 ,e2 = 1. This is clearly an extremal weight. So it suffices to show that no cluster variable has this g-vector. For this, we need a result in [10], which says that the g-vector of a cluster variable is a g-vector of a rigid presentation in the Jacobian algebra. The rigidity is characterized by the vanishing of the E-invariant introduced in [13, 10]. It is not hard to show that E(f, f ) := HomK b (J) (f, f [1]) 6= 0 where f is a general presentation in HomJ (P ([g]+ ), P ([g]− )). 9. Epilogue 9.1. Cluster Structure of Base Affine Spaces. We denote the ice quiver obtained from ∆2Q by deleting neutral frozen vertices by ∆♯Q . Let M♯Q := {mf }f ∈(∆♯ )0 . Q We define the weight configuration σ ♯Q on ∆♯Q by σ ♯Q (f ) = (e(f ), f). Recall from Lemma 6.2 that the degree of mf is ̟(σ ♯Q (f )). As a side result, we will show that the ring of regular functions on the base affine space A := U − \G is the graded upper cluster algebra C(∆♯Q , M♯Q ; ̟(σ ♯Q )). The proof is similar to but easier than that for Conf 2,1 so our treatment may be a little sketchy. Recall the open set G0 = U − HU of G. We have an open embedding ι0 : H × U = U − \G0 ֒→ U − \G. The localization of k[A] at all mf ’s for f positive is exactly k[H × U ]. In particular, k[A] is contained in the Laurent polynomial ring in M♯Q . Consider the open embedding ι = (ι1 , ι2 ) : H × A ֒→ Conf 2,1 , (h, U − g) 7→ (U − h, U − g, U ). Note that the map ι(IdH , ι0 ) : H × H × U → Conf 2,1 is the map i defined in Section 7.1. Let p be a rational inverse of ι, and p2 be the map p followed by the second component projection. We define the birational map r′ : U − \G → U − \G to be the composition r′ := p2 rι2 . The map r′ can be viewed as a variation of FominZelevinsky’s twist automorphism [21] on the big cell U − \G0 . Readers can check that they differ by a (fibrewise) rescaling along the toric fibre of H × U , but we do not need this fact. Let iu′ := ι0 |U , and ir′ := r′ iu′ . Then we have the following commutative diagram. The map ir′ is in fact regular because r′ is regular on the image of iu′ . r Conf 2,1 c●● ●●iu ●● ●● 1Q p2 Mm U  ✇ iu′ ✇✇✇ ✇✇  {✇✇ r′ U − \G / Conf 2,1 ; ir ✇✇✇ ✇ ✇ ✇✇ -✇ p2 q●● ●● ir′ ●● ●● #  / U − \G To finish the proof, we only need three maps iu′ , ir′ and r′ . Analogous to Theorem B.14, we have that the pullback (r′ )∗ is related to the sequence of mutations µr . More precisely, we have that µr (mf ) = (r′ )∗ (mf ) for any f mutable. TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 39 So analogous to Lemma 7.12, we have that i∗r′ (mf ) = (r′ )∗ (mf ) = µr (mf ) for f mutable. Obviously we also have that i∗r′ (mO+ ) = mO− and i∗r′ (mO− ) = 1. i i ♯ i Analogous to i∗u and i∗r , we have the maps i∗u′ : R(∆Q )0 → R(∆Q )0 and i∗r′ : ♯ r R(∆Q )0 → R(∆Q )0 . The map i∗u′ (resp. i∗r′ ) forgets the coordinates corresponding to the positive (resp. negative) frozen vertices. We have the following analog of Lemma 7.13. If m ∈ A has a well-defined g-vector g with respect to (M♯Q , ∆♯Q ), then so are the g-vectors of i∗u′ (m) and i∗r′ (m) with respect to (MQ , ∆Q ) and (MrQ , ∆rQ ). They are equal to i∗u′ (g) and i∗r′ (g) respectively. We restrict the iARt QP (∆2Q , WQ2 ) to the subquiver ∆♯Q . We denote the restricted potential by WQ♯ . Almost the same proof as in Theorem 5.8 shows that the set G(∆♯Q , WQ♯ ) of µ-supported g-vectors are given by the lattice points in the polyhedral cone G∆♯ . The cone is defined by the two sets of defining conditions of Q G∆2Q , namely, gHu ≥ 0 and gHr ≥ 0. By almost the same argument, we get the following analog of Lemma 7.14. For the weight configuration ̟(σ ♯Q ) we define the polytope G∆♯ (µ, λ) as in Definition 5.13. Q Lemma 9.1. The polytope G∆♯ (µ, λ) has lattice points no less than dim k[A]µ,λ . Q Theorem 9.2. Suppose that Q is trivially valued. Then the ring of regular functions on A is the graded upper cluster algebra C(∆♯Q , M♯Q ; ̟(σ ♯Q )). Moreover, (∆♯Q , WQ♯ ) is a cluster model. In particular, the weight multiplicity dim L(µ)̟(λ) is counted by lattice points in G∆♯ (µ, λ). Q Proof. We know from [7] that (∆♯Q , M♯Q ) is a seed satisfying the condition of Lemma 1.8. By Lemma 1.8, C(∆♯Q , M♯Q ; ̟(σ ♯Q )) is a graded subalgebra of k[A].   ♯ By Theorem 4.10, CW G∆♯ (µ, λ) ∩ Z(∆Q )0 is a linearly independent set in Q k[A]µ,λ for any µ, λ. But Lemma 9.1 implies that they actually span the vector space k[A]µ,λ . We conclude that C(∆♯Q , M♯Q ; ̟(σ ♯Q )) is equal to k[A], and (∆♯Q , WQ♯ ) is a cluster model.  Remark 9.3. For Q of type A, the theorem was proved in [15, Example 3.10] using the technique of projection. The map p2 is induced from a semi-orthogonal decomposition of the module category of a triple flag quiver. In general, the cluster algebra C(∆♯Q ) is also strictly contained in its upper cluster algebra. The example is still given by the Q of type D4 . We take the same g-vector as in Example 8.3. Remark 9.4. Let K 2 Q := K 2 (proj -Q) be the homotopy category of C 2 Q as in [10]. Then the ice quiver ∆♯Q can be obtained from the ARt quiver ∆(K 2 Q) by freezing the negative and positive vertices. There are two other seeds mutation equivalent to this seed via µl and µr . Their quivers are obtained from ∆2Q by deleting the positive and negative frozen vertices respectively. 9.2. Remark on the Non-simply Laced Cases. Suppose that G is not simply laced, or equivalently Q is not trivially valued. To prove Conjecture 8.2, let us exam the arguments in proving Theorem 8.1. We find that there are two missing 40 JIARUI FEI ingredients. One is the straightforward generalization of Theorem B.14, and the other is an analogous cluster character for species with potentials. Although the proof of Theorem B.14 depends on a result involving preprojective algebras, the author has a much longer proof for all cases using only a conjectural generalization of Lemma B.3 to the species with potentials. For the missing cluster character, as proved in [14] for the usual QPs, the existence of such a map to the upper cluster algebra is also equivalent to Lemma B.3. The argument there can be generalized. So the upshot is that the only missing part for Conjecture 8.2 is certain generalization of Lemma B.3. We expect that such a result will appear in the subsequent paper of [37]. B. The Twisted Cyclic Shift via Mutations B.1. Mutation of Representations, g-vectors, and F -polynomials. We review some material from [13]. Let (∆, W ) be a QP as in Section 4.1. The mutation µu of (∆, W ) at a vertex u is defined as follows. The first step is to define the e W f ). We put ∆ e 0 = ∆0 and ∆ e 1 is the union of following new QP µ eu (∆, W ) = (∆, three different kinds • all arrows of ∆ not incident to u, • a composite arrow [ab] from t(a) to h(b) for each a, b with h(a) = t(b) = u, • an opposite arrow a∗ (resp. b∗ ) for each incoming arrow a (resp. outgoing arrow b) at u. e is the result of first two steps in Definition 1.3. The new potential Note that this ∆ e is given by on ∆ X f := [W ] + b∗ a∗ [ab], W h(a)=t(b)=u where [W ] is obtained by substituting [ab] for each words ab occurring in W . Finally e W f ). we define (∆′ , W ′ ) = µu (∆, W ) as the reduced part ([12, Definition 4.13]) of (∆, For this last step, we refer readers to [12, Section 4,5] for details. Now we start to define the mutation of decorated representations of (∆, W ). Consider the resolution of the simple module Su (B.1) (B.2) ··· → M h(a)=u a (∂[ab] )b Pt(a) −−−−−→ (a)a 0 → Su → Iu −−−→ M h(a)=u M t(b)=u b (b) Ph(b) −−→ Pu → Su → 0, a (∂[ab] )b It(a) −−−−−→ M t(b)=u Ih(b) → · · · , where Iu is the indecomposable injective representation of J(∆, W ) corresponding to a vertex u. We thus have the triangle of linear maps with βu γu = 0 and γu αu = 0. ♣♣8 M (u) αu ♣♣♣ L ♣♣♣ ♣♣♣ o h(a)=u M (t(a)) γu ◆◆◆ ◆◆◆βu ◆◆◆ ◆◆ L & t(b)=u M (h(b)) TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 41 f = (M f, M f+ ) of µ We first define a decorated representation M eu (∆, W ). We set f(v) = M (v), M f+ (v) = M + (v) (v 6= u); M Ker βu f(u) = Ker γu ⊕ Im γu ⊕ Ker αu ⊕ M + (u), M f+ (u) = M . Im βu Im γu Ker βu ∩ Im αu f(a) = M (a) for all arrows not incident to u, and M f([ab]) = M (ab). It We then set M ∗ ∗ f(a ), M f(b ) making M f a representation is defined in [12] a choice of linear maps M e f of (∆, W ). We refer readers to [13, Section 10] for details. Finally, we define f M′ = µu (M) to be the reduced part ([12, Definition 10.4]) of M. Recall the g-vector form Definition 4.2. We can also define the dual g-vectors using the injective presentations. Let M be a representation of J(∆, W ) with ∨ ∨ minimal injective presentation 0 → M → I(β− ) → I(β+ ), then the dual g-vector ∨ ∨ ∨ ∨ g (M ) of M is the reduced weight vector g = β+ − β− . The definition can also be extended to all decorated representations similar to g-vectors. Recall the y-variables yu = x−bu as in Section 1.3. The seed mutation of Definition 1.4 induces the y-seed mutation. We recall the mutation rule [23, (3.8)]. Let (y′ , ∆′ ) := µu (y, ∆) and B(∆) = (bu,v ), then ∆′ = µu (∆) and ( yu−1 if v = u; ′ (B.3) yv = [−bu,v ]+ bu,v (yu + 1) yv yu if v 6= u. Definition B.1 ([13]). We define the dual F -polynomial of a representation M by X ∨ (B.4) FM (y) = χ(Gre (M ))ye , e where Gre (M ) is the variety parameterizing e-dimensional subrepresentations of M . P Remark B.2. The F -polynomial of M is FM (y) = e χ(Gre (M ))ye . We only need the dual version in B.3. Our dual g-vector and dual F -polynomial is the g-vector and F -polynomial in [23, 13]. Here is the key Lemma in [13]. Lemma B.3. Let M be an arbitrary representation of a nondegenerate QP (∆, W ), and let M′ = µu (M), then (1) The F -polynomials of M and M ′ are related by (B.5) ∨ ∨ ′ ∨ ∨ ′ (yu + 1)−β− (u) FM (y) = (yu′ + 1)−(β− ) (u) FM ′ (y ). (2) The g-vector and g∨ -vector of M and M′ are related by ( −g(u) if u = v; ′ (B.6) g (v) = g(v) + [bv,u ]+ g(u) + bv,u β− (u) if u 6= v, ( −g∨ (u) if u = v; ∨ ′ (B.7) (g ) (v) = ∨ ∨ ∨ g (v) + [−bv,u ]+ g (u) − bv,u β− (u) if u 6= v. Remark B.4. If M is g-coherent (that is min(β+ (u), β− (u)) = 0 for all vertices u, or equivalently, β+ = [g]+ and β− = [−g]+ ), then (B.6) reads   if u = v; −g(u) (B.8) g′ (v) := g(v) + bv,u [g(u)]+ if bv,u ≥ 0;   g(v) + bv,u [−g(u)]+ if bv,u ≤ 0. 42 JIARUI FEI It is known [14] that if M corresponds to a general presentation, then g(M) is gcoherent. If M is obtained from positive simple (0, Su ) via a sequence of mutations, then M corresponds to an indecomposable rigid presentation. In particular, it is general [10]. B.2. The Twisted Cyclic Shift via Mutations. We recall the twisted cyclic ∨ shift l on Conf 2,1 induced by (A1 , A2 , A∨ 3 ) 7→ (A2 , A3 w0 , (A1 w0 ) ). To understand this map, we introduce a “half” of this map induced by ∨ (A1 , A2 , A∨ 3 ) 7→ (A3 , A1 w0 , A2 ). √ √ 2 We denote this map by l. It is clear that l = ( l) . Let us recall a sequence of mutations constructed in [26, Section 13]. The sequence is originally defined for the ice quiver ∆Q in terms of reduced expression of w0 . We now translate it into our setting. Recall that we can label the non-neutral vertices of ∆2Q by a pair (i, t) (before Lemma 6.3). Let 2 ti := t∗ (O+ i ) = max{t | (i, t) ∈ ∆Q }. We first assume that the vertices of Q are ordered such that i < j if (i, j) ∈ Q1 . Then we totally order the mutable vertices of ∆2Q by the relation that (i, t) < (i′ , t′ ) if t < t′ , or t = t′ and i < i′ . Starting from the minimal vertex (1, 1) in the ascending order just defined, we perform a sequence of mutations µi,t for each (mutable) vertex of ∆2Q . For the vertex (i, t), the sequence of mutations is defined to be µi,t := µi,ti −t · · · µi,2 µi,1 . So the whole sequence of mutations µ√l := · · · µ2,1 µ1,1 . Let π be the permutation on (∆2Q )0 defined by (i, t) 7→ (i, ti − t) if (i, t) is mutable, + − + O− i 7→ Idi , Oi 7→ Oi∗ , Idi 7→ Oi . It is clear that π is an involution on the set of mutable vertices. When applying it to the quiver ∆2Q or its B-matrix, we view π as a relabelling of the vertices. For a (trivially valued) Dynkin quiver Q, let ΠQ be its associated preprojective algebra. Recall that the module category mod ΠQ is a Frobenius category. Let modΠQ be its stable category. Recall that such a stable category is naturally triangulated with the shift functor given by the (relative) inverse Syzygy functor Ω−1 . Readers can find these standard terminology in, for example, [26]. We will write HomΠQ L for HommodΠQ . In [25] the authors defined a tilting module V := u∈(∆Q )0 Vu in mod ΠQ (it is denoted by IQ in [25]). They proved that the quiver of the endomorphism algebra EndΠQ (V )op is exactly ∆Q with its frozen vertices of corresponding to the projective-injective objects in mod ΠQ . Moreover, it is known [4] that the stable endomorphism algebra EndΠQ (V )op is the Jacobian algebra J µ (see Remark ??). For each rigid ΠQ -module M , we can associate a function ϕM ∈ k[U ] as in [24]. This function turns out to be a cluster variable for the standard seed (∆Q , MQ ). In [24] the mutation of maximal rigid modules of ΠQ is defined so that it is compatible with the seed mutation. In particular, it is compatible with the quiver mutation, so the quiver of EndΠQ (µu (V )) is exactly µu (∆Q ). Proposition B.5. [26, Proposition 13.4] The sequence of mutations µ√l takes the modules Vπ(u) to Ω−1 (Vu ) for u mutable in ∆Q . TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 43 Corollary B.6. Identifying ∆Q as a subquiver of ∆2Q , we have that µ√l (∆Q ) and π(∆Q ) have the same restricted B-matrix. Proof. Since Ω is an autoequivalence, µ√l (∆Q ) and π(∆Q ) have the same mutable part. Let µi,t := µi,t · · · µ1,1 . In [26, Section 13.1], the authors described the each mutated quiver after applying each µi,t . It follows easily from their description that after forgetting arrows between frozen vertices, µ√l (∆Q ) is a subquiver of ∆2Q , − √  whose frozen vertex O+ i is identified with the frozen vertex Oi of µ l (∆Q ). Let (∆µQ , WQµ ) be the restriction of (∆Q , WQ ) to its mutable part. Next, we compute the g-vectors of cluster variables after applying µ√l to the seed (∆µQ , x). The g-vector of the initial cluster variable xi,t at the vertex (i, t) is the unit vector µ ei,t ∈ Z(∆Q )0 . By [13, Theorem 5.2], the g-vector of µ√l (xi,t ) is nothing but √ (0, Si,t ), where (0, Si,t ) is the positive simple representation of the g-vector of µ−1 l µ µ µ√l (∆Q , WQ ). √ (ei,t ) = −ei,ti −t for each ei,t . Here, we view ei,t Lemma B.7. We have that µ−1 l as the g-vector of the simple representation (0, Si,t ) of µ√l (∆µQ , WQµ ). Proof. It is equivalent to show that µ√l (−ei,ti −t ) = ei,t . Let µi,t := µi,t · · · µ1,1 . We study the mutated g-vector µi,t (−ei,ti −t ) for each (i, t). From the description of mutated quiver µi,t (∆Q ) [26, Section 13.1] and the formula (B.8), we find that for j 6= i, each µj,s does not play any role in µ√l (−ei,ti −t ). In other words, it suffices to compute µ√l (−ei,ti −t ) on the full linear subquiver with vertices (i, t) for t = 1, · · · , ti − 1. Using the formula (B.8), we can easily prove by induction that ( ei,ti −s − ei,ti −s−t if s + t < ti , µi,s (−ei,ti −t ) = ei,t otherwise. We conclude from this that µ√l (−ei,ti −t ) = ei,t .  π Let g√ be the set of g-vectors given by l  X π e(f )(i)eO− + f+ (i)eO+ + f− (i)eIdi . (f ) = −ef + (B.9) g√ l i i i∈Q0 Let σ π√l be the weight configuration defined by σ π√l (f ) = (f+ , e(f )∗ , f− ), where f 7→ f ∗ is the involution on ZQ0 induced by ei 7→ ei∗ , so −w0 (̟(f)) = ̟(f ∗ ). π √ (ef ) = g√ (π(f )). Here, we Corollary B.8. For any f mutable, we have that µ−1 l l √ view ef as the g-vector of the simple representation (0, Sf ) of µ l (∆2Q , WQ2 ). So µ√l (σ 2Q )(f ) = σ π√l (π(f )). −1 (e ), which is Proof. In Lemma B.7 we have obtained the principal part of µ√ l f −eπ(f ) . Now we are going to recover its extended part from its principal part. We recall from [14, Proposition 5.15] that being µ-supported for a g-vector is mutation invariant. The general presentation with g-vector equal to −ef corresponds to the indecomposable projective representation P f . By Lemma 5.2 + ∼ P f (O+ i ) = HomQ (P+ , Pi ). So to make the g-vector −ef not supported on Oi , 44 JIARUI FEI we must add at least f+ (i)eO+ to it. Similarly we must add at least f− (i)eIdi and i e(f )(i)eO− as well. On the other hand, it is easy to check that (B.9) is µ-supported. i So adding any additional positive component on the frozen vertices will make the general presentation of this g-vector decomposable, which is not the case. Hence, π (π(f )). µ√l (ef ) = g√ l For the last statement, recall that σ 2Q (f ) = (e(f ), f− , f+ ). Then π g√ (f )σ 2Q = −(e(f ), f− , f+ ) + (e(f ), e(f )∗ , 0) + (f+ , 0, f+ ) + (0, f− , f− ) l = (f+ , e(f )∗ , f− ). π (π(f ))σ 2Q = σ π√l (π(f )). Hence, µ√l (σ 2Q )(f ) = µ√l (ef )σ 2Q = g√ l  Corollary B.9. µ√l (∆2Q ) and π(∆2Q ) have the same restricted B-matrix. Proof. By Corollary B.6, µ√l (∆2Q ) and π(∆2Q ) have the same mutable part as well. It remains to show that the blocks of their restricted B-matrices corresponding to positive and neutral frozen vertices are equal. But this follows from an easy linear 2 algebra consideration. Indeed, let B√ and B√l be the restricted B-matrices of l 2 2 2 µ√l (∆Q ) and µ√l (∆Q ) and denote σ √l := µ√l (σ 2Q ). We have that B√ σ 2√l = 0. l  √ σ l 2 and σ 2√l in blocks: (B√l , C√l ) and σ+,Id We write B√ , where σ +,Id = ( I0 I0 II ) l contains weights corresponding to the positive and neutral frozen vertices. We have that σ √l  = B√l σ √l + C√l σ +,Id = 0. (B.10) (B√l , C√l ) σ+,Id Since σ +,Id is of full rank, this linear equation is overdetermined (for solving C√l ). We have seen that B√l = Bπ(∆Q ) and σ √l (f ) = σ π√l (π(f )), then it is clear that  C√l = π(C) satisfies (B.10), where C is the block in B∆2Q = (B∆Q , C). Definition B.10. Let µπ√l := π −1 µ√l π. We define µl := µπ√l µ√l and µr = µ2l . Corollary B.11. µl (∆2Q ) and π 2 (∆2Q ) have the same restricted B-matrix. More∗ over, µl (σ 2Q )(f ) = (f− , f+ , e(f )∗ ) for f mutable. Note that π 2 fixes the mutable vertices and shuffles the frozen vertices of ∆2Q + + − O− i 7→ Oi , Oi 7→ Idi∗ , Idi 7→ Oi . Corollary B.12. We have that µ3l (B∆2Q , x) = (B∆2Q , x). Proof. Since π 6 = Id, the fact that µ3l (B∆2Q ) = B∆2Q follows from Corollary B.11. As pointed in [25], it is well-known [1] that the functor Ω is also 6-periodic. [27, Proposition 6.3] says that the mutation of maximal rigid modules of ΠQ is compatible with the mutation of the corresponding QP-representations (see remarks before Proposition B.5). So by Proposition B.5 the positive simples (0, Sf ) of (∆Q , WQ ) are invariant under µ3l . The invariance obviously extends to (∆2Q , WQ2 ). The desired result follows.  By experiment with [31] we believe the following conjecture. Conjecture B.13. Corollary B.9 (and thus Corollary B.11, B.12 and Theorem B.14 below) hold for non-trivially valued Q as well. TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 45 2 ) induced by We define a rational self-map µl of C(∆2Q , SQ ( sf 7→ µl (sf ) if f is mutable, sO− 7→ sO+ , sO+ 7→ sIdi∗ , sIdi 7→ sO−∗ . i i i i Since our main theorem (Theorem 8.1) use the following theorem, we cannot assume 2 that C(∆2Q , SQ ) = k[Conf 2,1 ]. But from the proof of Theorem 8.1, we have that 2 2 2 C(∆Q , SQ ) ⊆ k[Conf 2,1 ]. So the pullback l∗ is defined on C(∆2Q , SQ ). It is easy to see that sf is in fact in Q[Conf 2,1 ], so the containment holds over Q. Moreover, 2 2 C k (∆2Q , SQ ) can be obtained from the base change k ⊗Q C Q (∆2Q , SQ ). The cluster mutation and twisted cyclic shift are clearly defined over Q as well. Theorem B.14. The map µl is equal to the pullback l∗ of the twisted cyclic shift. Proof. We need to show that µl (ϕ) = ϕl. It suffices to prove the statement for ϕ’s in some extended cluster because of the definition of the upper cluster algebra. The 2 2 cluster we shall take is µ2l (SQ ), which is the same as µ−1 l (SQ ) by Corollary B.12. The statement is clearly true for the coefficient variables. So let us assume that ϕ = µ2l (sf ) for some mutable f ∈ ind(C 2 Q). ∨ We need to show that µl (ϕ)(A1 , A2 , A∨ 3 ) = ϕ(A2 , A3 w0 , (A1 w0 ) ). We argue by ∗ ∗ multidegrees. By Corollary B.11, the degree of ϕ is (f+ , e(f ), f− ). Then according to (6.1) and (6.2), the degree of ϕ(A2 , A3 w0 , (A1 w0 )∨ ) is (e(f ), f− , f+ ), which is also the degree of µl (ϕ) = sf . So by Lemma 7.4, we must have that µl (ϕ) = cϕl for some c ∈ k. To show c = 1, we can consider the relevant constructions over Q by the above remark. We note that µl clearly commutes with l so ϕ = µ2l µl (ϕ) = µ2l (cϕl) = cµ2l (ϕ)l = c2 µl (ϕl2 ) = c3 ϕl3 = c3 ϕ. We must have that c = 1.  B.3. An Algorithm. In this section, we present an algorithm to find all subrepresentations of Tv defined in Section 5.1. As said in the introduction, only few Tv ’s for type E7 and E8 need this algorithm. With a little more effort, one can show that Corollary B.11 can be strengthened to µl (∆2Q ) = π 2 (∆2Q ) for any orientation of type Dn and E7 , E8 , and for some orientations of type An and E6 . For any particular case whether µl (∆2Q ) = π 2 (∆2Q ) holds can be checked by computer. We first observe that by the description of Tv ’s in Theorem 5.3, all subrepresentations of TO− are known. In particular, we have the dual F -polynomial of i TO− . Even better, TO− can be mutated from the positive simple (0, SO− ) of the i i i  (unfrozen) QP µi ∆(C 2 Q), WQ2 via µ−1 i , where µi := µi,0 µi,0 and the bold µi,0 is defined in the beginning of Appendix B.2. Indeed, it can be easily checked Pby (B.7) that the g∨ -vector of such a mutated representation is exactly eO− − i→j eO− . j i Since TO− is the cokernel of a general presentation of such a weight, our claim i follows from Remark B.4. The idea of the algorithm is that we can generate the dual F -polynomial of TIdi∗ (resp. TO+ ) from that of TO− through the sequence of i i mutations µl (resp. µr ) just defined. Proposition B.15. If µl (∆2Q ) = π 2 (∆2Q ), then µl (TO− ) = TIdi∗ and µr (TO− ) = i i TO+ . Here, we view TIdi∗ and TO+ as representations of the original QP via the i i automorphisms π −2 and π 2 . 46 JIARUI FEI Proof. It is clear from (B.7) that the g∨ -vector eO− − i P i→j eO− is unchanged under j µl . If = π by Corollary B.11 the mutated g∨ -vector can be view as eIdi∗ − i→j eIdj∗ for the original QP (∆2Q , WQ2 ) via the automorphism π −2 . Note that i → j if and only ifP i∗ → j ∗ . Clearly TIdi∗ is the cokernel of the general presentation of weight eIdi∗ − i∗ →j ∗ eId∗j . Hence, µl (TO− ) = TIdi∗ . The argument i for µr (TO− ) = TO+ is similar.  µl (∆2Q ) P i 2 (∆2Q ), i By the positivity of the cluster variables [30, 35], the coefficients of F -polynomials of these cluster variables are all positive. So we can compute FT∨Id ∗ (resp. FT∨ + ) i by applying µl (resp. µr ) to FT∨ subrepresentations of Tv ’s. − O i O i using the formula (B.5). In this way, we find all Acknowledgement The author would like to thank the NCTS (National Center for Theoretical Sciences) and Shanghai Jiao Tong University for the financial support. He would also like to thank Professor Bernard Leclerc for some interesting discussion during the Taipei Conference in Representation Theory Jan 2016. References 1. M. Auslander, I. Reiten, D Tr-periodic modules and functors, Representation theory of algebras, (Cocoyoc, 1994), CMS Conf. Proc., vol. 18, Amer. Math. Soc., Providence, RI, 1996, pp. 39–50. 2. M. Auslander, I. Reiten, S. O. Smalø, Representation theory of Artin algebras, Corrected reprint of the 1995 original. Cambridge Studies in Advanced Mathematics, 36. Cambridge University Press, Cambridge, 1997. 3. I. Assem, D. Simson, A. Skowroński, Elements of the Representation Theory of Associative Algebras, London Mathematical Society Student Texts 65, Cambridge University Press, 2006. 4. A. Buan, O. Iyama, I. Reiten, D. Smith, Mutation of cluster-tilting objects and potentials, Amer. J. Math. 133 no.4 (2011) 835–887. 5. A. Berenstein, A. Zelevinsky, Triple multiplicities for sl(r+1) and the spectrum of the exterior algebra of the adjoint representation, J. Alg. Comb., 1 (1992), 7–22. 6. A. Berenstein, A. Zelevinsky, Tensor product multiplicities, canonical bases and totally positive varieties, Invent. Math. 143 (2001), 77–128. 7. A. Berenstein, S. Fomin, A. Zelevinsky, Cluster algebras. III. Upper bounds and double Bruhat cells, Duke Math. J. 126 (2005), no. 1, 1–52. 8. G. I. Cerulli, E. Feigin, M. Reineke, Desingularization of quiver Grassmannians for Dynkin quivers, Adv. Math. 245 (2013), 182–207. 9. V. Dlab, C. M. Ringel, Indecomposable representations of graphs and algebras, Mem. Amer. Math. Soc. 6 (1976), no. 173, v+57. 10. H. Derksen, J. Fei, General presentations of algebras, Adv. Math. 278 (2015), 210–237. 11. H. Derksen, J. Weyman, Semi-invariants of quivers and saturation for Littlewood-Richardson coefficients, J. Amer. Math. Soc. 13 (2000), no. 3, 467–479. 12. H. Derksen, J. Weyman, A. Zelevinsky, Quivers with potentials and their representations I, Selecta Math. (N.S.) 14 (2008), no. 1, 59–119. 13. H. Derksen, J. Weyman, A. Zelevinsky, Quivers with potentials and their representations II, J. Amer. Math. Soc. 23 (2010), no. 3, 749–790. 14. J. Fei, Cluster algebras and semi-invariant rings I. Triple Flags, Proc. Lond. Math. Soc. (3) 115 (2017), no. 1, 1–32. 15. J. Fei, Cluster algebras and semi-invariant rings II. Projections, Math. Z. 285 (2017), no. 3-4, 939–966. 16. J. Fei, Cluster algebras, invariant theory, and Kronecker coefficients I, Adv. Math. 310 (2017), 1064–1112. 17. J. Fei, Polyhedral Models for Tensor Product Multiplicities to appear Contemp Math. TENSOR PRODUCT MULTIPLICITIES VIA UPPER CLUSTER ALGEBRAS 47 18. V. Fock, A. Goncharov, Moduli spaces of local systems and higher Teichmüller theory, Publ. Math. Inst. Hautes Études Sci., 103 (2006) 1–211. 19. V. Fock, A. Goncharov, Cluster ensembles, quantization and the dilogarithm, Ann. Sci. Éc. Norm. Supér. (4) 42 (2009), no. 6, 865–930. 20. S. Fomin, P. Pylyavskyy, Tensor diagrams and cluster algebras, arXiv:1210.1888. 21. S. Fomin, A. Zelevinsky, Double Bruhat cells and total positivity, J. Amer. Math. Soc. 12 (1999), no. 2, 335–380. 22. S. Fomin, A. Zelevinsky, Cluster algebras. I. Foundations, J. Amer. Math. Soc. 15 (2002), no. 2, 497–529. 23. S. Fomin, A. Zelevinsky, Cluster algebras. IV. Coefficients, Compos. Math. 143 (2007), no. 1, 112–164. 24. C. Geiss, B. Leclerc, J. Schröer, Rigid modules over preprojective algebras, Invent. Math. 165 (2006), 589–632. 25. C. Geiss, B. Leclerc, J. Schröer, Auslander algebras and initial seeds for cluster, Journal of the London Mathematical Society (2) 75 (2007), 718–740. 26. C. Geiss, B. Leclerc, J. Schröer, Kac-Moody groups and cluster algebras, Adv. Math. 228 (2011), no. 1, 329–433. 27. C. Geiss, B. Leclerc, J. Schröer, Generic bases for cluster algebras and the Chamber ansatz, J. Amer. Math. Soc. 25 (2012), no. 1, 21–76. 28. A. Goncharov, L. Shen, Geometry of canonical bases and mirror symmetry, Invent. Math. 202 (2015), no. 2, 487–633. 29. R. Goodman, N. R. Wallach, Symmetry, representations, and invariants, Graduate Texts in Mathematics, 255. Springer, Dordrecht, 2009. 30. M. Gross, P. Hacking, S. Keel, M. Kontsevich, Canonical bases for cluster algebras, arXiv:1411.1394. 31. B. Keller, Quiver mutation in Java, http://www.math.jussieu.fr/˜keller/quivermutation/. 32. B. Kostant, A formula for the multiplicity of a weight, Trans. Amer. Math. Soc., 93, (1959), No. 1, 53–73. 33. F. Knop, H. Kraft, T. Vust, The Picard group of a G-variety. Algebraische Transformationsgruppen und Invariantentheorie, 77–87, DMV Sem., 13, Birkhäuser, Basel, 1989. 34. A. Knutson, T. Tao, The honeycomb model of GLn (C) tensor products. I. Proof of the saturation conjecture, J. Amer. Math. Soc. 12 (1999), no. 4, 1055–1090. 35. K. Lee, R. Schiffler, Positivity for cluster algebras, Ann. of Math. (2) 182 (2015), no. 1, 73–125. 36. M. A. A. van Leeuwen, A. M. Cohen, B. Lisser, “LiE, a package for Lie group computations”, http://www-math.univ-poitiers.fr/˜maavl/LiE/. 37. D. Labardini-Fragoso, A. Zelevinsky, Strongly primitive species with potentials I: mutations, Boletn de la Sociedad Matemtica Mexicana, 1–69. 38. P. Littelmann, A Littlewood-Richardson rule for symmetrizable Kac-Moody algebras, Invent. math. 116 (1994), 329–346. 39. G. Lusztig, Canonical basis arising from quantum canonical algebras, J. Amer. Math. Soc. 3, (1990) no 2, 447–498. 40. K.P. Partasarathy, R. Ranga Rao, V.S. Varadarajan, Representations of complex semisimple Lie groups and Lie algebras, Ann. Math. 85 (1967), 383–429. 41. P. Plamondon, Generic bases for cluster algebras from the cluster category, Int. Math. Res. Notices (2013) 2013 (10): 2368–2420. 42. V. L. Popov, E. B. Vinberg, Invariant theory, in: Algebraic geometry. IV, Encyclopaedia of Mathematical Sciences, vol. 55, Springer-Verlag, Berlin, 1994, 123–284. 43. B. Shapiro, M. Shapiro, A. Vainshtein, A. Zelevinsky, Simply-laced Coxeter groups and groups generated by symplectic transvections, Michigan Math. J. 48 (2000), 531–551. 44. A. Zelevinsky, From Littlewood-Richardson coefficients to cluster algebras in three lectures, Symmetric functions 2001: surveys of developments and perspectives, 253–273, NATO Sci. Ser. II Math. Phys. Chem., 74, Kluwer Acad. Publ., Dordrecht, 2002. Shanghai Jiao Tong University, Shanghai, China E-mail address: [email protected]
0math.AC
List of figures Fig.1 Auxiliary support function φ ( x ) for the function f(x). Fig.2 A circuit realization of the Chebyshev filter. Fig.3 A circuit realization of the passband filter. Fig.4 Possible locations for the roots of the equation φ i ( x) = 0 . The case when xn+1 is calculated by the formula (29). Fig.5 (a)-(b) Possible locations for the roots of the equation φ i ( x) = 0 . The case when xn+1 is calculated by the formula (30). Fig.6 (a)-(b) Possible locations for the roots of the equation φ i ( x) = 0 . The case when xn+1 is calculated by the formula (31). Fig.7 (a)-(b) Flow-chart for the algorithms. Fig.8 Plot of function (38) for determining the cutoff frequency for the Chebyshev filter. Fig.9 Plot of function (39) for determining the cutoff frequency for the passband filter. Fig.1 L Vin C R C R Vout Fig.2 I in R1 R2 C1 L1 L2 C2 R3 C3 L 3 Vout Fig.3 zi-1 xi-1 | x . y' i | n+1 xi φi ( y' ) i zi Fig.4 z i-1 φi ( y' ) i xi-1 Λ . xi xn+1 | y' | xi i φi ( Λ x) i a) z i-1 φi ( y' ) i .x n+1 | xi-1 yi xi | y'i φi ( y ) i zi b) Fig.5 z i-1 φi ( y ) i φi ( y' ) i φ (Λ x i) i xi-1 | y' i y Λ x . x n+1 | i | xi i zi a) z i-1 φi ( y' ) i φi ( y ) i | xi-1 y'i . y i xn+1 xi | zi b) Fig.6 start 1 order the points x,..,x k calculate mi Initiate the index sets Main cycle xi -x i −1 yes ≤σ no calculate f(x n+1) n=n+1 calculate f ' (x n+1) no f(xn+1 )<  yes b n+1 = x n+1 stop Fig.7 a) start of the main cycle yes i>k no calculate yi and y'i yes M'i (yi' ) . M' (y )<0 i calculate Ri = Mi (hi ) according (18) no i calculate Ri = M(h ) i i according (21) i in I no h =x i i yes R ≤0 i yes i in Y' i in Y no i=i+1 M(y' ) ≤0 i no M'i (yi' ) . M' (y ) ≥0 i i yes Find i according to (27) and xn+1 according to (28) calculate x n+1 according to (29) i no no M( x )> 0 i i M(y )> 0 i i yes calculate xn+1 according to (30) end of the main cycle Fig. 7 b) yes no yes calculate xn+1 according to (31) F( Š ) 0 Š Fig.8 F( Š ) 0 Š Fig.9 List of Tables Tab.I Characteristics of the functions utilized for numerical experiments. Tab.II Comparison between grid technique and the methods using smooth auxiliary functions. N° Function f(x) Number of roots 1 −0.5x 2 ln( x) + 5 1 3.0117 3 −x FRL Number of local extrema 2 −e sin(2πx) + 1 - - 13 3 − x ⋅ sin( x) + 1 3 1.17479 4 4 xsin( x) + sin(10x / 3) + ln( x) − 0.84 x + 13 . 2 2.96091 6 5 x + sin(5x) 2 0.82092 13 6 − x ⋅ sin( x) + 5 - - 4 7 sin( x) cos( x) − 15 . sin ( x) + 12 . 4 1.34075 7 8 2 cos( x) + cos(2 x) + 5 - - 6 2 3.1416 4 9 2 2 ⋅ sin( x) ⋅ e −x 10 ( 3x − 1.4) sin(18x) + 1.7 34 1.26554 42 11 2 1.36465 3 12 ( x + 1) / x − 7.1 ⎧sin(5x) + 2 x≤π ⎨ x>π ⎩5sin( x) + 2 2 3.55311 8 13 esin( 3x ) - - 9 2 4.78308 15 2 3.281119 4 14 15 3 2 5 ∑ k cos[ ( k + 1) x + k] + 12 k =0 2( x − 3) 2 − e x / 2 + 5 - - 4 17 −esin ( x) + 4 xsin 2 ( x) 4 3.141128 6 18 cos( x) − sin(5x) + 1 6 1.57079 13 19 − x − sin( 3x) + 1.6 3 1.96857 9 2 1.14071 4 16 20 cos( x ) + 2 cos( 2 x ) e −x Tab.I N° Function Grid A1 A2 1 4135 5 5 2 10000 31 34 3 1295 6 5 4 4060 12 7 5 5470 7 11 Tab.IIa Two methods for solving optimization problems arising in electronic measurements and electrical engineering N Function Grid A1 A2 6 10000 10 9 7 1678 5 6 8 10000 36 24 9 4326 15 10 10 1567 55 12 Tab.IIb Two methods for solving optimization problems arising in electronic measurements and electrical engineering N Function Grid A1 A2 11 1713 69 60 12 4931 13 6 13 10000 99 39 14 6740 23 18 15 4531 9 9 Tab.IIc Two methods for solving optimization problems arising in electronic measurements and electrical engineering N Function Grid A1 A2 16 10000 7 12 17 4325 20 17 18 2016 11 10 19 2601 12 12 20 7413 6 6 Average 5119.15 22.55 16.17 Tab.IId Two methods for solving optimization problems arising in electronic measurements and electrical engineering
5cs.CE
The algebra of interpolatory cubature formulæ for generic nodes arXiv:1207.3143v3 [math.ST] 13 Mar 2013 Claudia Fassino · Giovanni Pistone · Eva Riccomagno Submitted November 14, 2017 Abstract We consider the classical problem of computing the expected value of a real function f of the d-variate random variable X using cubature formulæ. We use in synergy tools from Commutative Algebra for cubature rulæ, from elementary orthogonal polynomial theory and from Probability. Keywords Design of experiments · Cubature formulæ · Algebraic Statistics · Orthogonal polynomials · Evaluation of expectations 1 Introduction Consider the classical problem of computing the expected value of a real function f of the d-variate random variable X as a linear combination of its values f (z) at a finite set of points z ∈ D ⊂ Rd . The general cubature problem is: determine classes of functions f : Rd → R, finite set of n nodes D ⊂ Rd and positive weights {λz }z∈D such that E( f (X)) = Z Rd f (x) d λ (x) = ∑ f (z)λz (1) z∈D where λ is the probability distribution of the random vector X. In the univariate case, d = 1, the set D is the set of zeros of a node polynomial, e.g. the n-th orthogonal polynomial for λ , see e.g. [5, Sec. 1.4]. Not much is known in the multivariate case, unless the set of nodes is product of one-dimensional sets. Claudia Fassino · Eva Riccomagno Dipartimento di Matematica, Università di Genova, Via Dodecaneso 35, 16141 Genova, Italy E-mail: [email protected] E-mail: [email protected] Giovanni Pistone Collegio Carlo Alberto, Via Real Collegio 30, 10024 Moncalieri, Italy E-mail: [email protected] A similar setting appears in statistical Design of Experiment (DoE) where one considers a finite set of treatments D and the experimental outputs as function of the treatment. The set of treatments and the set of nodes are both described efficiently as zeros of systems of polynomial equations, i.e. as what is called a 0-dimensional variety in Commutative Algebra. This framework is systematic for Algebraic Statistics where tools from modern Computational Commutative Algebra are used to address problems in statistical inference and modelling, see e.g. [3, 6, 10]. In DoE the set D is called a design and the affine structure of the ring of real functions on D is analyzed in detail because it represents the set of real responses to treatments in D. However, in the algebraic setting the euclidean structure, such as the computation of mean values, is missing. In algebraic design of experiment the computation of mean values has been obtained by considering very special sets called factorial designs, e.g. {+1, −1}d , see e.g. [4] and [10, Ch. 5]. Note that {+1, −1} is the zero set of the polynomial x2 − 1. The purpose of the present paper is to discuss how the above comes together by considering orthogonal polynomials. In particular, we consider algorithms from Commutative Algebra for the cubature problem in (1) by mixing tools from elementary orthogonal polynomial theory and Probability. Vice versa, Formula (1) provides an interesting interpretation of the RHS term as expected value. We proceed by steps of increasing degree of generality. In Section 2 we consider the univariate case and take λ to admit an orthogonal system of polynomials. Let g(x) = ∏z∈D (x − z) and by univariate division given a polynomial p there exist unique q and r such that p = q g + r and r has degree smaller than the number of points in D, that is the degree of g. Furthermore, r can be written as ∑z∈D r(z)lz (x) where lz is the Lagrange polynomial for z ∈ D. Then we show that 2 1. the expected values of p and r coincide if and only if the n-coefficients of the Fourier expansion of q with respect to the orthogonal polynomial system is zero, 2. the weights λz in (1) are the expected values of the Lagrange polynomials lz , for z ∈ D. The case when the design D is a proper subset of the zero set of the n-th orthogonal polynomial is developed in Section 4. In Section 3, λ is a standard Gaussian probability law and D the zero set of the n-th Hermite polynomial Hn . By applying Stein-Markov theory we give a representation of some Hermite polynomials, including those of degree 2n − 1, as sum of an element in the polynomial ideal generated by Hn and of a reminder. See Theorem 4 and the following discussion, in particular Equation (8) which, unsurprisingly, is reminiscent of other formulæ for iterated Itō integrals e.g. [9, Eq. (6.4.17)]. The point is to describe a ring structure of the space generated by Hermite polynomials up to a certain order. This ring structure is essentially the aliasing on functions induced by limiting observations to D. The particular form of the recurrence relationship for Hermite polynomials makes this possible and we suspect that the study of the ring structure over D for other systems of orthogonal polynomials will require different tools from those we use here. This result implies a system of equations in Theorem 5 (extended to the multidimensional case in Section 5) which gives an implicit description of design and weights via two polynomial equations. We envisage applicability of this in the choice of D for suitable classes of functions but have not developed this here. Section 6 contains our most general set-up: we restrict ourselves to product probability measures on Rd but consider any set of n distinct points in Rd . Then a BuchbergerMöller type of algorithm is provided that works exclusively over vector-space generated by orthogonal polynomials up to a suitable degree. It gives a generating set of the vanishing ideal of D expressed in terms of orthogonal polynomials. This is used to determine sufficient and necessary conditions on a polynomial function f for which (1) holds for suitably defined weights. Furthermore, exploiting the Fourier expansion of a Gröbner basis of the vanishing ideal of D, some results about the exactness of the cubature formulæ are shown. Of course it will be of interest to determine generalisations of our results to the cases where λ is not a product measure and still admits an orthogonal system of polynomials. 1.1 Basic commutative algebra We start with some notation on polynomials: R[x] is the ring of polynomials with real coefficients and in the d-variables (or indeterminate) x = (x1 , . . . , xd ); for α = (α1 , . . . , αd ) a d- C Fassino et al. dimensional vector with non-negative integer entries, xα = α x1α1 . . . xd d indicates a monomial; τ indicates a term-ordering on the monomials in R[x]. If d = 1 there is only one term ordering, this is not the case for d ≥ 2. Designs of product form share some commonalities with the one dimension case. Because of this, term orders are not much used in standard quadrature theory. We will see that refining the division partial order to a proper term-order is actually relevant in some, but not all, multivariate cases. The total degree of the monomials xα is ∑di=1 αi . The symbol R[x]k indicates the set of polynomials of total degree k and R[x]≤k the vector space of all polynomials of at most total degree k. Let D be a finite set of distinct points in Rd , λ a probability measure over Rd and X a real-valued random vector with probability distribution λ so that the expected value of the random variable f (X) is E( f (X)) = R f (x)d λ (x). Given a term ordering τ , let f1 , . . . , ft ∈ R[x] form a Gröbner basis with respect to τ (see [2, Ch. 2]) of the ideal I (D) of polynomials vanishing over D. For each p ∈ R[x] there exist hi , i = 1, . . . ,t and an unique r ∈ R[x] such that t p(x) = ∑ hi (x) fi (x) + r(x), (2) i=1 and r has its largest term in τ not divisible by the largest term of fi , i = 1, . . . ,t. Note that the hi (x) ∈ R[x] are not necessarily unique. For all p ∈ R[x], the polynomial r above is referred to as reminder or normal form. It is often indicated with the symbol NFτ (p, { f1 , . . . , ft }), or the shorter version NF(p), while h f1 , . . . , ft i indicates the polynomial ideal generated by f1 , . . . , ft . Moreover, monomials not divisible by the largest terms of fi , i = 1, . . . ,t, form a vector basis of monomial functions for the vector space L (D) of real functions on D. The polynomials g = ∑ti=1 hi fi and r in (2) are fundamental in the applications of Algebraic Geometry to finite spaces. Various general purpose softwares, including Maple, Mathematica, Matlab and computer algebra softwares, like CoCoA, Macaulay, Singular, allow manipulation with polynomial ideals, in particular compute reminders and monomial bases. The polynomial r(x) in (2) can be written uniquely as r(x) = ∑ p(z)lz (x) (3) z∈D where lz is the indicator polynomial of the point z in D, i.e for x ∈ D it is lz (x) = 1 if x = z and lz (x) = 0 if x 6= z. Equation (3) follows from the fact that {lz : z ∈ D} is a R-vector space basis of L (D). The expected value of the random polynomial function p(X) with respect of λ is E (p(X)) = E (g(X))+E (r(X)) = E (g(X))+ ∑ p(z) E (lz (X)) z∈D Algebra of Cubature by linearity. In this paper we discuss classes of polynomials p and design points D for which E (p(X)) = ∑ p(z) E (lz (X)) z∈D equivalently E (g(X)) = 0. In one dimension, the polynomial f vanishing over D and of degree n = |D| forms a Gröbner basis for I (D). Here |A| indicates the number of elements of a set A. Furthermore, r satisfies three main properties: 1. r is a polynomial of degree less or equal to n − 1, 2. p(x) = g(x) + r(x) = q(x) f (x) + r(x) for a suitable q ∈ R[x] and g, f ∈ h f i, 3. r(x) = p(x) if x is such that f (x) = 0. Such an x is in D. In Section 2 we consider the algebra of orthogonal polynomials in one variable. 3 Theorem 1 (Favard’s theorem) Let γn , αn , βn be sequences of real numbers and for n ≥ 0 let πn+1(x) = (γn x − αn )πn (x) − βn πn−1 (x) be defined recurrently with π0 (x) = 1, π−1 (x) = 0. The πn (x), n = 0, 1, . . . form a system of orthogonal polynomials if and only if γn 6= 0, αn 6= 0 and αn γn γn−1 > 0 for all n ≥ 0. If γn = 1 for all n then the system is of monic orthogonal polynomials. In the monic case, αk = hxπk , πk i hπk , πk i ∑ πek (x)πek (t) = k=0 In this section let d = 1 and D be the zero set of a polynomial which is orthogonal to the constant functions with respect to λ . We next recall the basics on orthogonal polynomials we use, see e.g. [5]. Let I be a finite or infinite interval of R and λ a posiR tive measure over I such that all moments µ j = I x j d λ (x), j = 0, 1, . . ., exist and are finite. In particular, each polynomial function is square integrable on I and the L2 (λ ) scalar product is defined by h f , giλ = Z I f (x)g(x) d λ (x) related inner product is defiWe consider only λ whose p nite positive, i.e. || f || = h f , f iλ > 0 if f 6= 0. In this case there is a unique infinite sequence of monic orthogonal polynomials with respect to λ and we denote them as π0 (x) = 1, π1 (x) = x + · · · , π2 (x) = x2 + · · · , . . . Furthermore we have πk ∈ R[x]k ; π0 , . . . , πk form a real vector space basis of R[x]≤k ; πk is orthogonal to all polynomials of total degree smaller than k; for p ∈ R[x] and n ∈ Z≥0 there exists unique cn (p) ∈ R, called n-th Fourier coefficient of p, such that p(x) = ∑+∞ n=0 cn (p)πn (x) and only a finite number of cn (p) are not zero. Since the inner product satisfies the shift property, i.e. hxp(x), q(x)iλ = hp(x), xq(x)iλ for p, q ∈ R[x] then the corresponding orthogonal polynomial system satisfies a three-term recurrence relationship. More precisely, all orthogonal polynomial systems on the real line satisfy a three-term recurrence relationships. Conversely, Favard’s theorem holds [11]. βk = hπk , πk i hπk−1 , πk−1 i hold true and therefore the norm of πn is computed from the β ’s as kπn k2 = βn βn−1 . . . β0 . For orthonormal polynomials πek = πk / kπk k the ChristoffelDarboux formulæ hold n−1 2 Orthogonal polynomials and their algebra and p πen (x)πen−1 (t) − πen−1(x)πen (t) βn x−t p  ′ (t)πen (t) ∑ πek (t) = βn πen′ (t)πen−1(t) − πen−1 n−1 (4) 2 k=0 Example 1 Inner products of the Sobolev type, namely hu, viS = hu, viλ0 + hu′ , v′ iλ1 · · · + hu(s) , v(s) iλs where λi are positive measures possibly having different support, do not satisfy the shift condition. Neither do the complex Hermitian inner products. Theorem 2 Let D = {x ∈ R : πn (x) = 0} be the zero set of the n-th orthogonal polynomial with respect to the distribution λ of the real random variable X. Consider the division of p ∈ R[x] by πn giving p(x) = q(x)πn (x) + r(x) as above. Then there exist weights λz , z ∈ D, such that the expected value of p(X) is E (p(X)) = cn (q) kπn kλ2 + ∑ p(z)λz z∈D where λz = E (lz (X)) and cn (q) is the n-th Fourier coefficient of the polynomial q. Remark 1 This theorem is a version of a well known result, see e.g. [5, Sec. 1.4]. We include the proof to underline a particular form of the error in the quadrature formula, to be used again in Theorem 3 and in Section 6. Applying Theorem 2 to p = 1 we have that 1 = ∑z∈D λz . Proof The set D contains n distinct points. For a univariate polynomial p, we can write uniquely p(x) = q(x)πn (x) + r(x) with deg(r) < n and deg(q) = max{deg(p) − n, 0}. Furthermore, the indicator functions in the expression r(x) = ∑z∈D p(z)lz (x) are the Lagrange polynomials for D: namely lz (x) = x−w , w∈D :w6=z z − w ∏ z ∈ D. 4 C Fassino et al. Proof From Hence we have E (p(X)) = E (q(X)πn (X)) + ∑ p(z) E (lz (X)) z∈D +∞ = ∑ ck (q) E (πk (X)πn (X)) + ∑ p(z)λz k=0 z∈D = cn (q) kπn k2λ + ∑ p(z)λz A particular case of Theorem 2 occurs if p has degree less than 2n. In this case q has degree at most n − 1 and cn (q) = 0. This shows that the quadrature rule with n nodes given by the zeros of πn and weights {λz }z∈D is a Gaussian quadrature rule and it is exact for all polynomial functions of degree smaller or equal to 2n − 1. For notes on quadrature rules see for example [5, Ch. 1]. Example 2 (Identification) For f polynomial of degree N ≤ 2n − 1 we can write f (x) = ∑Nk=0 ck ( f )πk (x). The constant term is given by ∑ and for all i such that N + i ≤ 2n − 1 ||πi ||λ2 ci ( f ) = E( f (X)πi (X)) = ∑ f (z)πi (z)λz z∈D In particular, if deg f = n − 1 then all coefficients in the Fourier expansion of f can be computed with an evaluation on D. In general for a polynomial of degree N possibly larger than 2n − 1, Theorem 2 gives the Fourier expansion of its reminder by πn , indeed ∑ f (z)πi (z)λz = ∑ NF( f πi )(z)λz z∈D = E(NF( f (X)πi (X))) = ||πi ||2λ ci (NF( f )) Theorem 3 below generalises Theorem 2 to a generic finite set of n distinct points in R, say D. As above, the inx−w dicator function of z ∈ D is lz (x) = ∏w∈D :w6=z . Let z−w g(x) = ∏z∈D (x − z) be the unique monic polynomial vanishing over D and of degree n. Write a polynomial p ∈ R[x] uniquely as p(x) = q(x)g(x) + r(x) and consider the Fourier expansions of q and g: q(x) = ∑+∞ k=0 ck (q)πk (x) and g(x) = n c (g) π (x). ∑k=0 k k Theorem 3 With the above notation, +∞ Eλ (p(X)) = ∑ ck (q)ck (g) kπk kλ2 + ∑ p(z)λz . k=0 +∞ n k=0 j=0 ∑ ck (q)πk (x) ∑ c j (g)π j (x) + ∑ p(z)lz z∈D we have +∞ n E (p(X)) = ∑ ∑ ck (q)c j (g) E (πk (X)π j (X)) + ∑ p(z)λz k=0 j=0 n = z∈D ∑ ck (q)ck (g)||πk ||λ2 + ∑ p(z)λz k=0 z∈D ⊓ ⊔ and this proves the theorem. The condition in Theorem 3 is linear in the Fourier coefficients of q, which is found easily from f by polynomial division. The first |D| Fourier coefficients of q appearing in the conditions of the theorem are determined by solving the system of linear equations M [ck (q)]k=0,...,|D |−1 = [q(z)]k=0,...,|D |−1 f (z)λz z∈D z∈D = z∈D as E (πk (X)πn (X)) = δk,n with δk,n = 1 if k = n and 0 otherwise. ⊓ ⊔ c0 ( f ) = E( f (X)) = p(x) = q(x)g(x) + r(x) z∈D (5) where M = [πk (z)]z∈D ,k=0,...,|D |−1 is the design/evaluation matrix for the first |D| orthogonal polynomials. Theorem 3 can be used in two ways at least. If p is known, the condition in the theorem can be checked to verify if the expected value of p can be determined by Gaussian quadrature rule with nodes D and weights X −w λz = E ∏ z−w w∈D :w6=z ! ∑n−1 α (z, k) E X k = 0 n−1 ∑0 α (z, k)zk  for z ∈ D, where α (z, k) is the k-th symmetric function of the polynomial πn (x)/(x − z). The Fourier coefficients of g can be computed analogously to those of q adapting Equation (5). If p is an unknown polynomial and p(x) = ∑α pα xα for a finite number of non-zero, unknown, real coefficients pα , Theorem 3 characterizes all the polynomials for which the Gaussian quadrature rule is exact, namely Eλ (p(X)) = ∑z∈D p(z)λz . Furthermore, the characterization is a linear expression in the unknown pα . This is because in Equation (5) the q(z) are linear combinations of the coefficients of p. In Section 3 we shall specialise our study to Hermite polynomials, while in Section 5 we shall generalise Theorem 3 to higher dimension. To conclude this section, we discuss the remainder r vs the orthogonal projection. Remark 2 Let p(x) ∈ R[x] and write p(x) = q(x)πn (x)+r(x) where r has degree less than n. Then Algebra of Cubature 5 1. q is the unique polynomial such that p − qπn is orthogonal to all πm with m ≥ n. This is a rephrasing of the characteristic property of the remainder: r belongs to R[x]≤n−1 if, and only if, r is orthogonal to all πm with m ≥ n. Should it exist two such q’s, then (q1 − q2 )πn would be in the same space, hence null. 2. If deg(p) = n, then r is the orthogonal projection of p on R[x]≤n−1 . In fact, q is the leading coefficient of p, therefore p − r is a multiple of πn and indeed orthogonal to R[x]≤n−1 . 3. If deg(p) = N ≥ n, then the orthogonal projection of p on R[x]≤n−1 differs from r, unless the projection of qπn is zero. Example 3 Substituting the Fourier expansions of q and p in the division above, for m ≥ n we find that the m-th coefficient in the Fourier expansion of p can be written as E (p(X)πm (X)) = E (q(X)πn (X)πm (X)) +∞ +∞ ∑ ck (p) E (πk (X)πm (X)) = ∑ c j (q) E (π j (X)πn(X)πm (X)) j=0 k=0 cm (p)||πm || = 2 +∞ ∑ c j (q) E (π j (X)πn(X)πm (X)) j=0 For Hermite polynomials it can be simplified by e.g. using the product formula in Theorem 4 of Section 3. 3 Hermite polynomials There is another way to look at the algebra of orthogonal polynomials that we discuss here in the case of Hermite polynomials. The reference measure λ is the normal√dis2 tribution and d λ (x) = w(x) dx, with w(x) = e−x /2 / 2π , x ∈ R. For a real valued, differentiable function f , define  d 2 2 d  δ f (x) = x f (x) − f (x)e−x /2 , f (x) = −ex /2 dx dx d , and consider Z ∼ λ . The following identity holds dxn (6) 2 H0 = 1 H1 (x) = x H2 (x) = x2 − 1 H3 (x) = x3 − 3x H4 (x) = x4 − 6x2 + 3 H5 (x) = x5 − 10x3 + 15x 2 2 2. Hn (x) = (−1)n ex /2 d n (e−x /2 ) (Rodrigues’ formula) 3. d δ − δ d is the identity operator. From this the relationn! ships dHn = nHn−1 , d m Hn = Hn−m for m ≤ n and the m! three-term recurrence relationship Hn+1 = xHn − nHn−1 (7) are deduced. 4. Hermite polynomials are orthogonal with respect to the standard normal distribution λ . Indeed from Equation (6) we have E (Hn (Z)Hm (Z)) = n!δn,m where δn,m = 0 if n 6= m and δn,m = 1 if n = m. We already mentioned that {Hn (x) : n ≤ d} spans R[x]≤d and that Hn is orthogonal to any polynomial of degree different from n. The ring structure of the space generated by the Hermite polynomials is described in Theorem 4. Theorem 4 The Fourier expansion of the product Hk Hn is n∧k    n k i!Hn+k−2i Hk Hn = Hn+k + ∑ i i=1 i hHk Hn , ψ i = hδ n 1, Hk ψ i n E (φ (Z) δ n ψ (Z)) = E (d n φ (Z) ψ (Z)) 1. The first Hermite polynomials are Proof Note that hφ , ψ i = E(φ (Z)ψ (Z)) is a scalar product on the obvious space and let n ≤ k with Z ∼ N (0, 1) and ψ , φ square integrable functions for which identity (6) holds. Then 3.1 Stein-Markov operators for standard normal distribution dn = (δ f )(x + a) holds for a ∈ R and the constant one is mapped into x. The n-th Hermite polynomial can be defined as Hn (x) = δ n 1. Direct computation using δ proves the following wellknown facts: if φ , ψ are such that limx→±∞ φ (x)ψ (x)e−x /2 = 0 and are square integrable, see [7, Ch. V Lemma 1.3.2 and Proposition 2.2.3]). Polynomials satisfy these conditions and δ is also called the Stein-Markov operator for the standard normal distribution. It is a shift-equivariant linear operator on R[x], namely δ f (x + a) = (x + a) f (x + a) − f ′ (x + a) =   n i = h1, d (Hk ψ )i = ∑ h1, d Hk d n−i ψ i i i=0   n n i d Hk d n−i ψ i = h1, Hk d n ψ i + ∑ h1, i i=1 n   n = hHn+k , ψ i + ∑ k(k − 1) . . . (k − i + 1)hHn+k−2i, ψ i i=1 i n    n k = hHn+k , ψ i + h ∑ i!Hn+k−2i , ψ i i i=1 i n n ⊓ ⊔ 6 C Fassino et al. Example 4 (Aliasing) As an application of Theorem 4, observe that the three-term recurrence relation for Hermite polynomials, Equation (7) Corollary 1 Let Dn = {x : Hn (x) = 0} and p ∈ R[x]. Let p(x) = q(x)Hn (x) + r(x) with the degree of r smaller than n and let Z ∼ N (0, 1). Then Hn+1 = xHn − nHn−1 E (p(Z)) = evaluated on the zeros of Hn (x), say Dn , becomes Hn+1 (x) ≡ −nHn−1(x) where ≡ indicates that equality holds for x ∈ Dn . n+k In general let Hn+k ≡ ∑n−1 j=0 h j H j be the Fourier expansion of the normal form of Hn+k at Dn , where we simplified the notation for the Fourier coefficients. Substitution in the product formula in Theorem 4 gives the formula to write hn+k in terms of Fourier coefficients of lower order Hermite j polynomials: n k  (8) NF(Hn+k ) ≡ − ∑n∧k i=1 i i i! NF(Hn+k−2i )   n−1 n+k−2i n∧k n k ≡ − ∑i=1 i i i! ∑ j=0 h j Hj ∑ p(z)λz if and only if E (d n q(Z)) = 0 z∈Dn with λz = E (lz (Z)) and lz (x) = ∏w∈D :w6=z x−w , z ∈ Dn . z−w Proof From Equation (6) we have E (q(Z)Hn (Z)) = E (q(Z)δ n 1) = E (d n q(Z)) Now by the same steps followed in the proof of Theorem 2 we conclude that E (p(Z)) = E (d n q(Z)) + ∑ p(z)λz z∈Dn Equating coefficients gives a closed formula n∧k  hn+k =−∑ j i=1   n k i!hn+k−2i j i i In Table 1 the normal form of Hk+n with respect to Hn is written in terms of Hermite polynomials of degree smaller than n. For example, Hn+3 (x) = −n(n − 1)(n − 2)Hn−3(x) + 3nHn−1(x) for those values of x such that Hn (x) = 0. ⊓ ⊔ 3.2 Algebraic characterisation of the weights Theorem 5 gives two polynomial equations whose zeros are the design points and the weights. This is a particular case of Formula (1.17) in [12]. We provide the proof to highlight the algebraic nature of the result and of its proof. Theorem 5 Let D = {x : Hn (x) = 0}. k 1 2 3 4 5 6 Hn+k ≡ −nHn−1 −n(n − 1)Hn−2 −n(n − 1)(n − 2)Hn−3 + 3nHn−1 −n(n − 1)(n − 2)(n − 3)Hn−4 + 8n(n − 1)Hn−2 n! − (n−5)! Hn−5 + 5nHn−1 + 15n(n − 1)(n − 2)Hn−3 n! − (n−6)! Hn−6 + 24n(n − 1)(n − 2)(n − 3)Hn−4 +10n(n − 1)(2n − 5)Hn−2 Table 1 Aliasing of Hn+k , k = 1, . . ., 6 over D = {Hn (x) = 0} Example 5 Observe that if f has degree n + 1 equivalently k = 1 then n−1 f= ∑ ci ( f )Hi + cn( f )Hn + cn+1( f )Hn+1 i=0 n−2 ≡ ∑ ci ( f )Hi + (cn−1( f ) − ncn+1( f )) Hn−1 i=0 and all coefficients up to degree n − 2 are “clean”. We give another proof of Theorem 2 for Hermite polynomials. 1. There exists only one polynomial λ of degree n − 1 such that λ (x) = λx for all x ∈ D, −2 2. furthermore λx = (n−1)! n Hn−1 (x). Equivalently 3. the polynomial λ satisfies   Hn (x) = 0 2 (x) = (n − 1)!  λ (x)Hn−1 n Proof 1. The univariate polynomial λ is the interpolation polynomial of the values λz ’s at the n distinct points in Dn and hence it exists, unique of degree n − 1. 2. Observe that for√Hermite polynomials αn = 0, βn = n, en (x) = Hn (x)/ n! and H en′ (x) = √nH en−1 (x). SubstituH tion in the Christoffel-Darboux formulæ and evaluation at Dn = {x1 , . . . , xn } give n−1 ∑ Hek (xi )Hek (x j ) = 0 if i 6= j k=0 n−1 ∑ Hek (xi )2 = nHen−1(xi )2 k=0 In matrix form Equations (9) becomes en−1 (xi )2 : i = 1, . . . , n) Hn Htn = n diag(H (9) Algebra of Cubature 7 h i e j (xi ) where Hn is the square matrix Hn = H and diag indicates a diagonal matrix. invertible and provided in Section 3.3. The polynomial system to be conis i=1,...,n; j=0,...,n−1 sidered Observe that Hn is 0 = H3 (x) = x3 − 3x 2/3 = λ (x)H22 = (θ0 + θ1 x + θ2 x2 )(x2 − 1)2 t −1 e −2 (xi ) : i = 1, . . . , n) H−1 diag(H n = Hn n n−1 Now, let f be a polynomial of degree at most n − 1, that is a typical remainder by division for Hn , then f (x) = e ∑n−1 j=0 c j H j (x). Write f = Hn c where f = [ f (xi )]i=1,...,n and c = [c j ] j . Furthermore note that c= = cj = t −1 e −2 (xi ) : i = 1, . . . , n) f H−1 diag(H n f = Hn n n−1 e −2 (xi ) f (xi ) : i = 1, . . . , n) Htn n−1 diag(H n−1 1 n e e −2 (xi ) H j (xi ) f (xi )H n−1 n i=1 ∑ (10) (11) The expected value of lk (Z) is   e j (x) = ck0 λk = E (lk (Z)) = ∑ ck j E H n−1 j=0 Substitution in Equation (11) for j = 0 gives (n − 1)! −2 1 e −2 λk = H (xk ) = Hn−1 (xk ) n n−1 n This holds for all k = 1, . . . , n. 3. The system of equations is a rewriting of the previous parts of the theorem because the first equation Hn (x) = 0 states that only values of x ∈ Dn are to be considered and the second equation is what we have just proven. ⊓ ⊔ Item 2 in Theorem 5 states that the weights are strictly positive. Theorem 2 applied to the constant polynomial p(x) = 1 shows that they sum to one. In other words, the mapping that associates z to λz , z ∈ Dn , is a discrete probability density. Theorem 2 states that the expected value of the polynomial functions of Z ∼ N (0, 1) for which cn (q) = 0, is equal to the expected value of a discrete random variables X given by Pn (X = xk ) = E (lk (Z)) = λk , k = 1, . . . , n n E (p(Z)) = 2/3 = λ (x)H22 = θ0 + θ1 4x + (θ0 + 4θ2)x2 (12) Coefficients in Equation (12) are equated to give λ (x) = 23 − x2 6. Apply this to the k-th Lagrange polynomial, f (x) = lk (x), e whose Fourier expansion is f (x) = ∑n−1 j=0 ck j H j (x). Using lk (xi ) = δik in Equation (10), obtain 1e e −2 (xk ) ck j = H j (xk )H n−1 n where λ (x) = θ0 + θ1 x + θ2 x2 . The degree of λ (x)H22 is reduced to 2 by using x3 = 3x ∑ p(xk )λk = En (p(X)) k=1 Example 6 For n = 3 the polynomial λ in Theorem 5 can be determined by-hand. For larger values of n an algorithm is In some situations, e.g. the design of an experimental plan or of a Gaussian quadrature rule, the exact computation of the weights might not be necessary and λ (x) is all we need. When the explicit values of the weights are required, the computation has to be done outside a symbolic computation setting as we need to solve, e.g., H3 (x) = 0 to √ √ get D3 = {− 3, 0, 3} and evaluate λ (x) to find λ−√3 = √ λ (− 3) = 61 = λ√3 and λ0 = λ (0) = 32 . Example 7 Let a positive integer N be given and let k and n be positive integers such that kn < 2N then   E Hn (Z)k = ∑ z∈DN (n − 1)! Hn (z)k 2 (z) n HN−1 by Theorem 5. The issue is then the evaluation of HN−1 and Hn at the zeros of HN , for which the recurrence relationship can be used when the values are not tabulated. 3.3 Code for the weighing polynomial The polynomial λ (x) in Theorem 5 is called the weighing polynomial. Table 2 gives a code written in the specialised software for symbolic computation called CoCoA [1] to compute the Fourier expansion of λ (x) based on Theorem 5. Line 1 specifies the number of nodes N. Line 2 establishes that the working environment is a polynomial ring whose variables are the first (N − 1)-Hermite polynomials plus an extra variable w which encodes the weighing polynomial; here it is convenient to work with a elimination termordering of w, called Elim(w), so that the variable w will appear as least as possible. Lines 3, 4, 5 construct Hermite polynomials up-to-order N by using the recurrence relationships (7). Specifically they provide the expansion of H j over Hk with k < j for k = 0, . . . , N − 1. Line 6 states that HN = H1 HN−1 −(N −1)HN−2 = 0, ‘giving’ the nodes of the quadrature. Line 7 is the polynomial in the second equation in the system in Item 3 of Theorem 5 and ‘gives’ the weights. There are N equations which are collected in an ideal whose Gröbner basis is computed in Line 8. In our application it is 8 interesting that the Gröbner bases contains a polynomial in which w appears alone as a term of degree one. This element of the Gröbner basis relates explicitly to the desired weighing polynomial w to the first (n − 1)-Hermite polynomials. L1 N:=4; L2 Use R::=Q[w,h[1..(N-1)]], Elim(w); L3 Eqs:=[h[2]-h[1]*h[1]+1]; L4 For I:=3 To N-1 Do L5 Append(Eqs,h[I]-h[1]*h[I-1]+(I-1)*h[I-2]) EndFor; L6 Append(Eqs,h[1]*h[N-1]-(N-1)*h[N-2]); L7 Append(Eqs,N*w*h[N-1]^2-Fact(N-1)); L8 J:=Ideal(Eqs); GB_J:=GBasis(J); Last(GB_J); L9 3w + 1/4h[2] - 5/4 Table 2 Computation of the Fourier expansion of the weighing polynomial using Theorem 5 C Fassino et al. polynomial of degree at most 2n − 1. Then from Theorem 2 we have E((p1F )(X)) = ∑ p(z)λz = En (p(Y )1F (Y )) z∈F = En (p(Y )|Y ∈ F ) Pn (Y ∈ F ) where X is a random variable with probability law λ and Y is a discrete random variable taking value z ∈ F with probability Pn (Y = z) = λz . The first equality follows from the fact that p(x)1F (x) is zero for x ∈ D \ F and the last equality from the definition of conditional expectation. Another approach is to consider the polynomial whose zeros are the elements of F , say ωF (x) = ∏ (x − z). Now z∈F consider the Lagrange polynomials for F , namely lzF (x) = x−w for z ∈ F . ∏ w6=z z − w w∈F Line 9 in Table 2 gives the polynomial obtained for N = 4, as set in Line 1, namely   1 5 1 6 − x2 λ (x) = − H2 (x) + = 4 4 3 12 p √ The√ nodes are ± 3 ± 6 and the values of the weights are 3± 6 12 , showing that both nodes and weights are algebraic numbers but not rational numbers. On a Mac OS X with an Intel Core 2 Duo processor (at 2.4 GHz) using CoCoA (release 4.7) the result is obtained for N = 10 in Cpu time = 0.08, User time = 0; for N = 20 in Cpu time = 38.40, User time = 38; for N = 25 in Cpu time = 141.28, User time = 142 and for N = 30 in Cpu time = 5132.71, User time = 5186. Observe that this computations can be done once for all and the results stored. Observe furthermore that in Line 8 the CoCoA commands GB J:=GBasis(J); Last(GB J); could be substituted by NF(w,J) . This does not improve on computational cost as NF requires the computation of a Gröbner basis and a reduction. As a minor point we observe that the symbol w would not appear in Line 9. Lemma 1 Let F ⊂ Dn . The Lagrange polynomial for z ∈ F is the remainder of the Lagrange polynomial for z ∈ Dn with respect to ωF (x), namely lzF (x) = NF (lz (x), hωF (x)i) Proof There exists unique NF(lz )(x), polynomial of degree small than m, such that lz (x) = q(x)ωF (x) + NF(lz )(x) Furthermore, for a ∈ F we have lz (a) = NF(lz )(a) = δz,a = lzF (a). The two polynomials laF (x) and NF(lz )(x) have degree smaller than m and coincide on m points, by interpolation they must be equal. ⊓ ⊔ For a polynomial p of degree N, write p(x) = q(x)ωF (x)+ r(x) with f (z) = r(z) if z ∈ F and r(x) = ∑z∈F p(z)lzF (x). m Let q(x) = ∑N−m j=0 b j π j (x) and ωF (x) = ∑i=0 ci πi (x) as ωF has degree m. Then ! N−m E (p(X)) = E ∑ j=0 4 Fractional design In this section we return to the case of general orthogonal polynomials, {πn }n , and positive measure, d λ . We assume that the nodes are a proper subset F of Dn = {x ∈ R : πn (x) = 0} with m points, 0 < m < n. We work within two different settings, in one the ambient design Dn is considered while in the other one it is not. Consider the indicator function of F as subset of Dn , namely 1F (x) = 1 if x ∈ F and 0 if x ∈ Dn \ F . It can be represented by a polynomial of degree n because it is a function defined over Dn [2, 10]. Let p be a polynomial of degree at most n − 1 so that the product p(x)1F (x) is a b0 c0 ||π0 ||2λ m b j π j (X) ∑ ci πi (X) + E (r(X)) = + b1 c1 ||π1 ||2λ i=0 + ··· + b(N−m)∧m c(N−m)∧m ||π(N−m)∧m ||λ2 + ∑ p(z)λzF z∈F where λzF = E (NF(lz (X), hωF (X)i), z ∈ F . Note that the error of the Gaussian quadrature rule, b0 c0 ||π0 ||2λ + b1 c1 ||π1 ||2λ + · · ·+ b(N−m)∧m c(N−m)∧m ||π(N−m)∧m ||2λ is linear in the Fourier coefficients b j , and also in the Fourier coefficients c j of the node polynomial. This is generalised in Section 5. If the fraction F coincides with the ambient design Dn and hence contains n points and if p is a polynomial Algebra of Cubature 9 of degree at most 2n − 1, then we obtain the well known result of zero error because (N − n) ∧ n ≤ n − 1 and the only non-zero Fourier coefficient of the node polynomial πn is of order n. In general one should try to determine pairs of F and sets of polynomials for which the absolute value of the errors is minimal. Theorem 6 It holds Eλ d ( f (X1 , . . . , Xd )) = d ∑ Eλ d−1 (ck (qk )(X−k ))||πk ||λ2 + k=1 ∑ f (x1 , . . . , xd )λxn11 . . . λxndd (x1 ,...,xn )∈Dn1 ...nd 5 Higher dimension: zero set of orthogonal polynomials as design support In this section we return to the higher dimensional set-up of Section 1.1 but we restrict ourselves to consider the product measure λ d = ×di=1 λ and X1 , . . . , Xd independent random variables each one of which is distributed according to the probability law λ . As design we take a product grid of zeros of orthogonal polynomials with respect to λ , more precisely our design points or interpolation nodes are Dn1 ,...,nd = o n x ∈ Rd : πn1 (x1 ) = πn2 (x2 ) = . . . = πnd (xd ) = 0 where πnk is the orthogonal polynomial with respect to λ of degree nk . The Lagrange polynomial of the point y = (y1 , . . . , yd ) ∈ n Dn1 ,...,nd is defined as ly (x1 , . . . , xd ) = ∏dk=1 lykk (xk ), the apex nk indicates that l nk (x ) is the univariate Lagrange polynoyk k mial for yk ∈ {xk : πnk (xk ) = 0} = Dnk ⊂ R. The Span ly : y ∈ Dn1 ,...,nd is equal to the linear space generated by the monomials whose exponents lie on the integer grid {0, . . . , n1 − 1} × . . . × {0, . . . , nd − 1}. Any polynomial f ∈ R[x] can be written as d f (x1 , . . . , xd ) = ∑ qk (x1 , . . . , xd )πnk (xk ) + r(x1, . . . , xd ) k=1 where r is unique, its degree in the variable xk is smaller than nk , for k = 1, . . . , d, and belongs to that Span. The coefficients of the Fourier expansion of qk with respect to the variable xk are functions of x1 , . . . , xd but not of xk . Let x−k denote the (d − 1)-dimensional vector obtained from (x1 , . . . , xd ) removing the k-th component and write f (x1 , . . . , xd ) = d +∞ k=1 j=0 ! ∑ ∑ c j (qk )(x−k )π j (xk ) πnk (xk ) + r(x1 , . . . , xd ) Only a finite number of c j (qk )(x−k ) are not zero. From the independence of X1 , . . . , Xn , the expected value of the Lagrange polynomial ly is d d  n Eλ d (ly (X1 , . . . , Xd )) = ∏ Eλ lynkk (Xk ) = ∏ λk k n where λk k k=1 k=1  nk lyk (Xk ) is the expected value of a univariate =E random Lagrange polynomial as in the previous sections. Proof The proof is very similar to that of Theorem 2 and we do it for d = 2 only. In a simpler notation the design is the n × m grid given by Dnm = {(x, y) : πn (x) = 0 = πm (y)} and X and Y are independent random variables distributed according to λ . The polynomial f is decomposed as f (x, y) = q1 (x, y)πn (x)+ q2 (x, y)πn (y)+ ∑ f (a, b) lan (x)lbm (y) = (a,b)∈Dn,m +∞ +∞ j=0 j=0 ∑ c j (q1 )(y)π j (x) πn (x) + ∑ c j (q2 )(x)π j (y) πn(y)+ ∑ f (a, b) lan (x)lbm (y) (a,b)∈Dn,m Taking expectation, using independence of X and Y and orthogonality of the πi , we have Eλ 2 ( f (X,Y )) = Eλ (cn (q1 )(Y )) ||πn ||2λ + Eλ (cm (q2 )(X)) ||πm ||λ2 + ∑ f (a, b)λan λbm (a,b)∈Dn,m ⊓ ⊔ Note in the proof above that a sufficient condition for Eλ (cn (q1 )(Y )) being zero is that f has degree in x smaller then 2n − 1, similarly for Eλ (cm (q2 )(X)). We retrieve the well-known results that if for each i the degree in xi of f is smaller than 2ni − 1, then Eλ d ( f (X1 , . . . , Xd )) = ∑ f (x1 , . . . , xd )λxn11 . . . λxndd (x1 ,...,xn )∈Dn1 ...nd In the Gaussian set-up, by Theorem 5 applied to each variable, weights and nodes satisfy the polynomial system   =0 Hn1 (x1 )    (n1 − 1)!  2   λ1 (x1 )Hn1 −1 (x1 ) =   n1  .. (13) .     Hnd (xd ) =0    (nd − 1)!    λd (xd )Hnd −1 (xd )2 = nd For the grid set-up of this section and for the Gaussian case, in analogy to Example 4 some Fourier coefficients of polynomials of low enough degree can be determined exactly from the values of the polynomials on the grid points as shown in Example 8 below. 10 C Fassino et al. 1(0,0)∈F (x, y) = 1(√3,√3)∈F (x, y) = 1(√3,−√3)∈F (x, y) = 1(−√3,√3)∈F (x, y) = 1(−√3,−√3)∈F (x, y) = 1 2 H0 − H2 (y) 3 3 1 1√ H0 + 3H1 (x) + 12 12 1 1√ 3H1 (x) + H0 − 12 12 √ 1 1 H0 + 3H1 (x) − 12 12 √ 1 1 H0 − 3H1 (x) − 12 12 1 1√ 1 3H1 (y) + H1 (x)H1 (y) + H2 (y) 12 12 12 1√ 1 1 3H1 (y) − H1 (x)H1 (y) + H2 (y) 12 12 12 1 1√ 1 3H1 (y) − H1 (x)H1 (y) + H2 (y) 12 12 12 1 1√ 1 3H1 (y) + H1 (x)H1 (y) + H2 (y) 12 12 12 Table 3 Indicator functions for Example 9 Example 8 Consider a square grid of size n, Dnn , and a polynomial f of degrees in x and in y smaller than n, the Hermite polynomials and the standard normal distribution. Then we can write n−1 f (x, y) = ∑ As both the degree in x of f Hk and the degree in y of f Hh are smaller than 2n − 1, we have E ( f (Z1 , Z2 )Hk (X1 )Hh (X2 )) = chk ||Hk (X1 )||2 ||Hh (X2 )||2 1 ckh = ∑ f (x, y)Hk (x)Hh (y)λx λy k!h! (x,y)∈D i = 1, 2, 3 r(x, y) = ∑ f (a, b)1(a,b)∈F (x, y) (a,b)∈F Each indicator function 1(a,b)∈F belongs to nn Note if f is the indicator function of a fraction F ⊂ Dnn then 1 ∑ Hk (x)Hh (y)λx λy k!h! (x,y)∈F E(gi (Z1 , Z2 )qi (Z1 , Z2 )) = 0, for Z1 and Z2 independent normally distributed random variables. Write r as a linear combination of the indicator functions of the points in F , i.e. ci j Hi (x)H j (y) i, j=0 ckh = with ai , θ j ∈ R for i = 0, . . . , 5 and j = 1, . . . , 3, then Span (H0 , H1 (x), H1 (y), H1 (x)H1 (y), H2 (y)) and are shown in Table 3. Their expected values are given by the H0 -coefficients. Furthermore, by linearity with 0 ≤ h, k < n Example 9 deals with a general design and introduces the more general theory of Section 6. Example 9 Let F be the zero set of  g = x2 − y2 = H2 (x) − H2 (y) = 0   1 g2 = y3 − 3y = H3 (y) = 0   g3 = xy2 − 3x = H1 (x) (H2 (y) − 2H0) = 0 E( f (Z1 , Z2 )) = E(r(Z1 , Z2 )) = ∑ f (a, b) E(1(a,b)∈F (Z1 , Z2 )) (a,b)∈F and we can conclude f (0, 0) + E( f (Z1 , Z2 )) = E(r(Z1 , Z2 )) = 2 √ √ namely F is given by the five points (0, 0), (± 3, ± 3). √ √ √3 √ √ √ √ √ Write a polynomial f ∈ R[x, y] as f = ∑ qi gi +r where r(x, y) = f ( 3, 3) + f ( 3, − 3) + f (− 3, 3) + f (− 3, − 3) 12 f (x, y) for (x, y) ∈ F and r ∈ Span (H0 , H1 (x), H1 (y), H1 (x)H1 (y), H2 (y)) = Span 1, x, y, xy, y If, furthermore, f is such that q1 (x, y) = a0 + a1 H1 (x) + a2 H1 (y) + a3H1 (x)H1 (y) q2 = θ1 (x) + θ2 (x)H1 (y) + θ3 (x)H2 (y) q3 = a4 + a5 H1 (y) The key points in Example 9 are  2 . 1. determine the class of polynomial functions for which E(gi (Z1 , Z2 )qi (Z1 , Z2 )) = 0 and 2. determine the H0 -coefficients of the indicator functions of the points in F . In Section 6 we give algorithms to do this for any fraction F. Algebra of Cubature 11 Theorem 7 1. For d = 1 and k ∈ Z≥0 , in the notation of Equation (14) we have that 6 Higher dimension: general design support In the previous sections we considered particular designs whose sample points were zeros of orthogonal polynomials. In the Gaussian case we exploited the ring structure of the set of functions defined over the design in order to obtain recurrence formula and to write Fourier coefficients of higher order Hermite polynomials in terms of those of lower order Hermite polynomials (Example 4). Also we deduced a system of polynomial equations whose solution gives the weights of a quadrature formula. The mathematical tools that allowed this are Equation (6) and the particular structure it implies for Hermite polynomials on the recurrence relation for general, orthogonal polynomials πk+1 (x) = (γk x − αk )πk (x) − βk πk−1 (x) x∈R (14) with γk , αk 6= 0 and αk γk γk−1 > 0 (cf. Theorem 1). In this section we switch focus and consider a generic set of points in Rd as a design, or nodes for a cubature formula, and a generic set of orthogonal polynomials. We gain something and lose something. The essential computations are linear: such is the computation of a Gröbner basis for a finite set of distinct points [8]; the Buchberger Möller type of algorithm in Table 4 is based on finding solutions of linear systems of equations; in Section 6.1 we give a characterisation of polynomials with the same expected values which is a linear expression of some Fourier coefficients and a square free polynomial of degree two in a larger set of Fourier coefficients (see Equation 18). Given a set of points and a term-ordering the algorithm in Table 4 returns the reduced Gröbner basis of the design ideal expressed as linear combination of orthogonal polynomial of low enough degree. It does so directly; that is, it computes the Gröbner basis by working only in the space of orthogonal polynomials. We lose the equivalent of Theorem 4 for Hermite polynomials, in particular we do not know yet how to impose a ring structure on Span(π0 , . . . , πn ) for generic orthogonal polynomials π and we miss a general formula to write the product πk πn as linear combination of πi with i = 0, . . . , n ∧ k, n + k, which is fundamental for the aliasing structure discussed for Hermite polynomials. For multivariate cubature formulæ we refer e.g. to [13] and [14] which, together with [8], are basic references for this section. For clarity we repeat some basics and notation. Let λ be a one-dimensional probability measure and {πn }n∈Z≥0 be its associated orthogonal polynomial system. To a multi-index α = (α1 , . . . , αd ) ∈ Zd≥0 we associate the α monomial xα = xα1 1 · · · xd d and the product of polynomials πα (x) = πα1 (x1 ) . . . παd (xd ). Note that {πα }α is a system of orthogonal polynomials for the product measure λ d . Theorem 7 describes the one-to-one correspondence between the xα ’s and the πα (x)’s. xk = k ∑ c j (xk )π j (x) j=0 where c0 (x0 ) = 1, c−1 (x0 ) = c1 (x0 ) = 0, and, for k = 1, 2, . . . and j = 0, . . . , k − 1 c−1 (xk ) = ck+1 (xk ) = 0 c j−1 (xk−1 ) c j (xk−1 )α j c j+1 (xk−1 )β j+1 + + γ j−1 γj γ j+1 1 ck (xk ) = γ0 . . . γk−1 c j (xk ) = 2. For d > 1, the monomial xα is a linear combination of πβ , with β ≤ α component wise, and vice versa. In formulæ πα = ∑ a β xβ and β ≤α xα = ∑ bβ πβ (15) β ≤α where β ≤ α holds component wise. Proof The proof of Item 1 is by induction and that of Item 2 follows by rearranging the coefficients in the product. They are given in Appendix 8. ⊓ ⊔ Example 10 If π j is the j-th Hermite polynomial, then Item 1 of Theorem 7 gives the well known result c j (xk ) = 0   k c j (xk ) = (k − j − 1)!! j if k + j is odd if k + j is even Direct application of Theorem 7 is cumbersome and we need only to characterise the polynomial functions for which the cubature formula is exact. So we proceed by another way. The finite set of distinct points D ⊂ Rd is associated to its vanishing polynomial ideal I (D) = { f ∈ R[x] : f (z) = 0 for all z ∈ D} Let LTσ ( f ) or LT ( f ) denote the largest term in a polynomial f with respect to a term-ordering σ . Let [ f (z)]z∈D be the evaluation vector of the polynomial f at D and for a finite set of polynomials G ⊂ R[x] let [g(z)]z∈D ,g∈G be the evaluation matrix whose columns are the evaluation vectors at D of the polynomials in G. In DoE often this matrix is called the X-matrix of D and G. As mentioned at the end of Section 1.1, the space L (D) of real valued functions defined over D is a linear space and particularly important vector space bases can be constructed as follows. Let LT (I (D)) = hLTσ ( f ) : f ∈ I (D)i. If G is the σ -reduced Gröbner basis of I (D), then LT (I (D)) = hLTσ ( f ) : f ∈ Gi. Now we can define two interesting vector 12 C Fassino et al. space bases of L (D). Let L = {α ∈ Zd≥0 : xα 6∈ LT (I (D))} and define B = {xα : α ∈ L} and OB = {πα : α ∈ L} Example 11 For d = 2 and L = {(0, 0), (1, 0), (0, 1), (2, 0)} we have B = {1, x, y, x2 } and OB = {1, π1(x), π1 (y), π2 (x)}, since π0 (x) = π0 (y) = 1. The sets L, B and OB depend on σ . It is well known that if t ∈ B and r divides t, then r ∈ B; it follows that if α ∈ L and β ≤ α component wise then also β belongs to L and πβ to OB. Note that σ induces a total ordering also on the orthogonal polynomials: πα <σ πβ if and only if xα <σ xβ ; analogously α <σ β if and only if xα <σ xβ for each α , β ∈ Zd≥0 . Here we used the same symbol to indicate related orderings over the α ’s, the xα ’s and the πα ’s. Further, given α ≤ β componentwise, since xα divides xβ and since 1 ≤σ xβ −α , we have xα ≤σ xβ , that is α ≤σ β . Now, given a term-ordering σ , any g ∈ G can be uniquely written as its leading term, xα = LT (g), and tail which is a linear combination of terms in B preceding LT (g) in σ , that is g = xα + ∑β ∈L,β <α α aβ xβ with aβ ∈ R. Theorem 8 provides an alternative to the classical method of rewriting a polynomial f in terms of orthogonal polynomials. It does so by applying Theorem 7 and by substituting each monomial in f . Theorem 8 gives linear rules to write the elements of G and the remainder of a polynomial divided by G as linear combinations of orthogonal polynomials of low enough order. The proof is in Appendix 8. Theorem 8 1. Span(B) = Span(OB); 2. Let G be the reduced σ -Gröbner basis of I (D). Each g ∈ G with LT (g) = xα is uniquely written as g = πα − ∑ β ∈L,β <σ α bβ πβ where b = [bβ ]β ∈L,β <σ α solves the linear system   πβ (z) z∈D ,β ∈L,β < α b = [πα (z)]z∈D ; in words the coefficient matrix is the evaluation matrix over D of the orthogonal polynomials πβ with xβ in tail of g and the vector of constant terms is the evaluation vector of πα . 3. Let p ∈ R[x] be a polynomial and [p(z)]z∈D its evaluation vector. The polynomial p∗ defined as ∑ aβ πβ   ρ = [πα (z)]z∈D − πβ (z) z∈D,β ∈L b Step 4 1. If kρ k > 0, then include α in L, and include in M those elements of {x1 xα , . . ., xd xα } which are not multiples of an element in M or of LT (g), g ∈ G. Return to Step 2. 2. If kρ k = 0, then include in G the polynomial ∑ bβ πβ β ∈L where the values bβ , β ∈ L, are the components of the solutions b of the linear system in Step 3. Delete from M all multiples of xα . Table 4 Buchberger-Möller algorithm using orthogonal polynomials (16) β ∈L where a = [aβ ]β ∈L solves the linear system   πβ (z) z∈D ,β ∈L a = [p(z)]z∈D Input: a set D of distinct points in Rd , a term-ordering σ and any vector norm || · ||. Output: the reduced σ -Gröbner basis G of I (D ) as linear combination of orthogonal polynomials and the set L. d }, OB = [1], G = [ ] and M = [x , . . ., x ]. Step 1 Let L = {0 ∈ Z≥0 1 d Step 2 If M = [ ] stop; else set xα = minσ (M) and delete xα from M. Step3 Solve in b the overdetermined linear system  πβ (z) z∈D,β ∈L b = [πα (z)]z∈D and compute the residual g = πα − σ p∗ = Theorem 8 provides a pseudo-algorithm to compute a Gröbner basis for I (D) and the interpolating polynomial at D in terms of orthogonal polynomials of low order directly from D and OB. Table 4 gives the algorithm which is a variation of the Buchberger-Möller algorithm [8]. It starts with a finite set of distinct points D and a term-ordering σ and it returns L and the expressions g = πα − ∑β ∈L,β <σ α bβ πβ for g in the reduced σ -Gröbner basis of I (D). It does so by performing linear operations. If the real vector [p(z)]z∈D is assigned, then the expression p∗ = ∑β ∈L aβ πβ can now be found using Item 3 in Theorem 8. This permits to rewrite every polynomial p ∈ R[x] as a linear combination of orthogonal polynomials. The algorithm in Table 4 returns the σ -reduced Gröbner basis as linear combination of orthogonal polynomials. It performs operations only with orthogonal polynomials and in particular it does not involve at any step the xα monomials. This is computationally faster than first computing a classical Gröbner basis in the xα and next substituting the πα . Furthermore working with only one vector space basis, and not switching between the xα and the πα , is conceptually appealing. is the unique polynomial belonging to Span(OB) such that p∗ (z) = p(z) for all z ∈ D. Summarising: given a function f , a finite set of distinct points D ⊂ Rd and a term-ordering σ , a probability product measure λ d over Rd , its system of product orthogonal polynomials, and a random vector with probability distribution λ d , then the expected value of f with respect to λ d can be approximated by Algebra of Cubature 13 1. computing L with the algorithm in Table 4 and 2. determining, by solving the linear system   πβ (z) z∈D ,β ∈L a = [ f (z)]z∈D , the unique polynomial p∗ such that p∗ (z) = f (z) for all z ∈ D. The polynomial p∗ is expressed as linear combination of orthogonal polynomials. 3. The coefficient a0 of π0 is the required approximation. Recall that p∗ (x) = ∑z∈D f (z)lz (x) is a linear combination of the indicator functions of the points in D (Lagrange polynomials) and hence a0 = ∑z∈D f (z) E (lz (X)). In particular, E (lz (X)) = λz , z ∈ D, can be computed by applying the above to f = lz . Notice however that as λ d is a product measure, the λz can be obtained from the one-dimensional ones as noticed before Theorem 6. It would be interesting to generalise this section to non-product measures. Here an algorithm has been provided to approximate the expected value of polynomials. Next the set of polynomials whose expected value coincides with the value of the cubature formula is characterised. In Section 6.1 we provide a characterisation of the full set via linear relationships that Fourier coefficients of suitable polynomials have to satisfy, while in Section 6.2 a possibly proper subset has been characterised via a simple condition on the total degree of the polynomials. 6.1 Characterisation of polynomial functions with zero expectation In this section we characterise the set of polynomials whose expected value coincides with the value of the cubature formula. As mentioned in Section 1.1 given D ∈ Rd , its vanishing ideal I (D), a term-ordering σ and the Gröbner basis G of I (D) with respect to σ , then any polynomial f ∈ R[x] can be written as f (x) = ∑ qg(x)g(x) + rσ (x) We do this in two ways. First we study the Fourier expansion of the elements of E0 , next we present some results about the degree of the elements belonging to Eσ . The elements of E0 are characterized in Theorem 9. Note that if f ∈ R[x] is such that f = p + rσ with p ∈ E0 and rσ ∈ Span(B) then by linearity and independence Eλ ( f ) = ∑ (z1 ,...,zn )∈D f (z1 , . . . , zn )λzn11 · · · λzndd Theorem 9 Let λ be a product probability measure with product orthogonal polynomials πα (x), α ∈ Zd≥0 and let X be a random vector with distribution λ d . Let D ⊂ Rd be a set of distinct points, σ a term-ordering, G the σ -reduced Gröbner basis of I (D) whose elements as linear combinations of orthogonal polynomials. Thus for g ∈ G and xα = LT (g) we write g = πα − ∑ α >σ β ∈L cβ (g)πβ where α >σ β ∈ L stands for α >σ β and β ∈ L. Let p = ∑g∈G qg g ∈ I (D) for suitable qg ∈ R[x], and consider the Fourier expansion of each qg , g ∈ G, qg = ∑ cβ (qg )πβ (17) β ∈Zd≥0 Then Eλ (p(X)) = 0 if and only if ∑ kπα kλ2 cα (qg ) − ∑ ∑ g∈G g∈G α >σ β ∈L kπβ kλ2 cβ (qg )cβ (g) = 0 (18) Proof The key observation is that Eλ (πm πn ) = 0 if n 6= m and then linearity of Eλ is used. The proof can be found in Appendix 8. ⊓ ⊔ Importantly, only terms of low enough Fourier order in Equation (17) matter for the computation of the expectation. Example 12 Consider Z1 and Z2 two independent standard normal random variables and hence the Hermite polynomials. Consider also the five point design g∈G where rσ (x) is unique in Span(B) such that rσ (z) = f (z) for all z ∈ D and can be written as rσ (x) = ∑z∈D f (z)lz (x), where the lz ’s are the product Lagrange polynomials in Section 5. Theorem 8 states how to write rσ over OB. If f ∈ R[x] is such that Eλ ( f (X)) = Eλ (rσ (X)) then we have Eλ ( f (X) − rσ (X)) = 0, where f − rσ ∈ I (D). Hence to study the set Eσ = { f ∈ R[x] : Eλ ( f (X)) = Eλ (rσ (X))} D = {(−6, −1), (−5, 0), (−2, 1), (3, 2), (10, 3)} and the σ =DegLex term-ordering over the monomials in R[x, y]. The algorithm in Table 4 gives G = {g1, g2 , g3 } where g1 = H2 (y) − H1 (x) + 2H1(y) − 4 g2 = H2 (x)H1 (y) − 9H2(x) + 47H1(x)H1 (y) − 123H1(x) +271H1(y) − 399 g3 = H3 (x) − 47H2(x) + 300H1(x)H1 (y) − 845H1(x) +2040H1(y) − 2987 is equivalent to characterize the set and L = {(0, 0), (1, 0), (0, 1), (1, 1), (0, 2)}, that is E0 = {p ∈ I (D) : Eλ (p(X)) = 0} OB = {1, H1 (y), H1 (x), H1 (x)H1 (y), H2 (x)} 14 C Fassino et al. (1) c0 = −34 (2) c0 = 0 (3) c0 (1) c(1,0) = 0 (2) c(2,0) = 0 (3) c(2,0) = 10 =2 (1) c(0,1) = −2 (2) c(1,1) = 1 (3) c(1,1) =0 (1) c(0,0) = −8 (2) c(1,0) = −2 (3) c(1,0) = −5 (2) c(0,1) = 1 (3) c(0,1) =1 (2) c(0,0) = 2 (3) c(0,0) = 5863/2987 Table 5 A solution for (18) By Theorem 9 for the purpose of computing its expectation a polynomial p = q1 g1 + q2 g2 + q3g3 ∈ I (D) can be simplified to have the form   (1) (1) (1) (1) p = c0 H2 (y) + c(1,0)H1 (x) + c(0,1)H1 (y) + c(0,0) g1 +  (2) (2) (2) c0 H2 (x)H1 (y) + c(2,0)H2 (x) + c(1,1)H1 (x)H1 (y)+   (3) (2) (2) (2) c(1,0) H1 (x) + c(0,1)H1 (y) + c(0,0) g2 + c0 H3 (x)+ (3) (3) (3) c(2,0) H2 (x) + c(1,1)H1 (x)H1 (y) + c(1,0)H1 (x)+ (3) (3) c(0,1) H1 (y) + c(0,0) and furthermore by Equation (18) (1) (1) (1) (1) (2)  (2) g3 (2) (2) (3) (2) (3) − 123c(1,0) + 271c(0,1) − 399c(0,0) + c0 3! − 47c(2,0)2! (3) (3) (3) (3) + 300c(1,1) − 845c(1,0) + 2040c(0,1) − 2987c(0,0) =0 In practice, for i = 1, 2, 3, put coefficients of gi and qi in two vectors, multiply them component wise and sum the result. There are infinite polynomials that satisfy the above equations, one such polynomial is given in Table 5. The above equation involves only a finite number of Fourier coefficients, namely cβ (qg )’s is relevant for the equation if and only if β ∈ L and xβ is smaller in σ than the leading terms of g ∈ G. Hence if to qg we add a polynomial of the form ∑β >α or β 6∈L cβ Hβ we still obtain a zero mean polynomial. That is, we can modify qg by adding high enough terms without changing the mean value. For example by adding H4 (x) to q1 and H4 (y) to q2 we obtain the following zero mean polynomial p(Z1 , Z2 ) = Z12 Z25 + 10Z16 + Z14 Z22 − 9Z12Z24 + 47Z1Z25 − 469Z15+ 3002Z14Z2 + Z13 Z22 − 6Z12Z23 − 123Z1Z24 + 270Z25− 8614Z14 + 20990Z13Z2 + 96Z12Z22 − 282Z1Z23 − 424Z24 Here we adopt another viewpoint and characterise the set Eσ . Instead of studying the Fourier expansion of its polynomials, we focus our attention on their degree. Given a degree compatible term ordering σ , we show how to compute the maximum degree s such that R[x]≤s is in Eσ , that is the degree of the cubature formula with nodes D. Our strategy is based on the definition of s-orthogonal polynomials [14]. Definition 1 A polynomial g ∈ R[x] is s-orthogonal if s ∈ Z>0 is the maximum integer such that f g ∈ R[x]≤s c0 2! − c(1,0) + 2c(0,1) − 4c(0,0) + c0 2! − 9c(2,0)2! + 47c(1,1) (2) 6.2 On exactness of cubature formulæ − 87898560/2987Z13 − 6700Z12Z2 + 1389Z1Z22 − 1690Z23+ 71785814 2 218275468 Z1 − Z1 Z2 + 4845Z22+ 2987 2987 5937584 5931425 307862660 Z1 − Z2 − 2987 2987 2987 implies Eλ ( f (X)g(X)) = 0 Furthermore, a set G of polynomials is s-orthogonal if each g ∈ G is s(g)-orthogonal and s = ming∈G s(g). Theorem 10 reformulates and summarizes two theorems about the degree of a cubature formula presented in [13] and [14] (where H-bases are considered). Theorem 10 Given a set D and a degree compatible term ordering σ the following conditions are equivalent. 1. R[x]≤s ⊂ Eσ ; 2. Eλ ( f (X)) = 0 for all f ∈ I (D) ∩ R[x]≤s ; 3. the σ -Gröbner basis G of I (D) is s-orthogonal. Proof 1 ⇒ 2. Let f ∈ I (D) ∩ R[x]≤s . Since by hypothesis R[x]≤s ⊂ Eσ , then f ∈ Eσ that is Eλ ( f (X)) = Eλ (rσ (X)). Since f ∈ I(D), we have rσ (x) ≡ 0 and so Eλ ( f (X)) = 0. 2 ⇒ 3. For each g ∈ G let f be such that f g ∈ R[x]≤s . Since f g ∈ I (D) ∩ R[x]≤s then Eλ ( f (X)g(X)) = 0 and so g is s-orthogonal. 3 ⇒ 1. For p ∈ R[x]≤s , we have that p= ∑ gqg + rσ g∈G where each gqg is such that LTσ (gqg ) ≤ LTσ (p) and so, since σ is degree compatible, deg(gqg ) ≤ deg(p) ≤ s. It follows that, since G is s-orthogonal, Eλ (gqg ) = 0. By linearity we obtain Eλ (p(X)) = Eλ (rσ (X)), that is p ∈ Eσ . ⊓ ⊔ Remark 3 The maximum integer s such that Eλ ( f (X)) = 0 for each f ∈ I (D) ∩ R[x]≤s is the degree of the cubature formula with nodes D and with respect to σ . Algebra of Cubature 15 Theorem 10 shows that the maximum integer s such that all polynomials of total degree s are in Eσ coincides with the maximum s such that G is s-orthogonal. Hence we focus our attention on the s-orthogonality of the elements of G. Theorem 11 Each polynomial g ∈ R[x] is s-orthogonal with deg(g) − 1 ≤ s < 2 deg(g). Proof Since f g ∈ R[x]≤(deg(g)−1) if and only if f is the identically zero polynomial, then f g ∈ R[x]≤(deg(g)−1) implies Eλ ( f (X)g(X)) = 0, that is g is always s-orthogonal with s ≥ deg(g) − 1. Moreover, s < 2 deg(g); in fact, g2 belongs to R[x]≤2 deg(g) and, from the orthogonality of the polynomials πα , we have Eλ (g(X)g(X)) = ∑(cα (g))2 kπα k2 α As g is not identically zero, then Eλ (g(X)g(X)) 6= 0 and so g cannot be 2 deg(g)-orthogonal. ⊓ ⊔ The following theorem shows how to detect the s-orthogonality of a polynomial analysing its Fourier coefficients. Theorem 12 Let g = ∑α cα (g)πα the Fourier expansion of a polynomial g ∈ R[x]. 1. If c0 (g) 6= 0, the polynomial g is (deg(g)−1)-orthogonal. 2. If c0 (g) = 0, the polynomial g is s-orthogonal where s is such that cα (g) = 0 for all α s.t. ∑di=1 αi ≤ s − deg(g). Proof 1. Let f be a constant polynomial, f ≡ 1. We have that f g ∈ R[x]≤deg(g) implies Eλ ( f (X)g(X)) = Eλ (g(X)) = c0 (g) 6= 0 and so g is not deg(g)-orthogonal. From Theorem 11 we conclude that g is (deg(g) − 1)-orthogonal. 2. Let s < 2 deg(g) and let f be a polynomial such that f g ∈ R[x]≤s . Since deg( f ) ≤ s − deg(g) the Fourier expansion of f is such that f = ∑|α |≤s−deg(g) cα ( f )πα . From the orthogonality of the polynomials πα and from s−deg(g) < deg(g) we have Eλ ( f (X)g(X)) = ∑ |α |≤s−deg(g) cα ( f )cα (g)kπα k2 The generality of f implies that Eλ ( f (X)g(X)) = 0 only if cα (g) = 0 for each α such that |α | ≤ s − deg(g). ⊓ ⊔ Corollary 2 Given a finite set of distinct points D ∈ Rd and a degree compatible term ordering σ on R[x], let G be the reduced σ -Gröbner basis of I (D). Then the maximum integer sc such that R[x]≤sc is in Eσ is sc = min s(g) g∈G where s(g) is such that the Fourier expansion of each g ∈ G is given by g = ∑ cα (g)πα α with cα (g) = 0 forall α such that ∑di=1 αi ≤ s(g) − deg(g). Proof By Theorem 10 the thesis follows if G is sc -orthogonal. But this is true since, from Theorem 12, we have that each g ∈ G is s(g)-orthogonal. ⊓ ⊔ Example 13 Given the set of points √ √ D = {(−1, 0), (−1, −2), (1, −1+ 3), (1, −1− 3), (2, 1)} and the degree compatible term ordering σ =DegLex, then by e.g. the algorithm in Table 4 the reduced σ -Gröbner basis G of I (D) can be written as   g1 = H2 (y) − H1(x) + 2H1(y) G = g2 = H2 (x)H1 (y) − H2(x)  g3 = H3 (x) − 2H2(x) + 2H1(x) Since for each g ∈ G, s(g) is such that cα (g) = 0 for each α s.t. ∑di=1 αi ≤ s(g) − deg(g), we have that s(g1 ) = 2, s(g2 ) = 4 and s(g3 ) = 3. It follows that G is sc -orthogonal, with sc = min{s(g1 ), s(g2 ), s(g3 )} = 2 and so, from Corollary 2 it follows that the maximum integer s such that R[x]≤s ⊂ Eσ is 2. Example 14 For D of Example 12 and σ =DegLex, the reduced σ -Gröbner basis G = {g1 , g2 , g3 } of I (D) is such that s(g1 ) = 1, s(g2 ) = 2 and s(g3 ) = 3 since c0 (g) 6= 0 for all g ∈ G. It follows that G is 1-orthogonal and so the cubature formula w.r.t. σ and D is exact for all the polynomials with degree 0 and 1. Nevertheless let us remark that the cubature formula is exact for a much larger class of polynomials as shown in Theorem 9 and in Example 12. 7 Conclusion In this paper we mixed tools from Computational Commutative Algebra, orthogonal polynomial theory and Probability to address the recurrent statistical problem of estimation of mean values of polynomial functions. Our work shares great similarity with applications of computational algebra to design and analysis of experiments, which inspired us with a non-classical viewpoint to cubature formulae. We obtained two main results. In the Gaussian case we obtained a system of polynomial equations whose solution gives the weights of a quadrature formula (Theorem 13). For a finite product measure which admits an orthogonal system of polynomials, we characterise the set of polynomials with the same mean value. This depends substantially from Equation (18) and it is in terms of Fourier coefficients of particular polynomials obtained by adapting Gröbner basis theory. 16 C Fassino et al. 8 Appendix: proofs Theorem 7: d Proof 1. The proof is by induction on the monomial degree k. From the three terms recurrence formula π j+1 = (γ j x − α j )π j − β j π j−1 we have xπ j = π j+1 α j βj + π j + π j−1 γj γj γj For k = 0 we have x0 = π0 (x) = c0 (x0 )π0 . For k = 1 from the three terms recurrence formula we have π1 α0 + π0 = c1 (x)π1 + c0 (x)π0 x = xπ 0 = γ0 γ0 In the inductive step the thesis holds for k and we prove it for k + 1. From the three terms recurrence formula we have xk+1 =xxk = k ∑ c j (xk )xπ j j=0  π j+1 α j βj = ∑ c j (x ) + π j + π j−1 γ γ γj j j j=0 k k  k k−1 c j−1 (xk ) αj β j+1 π j + ∑ c j (xk ) π j + ∑ c j+1 (xk ) πj γj γ j+1 j=1 γ j−1 j=0 j=0  k−1  c j−1 (xk ) αj β j+1 πj + c j (xk ) + c j+1(xk ) =∑ γ j−1 γj γ j+1 j=1 k+1 =∑ ck−1 (xk ) c (xk ) πk + k πk+1 γk−1 γk c1 (xk )β1 c0 (xk )α0 c (xk )αk π0 + k πk + π0 + γ0 γk γ1 + k−1 = ∑ c j (xk+1 )π j + ck+1(xk+1 )πk+1 j=1  ck−1 (xk ) ck (xk )αk + γk−1 γk  πk   c0 (xk )α0 c1 (xk )β1 + + π0 γ0 γ1 + This concludes the proof of the first part of the theorem. To prove the second part we apply what we just proved α and unfold the multiplication. Given xα = xα1 1 · · · xd d , the polynomial πα = πα1 (x1 ) · · · παd (xd ) is the product of d univariate polynomials πα j each of degree α j in x j , j = 1, . . . , d. Clearly if α j = 0 then πα j = 1 and x j does not divide xα . Furthermore we have d αj xβ with β ≤ α component wise. Vice versa, applying the first part of the theorem we have " # πα = ∏ ∑ dk xkj ( j) j=1 k=0 We deduce that πα is a linear combination of xα and of the power products which divide xα , that is of power products αk d α xα = ∏ xk k = ∏ k=1 k=1 α ∑ c jk (xk k )π jk (xk ) jk =0 and commuting product with sum shows that xα is a linear combination of products of πβi (xi ) where β = (β1 , . . . , βd ) is such that β ≤ α component wise, that is xβ divides xα . ⊓ ⊔ Theorem 8: Proof Recall that B and OB are defined in terms of a common set L of d-dimensional vectors with non-negative integer entries satisfying the property of ‘factor-closeness’, that is if (α1 , . . . , αd ) ∈ L and if βi ≤ αi for all i = 1, . . . , d then (β1 , . . . , βd ) ∈ L. 1. If xα ∈ B for some α , then from Theorem 7 xα = ∑β ≤α bβ πβ follows. Since β ≤ α then β ∈ L and so each πβ ∈ OB: we have that xα belongs to Span(OB). The vice versa is proved analogously. 2. The matrix [πβ (z)]z∈D ,β ∈L is a square matrix since L has as many elements as D and has full rank. The linear independence of the columns of this matrix follows from the fact that each linear combination of its columns corresponds to a polynomial in Span(OB) which coincides with Span(B). Any polynomial g ∈ G can be written as g = xα − ∑ cβ xβ α >σ β ∈L where xα = LT (g) is a multiple by some x j of an element of B. By Theorem 7 we have g= ∑ aγ (g) γ ≤α πγ − ∑ α >σ β ∈L cβ ∑ dγ (g) γ ≤β πγ The polynomial πα appears only in the first sum with the coefficient 1. For the other terms in the first sum observe that as γ < α then γ ∈ L and also γ <σ α . Analogously, for the second sum consider γ ≤ β < α ; since β ∈ L then γ ∈ L and since γ < α then γ <σ α . And so, with obvious notation, g = πα − ∑ α >σ β ∈L bβ πβ Since g(z) = 0 for z ∈ D, then the vector b = [bβ ]β of the coefficients in the identity above solves the linear system [πβ (z)]z∈D ,α >σ β ∈L b = [πα (z)]z∈D . Furthermore, since [πβ (z)]z∈D ,α >σ β ∈L is a full rank matrix, then b is the unique solution of such a system. 3. Let p∗ = ∑β ∈L aβ πβ be the polynomial whose coefficients are the solution of the linear system [πβ (z)]z∈D ,β ∈L a = [p(z)]z∈D . Algebra of Cubature 17 Such a polynomial obviously interpolates the values p(z), z ∈ D, and, since the columns of [πβ (z)]z∈D ,β ∈L are the evaluation vectors of the elements of OB at D, it belongs to Span(OB). We conclude that p∗ is the unique polynomial belonging to Span(OB) which interpolates the values p(z), z ∈ D. ⊓ ⊔ Theorem 9: Proof As G is a basis of I (D), then for every p ∈ I (D) and g ∈ G there exist qg ∈ R[x] such that p = ∑g∈G qg g. Since by linearity ! Eλ ∑ qg g = g∈G ∑ Eλ (qg g) g∈G the thesis follows once we show that, for each g ∈ G and xα = LT (g) Eλ (qg g) = kπα k2λ cα (qg ) − ∑ α >σ β ∈L cβ (qg )cβ (g)kπβ kλ2 holds. From Equation (17) we have qg g = ∑ cβ (qg )πβ g β and we substitute the Fourier expansion of g given in Theorem 8 g = πα − ∑ α >σ β ∈L cβ (g)πβ In computing the expectation we use the fact that Eλ (πh πk ) = 0 for different h and k. Then the expectation of cβ (qg )πβ g vanishes if β >σ α or β <σ α , β ∈ / L, the expectation of cα (qg )πα g gives kπα k2λ cα (qg ) and, if α >σ β ∈ L, the ex⊓ ⊔ pectation of cβ (qg )πβ g gives −cβ (qg )cβ (g)kπβ k2λ . Acknowledgments G. Pistone is supported by de Castro Statistics Initiative, Collegio Carlo Alberto, Moncalieri Italy. E. Riccomagno worked on this paper while visiting the Department of Statistics, University of Warwick, and the Faculty of Statistics at TU-Dortmund on a DAAD grant. Financial support is gratefully acknowledged. The authors thank Prof. H. P. Wynn, Prof. G. Monegato (Politecnico di Torino) and Prof. Dr. Hans Michael Möller (Technische Universität – Dortmund) for their useful suggestions. References [1] CoCoATeam, CoCoA: a system for putations in Commutative Algebra. http://cocoa.dima.unige.it. doing ComAvailable at [2] David Cox, John Little, and Donal O’Shea, Ideals, varieties, and algorithms, Third, Undergraduate Texts in Mathematics, Springer, New York, 2007. [3] Mathias Drton, Bernd Sturmfels, and Seth Sullivant, Lectures on Algebraic Statistics, Birkhäuser, Basel, 2009. [4] Roberto Fontana, Giovanni Pistone, and Maria Piera Rogantin, Classification of two-level factorial fractions, J. Statist. Plann. Inference 87 (2000), no. 1, 149–172. [5] Walter Gautschi, Orthogonal polynomials: computation and approximation, Numerical Mathematics and Scientific Computation, Oxford University Press, New York, 2004. [6] Paolo Gibilisco, Eva Riccomagno, Maria Piera Rogantin, and Henry P. Wynn (eds.), Algebraic and geometric methods in statistics, Cambridge University Press, Cambridge, 2010. [7] Paul Malliavin, Integration and probability, Graduate Texts in Mathematics, vol. 157, Springer-Verlag, New York, 1995. [8] H. M. Möller and B. Buchberger, The construction of multivariate polynomials with preassigned zeros, Computer algebra (Marseille, 1982), 1982, pp. 24–31. [9] Giovanni Peccati and Murad S. Taqqu, Wiener Chaos: Moments, Cumulants and Diagrams, Bocconi & Springer-Verlag, Italia, 2011. [10] Giovanni Pistone, Eva Riccomagno, and Henry P. Wynn, Algebraic statistics, Monographs on Statistics and Applied Probability, Chapman & Hall/CRC, Boca Raton, FL, 2001. [11] Wim Schoutens, Orthogonal polynomials in Stein’s method, J. Math. Anal. Appl. 253 (2001), no. 2, 515–531. [12] A. H. Stroud and Don Secrest, Gaussian quadrature formulas, Prentice-Hall Inc., Englewood Cliffs, N.J., 1966. [13] Yuan Xu, Polynomial interpolation in several variables, cubature formulæ, and ideals, Adv. Comput. Math. 12 (2000), no. 4, 363– 376. [14] H. Michael Möller, On the construction of cubature formulae with few nodes using Groebner bases, Numerical integration (Halifax, N.S., 1986), NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., vol. 203, Reidel, Dordrecht, 1987, pp. 177–192.
0math.AC
Depression Scale Recognition from Audio, Visual and Text Analysis Shubham Dham∗ , Anirudh Sharma† , Abhinav Dhall‡ arXiv:1709.05865v1 [cs.CV] 18 Sep 2017 Department of Computer Science and Engineering, Indian Institute of Technology (IIT) Ropar, India Email: ∗ [email protected] , † [email protected] , ‡ [email protected] Abstract—Depression is a major mental health disorder that is rapidly affecting lives worldwide. Depression not only impacts emotional but also physical and psychological state of the person. Its symptoms include lack of interest in daily activities, feeling low, anxiety, frustration, loss of weight and even feeling of selfhatred. This report describes work done by us for Audio Visual Emotion Challenge (AVEC) 2017 during our second year BTech summer internship. With the increase in demand to detect depression automatically with the help of machine learning algorithms, we present our multimodal feature extraction and decision level fusion approach for the same. Features are extracted by processing on the provided Distress Analysis Interview Corpus-Wizard of Oz (DAIC-WOZ) database. Gaussian Mixture Model (GMM) clustering and Fisher vector approach were applied on the visual data; statistical descriptors on gaze, pose; low level audio features and head pose and text features were also extracted. Classification is done on fused as well as independent features using Support Vector Machine (SVM) and neural networks. The results obtained were able to cross the provided baseline on validation data set by 17% on audio features and 24.5% on video features. Keywords—AVEC 2017, SVM, Depression, Neural network, RMSE, MAE, fusion, speech processing. I. I NTRODUCTION Depression is a common mood disorder which affect the lives of the individual suffering from it. It is a world wide problem affecting innumerable lives. Depressed people are more prone to anxiety, sadness, loneliness, hopelessness, and are frequently worried and disinterested. They find it hard to concentrate on their work, communicate with people and even become introvert [18]. The may suffer from insomnia, restlessness, loss of appetite and may have suicidal thoughts. To be diagnosed with depression, the symptoms must be present for at least two weeks [1]. Hence detection of depression is a major issue. The present techniques rely on clinicians review of the patient in person. Those methods are subjective, done on interview and depend on reports by the patient. With rise in the depression, some automatic and reliable means of depression recognition is required. Efforts are being made in this direction to assess and detect depression through computer vision and machine learning. Depression detection can be done through audio and video recordings of the patient. Depressed people behave different from normal people which can be detected in audio and visual recordings of the patient. Studies have shown that depressed people tend to avoid eye contact, engage less in verbal communication, speak anxiously in short phrases and monotonously [2] [3] [4]. AVEC provides opportunity to teams from around the globe to come forth and develop a program for identifying depression using the dataset provided. Taking motivation from AVEC 2016 [5], our team participated in AVEC 2017 [6], which has similar dataset as AVEC 2016. Hence in this paper we present our methods of visual, audio and text feature extraction, followed by decision level fusion which help identify depressed subjects with appreciable accuracy as evident in results obtained. II. R ELATED W ORK Recent researches have been very prominent in detecting depressed people. Previous year’s AVEC depression sub challenge have brought papers with decent accuracies as depicted in the results published. Le Yang et al. [23] during AVEC 2016 achieved quite plausible accuracy through Decision Tree classification and multimodal fusion of audio-visual and text features. Asim Jan et al. [7] in AVEC 2014 competition used Motion History Histogram for depression recognition by extracting Local Binary Pattern (LBP) and Edge Orientation Histogram features (EOF). Low Lu-Shih et al. [8] proposed description of depression recognition using acoustic speech analysis. The five acoustic feature categories used were prosodic, cepstral, spectral, glottal and Teager energy operator (TEO) based features, with TEO observed to perform best. The paper introduced benefits of Gaussian Mixture Model (GMM) clustering in depression recognition. Douglas Sturim et al. [9] also brought forward GMM clustering in classification on speech data. Also other features such as head pose, blinking rate, and even textual features have been used for depression recognition. Head pose and facial movement is a discriminative feature to demarcate depressed people from non-depressed. Depressed people show less nodding [3] , with more likeliness to position head down than non-depressed people [4], avoid eye contact [3]. Alghowinem et al. [10] used head pose and movement features associated with the face to perform classification with SVM on depression recognition; and concluded that head movements of depressed people are different than that of normal person. Inspired by the present deep learning performance in detecting emotion, Tzirakis et al. [11] in AVEC 2016 presented a deep learning based approach in assessing emotion as well depression state of the person. They used Convolution Neural Network (CNN) on audio and Deep residual Network (resNet) of 50 layers on visual data. They concluded that their deep learning approach achieved relatively better results than other methods at the time. Pampouchidou et al. [12] in AVEC 2016 extracted features from the recorded verbal communication recorded in the transcript file provided in the data. Arousalvalence rating of the words which negative emotion were taken as a feature and they concluded that its removal had a negative impact on the overall accuracy obtained on the trained fused model. Also Cummins et al. [19] during AVEC 2013, presented multimodal approach by fusing audio and visual modalities with different techniques to detect depression. Hence all the recent researches are emphasizing on the need of depression recognition in some automatic and fruitful way to provide help to clinicians and patients. Therefore in this paper we also present our approach to detect the depression. III. F EATURE E XTRACTION A. Video Features The challenge organizers did not share the raw video recordings, however 2-D, 3-D facial landmarks, Action Units (AUs), gaze and pose features were provided. Therefore we performed preprocessing on the 2-D facial landmarks and obtained features in the form of head features, distance and blink rate. Statistics derived from the provided features of emotion, AUs, gaze, and pose were also used as complementary features. 1) Head Features: The Head pose features provided by the organizers were not used as they did not depict temporal information, i.e features did not convey the change with respect to time. Hence Head Motion was judged by horizontal and vertical motion of certain facial points, i.e., points 2, 4, 14, and 16 [Fig. 1] [12]. The points selected were those that are minimally involved in facial movements and expressions such as blinking, smiling and other facial expressions, such that their motion prominently represent head motion. For each of the above facial points, change in their position was calculated between every consecutive frame. This change was measured in both horizontal and vertical direction as well as net magnitude. Then statistical features were calculated as mean, median and mode of displacement in horizontal, vertical direction and magnitude of displacement, and velocity in horizontal, vertical direction and magnitude of velocity. regarding effective state of the person and hence one may find correlation among them. This was implemented as features extracted on the facial regions, namely, eyes, mouth, eyebrows and head. The data was sub-sampled as changes between two consecutive frames would be negligible. Hence one in every three frames was used. Initially affine transformation was applied on the facial points over all the frames to get rid of unwanted scale, translation and rotational variance. The reference points used were taken from a particular frame, because the points for that frame when plotted gave a perfectly aligned face with respect to the frame. And this was observed for all the videos. For left eye, distance between points {37, 40} was used as horizontal distance, and between {38, 42} for vertical distance. Similarly for right eye, {43, 46}, {44, 48} for horizontal and vertical distance respectively. For mouth, points {52, 58} were used for vertical distance. But for horizontal distance, average of the distance between two pair of points {55, 49}, {65, 61} was used. For head, horizontal distance was the average of distance between two pair of points-{2, 16} and {4, 14}, and vertical distance was the average of {22, 8} and {23, 10}. Since the motion of the two eyebrows occur simultaneously, hence their horizontal and vertical motion was calculated together. For the horizontal motion, average of the distance between two pair of points{22, 23} and {27, 18} was used. The vertical motion of the eyebrows was judged with respect to a reference point that hardly moves in the facial expressions, i.e., nose tip. Hence for vertical motion, average of the distance between pairs{31, 25}, {31, 20} was used. Hence these distances were calculated for all the frames and were stored in a vector. Hence this resulted in a total of 10 vectors per video. These vectors were individually normalized by its sum to account for the different type of faces, i.e., a person may have long face or small eyes, but it should not affect our results. In this way normalization removed the unwanted effects that the characteristics of a person’s face may have. The resultant vectors produced were further processed using Gaussian Mixture Model (GMM) for producing bag of words. Then the fisher vectors [20] [21] were extracted for each video from the clusters produced using GMM. GMM model gives the probability of points belonging to a cluster. According to the model the, probability is normally distributed around the mean/centroid of the cluster. The resultant vectors produced from above approach were further classified into 64 clusters using K-means. Then the clusters formed by K-means were used as inputs for the initialization of the EM (expectation maximization) for forming the GMM. Fig. 1: Facial Landmarks. 2) Distance: Head motion and facial expressions combined can give substantial information regarding the behavior of a person. Their temporal information may convey information 3) Emotions, AUs, Gaze and Pose: Statistical features, namely minimum, maximum, mean, mode, median, range, mean deviation, variance, standard deviation, skewness, and kurtosis were calculated for pre extracted AU, gaze and pose features provided in the given data set. These statistical features were also used for depression analysis. 4) Blink Rate: Blink rate was calculated using the 2D facial landmarks. First, the region of points enclosing the eye (points 37-42, of left eye) were taken. For the entire number of frames given, the area of the polygon made by those points was calculated. This way, the data eye area VS frame data was obtained. The given data was plotted to obtain an idea of blink rate. Now to obtain approximate area of open eye, mode of the total number of areas (per frame) was obtained and considered as area of open eye. For the close eye area, the minimum of the area of random 1000 frames was taken. A blink was considered if the area covered by the eye points is less than 90 percent of the area with eye opened. This way the number of blinks were calculated over the entire number of frames. The blink frequency was calculated by dividing the number of blinks by corresponding duration of the interview. The sample of the graph of eye area VS frames is shown below. Low Level Descriptors normalized F0, NAQ, QOQ, H1H2, PSP, MDQ, peak Slope, Rd, Rd conf, MCEP 0-24, HMPDM 1-24, HMPDD 1-12 Statistical features mean, min, skewness, kurtosis, standard deviation, median, peak-magnitude to root-mean-square ratio, root mean square level, interquartile range TABLE I: Statistical descriptors calculated from the pre-extracted audio features. the two sets of features were concatenated and used for further analysis. C. Text Features Fig. 2: eye area VS frames. 5) Other Features: For visual features, Motion History image that converts the motion into grayscale image, with the most recent movement shown by pixels with highest grayscale value and earliest by pixels with least grayscale value [13], was also computed on the facial 2D landmarks. The in between motion depicted by gradually increasing grayscale value, with 255 for most recent activity. HOG and LBP features were extracted on the resulting MHI formed. But the results obtained from classification on the HOG and LBP features extracted, were vague and inaccurate. Hence this approach is not discussed in detail. B. Audio Features The audio data provided by AVEC consisted of an audio file of entire interview with the participant, pre-extracted features using the COVAREP toolbox at 10-ms intervals over the entire recording (F0, NAQ, QOQ, H1H2, PSP, MDQ, peak Slope, Rd, Rd conf, MCEP 0-24, HMPDM 1-24, HMPDD 1-12, and Formants 1-3), Transcript file containing speaking times and values of participant and the virtual interviewer, and a formant file. The given audio file was processed so as to obtain the voice of the participant only. Hence from the audio file, the voice of the participant was isolated using the speaking times giveb in transcript file. Two sets of features were calculated for audio modality. The first set consists of statistical features for low level descriptors as shown in the following Table I. These features were calculated using the inbuilt functions in matlab. These low level descriptors have been pre extracted and provided with the data set in the covarep file. The second set of audio features consisted of Discrete Cosine Transform (DCT) coefficients for each descriptor in the first column of Table 1. The first 10 values of the DCT were retained to reduce the complexity of processing data. Finally, Features were also extracted from the verbal responses of the participant as given in the transcript file. Total number of sentences, words spoken by the participant, average words spoken in each sentence, ratio of the laughter count to the total number of words spoken by participant, ratio of the depression related words to the total number of words spoken by the participant, are the text features extracted. The total number of sentences and words spoken were normalized by duration of the video. Referring to the mentioned paper [14], slow and less amount of speech, longer speech pauses and brief answers are manifestations of depression. For the negative/depression related words, a dictionary of about 720 words was constructed manually using online resources [15], [16]. Another seven features were extracted using the Affective Norms for English Words ratings (ANEW) [17]. Mean and Standard Deviation of pleasure, arousal and dominance ratings, word frequency for each word spoken by participant were stored. Eventually mean of these features were taken over all the words, giving total of seven features. IV. C LASSIFICATION SVM and Neural Network were used as classification algorithms. SVM classifier was applied separately on all the above extracted features. For each of the above features, eight models were trained, with each model giving score 0 ,1, 2 or 3 as intensities of PHQ8 NoInterest, PHQ8 Depressed, PHQ8 Sleep, PHQ8 Tired, PHQ8 Appetite, PHQ8 Failure, PHQ8 Concentrating and PHQ8 Moving. Results of all the eight models were added to obtain the final predicted PHQ8 score [22] out of 24. Neural Network was applied only on the Fisher vectors calculated because the dimension of other feature vectors for a video were very small and hence application of layers to such small dimension did not make sense. Also it was checked experimentally and the results were not good. Hence for Fisher vectors, similar to SVM, 8 classification networks were trained, whose sum gave the net score out of 24. Another regression neural network model was also trained that gave the combined output of all the 8 PHQ8 labels. The model was constructed with 8 nodes as output of the last layer with each node giving value between 0-3 supposedly. The score of all 8 nodes obtained was first rounded to nearest integer value. Hence the sum of all the 8 nodes gave score out of 24 that was used to judge depression. In this model, interlink was kind of set between different PHQ8 labels and was expected to give better results. But the network with 6 layers gave best RMSE 6.85. V. E XPERIMENTAL R ESULTS A. Support Vector Machine (SVM) Initially the default parameters were used for the SVM classifier. But the training accuracy obtained with default parameters was very less. Hence a K-fold cross validation was applied to obtain the optimum values of cost and gamma for the classifier for each of the features extracted above. Cross validation performed was 5 fold, applied using the python script. With cross validation, model was able to fit the training data for a number of values of cost and gamma. Eventually those values of cost and gamma were selected for which accuracy on development set was maximum. The best results along with SVM parameters, obtained for each of the above, i.e , audio, text, head pose and fisher vectors on validation dataset are displayed in Table 2. The results for other modalities are not listed as they were not even close to the baseline. Best RMSE on validation set obtained for Histogram of Oriented Gradients (HOG) on MHI was 7.72. Similar RSME was obtained for Local Binary Pattern (LBP), statistical features of AU, gaze pose and blink rate. features Audio Text Head fisher RMSE 6.32 6.45 6.39 6.46 MAE 4.4 5.0 5.08 4.91 kernel Linear Radical Cyclic radical cost 9 9 5 5 gamma -7 -7 1 1 TABLE II: Results obtained with SVM on validation set. All these results are on the development set. The results obtained were better than the baseline (Table III) on the development set. For test set, only one submission was made till date, whose result, as mentioned in the Table is not that accurate. Partition Development Development Development Test Test Test Modality Audio Video Audio-Video Audio Video Audio-Video RMSE 6.74 7.13 6.62 7.78 6.97 7.05 MAE 5.36 5.8 5.52 5.72 6.12 5.66 TABLE III: Provided Baseline Results. B. Neural Networks For all the eight models, a total of 6 layers neural network was used. Dropout was added between most layers to prevent overfitting. The range of dropout was kept between 0.20.5. The number of dropouts and dimensions of layers were manually found so as to keep the RMSE and MAE of each of the network minimum. For all the networks, ‘Adam’ optimizer was used over ‘SGD’ as it was faster as well as efficient. The best RMS and MAE for both optimizers is given in Table IV. Adam SGD RMSE 6.342 8.473 MAE 5.085 6.142 TABLE IV: Neural Network classification Result on validation set. With the results obtained on the features from both the approaches, SVM and Neural Networks, Decision Level Fusion was applied on the results of different modalities. Outputs of Audio and Text were fused together (Table 4) and Head pose and Fisher together (Table 5). This was so done as the dimensionality of Audio-Text and Fisher-Head pose were close. Since RMSE for each of them were nearly same, so the weights were decided experimentally as mentioned in the Table V and Table VI. Weight(aud) 0.5 0.6 0.7 0.8 0.4 0.3 0.2 Weight(text) 0.5 0.4 0.3 0.2 0.6 0.7 0.8 RMSE(dev) 5.593 5.789 5.901 6.009 5.684 5.786 5.916 MAE(dev) 4.3714 4.314 4.3714 4.285 4.428 4.457 4.600 RMSE(test) 7.631 MAE(test) 6.2766 TABLE V: Fusion of Audio and text features. Weight(head) 0.5 0.6 0.7 0.8 0.4 0.3 0.2 Weight(fisher) 0.5 0.4 0.3 0.2 0.6 0.7 0.8 RMSE 5.744 5.789 5.901 6.009 5.855 5.87 6.00 MAE 4.3714 4.314 4.3714 4.285 4.514 4.514 4.657 TABLE VI: Fusion of Head pose and fisher features on validation set. All these results are on the development set. On the training set, all models fit quite accurately and gave very less RMSE and MAE for all the cases. Eventually equal weights were given to both the modalities as it gave least RMSE for both the fused models. Also on fusion of all the four modalities with equal weights given to all, gave the best RMSE and MAE. (Table VII) audio 0.25 text 0.25 fisher 0.25 head 0.25 RMSE 5.4143 MAE 4.1714 TABLE VII: Equal weight to results of all four modalities on validation set. Another fusion technique applied was to maximize the outputs of both the modalities. With each output compared for both the modalities, maximum among them was taken. This technique was applied across all four modalities also. (Table VIII) Features Audio-text fisher-head Audio-text-fisherhead RMSE 5.983 5.3799 5.3586 MAE 4.3714 4.6571 4.3714 TABLE VIII: Fusion Classification results on validation set. The results given for test were after training on both train and development sets. VI. C ONCLUSION Behavior of a depressed person shows relative change in terms of speech pattern, facial expressions and head movement when compared to a non-depressed person. In this paper, we introduced depression recognition task through visual, audio and text features using SVM and neural networks as classifiers. GMM clustering and fisher vectors were calculated on the relative distance of the facial regions. Facial Regions used in recording the relative distance of certain points are the ones mostly involved in facial expressions like smiling, laughing, and other visible emotions. Head pose, statistical descriptors on gaze, pose and blinking rate were also calculated. Verbal responses of the person coded in the form of text (sentences, words, negative words) and audio (low level) features hold the information regarding the behavior of the person. The features extracted were trained on SVM. Results from audio, fisher vectors and text features, both individually and combined outperformed the baseline results on validation data set (Table 7). Fisher vector features were also classified using Neural Networks and also crossed the baseline results on validation data set (Table 7). Further decision fusion in the form of mean and maximizing the outputs was purely experimental. Results obtained on maximizing the outputs of all four modalities trained on SVM and taking mean of the outputs of different modalities were the best. The better accuracy obtained may have removed unwanted variance on the outputs giving desired results. Since the fused outputs improved accuracies, future work would be devoted towards fusing the outputs in more generic manner. Also the models that gave results overfit on the training data, hence parameters search would be to prevent overfitting. [5] Valstar, Michel, et al. ”Avec 2016: Depression, mood, and emotion recognition workshop and challenge.” Proceedings of the 6th International Workshop on Audio/Visual Emotion Challenge. ACM, 2016. [6] Ringeval, Fabien, et al. ”AVEC 2017Real-life Depression, and Affect Recognition Workshop and Challenge.” (2017). [7] Jan, Asim, et al. ”Automatic depression scale prediction using facial expression dynamics and regression.” Proceedings of the 4th International Workshop on Audio/Visual Emotion Challenge. ACM, 2014. Low, L. ”Detection of clinical depression in adolescents’ using acoustic speech analysis.” (2011). Sturim, Douglas, et al. ”Automatic detection of depression in speech using gaussian mixture modeling with factor analysis.” Twelfth Annual Conference of the International Speech Communication Association. 2011. Alghowinem, Sharifa, et al. ”Head pose and movement analysis as an indicator of depression.” Affective Computing and Intelligent Interaction (ACII), 2013 Humaine Association Conference on. IEEE, 2013. Tzirakis, Panagiotis, et al. ”End-to-End Multimodal Emotion Recognition using Deep Neural Networks.” arXiv preprint arXiv:1704.08619 (2017). Pampouchidou, Anastasia, et al. ”Depression Assessment by Fusing High and Low Level Features from Audio, Video, and Text.” Proceedings of the 6th International Workshop on Audio/Visual Emotion Challenge. ACM, 2016. Ptucha, Raymond, and Andreas Savakis. ”Towards the Usage of Optical Flow Temporal Features for Facial Expression Classification.” Advances in Visual Computing (2012): 388-397. Ellgring, Heiner. Non-verbal communication in depression. Cambridge University Press, 2007. mr. L. (Maple Canada) Depression Vocabulary, Depression Word List. Ghosh, Sayan, Moitreya Chatterjee, and Louis-Philippe Morency. ”A multimodal context-based approach for distress assessment.” Proceedings of the 16th International Conference on Multimodal Interaction. ACM, 2014. Bradley, Margaret M., and Peter J. Lang. Affective norms for English words (ANEW): Instruction manual and affective ratings. Technical report C-1, the center for research in psychophysiology, University of Florida, 1999. Katon, Wayne, and Mark D. Sullivan. ”Depression and chronic medical illness.” J Clin Psychiatry 51.Suppl 6 (1990): 3-11. Cummins, Nicholas, et al. ”Diagnosis of depression by behavioural signals: a multimodal approach.” Proceedings of the 3rd ACM international workshop on Audio/visual emotion challenge. ACM, 2013. Oneata, Dan, Jakob Verbeek, and Cordelia Schmid. ”Action and event recognition with fisher vectors on a compact feature set.” Proceedings of the IEEE international conference on computer vision. 2013. Dhall, Abhinav, and Roland Goecke. ”A temporally piece-wise fisher vector approach for depression analysis.” Affective Computing and Intelligent Interaction (ACII), 2015 International Conference on. IEEE, 2015. Kroenke, Kurt, et al. ”The PHQ-8 as a measure of current depression in the general population.” Journal of affective disorders 114.1 (2009): 163-173. Yang, Le, et al. ”Decision Tree Based Depression Classification from Audio Video and Language Information.” Proceedings of the 6th International Workshop on Audio/Visual Emotion Challenge. ACM, 2016. [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] VII. ACKNOWLEDGEMENT We would like to express our gratitude to our teacher and organizers of the AVEC 2017, who gave us the golden opportunity to participate in this project (AVEC 2017). The project helped us inculcate better understanding of computer vision, machine learning and even basic concepts of deep learning. [18] [19] [20] R EFERENCES [1] [2] [3] [4] Depression article on https://www.nimh.nih.gov/health/topics/depression/index.shtml J. Pedersen et al., An ethological description of depression, Acta psychiatrica scandinavica, vol. 78, no. 3, pp. 320330, 1988. L. Fossi, C. Faravelli, and M. Paoli, The ethological ap- proach to the assessment of depressive disorders, The Jour- nal of nervous and mental disease, vol. 172, no. 6, pp. 332 341, 1984. P. Waxer et al., Nonverbal cues for depression. Journal of Abnormal Psychology, vol. 83, no. 3, p. 319, 1974. [21] [22] [23]
1cs.CV
Efficiently Decodable Non-Adaptive Threshold Group Testing Thach V. Bui∗ , Minoru Kuribayashi‡ , Mahdi Cheraghchi§, and Isao Echizen∗† arXiv:1712.07509v3 [cs.IT] 30 Jan 2018 ∗ SOKENDAI (The Graduate University for Advanced Studies), Hayama, Kanagawa, Japan [email protected] ‡ Graduate § Department of † National Institute School of Natural Science Computing, Imperial of Informatics, and Technology, College London, UK Tokyo, Japan Okayama University, [email protected] [email protected] Okayama, Japan [email protected] Abstract We consider non-adaptive threshold group testing for identification of up to d defective items in a set of N items, where a test is positive if it contains least and negative otherwise. The  2 ≤uu ≤ d defective items,   at  2 2 1 d + log u log · d log N tests with probability defective items can be identified using t = O d2d−u2 e(d−u) u ǫ u u   2   · d log Nd + u log ud · d2 log N tests with probability 1. at least 1 − ǫ for any ǫ > 0 or t = O d2d−u2 e(d−u) u t The decoding time is O(d2 log N ) × poly(d2 log N ). This result significantly improves the best known results for decoding non-adaptive threshold group testing: O(N log N + N log 1ǫ ) for probabilistic decoding, where ǫ > 0, and O(N u log N ) for deterministic decoding. I. I NTRODUCTION The goal of combinatorial group testing is to identify at most d defective items among a population of N items (usually d is much smaller than N). This problem dates back to the work of Dorfman [1], who proposed using a pooling strategy to identify defectives in a collection of blood samples. In each test, a group of items are pooled, and the combination is tested. The result is positive if at least one item in the group is defective and is otherwise negative. Damaschke [2] introduced a generalization of classical group testing known as threshold group testing. In this variation, the result is positive if the corresponding group contains at least u defective items, where u is a parameter, is negative if the group contains no more than ℓ defective items, where 0 ≤ ℓ < u, and is arbitrary otherwise. When u = 1 and ℓ = 0, threshold group testing reduces to classical group testing. We note that ℓ is always smaller than the number of defective items. Otherwise, every test yields a negative outcome, then no information is enhanced from tests. There are two approaches for the design of tests. The first is adaptive group testing in which there are several testing stages, and the design of each stage depends on the outcomes of the previous stages. The second is non-adaptive group testing (NAGT) in which all tests are designed in advance, and the tests are performed in parallel. NAGT is appealing to researchers in most application areas, such as computational and molecular biology [3], multiple access communications [4] and data steaming [5] (cf. [6]). The focus of this work is on NAGT. In both threshold and classical group testing, it is desirable to minimize the number of tests and, to efficiently identify the set of defective items (i.e., have an efficient decoding algorithm). For both testings, one needs Ω(d log N) tests to identify all defective items [6], [7], [8] using adaptive schemes. In adaptive schemes, the decoding algorithm is usually implicit in the test design. The number of tests and the decoding time are significantly different between classical non-adaptive (CNAGT) and non-adaptive threshold group testing (NATGT). 1 In CNAGT, Porat and Rothschild [9] first proposed explicit nonadaptive constructions using O(d2 log N) tests. However, there is no efficient (sublinear-time) decoding algorithm associated with their schemes. For exact identification, there are explicit schemes allowing defective items be identified using poly(d, log N) tests in time poly(d, log N) [10], [11] (the number of tests can be as low as O(d1+o(1) log N) if false positives are allowed in the reconstruction). To achieve a nearly optimal number of tests in adaptive group testing and with low decoding complexity, Cai et al. [12] proposed using probabilistic schemes that need O(d log d · log N) tests to find the defective items in time O(d(log N + log2 d)). In  threshold group testing, Damaschke [2] showed that the set of positive items can be identified with N tests with up to g false positives and g false negatives, where g = u − ℓ − 1 is the gap parameter. u Cheraghchi [13] showed that it is possible to find the defective items with O(dg+2 log d · log(N/d)) tests, and that this trade-off is essentially optimal. Recently, De Marco et al. [14] improved this bound to O(d3/2 log(N/d)) tests under the extra assumption that the number of defective items is exactly d, which is rather restrictive in application. Although the number of tests has been extensively studied, there have been few reports that focus on the decoding algorithm as well. Chen andFu [15] proposed schemes  based  d−u d u d N on CNAGT for when g = 0 that can find the defective items using O u d log d tests in d−u  √ 1 u time O(N log N). Chan et al. [16] presented a randomized algorithm with O log ǫ · d u log N tests to find the defective items in time O(N log N + N log 1ǫ ) given that g = 0 and u = o(d). The cost of these decoding schemes increases with N. Our objective is to find an efficient decoding scheme to identify at most d defective items in NATGT when g = 0. Contributions: In this paper, we consider the case where g = 0, i.e., ℓ = u − 1 (u ≥ 2), and call this model u-NATGT. We first propose an efficient scheme for identifying at most d defective items in t NATGT in time O(d2 log × poly(d2 log N), where t is the number of tests. Our main idea is to create at N) least a specified number of rows in the test matrix such that the corresponding test in each row contains exactly u defective items and such that the defective items in the rows are the defective items to be identified. We “map” these rows using a special matrix constructed from a disjunct matrix (defined later) and its complementary matrix, thereby converting the outcome in NATGT to the outcome in CNAGT. The defective items in each row can then be efficiently identified. Although Cheraghchi [13] and De Marco et al. [14] proposed nearly optimal bounds on the number of tests, there are no decoding algorithms associated with their schemes. On the other hand, the scheme of Chen et al. [15] requires an exponential number of tests, which are much larger than the number of tests in our scheme. Moreover, the decoding complexity of their scheme is exponential in the number of items N, which is impractical. Chan et al. [16] proposed a probabilistic approach to achieve a small number of tests, which combinatorially can be better than our scheme. However, their scheme is only applicable when threshold u is much smaller than d (u = o(d)) and the decoding complexity remains high, namely O(N log N + N log 1ǫ ), where ǫ > 0 is the precision parameter. We present a divide and conquer scheme which we then instantiate via deterministic and randomized decoding. Deterministic decoding is a deterministic scheme in which all defective items can be found with probability 1. Randomized decoding reduces the number of tests; all defective items can be found t × poly(d2 log N). A with probability at least 1 − ǫ for any ǫ > 0. The decoding complexity is O(d2 log N) comparison with existing work is given in Table I. II. P RELIMINARIES For consistency, we use capital calligraphic letters for matrices, non-capital letters for scalars, and bold letters for vectors. All matrix and vector entries are binary. Here are some of the notations used: 1) N, d, x = (x1 , . . . , xN )T : number of items, maximum number of defective items, and binary representation of N items. 2) S = {j1 , j2 , . . . , j|S| }: the set of defective items; cardinality of S is |S| ≤ d. 2 TABLE I C OMPARISON WITH EXISTING WORK . Cheraghchi [13] De Marco et al. [14] Chen et al. [15] Chan et al. [16] Deterministic decoding Randomized decoding Number of tests (t) O(d2 log d · log Nd )   q O d2 · d−u · log Nd du     d−u  N d d u d log O u d−u d  √ 1 O log u log N · d u ǫ  2    · d log N O d2d−u2 e(d−u) + u log ud · d2 log N u d  2  u   2 d 1 O d2d−u2 e(d−u) u log + log · d log N u u ǫ Decoding complexity × Decoding type × × × O(N u log N ) O(N log N + N log t O(d2 log N) t O(d2 log N) Deterministic 1 ) ǫ Random 2 × poly(d log N ) Deterministic 2 × poly(d log N ) Random 3) ⊗, ⊙: operation related to u-NATGT and CNAGT, to be defined later. 4) T : t × N measurement matrix used to identify at most d defective items in u-NATGT, where integer t ≥ 1 is the number of tests. 5) G = (gij ): h × N matrix, where h ≥ 1. 6) M = (mij ): a k × N (d + 1)-disjunct matrix used to identify at most u defective items in u-NATGT and (d + 1) defective items in CNAGT, where integer k ≥ 1 is the number of tests. 7) M = (mij ): the k × N complementary matrix of M; mij = 1 − mij . 8) Ti,∗ , Gi,∗ , Mi,∗, Mj : row i of matrix T , row i of matrix G, row i of matrix M, and column j of matrix M, respectively. 9) xi = (xi1 , . . . , xiN )T , Si : binary representation of items and set of indices of defective items in row Gi,∗ . 10) diag(Gi,∗ ) = diag(gi1 , . . . , giN ): diagonal matrix constructed by input vector Gi,∗ . A. Problem definition We index the population of N items from 1 to N. Let [N] = {1, 2, . . . , N} and S be the defective set, where |S| ≤ d. A test is defined by a subset of items P ⊆ [N]. (d, u, N)-NATGT is a problem in which there are at most d defective items among N items. A test consisting of a subset of N items is positive if there are at least u defective items in the test, and each test is designed in advance. Formally, the test outcome is positive if |P ∩ S| ≥ u and negative if |P ∩ S| < u. We can model (d, u, N)-NATGT as follows: A t×N binary matrix T = (tij ) is defined as a measurement matrix, where N is the number of items and t is the number of tests. x = (x1 , . . . , xN )T is the binary representation vector of N items, where |x| ≤ d. xj = 1 indicates that item j is defective, and xj = 0 indicates otherwise. The jth item corresponds to the jth column of the matrix. An entry tij = 1 naturally means that item j belongs to test i, and tij = 0 means otherwise. The outcome of all tests is y = (y1 , . . . , yt )T , where yi = 1 if test i is positive and yi = 0 otherwise. The procedure to get the outcome vector y is called the encoding procedure. The procedure used to identify defective items from y is called the decoding procedure. Outcome vector y is     T1,∗ ⊗ x y1 def def .   =  ...  .. y =T ⊗x= (1) Tt,∗ ⊗ x yt P where ⊗ is a notation for the test operation in u-NATGT; namely, yi = Ti,∗ ⊗ x = 1 if N j=1 xj tij ≥ u, PN and yi = Ti,∗ ⊗ x = 0 if j=1 xj tij < u for i = 1, . . . , t. Our objective is to find an efficient decoding scheme to identify at most d defective items in (d, u, N)-NATGT. 3 B. Disjunct matrices When u = 1, u-NATGT reduces to CNAGT. To distinguish CNAGT and u-NATGT, we change notation ⊗ to ⊙ and use a k × N measurement matrix M instead of the t × N matrix T . The outcome vector y in ((1)) is equal to       WN M1,∗ ⊙ x y1 N j=1 xj ∧ m1j _ def def def   . .    . . y=M⊙x= Mj = ...  = (2) = . . WN j=1,x =1 j Mk,∗ ⊙ x yk j=1 xj ∧ mkj where ⊙ is the Boolean operator for vector multiplication in which multiplication is replaced with the W AND (∧) operator and addition is replaced with the OR (∨) operator, and yi = Mi,∗ ⊙x = N x ∧mij = j j=1 WN j=1,xj =1 mij for i = 1, . . . , k. W W W T The union of r columns of M is defined as follows: ri=1 Mji = ( ri=1 m1ji , . . . , ri=1 mtji ) . A column is said to not be included in another column if there exists a row such that the entry in the first column is 1 and the entry in the second column is 0. If M is a (d + 1)-disjunct matrix satisfying the property that the union of at most (d + 1) columns does not include any remaining column, x can always be recovered from y. We need M to be a (d + 1)-disjunct matrix that can be efficiently decoded, as in [11], [10], to identify at most d defective items in u-NATGT. A k × N strongly explicit matrix is a matrix in which the entries can be computed in time poly(k). We can now state the following theorem: Theorem 1. [10, Theorem 16] Let 1 ≤ d ≤ N. There exists a strongly explicit k × N (d + 1)-disjunct matrix with k = O(d2 log N) such that for any k × 1 input vector, the decoding procedure returns the set of defective items if the input vector is the union of at most d + 1 columns of the matrix in poly(k) time. C. Completely separating matrix We now introduce the notion of completely separating matrices which are used to get efficient decoding algorithms for (d, u, N)-NATGT. A (u, w)-completely separating matrix is defined as follows: Definition 1. An h × N matrix G = (gij )1≤i≤h,1≤j≤N is called a (u, w)-completely separating matrix if for any pair of subsets I, J ⊂ [N] such that |I| = u, |J| = w, and I ∩ J = ∅, there exists row l such that glr = 1 for any r ∈ I and gls = 0 for any s ∈ J. Row l is called a singular row to subsets I and J. When u = 1, G is called a w-disjunct matrix. This definition is slightly different from the one described by Lebedev [17]. It is easy to verify that, if a matrix is a (u, w)-completely separating matrix, it is also a (u, v)-completely separating matrix for any v ≤ w. Below we present the existence of such matrices. Theorem 2. Given integers 1 ≤ u, w < N, there exists a (u, w)-completely separating matrix of size h × N, where   eN (u + w)2  ew u e(u + w) h = (u + w) log · + u log +1 u+w u u(2u + w) u and e is base of the natural logarithm. Proof: An h × N matrix G = (gij )1≤i≤h,1≤j≤N is generated randomly in which each entry gij is assigned to 1 with probability of p and to 0 with probability of 1 − p. For any pair of subsets I, J ⊂ [N] such that |I| = u, |J| = w, the probability of a row is not singular is: 1 − pu (1 − p)w 4 (3) Then, the probability that there is no singular row to subsets I and J is: f (p) = (1 − pu (1 − p)w )h (4) Using union bound, the probability that any pair of subsets I, J ⊂ [N] such that |I| = u, |J| = w does not have a singular row, i.e., the probability that G is not a (u, w)-separating matrix, is:       u+w N u+w N (1 − pu (1 − p)w )h (5) f (p) = g(p, h, u, w, N) = u u+w u u+w To ensure that there exists G which is a (u, w)-separating matrix, one needs to find p and h such that u , we have: g(p, h, u, w, N) < 1. Choose p = u+w u+w !h  u u  u u w h f (p) = (1 − p (1 − p) ) = 1 − 1− (6) w u+w u+w !  u u  u 1− , where exp(x) = ex (7) ≤ exp −h w u+w      u u  u u w(2u + w)  u2 −u ≤ exp −h 1− = exp −h (8) ·e w (u + w)2 ew (u + w)2    x n x2 −x x We get (7) because 1 − x ≤ e for any x > 0 and (8) because 1 + n ≥ e 1 − n for n > 1, |x| ≤ n. Then we have: u+w  u     e(u + w) eN u+w N f (p) ≤ f (p) (9) g(p, h, u, w, N) = u+w u u u+w u+w  u    u u w(2u + w)  e(u + w) eN (10) · exp −h ≤ u+w u ew (u + w)2 <1 (11)  u+w  u     eN e(u + w) u u w(2u + w) ⇐⇒ (12) < exp h u+w u ew (u + w)2   (u + w)2  ew u eN e(u + w) · ⇐⇒ h > (u + w) log + u log (13) u+w u w(2u + w) u  k We got (9) because nk ≤ en and (10) by using (8). From (13), if we choose k   (u + w)2  ew u e(u + w) eN · + u log +1 (14) h = (u + w) log u+w u w(2u + w) u then g(p, h, u, w, N) < 1, i.e., there exists a (u, w)-completely separating matrix of size h × N. Suppose that G is an h × N (u, w)-completely separating matrix. If w is set to d − u, then every h × d submatrix, which is constructed by its d columns, is a (u, d − u)-completely separating matrix. This property is strict and makes the number of rows in G is high. To reduce the number of rows, we relax this property as follows: each h × d submatrix, which is constructed by d columns of G, is a (u, d − u)-completely separating matrix with high probability. The following corollary describes this idea in details. 5 Corollary 1. Let u, d, N be any given positive integers such that 1 ≤ u < d < N. For any ǫ > 0, there exists an h × N matrix such that each h × d submatrix, which is constructed by its d columns, is a (u, d − u)-completely separating matrix with probability at least 1 − ǫ, where u    ed 1 e(d − u) d2 u log + log h = 2 d − u2 u u ǫ and e is base of the natural logarithm. Proof: An h × N matrix G = (gij )1≤i≤h,1≤j≤N is generated randomly in which each entry gij is assigned to 1 with probability of ud and to 0 with probability of 1 − ud . Our task is now to prove that each h × d matrix G ′ , which is constructed by d columns of G, is a (u, d − u)-completely separating matrix with u   e(d−u) d2 1 ed probability at least 1 − ǫ for any ǫ > 0. Specifically, we prove that h = d2 −u2 + log u log u u ǫ is sufficient to achieve such G ′ . Similar to the proof in Theorem 2, the probability that G ′ is not a (u, d − u)-completely separating matrix at most ǫ is h  u      u u   u u  u d−u u d−u ed d 1− 1− (15) 1− ≤ exp −h d d u d d u  u u    (d − u)(d + u) ed u ≤ exp −h (16) u e(d − u) d2 ≤ǫ (17)  u u 2    d − u2 1 ed u (18) ⇐⇒ ≤ exp h ǫ u e(d − u) d2 u    d2 ed e(d − u) 1 ⇐⇒ h≥ 2 u log (19) + log d − u2 u u ǫ    n x n en k −x . (16) is derived because 1 + ≥ We get (15) because 1 − x ≤ e for any x > 0 and ≤ k n k  x2 x e 1 − n for n > 1, |x| ≤ n. This completes our proof. III. P ROPOSED SCHEME The basic idea of our scheme, which uses a divide and conquer strategy, is to create at least κ rows, e.g., i1 , i2 , . . . , iκ such that |Si1 | = · · · = |Siκ | = u and Si1 ∪ . . . ∪ Siκ = S. Then we “map” these rows by using a special matrix that enables us to convert the outcome in NATGT to the outcome in CNAGT. The defective items in each row can then be efficiently identified. We present a particular matrix that achieves efficient decoding for each row in the following section. A. When the number of defective items equals the threshold In this section, we consider a special case in which the number of defective items equals the threshold, i.e., |x| = u. Given a measurement matrix M and a representation vector of u defective items x (|x| = u), def def def def what we observe is y = M ⊗ x = (y1 , . . . , yk )T . Our objective is to recover y′ = M ⊙ x = (y1′ , . . . , yk′ )T from y. Then x can be recovered if we choose M as a (d + 1)-disjunct matrix described in Theorem 1. To achieve this goal, we create a measurement matrix:   M def (20) A= M where M = (mij ) is a k × N (d + 1)-disjunct matrix as described in Theorem 1 and M = (mij ) is the complement matrix of M, mij = 1 − mij for i = 1, . . . , k and j = 1, . . . , N. We note that M can 6 be decoded in time poly(k) = poly(d2 log N) because k = O(d2 log N). Let us assume that the outcome vector is z. Then we have:     M ⊗ x def y def def (21) z=A⊗x= = y M⊗x where y = M ⊗ x = (y1 , . . . , yk )T and y = M ⊗ x = (y 1 , . . . , y k )T . The following lemma shows that y′ = M ⊙ x is always obtained from z; i.e., x can always be recovered. Lemma 1. Given integers 2 ≤ u ≤ d < N, there exists a strongly explicit 2k × N matrix such that if there are exactly u defective items among N items in u-NATGT, the u defective items can be identified in time poly(k), where k = O(d2 log N). Proof: We construct the measurement matrix A in (20) and assume that z is the observed vector as in (21). Our task is to create vector y′ = M ⊙ x from z. One can get it using the following rules, where l = 1, 2, . . . , k: 1) If yl = 1, then yl′ = 1. 2) If yl = 0 and y l = 1, then yl′ = 0. 3) If yl = 0 and y l = 0, then yl′ = 1. We now prove the correctness of the above rules. Because yl = 1, there are at least u defective items in row Ml,∗ . Then, the first rule is implied. If yl = 0, there are less than u defective items in row Ml,∗. Because |x| = u, y l = 1, and the threshold is u, there must be u defective items in row Ml,∗. Moreover, since Ml,∗ is the complement of Ml,∗ , there must be no defective item in test l of M. Therefore, yl′ = 0, and the second rule is implied. If yl = 0, there are less than u defective items in row Ml,∗ . Similarly, if y l = 0, there are less than u defective items in row Ml,∗ . Because Ml,∗ is the complement of Ml,∗ , the number of defective items in row Ml,∗ or Ml,∗ cannot be equal to zero, since either yl would equal 1 or y l would equal 1. Since the number of defective items in row Ml,∗ is not equal to zero, the test outcome is positive, i.e., yl′ = 1. The third rule is thus implied. Since we get y′ = M ⊙x, M is a (d + 1)-disjunct matrix and u ≤ d, u defective items can be identified in time poly(k) by Theorem 1. Example: We demonstrate Lemma 1 by setting u = d = 2, k = 9, and N = 12 and defining a 9 × 12 2-disjunct matrix M with the first two columns as follows:         1 0 0 0 0 0 0 0 1 1 1 1 0 0 0 1 0  0 0 0 1 1 1 0 0 0 1 0 0  0         0 1  1 1 1 0 0 0 0 0 0 1 0 0  1         1 0  0 0 1 0 0 1 0 0 1 0 1 0  0   ′       0 0 1 0 0 1 0 0 1 0 0 1 0 0 (22) M=  , y =   , y =   , y = 1 0 1  1 0 0 1 0 0 1 0 0 0 1 0  0         0 1  0 1 0 1 0 0 0 0 1 0 0 1  0         1 0  0 0 1 0 1 0 1 0 0 0 0 1  0 0 1 1 0 0 0 0 1 0 1 0 0 0 1 0 Assume that the defective items are 1 and 2, i.e., x = [1, 1, 0, 0, 0, 0, 0, 0, 0]T ; then the observed vector is z = [yTWyT ]T . Using the three rules in the proof of Lemma 1, we obtain vector y′ . We note that y′ = M1 M2 = M ⊙ x. Using a decoding algorithm (which is omitted in this example), we can identify items 1 and 2 as defective items from y′ . 7 B. Encoding procedure To implement the divide and conquer strategy, we need to divide the set of defective items into small ⌉≥1 subsets such that defective items in those subsets can be effectively identified. We define κ = ⌈ |S| u as an integer, and create a h × N matrix G containing κ rows, denoted as i1 , i2 , . . . , iκ , with probability at least 1 − ǫ such that (i) |Si1 | = · · · = |Siκ | = u and (ii) Si1 ∪ . . . ∪ Siκ = S for any ǫ ≥ 0 where Si is the set of indices of defective items in row Gi,∗ . For example, if N = 6, the defective items are 1, 2, and 3, and G1,∗ = (1, 0, 1, 0, 1, 1), then S1 = {1, 3}. These conditions guarantee that all defective items will be included in the decoded set. To achieve such a G, for any |S| ≤ d, a pruning matrix G ′ of size h×d after removing N −d columns Gx for x ∈ [N]\S must be a (u, d−u)-completely separating matrix with high probability. From Definition 1, G ′ is also a (u, |S| − u)-completely separating matrix. Then, the κ rows are chosen as follows. We choose a collection of sets of defective items: Pl = {j(l−1)u+1 , . . . , jlu } for l = 1, . . . , κ − 1. P ′ is a set satisfying κ−1 ′ ′ P ′ ⊆ ∪κ−1 l=1 Pl and |P | = κu − |S|. Then we pick the last set as follows: Pκ = S \ ∪l=1 Pl ∪ P . From Definition 1, for any Pl , there exists a row, denoted il , such that gil x = 1 for x ∈ Pl and gil y = 0 for y ∈ S \ Pl , where l = 1, . . . , κ. Then, Sil = Pl and row il is singular to sets Sil and S \ Sil for l = 1, . . . , κ. Condition (i) thus holds. Condition (ii) also holds because ∪κl=1 Sil = ∪κl=1 Pl = S. The matrix G is specified in section IV. After creating the matrix G, we generate matrix A as in (20). Then the final measurement matrix T of size (2k + 1)h × N is created as follows:   G1,∗   M × diag(G1,∗ )  G1,∗   A × diag(G1,∗ ) M × diag(G1,∗ )      .. ..    T = (23) = . .         Gh,∗ Gh,∗    A × diag(Gh,∗ ) M × diag(Gh,∗ ) M × diag(Gh,∗ ) The vector observed using u-NATGT after performing the tests given by the measurement matrix T is     G1,∗ G1,∗ ⊗ x A × diag(G1,∗ )   A ⊗ x1      . .. ⊗x=  .. y =T ⊗x =  .        G G ⊗ x h,∗ h,∗ A × diag(Gh,∗ ) A ⊗ xh     G1,∗ ⊗ x y1   M ⊗ x1   y1  y1     M ⊗ x1  y1  z1    .  .  ..  =  ..  =  ..  =  .        G ⊗ x   y  y  h   h  h,∗ M ⊗ xh  yh  zh yh M ⊗ xh def (24) def where xi = diag(Gi,∗ ) × x, yi = Gi,∗ ⊗ x, yi = M ⊗ xi = (yi1 , . . . , yik )T , yi = M ⊗ xi = (y i1 , . . . , yik )T , and zi = [yiT yTi ]T for i = 1, 2, . . . , h. We note that xi is the vector representing the defective items corresponding to row Gi,∗ . If xi = (xi1 , xi2 , . . . , xiN )T , Si = {l | xil = 1, l ∈ [N]}. We thus have |Si | = |xi | ≤ d. Moreover, yi = 1 if and only if |xi | ≥ u. 8 C. The decoding procedure ′ ′ T The decoding procedure is summarized as Algorithm 1, where yi′ = (yi1 , . . . , yik ) is presumed to be M ⊙ xi . The procedure is briefly explained as follows: Line 2 enumerates the h rows of G. Line 3 checks if there are at least u defective items in row Gi,∗ . Lines 4 to 14 calculate yi′ , and Line 16 checks if all items in Gi are truly defective and adds them into S. Algorithm 1 Decoding procedure for u-NATGT Input: Outcome vector y, M, M, T . Output: The set of defective items S. 1: S = ∅. 2: for i = 1 to h do 3: if yi = 1 then 4: for l = 1 to k do 5: if yil = 1 then 6: yil′ = 1 7: end if 8: if yil = 0 and y il = 1 then 9: yil′ = 0 10: end if 11: if yil = 0 and y il = 0 then 12: yil′ = 1 13: end if 14: end for 15: Decode yi′ usingWM to get the defective set Gi . 16: if |Gi | = u and j∈Gi Mj ≡ yi then 17: S = S ∪ Gi . 18: end if 19: end if 20: end for 21: Return S. D. Correctness of the decoding procedure Our objective is to recover xi from yi and zi for i = 1, 2, . . . , h. Line 2 enumerates the h rows of G. We have that yi is the indicator that whether there are at least u defective items in row Gi,∗ . If yi = 0, it implies that there are less than u defective items in row Gi,∗ . Since we only focus on row Gi,∗ which has exactly u defective items, zi is not considered if yi = 0. Lines 3 does this task. When yi = 1, it implies that there are at least u defective items in row Gi,∗ . If there are exactly u defective items in this row, they are always identified as described in Lemma 1. Our task now is to prevent accusing false defective items by decoding yi′ . Lines 4 to 14 calculates yi′ from zi . We do not know that yi′ is the union of many columns in M, i.e., how many defective items are in row Gi,∗ . Therefore, our task is to decode yi′ using matrix M to get the defective set Gi , then validate whether all items in Gi are defective. There exists at least κ rows of G such that there are exactly u defective items in each row. And all defective items in these rows are the defective items we need to identify. Therefore, we only consider the case when the number of defective items obtained from decoding yi′ equals to u, i.e., |Gi | = u. Our task is now to prevent identifying false defective items, which is described in Line 16. There are two sets of 9 defective items corresponding to zi : the first one is the true set, which is Si and unknown, and the second one is Gi , which is expected to be Si (but not sure) and |Gi | = u. If Gi ≡ Si , we can always identify u defective items and the condition in line 16 always holds because of Lemma 1. We need to consider the case Gi 6≡ Si , i.e., there are more than u defective items in row Gi,∗ . We classify this case into two categories: 1) |Gi \ Si |W= 0: in this case, all elements in Gi are defective items. We do not need to consider whether j∈Gi Mj ≡ yi . If this condition holds, we receive the true defective items. If it does not hold, we do not take Gi into the defective W item set. 2) |Gi \ Si | = 6 0: in this case, we prove that j∈Gi Mj ≡ yi does not hold, i.e., none of elements in Gi is added to the defective item set. Let pick j1 ∈ Gi \ Si and j2 ∈ Gi \ {j1 }. Since |Si | ≤ d and M is a (d + 1)-disjunct matrix, there exists a row, denoted τ , such that mτ j1 = 1, mτ j2 = 0, and mτ x = 0 for x ∈ Si . In the other hand, because |Gi | = u and |Si | ≤ d, there is less than u defective items W ′ in row τ , i.e., yiτ = 0. Because x∈Gi mτ x =  W u ≤ |Si |, yiτW= 1. That implies yiτ = 0. However, W W W ′ miτ = 1 = 1 6= 0 = yiτ . Therefore, j∈Gi Mj 6≡ yi . x∈Gi \{j1 } mτ x x∈Gi \{j1 } mτ x Thus, line 16 eliminates all false defective items. Line 21 just returns the defective item set S. E. The decoding complexity Because T is constructed using G and M, the probability of successful decoding of y depends on these choices. Given an input vector yi′ , we get the set of defective items from decoding of M. The probability of successful decoding of y thus depends only on G. Since G has κ rows satisfying (i) and (ii) with probability at least 1 − ǫ, all |S| defective items can be identified in h × poly(k) time using t = h(2k + 1) tests with probability of at least 1 − ǫ for any ǫ ≥ 0. We summarize the divide and conquer strategy in the following theorem: Theorem 3. Let 2 ≤ u ≤ d < N be integers and S be the defective set. Suppose that an h × N matrix G contains κ rows, denoted as i1 , . . . , iκ , such that (i) |Si1 | = · · · = |Siκ | = u and (ii) Si1 ∪ . . . ∪ Siκ = S, where Sil is the index set of defective items in row Gil ,∗ . And suppose that an k × N matrix M is a (d + 1)-disjunct matrix that can be decoded in time A. Then a (2k + 1)h × N measurement matrix T , as defined in (23), can be used to identify at most d defective items in u-NATGT in time O(h × A). The probability of successful decoding depends only on the event that G has κ rows satisfying (i) and (ii). Specifically, if that event happens with probability at least 1−ǫ, the probability of successful decoding is also at least 1 − ǫ for any ǫ ≥ 0. IV. C OMPLEXITY OF PROPOSED SCHEME We specify the matrix G in Theorem 3 to get the desired number of tests and decoding complexity for identifying at most d defective items. Specifying G leads to two approaches on decoding: deterministic and randomized. Deterministic decoding is a deterministic scheme in which all defective items can be found with probability 1. It is achievable when every its h × d submatrices, which are constructed by its d columns, are (u, d − u)-completely separating matrices. Randomized decoding reduces the number of tests; all defective items can be found with probability at least 1 − ǫ for any ǫ > 0. It is achievable when each its h × d submatrix, which is constructed by its d columns, is (u, d − u)-completely separating matrix with probability at least 1 − ǫ. A. Deterministic decoding The following theorem states that there exists a deterministic algorithm for identifying all defective items by choosing G of size h × N to be a (u, d − u)-completely separating matrix in Theorem 2. 10 Theorem 4. Let 2 ≤ u ≤ d ≤ N. There exists a t × N matrix such that at most d defective items in t × poly(d2 log N), where u-NATGT can be identified in time O(d2 log N) u      N e(d − u) d d2 2 · d log N · d log + u log t=O d2 − u 2 u d u Proof: On the basis of Theorem 3, a t × N measurement matrix T is generated as follows: 1) Choose anh × N (u, d − u)-completely separating matrix G as in Theorem 2, where u  2 h = d2d−u2 e(d−u) + u log ed + 1. · d log eN u d u 2) Choose a k × N (d + 1)-disjunct matrix M as in Theorem 1, where k = O(d2 log N) and the decoding time of M is poly(k). 3) T is defined as in (23). Since G is a h × d (u, d − u)-completely separating matrix, for any |S| ≤ d, an h × d pruning matrix ′ G , which is created by removing N − d columns Gx for x ∈ [N] \ S, is also a (u, d − u)-completely separating matrix with probability 1. From Definition 1, G ′ is also a (u, |S| − u)-completely separating matrix. Then, there exists κ rows satisfying (i) and (ii) as described in section III-B. From Theorem 3, d defective items can be recovered using t = h × O(d2 log N) tests with probability at least 1, i.e., the probability 1, in time h × poly(k). B. Randomized decoding For randomized decoding, G is chosen such that the pruning matrix G ′ of size h×d created by removing N − d columns Gx of G for x ∈ [N] \ S is a (u, d − u)-completely separating matrix with probability at least 1 − ǫ for any ǫ > 0. This results is an improved number of tests and decoding time compared to Theorem 4: Theorem 5. Let 2 ≤ u ≤ d ≤ N. For any ǫ > 0, at most d defective items in u-NATGT can be identified using u      d d2 1 e(d − u) 2 t = O u log + log · d log N d2 − u2 u u ǫ tests with probability at least 1 − ǫ. The decoding time is t O(d2 log N ) × poly(d2 log N). Proof: On the basis of Theorem 3, a t × N measurement matrix T is generated as follows: u  2 e(d−u) 1 ed 1) Choose an h × N matrix G as in Corollary 1, where h = d2d−u2 + log . u log u u ǫ 2 2) Generate a k × N (d + 1)-disjunct matrix M using Theorem 1, where k = O(d log N) and the decoding time of M is poly(k). 3) Define T as (23). Let G be an h × N matrix as described in Corollary 1. Then for any |S| ≤ d, an h × d pruning matrix G ′ , which is created by removing N − d columns Gx for x ∈ [N] \ S, is a (u, d − u)-completely separating matrix with probability at least 1 − ǫ. From Definition 1, G ′ is also a (u, |S| − u)-completely separating matrix. Then, there exists κ rows satisfying (i) and (ii) as described in section III-B with probability at least 1 − ǫ. From Theorem 3, |S| defective items can be recovered using t = h × O(d2 log N) tests with probability at least 1 − ǫ in time h × poly(k). V. C ONCLUSION We introduced an efficient scheme for identifying defective items in NATGT. However, the algorithm works only for g = 0. Extending the results to g > 0 is left for future work. Moreover, it would be interesting to consider noisy NATGT as well, in which erroneous tests are present in the test outcomes. 11 VI. ACKNOWLEDGEMENT The first author thanks to SOKENDAI for supporting him via The Short-Stay Abroad Program 2017. R EFERENCES [1] R. Dorfman, “The detection of defective members of large populations,” The Annals of Mathematical Statistics, vol. 14, no. 4, pp. 436– 440, 1943. [2] P. Damaschke, “Threshold group testing,” in General theory of information transfer and combinatorics, pp. 707–718, Springer, 2006. [3] M. Farach, S. Kannan, E. Knill, and S. Muthukrishnan, “Group testing problems with sequences in experimental molecular biology,” in Compression and Complexity of Sequences 1997. Proceedings, pp. 357–367, IEEE, 1997. [4] J. Wolf, “Born again group testing: Multiaccess communications,” IEEE Transactions on Information Theory, vol. 31, no. 2, pp. 185– 191, 1985. [5] G. Cormode and S. Muthukrishnan, “What’s hot and what’s not: tracking most frequent items dynamically,” ACM Transactions on Database Systems (TODS), vol. 30, no. 1, pp. 249–278, 2005. [6] D. Du and F. Hwang, Combinatorial group testing and its applications, vol. 12. World Scientific, 2000. [7] H.-B. Chen and A. De Bonis, “An almost optimal algorithm for generalized threshold group testing with inhibitors,” Journal of Computational Biology, vol. 18, no. 6, pp. 851–864, 2011. [8] H. Chang, H.-B. Chen, H.-L. Fu, and C.-H. Shi, “Reconstruction of hidden graphs and threshold group testing,” Journal of combinatorial optimization, vol. 22, no. 2, pp. 270–281, 2011. [9] E. Porat and A. Rothschild, “Explicit non-adaptive combinatorial group testing schemes,” Automata, languages and programming, pp. 748–759, 2008. [10] H. Q. Ngo, E. Porat, and A. Rudra, “Efficiently decodable error-correcting list disjunct matrices and applications,” in International Colloquium on Automata, Languages, and Programming, pp. 557–568, Springer, 2011. [11] M. Cheraghchi, “Noise-resilient group testing: Limitations and constructions,” Discrete Applied Mathematics, vol. 161, no. 1, pp. 81–95, 2013. [12] S. Cai, M. Jahangoshahi, M. Bakshi, and S. Jaggi, “Grotesque: noisy group testing (quick and efficient),” in Communication, Control, and Computing (Allerton), 2013 51st Annual Allerton Conference on, pp. 1234–1241, IEEE, 2013. [13] M. Cheraghchi, “Improved constructions for non-adaptive threshold group testing,” Algorithmica, vol. 67, no. 3, pp. 384–417, 2013. [14] G. De Marco, T. Jurdziński, M. Różański, and G. Stachowiak, “Subquadratic non-adaptive threshold group testing,” in International Symposium on Fundamentals of Computation Theory, pp. 177–189, Springer, 2017. [15] H.-B. Chen and H.-L. Fu, “Nonadaptive algorithms for threshold group testing,” Discrete Applied Mathematics, vol. 157, no. 7, pp. 1581–1585, 2009. [16] C. L. Chan, S. Cai, M. Bakshi, S. Jaggi, and V. Saligrama, “Stochastic threshold group testing,” in Information Theory Workshop (ITW), 2013 IEEE, pp. 1–5, IEEE, 2013. [17] V. S. Lebedev, “Separating codes and a new combinatorial search model,” Problems of Information Transmission, vol. 46, no. 1, pp. 1–6, 2010. 12
7cs.IT
LYUBEZNIK NUMBERS AND INJECTIVE DIMENSION IN MIXED CHARACTERISTIC arXiv:1512.02298v1 [math.AC] 8 Dec 2015 DANIEL J. HERNÁNDEZ, LUIS NÚÑEZ-BETANCOURT, FELIPE PÉREZ, AND EMILY E. WITT Abstract. We investigate the Lyubeznik numbers, and the injective dimension of local cohomology modules, of finitely generated Z-algebras. We prove that the mixed characteristic Lyubeznik numbers and the standard ones agree locally for almost all reductions to positive characteristic. Additionally, we address an open question of Lyubeznik that asks whether the injective dimension of a local cohomology module over a regular ring is bounded above by the dimension of its support. Although we show that the answer is affirmative for several families of Z-algebras, we also exhibit an example where this bound fails to hold. This example settles Lyubeznik’s question, and illustrates one way that the behavior of local cohomology modules of regular rings of equal characteristic and of mixed characteristic can differ. 1. Introduction Since its introduction by Grothendieck in the 1960s, the theory of local cohomology has been an active area of research in commutative algebra, algebraic geometry, and neighboring fields. In this paper, we are especially motivated by the following observation: despite the fact that local cohomology modules of regular rings are typically large (e.g., not finitely generated), they often satisfy strong structural conditions. More precisely, consider the following conditions on a module M over a ring R. (1) The number of associated primes of M is finite. (2) The Bass numbers of M are finite. (3) inj. dimR M ≤ dim SuppR M .1 (4) The local cohomology of M with support in a maximal ideal of R is injective. In condition (3) above, we use inj. dimR M to denote the injective dimension of M (i.e., the length of a minimal injective R-resolution of M ), and dim SuppR M to denote the dimension of the support of M as a topological subspace of Spec R (i.e., the longest length of a chain of prime ideals in R whose terms are contained in the support of M .) It was shown by Huneke and Sharp in positive characteristic [HS93], and by Lyubeznik in equal characteristic zero [Lyu93] and in unramified mixed characteristic [Lyu00], that the local cohomology modules of a regular local ring with support in an arbitrary ideal satisfy conditions (1) and (2) above. More recently, it was shown by Bhatt, Blickle, Lyubeznik, Singh, and Zhang that the local cohomology modules of a smooth Z-algebra with support in an arbitrary ideal also satisfy condition (1) [BBL+ 14]. The extent to which local cohomology modules satisfy (3) and (4) is a subtle issue, and an overarching theme of this paper; we recall results regarding conditions (3) and (4) by Huneke and Sharp, Lyubeznik, and Zhao later in this introduction. The finiteness condition (2) for local cohomology is a crucial ingredient in the definition of the so-called Lyubeznik numbers, a family of integer-valued invariants associated to a local ring containing a field. In addition to encoding important properties of the ambient ring [Kaw00, Wal01, Kaw02, Lyu06, Zha07], Lyubeznik numbers also have interesting geometric and topological interpretations, including connections with singular and étale cohomology [GLS98, BB05] and with simplicial complexes [ÀMGLZA03, ÀMV14, AMY14]. It is worth pointing out that 1To avoid contemplating the dimension of ∅, we only consider property (3) for nonzero M . 1 2 D. J. HERNÁNDEZ, L. NÚÑEZ-BETANCOURT, F. PÉREZ, AND E. E. WITT properties (1), (3), and (4) play an important role in the proofs of these results. The finiteness property (2) is also used to define the mixed characteristic Lyubeznik numbers 2, a variant of the (standard) Lyubeznik numbers defined by the second and fourth authors [NBW13]. A major obstruction to understanding these invariants has been the uncertainty surrounding properties (3) and (4) in mixed characteristic. This article has two major goals: the first is to investigate the relationship between the (standard) Lyubeznik numbers and the mixed characteristic Lyubeznik numbers in the case in which they are both defined (that is, for local rings of positive characteristic), and the second is to investigate when local cohomology modules satisfy conditions (3) and (4). 1.1. A comparison of Lyubeznik numbers. Suppose that (R, m, k) is a local ring. If R contains a field, and may be realized as R ∼ = S/I, where S is a regular local ring containing a field, then the (standard) Lyubeznik number of R with respect to integers i, j ≥ 0 is defined as dim(S)−j (S); i.e., the ith Bass number of the local cohomology module HI dim(S)−j λi,j (R) = dimk ExtiS (k, HI (S)). In an attempt to study local rings of mixed characteristic (that is, rings of characteristic zero with positive characteristic residue fields), the second and fourth authors proposed the following analogous definition: if the residue field k of R has characteristic p > 0, and R may be realized as R ∼ = T /J, where T is an unramified regular local ring of mixed characteristic p > 0, then the mixed characteristic Lyubeznik number 3 of R with respect to integers i, j ≥ 0 is defined as ei,j (R) = dimk Exti (k, H dim(T )−j (T )). λ T J One may then extend these definitions to an arbitrary local ring (R, m, k) by defining ei,j (R) = λ ei,j (R). b and λ b λi,j (R) = λi,j (R) b can be realized in the specified ways follows from the Cohen structure theorems (The fact that R [Coh46]). In either case, the Lyubeznik numbers depend only on R and the indices i and j, but b as a quotient [Lyu93, NBW13]. not on any of the choices made realizing R (or R) It is important to note that, although the mixed characteristic Lyubeznik numbers were originally defined to study rings of mixed characteristic, both the standard and the mixed characteristic Lyubeznik numbers are defined for rings of (equal) characteristic p > 0. This observation motivates the following question. Question A. For which rings of positive characteristic is it true that the (standard) Lyubeznik numbers and the mixed characteristic Lyubeznik numbers agree? It is known that Question A has a positive answer for Cohen-Macaulay rings and rings of small dimension [NBW13, Corollary 5.3]. Interestingly enough, however, these types of Lyubeznik numbers need not always agree, as exhibited by a certain quotient of a power series over a field of characteristic two by a squarefree monomial ideal [NBW13, Remark 6.11]. In this article, we consider Question A from the perspective of reduction to positive characteristic. Recall that if R is a finitely generated Z-algebra, then we call Fp ⊗Z R = R/pR the reduction of R to characteristic p > 0. In this context, one may specialize Question A, and instead ask for which reductions of a given finitely generated Z-algebra do the standard and 2In [NBW13, NBWZ14], these invariants are referred to as Lyubeznik numbers in mixed characteristic. 3This terminology is not intended to imply that R must be of mixed characteristic in order to define λ ei,j (R) ei,j (R) arise from considering R (or R) b as a (this would be false), but is chosen to emphasize that the integers λ quotient of a regular unramified ring of mixed characteristic. LYUBEZNIK NUMBERS AND INJECTIVE DIMENSION 3 mixed characteristic Lyubeznik numbers agree locally. Before stating our general result in this direction, we present an illustrative example. Example 1.1 (Cohen-Macaulay rings). If a finitely-generated Z-algebra R is Cohen-Macaulay, then so are its reductions R/pR for all p ≫ 0 [HH99, Theorem 2.3.5]. In light of this, the aforementioned result [NBW13, Corollary 5.3] shows that if p ≫ 0 and A is the localization of ei,j (A) = λi,j (A) for all integers i, j ≥ 0. In fact, the vanishing R/pR at a prime ideal, then λ of local cohomology modules described in [PS73, Theorem III.4.1] implies that both of these invariants are zero unless j = dim(R). As suggested by this example, the standard and mixed characteristic Lyubeznik numbers agree locally for almost all reductions to characteristic p > 0 of a given finitely-generated Z-algebra. Theorem B (cf. Theorem 4.5). If R is a finitely generated Z-algebra, then there exists a finite set of prime integers W with the following property: If p is prime and not contained in W , and ei,j (A) = λi,j (A) for all integers i, j ≥ 0. A is the localization of Fp ⊗Z R at a prime ideal, then λ A fundamental tool in our proof of Theorem B is the fact that local cohomology modules of polynomial rings over Z have finitely many associated primes [BBL+ 14]. 1.2. The injective dimension of local cohomology modules. In light of the fact that the local cohomology of a regular ring containing a field satisfies the structural conditions (3) and (4), the following question of Lyubeznik [Lyu93, Lyu00] is natural. Question C (Lyubeznik). Given an ideal I of a regular local ring S, do the conditions (3) and (4) hold for nonzero HIj (S)? This question has been unresolved in the mixed characteristic case for more than two decades. To the best of our knowledge, the most significant step toward answering Question C is the result of Zhou [Zho98, Theorem 5.1], which says the following: If S is regular of unramified mixed characteristic, and HIj (S) is nonzero for some ideal I of S and integer j ≥ 0, then the injective dimension of HIj (S) is bounded above by dim SuppS HIj (S) + 1. Furthermore, if m is a maximal ideal of S, then the injective dimension of the iterated local cohomology module Hmi HIj (S) is at most one. In this article, we prove that the answer to the first part Question C is affirmative, in most cases, for Z-algebras that descend from Q-algebras. Theorem D (cf. Theorem 5.4). Given an ideal I of S = Z[x1 , . . . , xn ] with Q⊗Z HIj (S) 6= 0, there exists a finite set of primes W with the following property: If Q is a prime ideal in SuppS HIj (S) not lying over any prime in W 4, then inj. dimS HIj (S)Q ≤ dim SuppS HIj (S)Q . We also show that local cohomology modules satisfy condition (4) in several cases; see Proposition 5.10 for details. Moreover, although we have identified many cases with a positive answer to Question C, we also construct an example in which properties (3) and (4) fail to hold (cf. the example presented in [NBW13, Remark 6.11]). Theorem E (cf. Theorem 6.3). There exists an ideal I of a regular local ring (T, m) of mixed characteristic, and nonnegative integers i and j, for which inj. dimS HIj (T ) > dim SuppS HIj (T ), and such that the iterated local cohomology module Hmi HIj (T ) is not injective. This theorem establishes a negative answer to both parts of Question C, effectively settling it. Furthermore, it exhibits a previously-unknown way in which regular rings of equal characteristic and those of mixed characteristic can behave differently. 4The condition that Q ⊗ H j (S) 6= 0 and the finiteness of W imply that the set of all such prime ideals Q is a Z I non-empty open (i.e., dense) subset of SuppS HIj (S). See Subsection 5.2 for more details. 4 D. J. HERNÁNDEZ, L. NÚÑEZ-BETANCOURT, F. PÉREZ, AND E. E. WITT We conclude this introduction by mentioning one further result, which may be regarded as an extension of the recent noted work on smooth Z-algebras [BBL+ 14]; see Section 3 for the proof. Theorem F. Given ideals I1 , . . . , It of S = Z[x1 , . . . , xn ], and nonnegative integers j1 , . . . , jt , the t-fold iterated local cohomology module HIjtt · · · HIj11 (S) has only finitely many associated primes. We note that Theorem F, and the ideas behind the proof (namely, Corollary 3.11) play a key role in our proofs of Theorem B, Theorem D, and Proposition 5.10. 2. Local cohomology In this section, we recall some basic properties of local cohomology modules, and refer the reader to [BS98] for more details. 2.1. Koszul and local cohomology. The (cohomological) Koszul complex of an element f of a commutative ring R is the complex K • (f ; R) given by 0 −→ R −→ R −→ 0, where the nontrivial map is multiplication by f , the left-most copy of R is in degree zero, and the right-most copy in degree one. More generally, the Koszul complex of a finite sequence f = hf1 , · · · , fs i in R on an R-module M is defined to be the tensor product of complexes K • (f ; M ) := K • (f1 ; R) ⊗R · · · ⊗R K • (fs ; R) ⊗R M, where we regard M as a complex concentrated in degree zero. The k-th Koszul cohomology module of f is then defined to be H k (f ; R) := H k (K • (f ; R)). Given a non-negative integer t, the commutative diagram of R-modules (2.1.1) 0 /R ft /R /0 f 0 /R f t+1  /R /0 defines a map of complexes K • (f t ; R) → K • (f t+1 ; R), and tensoring maps of this form together, and then tensoring with M , produces a map K • (f t ; M ) → K • (f t+1 ; M ), (2.1.2) where f n is the sequence obtained by raising every term in f to the n-th power. We use K • (f ∞ ; M ) to denote the limit complex of this directed system, and we recall that there exists a canonical isomorphism K • (f ∞ ; M ) ∼ = Č • (f ; M ), where Č • (M ; f ) denotes the Čech complex of f on M . Taking homology in (2.1.2) produces a directed system of R-modules, and as direct limits commute with homology, we obtain a functorial isomorphism (2.1.3) lim H k (K • (f t ; M )) ∼ = H k (K • (f ∞ ; M )) ∼ = H k (Č • (f ; M )). t→∞ These R-modules depend only on the ideal a generated by the terms in f , and we refer to these isomorphic objects, denoted by Hak (M ), as the k-th local cohomology module of M with support in a. 2.2. Long exact sequences in local cohomology. Assume now that R is Noetherian. Given an ideal a of R and a short exact sequence 0 → M → N → P → 0 of R-modules, there is a functorial long exact sequence · · · → Hak (M ) → Hak (N ) → Hak (P ) → Hak+1 (M ) → · · · . Given f ∈ R and an R-module M , there is also a long exact sequence, functorial in M , k+1 k k k · · · → Ha+f R (M ) → Ha (M ) → Ha (Mf ) → Ha+f R (M ) → · · · , where the map from Hak (M ) to Hak (Mf ) ∼ = Hak (M )f is the natural localization map. LYUBEZNIK NUMBERS AND INJECTIVE DIMENSION 5 3. Finiteness of associated primes of iterated local cohomology modules The goal of this section is to prove Theorem F, which extends the recent result of Bhatt, Blickle, Lyubeznik, Singh, and Zhang on the finiteness of the associated primes of local cohomology [BBL+ 14, Theorem 1.2]. The proof itself appears in the last subsection, while in the earlier subsections, we survey some results from the theory of F -modules and D-modules. Our main reference for the theory of F -modules is [Lyu97]. 3.1. F -modules. Given a ring R of characteristic p > 0 and a non-negative integer e, let R(e) e denote the abelian group R considered as an R-bimodule via the rule r · x · s = rxsp . For the remainder of this subsection, we assume that R is regular, so that the functor F taking an R-module M to the (left) R-module F (M ) = R(e) ⊗R M is exact. An F -module consists of an R-module M and an R-module isomorphism M ∼ = F (M ), which we call the structure morphism of M . The isomorphism R → F (R) given by x 7→ x ⊗ 1 shows that R is an F -module, and in this article, we always regard R as an F -module in this way. A map between F -modules M ∼ = F (M ) and N ∼ = F (N ) is an R-linear map M → N that respects the structure morphisms (that is, so that the expected diagram commutes). Given a map α : M → F (M ) of R-modules, consider the following commutative diagram. α / F (M ) F (α)  M α F (α) / F 2 (M ) F 2 (α)  F (M ) / F 2 (M ) / ··· F 2 (α) F (α)  F 2 (α) / F 3 (M ) F 3 (α) / ··· If M denotes the direct limit of the first row of this diagram, the vertical arrows induce an isomorphism M → F (M ). In this case, we say that α generates the F -module M . 3.2. Local cohomology as an F -module. The localization of an F -module can be regarded as an F -module in such a way that the localization map is a map of F -modules [Lyu97, Example 1.2]. In particular, the Čech complex of an F -module is a complex of F -modules, and therefore, the (iterated) local cohomology of an F -module is also an F -module. In Proposition 3.3 below, which is an adaptation of [Lyu97, Proposition 1.11(b)], we describe another (isomorphic) F module structure on local cohomology. Before proceeding, we recall some basic facts: Remark 3.1 (Direct limits and tensor products). The natural numbers are a basic example of a filtered poset. As such, direct limits of N-directed systems satisfy certain desirable conditions. ∞ For example, if {Mi }∞ i=1 and {Ni }i=1 are two directed systems of R-modules, then     lim Ni ⊗R lim Mi ∼ = lim (Mi ⊗R Ni ). i→∞ i→∞ i→∞ The analogous identity for complexes of R-modules also holds [ILL+ 07, Theorem 4.28], and we use these identities without mention in the following remark. Remark 3.2 (Koszul complexes and the Frobenius functor). The standard F -module structure R ∼ = F (R) allows us to canonically identify K • (f p ; R) with F (K • (f ; R)), which gives us two different ways to regard the second row in the following commutative diagram 0 /R f /R /0. f p−1 0 /R fp  /R /0 6 D. J. HERNÁNDEZ, L. NÚÑEZ-BETANCOURT, F. PÉREZ, AND E. E. WITT In other words, there exists a commutative diagram of complexes K • (f ; R) ♦♦ ♦♦♦ ♦ ♦ ♦ w ♦♦ ♦ PPP PPP PPP PPP ( / F (K • (f ; R)). ∼ = K • (f p ; R) More generally, given a finite sequence f in R, the flatness of the Frobenius morphism allows us tensor sequences of this form together to obtain a commutative diagram K • (f ; R) ♦♦ ♦♦♦ ♦ ♦ ♦ w ♦♦ ♦ ◗◗◗ ◗◗◗ ◗◗◗ ◗◗◗ ( / F (K • (f ; R)). ∼ = K • (f p ; R) Next, fix an R-module M , and let M be the F -module generated by α : M → F (M ). Tensoring the vertical maps above with α produces (3.2.1) K • (f ; M ) ♠♠♠ ♠♠♠ ♠ ♠ ♠ v ♠♠ ♠ ◗◗◗ ◗◗◗ ◗◗◗ ◗◗◗ ( / F (K • (f ; M )), ∼ = K • (f p ; F (M )) and considering the directed systems that result from iterating these maps, we see that  e e lim F e (K • (f ; M )) ∼ = lim K • (f p , F e (M )) = lim K • (f p ; R) ⊗ F e (M ) e→∞ e→∞ e→∞     e • pe ∼ = lim K (f ; R) ⊗ lim F (M ) e→∞ e→∞   • pe = lim K (f ; R) ⊗ M e→∞  e ∼ = lim K • (f p ; R) ⊗ M = K • (f ∞ ; M ). e→∞ ∞ • K (f , M ) ∼ At the level of homology, the isomorphism = lime→∞ F e (K • (f ; M )) becomes Hak (M ) = H k (K • (f ∞ , M )) ∼ = lim H k (F e (K • (f ; M )) ∼ = lim F e (H k (f ; M )), e→∞ e→∞ where a is the ideal generated by the terms of f , and the last directed system is obtained by taking homology of the right-most map in (3.2.1). We summarize the content of Remark 3.2 below. Proposition 3.3 (cf. [Lyu97, Proposition 1.11(b)]). Fix a finite sequence f in R whose terms generate the ideal a of R. Fix an arbitrary R-linear homomorphism α : M → F (M ), and let β : H k (f ; M ) → F (H k (f ; M )) be the associated map induced by (3.2.1). If M is the F -module generated by α, and N is the F -module generated by β, then there exists an isomorphism of F -modules N ∼ = Hak (M ) such that H k (f ; M )  N / H k (f ; M ) ∼ =  / H k (M ) a commutes, where H k (f ; M ) → N is the map to the direct limit, H k (f ; M ) → H k (f ; M ) is the functorial map induced by the map to the direct limit M → M , and the right-most vertical map is the one given by the natural transformation from Koszul to local cohomology. LYUBEZNIK NUMBERS AND INJECTIVE DIMENSION 7 3.3. Iterated local cohomology. For the rest of this section, we work in the following context. Setup 3.4. Let S be a polynomial ring over Z. For every t ∈ N, fix a finite sequence f t in S generating an ideal at ⊆ S. For all d > 0 and u ∈ Zd , consider the compositions of functors Au ( ) = H u1 (f 1 ; ) ◦ · · · ◦ H ud (f d ; ) and B u ( ) = Hau11 ◦ · · · ◦ Haudd ( ) from the category of S-modules to itself. Note that the natural transformation from Koszul to local cohomology induces a natural transformation of functors Au → B u for every u ∈ Zd . Remark 3.5 (Induced functors on the reductions of S modulo a prime integer). Fix a prime integer p > 0, and set R = S/pS. If M is an R-module, then HIk (M ) is also an R-module, and k (M ). The analogous property also holds for Koszul is in fact isomorphic (as R-modules) to HIR cohomology, and it follows from these observations that, when restricted to the category of Rmodules, the functor Au (respectively, B u ) may be regarded as an iterated Koszul (respectively, local) cohomology functor on R. It is also worth pointing out that the natural transformation from Au → B u as functors of S-modules restricts to a natural transformation of functors of R-modules, and is compatible with the one given by considering Au and B u as iterated Koszul and local cohomology of R-modules. Finally, we note that the functor B u not only induces a functor of R-modules, but also induces a functor of FR -modules, where FR denotes the Frobenius functor on the polynomial ring R. The following technical result is centered on the behavior of iterated local and Koszul cohop mology with respect to the sequence 0 → S −→ S → S/pS → 0, with p > 0 a prime integer. This result plays an important role in the proof of Theored 3.10 later in this section, and in some sense, is what one needs to overcome the lack, in general, of long exact sequences associated to iterated cohomology. Lemma 3.6. Let Gu denote either the functor Au or B u . Fix a positive integer d and a prime p integer p > 0, and consider the exact sequence of S-modules 0 → S −→ S → S/pS → 0. If either p (1) the induced map Gu (S) −→ Gu (S) is injective for every 1 ≤ c ≤ d and u ∈ Zc , or (2) the induced map Gu (S) → Gu (S/pS) is surjective for every 1 ≤ c ≤ d and u ∈ Zc , then for all 1 ≤ c ≤ d and u ∈ Zc , we have an exact sequence p 0 → Gu (S) −→ Gu (S) → Gu (S/pS) → 0. Proof. Set R = S/pS. We proceed by induction on d. First, suppose that d = 1. In this case, the key point is that the functor Gu with u ∈ Z has an associated long exact sequence. Moreover, if multiplication by p on Gu (S) is injective for every u ∈ Z, or Gu (S) → Gu (R) is surjective for every u ∈ Z, then the long exact sequence in Gu induced by p 0 → S −→ S → R → 0 splits into short exact sequences, which establishes the lemma when d = 1. Next, suppose that the lemma is true for some positive integer d, and that either multiplication by p is injective on Gu (S) for every 1 ≤ c ≤ d+1 and u ∈ Zc , or that Gu (S) → Gu (R) is surjective for every 1 ≤ c ≤ d + 1 and u ∈ Zc . The inductive hypothesis then implies that (3.3.1) p 0 → Gu (S) −→ Gu (S) → Gu (R) → 0 is exact for every 1 ≤ c ≤ d and u ∈ Zc . Thus, to complete the proof, it suffices to show that (3.3.1) is exact when u is replaced with any w ∈ Zd+1 . Fix u ∈ Zd . As in the base case, the long exact sequence in either Koszul or local cohomology (depending on which iterated functor Gu represents) induced by (3.3.1) splits into short exact 8 D. J. HERNÁNDEZ, L. NÚÑEZ-BETANCOURT, F. PÉREZ, AND E. E. WITT p sequences of the form 0 → Gw (S) −→ Gw (S) → Gw (R) → 0, with w ∈ Zd+1 , by our assumptions on the maps induced by Gw . As every sequence of this form with w ∈ Zd+1 arises in this way, we can conclude the proof.  The following is an extension of Proposition 3.3 to the context of iterated cohomology. Proposition 3.7. Let R = S/pS denote the reduction of S modulo some prime integer p > 0. Fix an FR -module M ∼ = FR (M ) and u ∈ Zd , and let β : Au (M ) → FR (Au (M )) be the map 5 obtained by iterating the one induced by (3.2.1). If N is the FR -module generated by β, then there exists an isomorphism of FR -modules N ∼ = B u (M ) such that Au (M )  N ∼ = ∼ = / Au (M )  / B u (M ) commutes, where Au (M ) → N is the map into the direct limit, Au (M ) ∼ = Au (M ) is obtained by u applying A to the identification of M with the F -module generated by its structure morphism, and the right-most vertical map is the one given by the natural transformation from Au to B u . Proof. This follows from a straightforward induction on d ≥ 1 in which one repeatedly invokes Proposition 3.3, with α being the structure morphism M ∼  = F (M ). 3.4. F -modules and D-modules. Let T be a a ring, and let D(T ) denote the ring of Z-linear differential operators on T . Throughout this discussion, the term “D-module” always refer to a left D(T )-module. Note that a D-module is always T -module via restriction of scalars. The maps in the Čech complex of a sequence in T on a D-module M are D-linear, and so the local cohomology of M inherits a natural D-module structure from M [Lyu93, Example 2.1(iv)]. This also shows that given a map of D-modules, the associated (functorial) maps on local cohomology are again D-linear. It follows that the iterated local cohomology modules of a D-module are themselves D-modules, and that the functorial maps determined by iterated local cohomology are D-linear. There is another canonical D-module structure on iterated local cohomology in positive characteristic. Indeed, if T has characteristic p > 0 with Frobenius functor F , then the structure morphism M ∼ = F (M ) of an F -module M allows one to define a natural D(T )-module structure on M [Lyu97, Section 5]. Fortunately, these two D-module structures on (iterated) local cohomology (one coming from the Čech complex, and the other induced by the F -module structure) are isomorphic (see, e.g., [Lyu97, Example 5.1(b) and 5.2(c)]). When referring to (iterated) local cohomology in positive characteristic as a D-module, we always mean either of these two isomorphic D-module structures. The following result gives an important criterion for when a subset of given F -module generates it as D-module, and plays a crucial role in our proof of Proposition 3.9. Lemma 3.8 ([ÀMBL05, Corollary 4.4]). Suppose that T is a regular finitely generated algebra over an F -finite regular local ring of characteristic p > 0. If M is a finitely generated T -module, and a T -linear map M → F (M ) generates the F -module M , then the image of M in the direct limit M generates M as a D(T )-module. We now specialize to the context of Setup 3.4. Proposition 3.9. Fix u ∈ Zd and a prime integer p > 0. If the map Au (S) → Au (S/pS) induced by the map S → S/pS is surjective, then so is the induced map B u (S) → B u (S/pS). 5In this iterative process, we repeatedly apply the fact that the Frobenius functor is exact. LYUBEZNIK NUMBERS AND INJECTIVE DIMENSION 9 Proof. Set R = S/pS. Let α : R ∼ = F (R) be defined via x 7→ x⊗1, and let β : Au (R) → F (Au (R)) be the map given by iterating the one induced by 3.2.1. If N is the F -module generated by β, then by Proposition 3.7, there exists an isomorphism of F -modules N ∼ = B u (R) such that Au (R) ∼ = / Au (R) φ  N ∼ =  / B u (R) commutes, where φ denotes the map induced by the natural transformation from Au to B u . By Lemma 3.8, the image of Au (R) in N generates N as an D(R)-module. However, as the lower row is an isomorphism of F -modules, our earlier discussion implies that it is also an isomorphism of D(R)-modules, and consequently, the image of Au (R) under the composition Au (R) → N ∼ = B u (R), which equals Im(φ) by the commutativity of the diagram, generates B u (R) as a D(R)-module. Next, consider the following commutative diagram, which is obtained by applying the natural transformation Au → B u to the canonical surjection S → R. (3.4.1) / Au (R) Au (S) φ  B u (S) ψ  / B u (R) By our earlier discussion, the iterated local cohomology modules B u (S) and B u (R) are D(S)modules, and the map ψ is D(S)-linear. It follows that Im(ψ) is a D(S)-module that is killed by p, and is therefore a D(S)/pD(S) ∼ = D(R)-submodule of B u (R) (the preceding isomorphism + follows from the discussion in [BBL 14, Subsection 2.1]). We now combine these observations to complete the proof: By hypothesis, the top row in (3.4.1) is surjective, and therefore, Im(φ) ⊆ Im(ψ). However, as noted above, Im(φ) generates B u (R) over D(R), and Im(ψ) is a D(R)-submodule of B u (A), so that B u (R) = D(R) · Im(φ) ⊆ D(R) · Im(ψ) = Im(ψ), which allows us to conclude that ψ is surjective.  Theorem 3.10. Fix a positive integer d. If multiplication by a prime integer p > 0 is injective on Au (S) for every 1 ≤ c ≤ d and u ∈ Zc , then it is also injective on B u (S) for every 1 ≤ c ≤ d and u ∈ Zc . Proof. Set R = S/pS. According to Lemma 3.6, our hypothesis that multiplication by p is injective implies that Au (S) → Au (R) is surjective for every 1 ≤ c ≤ d and u ∈ Zc . It follows from Proposition 3.9 that B u (S) → B u (R) is also surjective for every 1 ≤ c ≤ d and u ∈ Zc , and applying Lemma 3.6 once more shows that multiplication by p on B u (S) is injective for every 1 ≤ c ≤ d and u ∈ Zc .  Corollary 3.11. Given a positive integer d, there are only finitely many prime integers p > 0 for which multiplication by p on is not injective on B u (S) for some u ∈ Zd . Proof. If p is a zero divisor on B u (S) for some u ∈ Zd , then it then follows from Theorem 3.10 that p must be a zero divisor Aw (S) for some 1 ≤ c ≤ d and w ∈ Zc , and therefore that p is contained in an associated prime of such a module. The fact that there are only finitely many such p > 0 then follows from the fact that there are only finitely many non-zero iterated Koszul cohomology modules of this form, and that each associated prime of each such module (of which there are only finitely many, since these modules are finitely generated over S) can contain at most one positive prime integer.  10 D. J. HERNÁNDEZ, L. NÚÑEZ-BETANCOURT, F. PÉREZ, AND E. E. WITT 3.5. The finiteness of associated primes of iterated local cohomology. We are now ready to prove Theorem F from the introduction. In the context of Setup 3.4, this theorem states that for every positive integer d and u ∈ Zd , the module B u (S) has finitely many associated primes. Proof of Theorem F. Fix a positive integer d and an element u ∈ Zd , and set B = B u (S). The unique map Spec(S) → Spec(Z) induces a map π : AssS (B) → Spec(Z). To show that AssS (B) is finite, it suffices to show that the image of π is finite, and that the fiber over any point in the image of π is also finite. To show that the image of π is finite, it suffices to show that it contains only finitely many positive prime integers. However, if a prime integer p > 0 is contained in the image of π, then p is contained in some associated prime of B, and is therefore a zero divisor on B. Corollary 3.11 then shows that there are only finitely many such prime integers. We now consider the fibers of π. The associated primes in π −1 (0) are in one-to-one correspondence with the associated primes of the localization Q ⊗Z B = B u (Q ⊗Z S), and there are only finitely many such primes by [Lyu93, Remark 3.7(i)]. On the other hand, the primes in π −1 (p) for some prime integer p > 0 are precisely the associated primes of B containing p, and there are  only finitely many such primes by [NB13, Theorem 1.2]. 4. On the agreement of (standard) and mixed characteristic Lyubeznik numbers In this section, we focus on Question C from the introduction, which is concerned with the equality of the (standard) and mixed characteristic Lyubeznik numbers in a natural context in which they are both defined (namely, for local rings of characteristic p > 0 obtained from some fixed finitely generated Z-algebra). The first subsection is dedicated to establishing some important results (the results established in this subsection are utilized in both the current and the next section). In the second subsection, we prove Theorem B (in fact, we prove a more precise statement in Theorem 4.5). We refer the reader to the introduction for the definition of the standard and mixed characteristic Lyubeznik numbers. 4.1. Preliminary lemmas. Throughout this subsection, we suppose that T is an unramified regular local ring of mixed characteristic p > 0, with residue field k. In what follows, we repeatedly use (without further reference) the fact that local cohomology modules of T and T /pT with support in arbitrary ideals have finite Bass numbers [Lyu00, HS93]. Lemma 4.1. If M is a T -module such that pM = 0, then ExtiT (k, M ) ∼ = ExtiT /pT (k, M ) ⊕ Exti−1 T /pT (k, M ) as modules over T /pT for every integer i. Proof. Let g be a finite sequence in T such that, together with p, forms a regular system of parameters for T . By definition, the complex K • (p, g; M ) equals   p K • (g; T ) ⊗T K • (p; M ) = K • (g; T ) ⊗T 0 → M −→ M → 0 , and therefore,   K i (p, g; M ) = K i (g; T ) ⊗T M ⊕ K i−1 (g; T ) ⊗T M , where the copy of M in the left-hand summand is in degree zero, and that in the right-hand summand is in degree one. By hypothesis, multiplication by p on M is the zero map, and given this, it follows that the differentials on K • (p, g; M ) preserve each summand in this decomposition. In fact, it is straightforward to verify that, up to a sign, the induced maps on each summand agree with the differentials on the complexes K i (g; M ) and K i−1 (g; M ), respectively. In other words, there is an isomorphism of complexes (4.1.1) K i (p, g; M ) ∼ = K i (g; M ) ⊕ K i−1 (g; M ). LYUBEZNIK NUMBERS AND INJECTIVE DIMENSION 11 By our choice of g, we may compute ExtiT (k, M ) as the cohomology of K i (p, g; M ). Moreover, the image of g modulo p forms a system of parameters for T /pT , and the Koszul complex of g modulo p on M , considered as a module over T /pT , can be identified with the complex K • (g; M ). As before, this complex can be used to compute the modules ExtiT /pT (k, M ), and the lemma then follows from (4.1.1).  Using Lemma 4.1, we can relate the Bass numbers of certain local cohomology modules of T with those of certain local cohomology modules of the reduction of T modulo p. Corollary 4.2. If a is an ideal of T such that multiplication by p is injective on Haj (T ) for every j ≥ 0, then for every pair of nonnegative integers i and j, j dimk ExtiT (k, Haj (T )) = dimk Exti−1 T /pT (k, Ha (T /pT )). Proof. Our hypothesis on p guarantees that the long exact sequence in local cohomology with p respect to a associated to the short exact sequence 0 → T −→ T → T /pT → 0 breaks into short exact sequences p 0 → Haj (T ) −→ Haj (T ) → Haj (T /pT ) → 0. In turn, each such short exact sequence in local cohomology induces a long exact sequence ··· / Exti−1 (k, Haj (T )) T GF @A j ❛❛❛0 ExtiT (k, Ha (T )) p / Exti−1 (k, Haj (T )) T p / Exti−1 (k, Haj (T /pT )) T / Exti (k, Haj (T )) T ED BC / ··· , and as multiplication by p is zero on each ExtiT (k, Haj (T )), we obtain short exact sequences (4.1.2) j i j j i−1 0 → Exti−1 T (k, Ha (T )) → ExtT (k, Ha (T /pT )) → ExtT (k, Ha (T )) → 0 for every pair of integers i and j. We are now ready prove the corollary by induction on i ≥ 0. If i = 0, then it is automatic j j i that Exti−1 T /pT (k, Ha (T /pT )) is zero, while the fact that ExtT (k, Ha (T )) is also zero follows from (4.1.2). Next, suppose that j dimk ExtiT (k, Haj (T )) = dimk Exti−1 T /pT (k, Ha (T /pT )) for some i ≥ 0 and all j ≥ 0. According to (4.1.2), j i j i j dimk Exti+1 T (k, Ha (T )) = dimk ExtT (k, Ha (T /pT )) − dimk ExtT (k, Ha (T )), and by Lemma 4.1 (with M = Haj (T /pT )), we also have that j dimk ExtiT (k, Haj (T /pT )) = dimk ExtiT /pT (k, Haj (T /pT )) + dimk Exti−1 T /pT (k, Ha (T /pT )). j Substituting the second identity into the first shows that dimk Exti+1 T (k, Ha (T )) equals j i j dimk ExtiT /pT (k, Haj (T /pT )) + dimk Exti−1 T /pT (k, Ha (T /pT )) − dimk ExtT (k, Ha (T )), and our inductive hypothesis implies that the two right-most terms are equal, leaving j i j dimk Exti+1 T (k, Ha (T )) = dimk ExtT /pT (k, Ha (T /pT )), which allows us to conclude the proof.  We now shift our attention to the study of certain local cohomology modules with support in ideals containing a prime integer p > 0. 12 D. J. HERNÁNDEZ, L. NÚÑEZ-BETANCOURT, F. PÉREZ, AND E. E. WITT j Lemma 4.3. If b is an ideal of T containing p, then Hbj−1 (Tp /T ) ∼ = Hb (T ) for every j ≥ 0. Furthermore, if a is an ideal of T such that multiplication by p is injective on Haj (T ) for every j ≥ 0, and b = a + pT , then for every j ≥ 0, there is a short exact sequence p 0 → Haj (T /pT ) → Hbj (Tp /T ) −→ Hbj (Tp /T ) → 0. Proof. Since bTp = Tp , the long exact sequence in local cohomology induced by the short exact j sequence 0 → T → Tp → Tp /T → 0 implies that Hbj−1 (Tp /T ) ∼ = Hb (T ) for all integers j ≥ 0, establishing the first claim. Next, let a be above, and set b = a + pT . Since multiplication by p is injective on Haj (T ), the localization map Haj (T ) → Haj (Tp ) is injective, and therefore, the long exact sequence · · · → Haj (T ) → Haj (Tp ) → Hbj+1 (T ) → · · · breaks into short exact sequences 0 → Haj (T ) → Haj (Tp ) → Hbj+1 (T ) → 0. (4.1.3) j As p is a unit in Tp , multiplication by p on Haj (Tp ) ∼ = Ha (T ) ⊗T Tp is surjective, and (4.1.3) then implies that multiplication by p must also be surjective on the quotient Hbj+1 (T ). It then follows from the first part of the lemma that multiplication by p is surjective on Hbj (Tp /T ), and consequently, the long exact sequence in local cohomology associated to the short exact sequence p 0 → T /pT → Tp /T −→ Tp /T → 0, in which the first homomorphism sends the class of x to the class of Hbj (T /pT ) Hbj (Tp /T ) p x p, induces short exact Hbj (Tp /T ) sequences 0 → → −→ → 0 for every j ≥ 0. Finally, because the expansions of a and b to T /pT agree, we may replace b with a in the first module in this short exact sequence.  Corollary 4.4. If a is an ideal of T such that multiplication by p is injective on Haj (T ) for every j ≥ 0, and b = a + pT , then for every pair of nonnegative integers i and j, dimk ExtiT (k, Hbj (T )) = dimk ExtiT /pT (k, Haj−1 (T /pT )). Proof. The short exact sequence given in Lemma 4.3 induces the long exact sequence ··· / Exti−1 (k, Haj (T /pT )) T GF @A j i ❜❜❜0 ExtT (k, Ha (T /pT )) / Exti−1 (k, H j (Tp /T )) T b p / Exti (k, H j (Tp /T )) T b p / Exti−1 (k, H j (Tp /T )) T b ED BC / Exti (k, H j (Tp /T )) T b / ··· . As multiplication by p is zero on k, it is also zero on every module in this sequence, which therefore breaks into short exact sequences j j i j i 0 → Exti−1 T (k, Hb (Tp /T )) → ExtT (k, Ha (T /pT )) → ExtT (k, Hb (Tp /T )) → 0. On the other hand, setting M = Haj (T /pT ) in Lemma 4.1 also shows that j ExtiT (k, Haj (T /pT )) ∼ = ExtiT /pT (k, Haj (T /pT )) ⊕ Exti−1 T /pT (k, Ha (T /pT )). Comparing these two descriptions of ExtiT (k, Hbj (T /pT )), and then inducing on i ≥ 0, show that dimk ExtiT /pT (k, Haj−1 (T /pT )) = dimk ExtiT (k, Hbj−1 (Tp /T )) for all i, j ≥ 0, and the corol lary then follows from the first isomorphism in Lemma 4.3. LYUBEZNIK NUMBERS AND INJECTIVE DIMENSION 13 4.2. On the agreement of Lyubeznik numbers. Theorem B is a statement about the Lyubeznik numbers of rings obtained from a polynomial ring over Z by first killing a prime integer p > 0, and then localizing at a prime ideal. Below, we set the notation needed to precisely describe this process (and this notation also appear in Theorem 4.5). Let S denote a polynomial ring over Z, fix an ideal I of S, and set R = S/I. Furthermore, fix a prime integer p > 0, and an ideal Q of S containing a prime integer p > 0 and I. The ideal Q expands to a prime ideal under the canonical map from S to each of the rings S/pS, R, and R/pR = S/(I + pS); abusing notation, we also use Q to denote the expansion of Q ⊆ S to these rings. We also set T = SQ , so that (T, QT, k) is a regular local ring of mixed characteristic p. In fact, because T /pT is the localization of a polynomial ring over Fp (and, in particular, a regular local ring), the prime p must be part of a minimal generating set for the maximal ideal of T (i.e., T is unramified). Finally, we set A = (R/pR)Q = SQ /(ISQ + pSQ ) = T /(IT + pT ). To prove Theorem B, it suffices to show that there exists a finite set of prime integers W for ei,j (A) for i, j ≥ 0 whenever p ∈ which λi,j (A) = λ / W . Below, we establish a slightly more refined statement. Theorem 4.5 (cf. Theorem B). Under the above notation, let W denote set of prime integers that are zero divisors on HIj (S) for some j ≥ 0 (which is a finite set by Corollary 3.11). If p ∈ /W and i, j ≥ 0, then ei,j (A) = λ ei+1,j+1 (RQ ). λi,j (A) = λ Proof. By definition, we have an exact sequence 0 → IT → T → RQ → 0 and 0 → IT + pT → T → A → 0 of T -modules, and an exact sequence of T /pT -modules 0 → (IT + pT )/pT → T /pT → A → 0. Set d = dim(T ). As the left-most term in the short exact sequence of T /pT -modules agrees with the expansion of IT to T /pT , these sequences imply that ei+1,j+1 (RQ ) = dimk Exti+1 (k, H d−j−1 (T )), λ T IT ei,j (A) = dimk Exti (k, H d−j (T )), and λ T IT +pT d−j−1 (T /pT )). λi,j (A) = dimk ExtiT /pT (k, HIT As p ∈ / W , multiplication by p is injective on HIj (S) for every j ≥ 0. As T is flat over S, j it follows that multiplication by p is also injective on HIT (T ) = T ⊗S HIj (S) for every j ≥ 0. In light of this, we may then combine Corollary 4.2 (with a = IT ) and our description of the ei+1,j+1 (RQ ) = λi,j (A). Similarly, we may apply Lyubeznik numbers above to conclude that λ ei,j (A) = λi,j (A). Corollary 4.4 (with a = IT ) to concluce that λ  5. Affirmative answers to Lyubeznik’s question In this section, we consider Lyubezink’s question (i.e., Question C from the introduction), which asks the following: Suppose I is an ideal of a regular local ring (S, m) such that HIj (S) 6= 0. • Is inj. dimS HIj (S) less than or equal to dim SuppS HIj (S)? • Is the iterated local cohomology module Hmi HIj (S) injective for every i ≥ 0? 14 D. J. HERNÁNDEZ, L. NÚÑEZ-BETANCOURT, F. PÉREZ, AND E. E. WITT In the first subsection (namely, in Theorem 5.4), we give a positive answer to the first question in an interesting case, and in the third subsection (namely, in Proposition 5.10), we do the same for the second question. In the second subsection, we discuss the condition that local cohomology over a finitely generated Z-algebra descends from local cohomology over Q. 5.1. Positive answers to the question on injective dimension. Remark 5.1 (On the support of certain local cohomology). Suppose that a is an ideal of a ring T . If j ≥ 0 and p > 0, then because multiplication by p is zero on Haj (T /pT ), (5.1.1) SuppT Haj (T /pT ) = SuppT Haj (T /pT ) ∩ V(pT ) = SuppT /pT Haj (T /pT ), where we have identified Spec(T /pT ) with V(pT ) ⊆ Spec T . If, in addition, multiplication by p is injective on every local cohomology module of T with support in a, then the exact sequence p 0 → Haj (T ) −→ Haj (T ) → Haj (T /pT ) → 0 implies further that SuppT Haj (T /pT ) ⊆ (SuppT Haj (T )) ∩ V(pT ). Furthermore, because every prime in (SuppT Haj (T )) ∩ V(pT ) must properly contain an associated prime of HaJ (T ) (the properness follows from the assumption that p is a nonzerodivisor on Haj (T ), and therefore cannot be contained in any associated prime), this shows that (5.1.2) dim SuppT Haj (T /pT ) ≤ dim(SuppT Haj (T ) ∩ V(pT )) ≤ dim SuppT Haj (T ) − 1. Lemma 5.2. Let (T, m) be an unramified regular local ring of mixed characteristic p > 0. If Q ( m is a prime ideal of T , and a is an arbitrary ideal of T , then for j ≥ 0, ExtiTQ (TQ /QTQ , Haj (TQ )) = 0 for i > dim SuppT Haj (T ). Proof. Any chain of prime ideals in SuppT Haj (T ) contained in Q can be extended by adding m. In other words, dim SuppT Haj (TQ ) ≤ dim SuppT Haj (T ) − 1, so that by [Zho98, Theorem 5.1], inj. dim Haj (TQ ) ≤ dim SuppT Haj (TQ ) + 1 ≤ dim SuppT Haj (T ). Thus, all Bass numbers of Haj (TQ ) vanish above dim SuppT Haj (T ).  Proposition 5.3. Let T be an unramified regular local ring of mixed characteristic p > 0, and let a be an ideal of T such that multiplication by p is injective on every local cohomology module of T with support in a. If b = a + pT and Hbj (T ) is nonzero, then inj. dimT Hbj (T ) ≤ dim SuppT Hbj (T ). Proof. Given such a nonzero local cohomology module Hbj (T ), we aim to show that all of its Bass numbers beyond the dimension of its support are zero. Lemma 5.2 shows that this holds for Bass numbers with respect to non-maximal prime ideals, so it remains to show that (5.1.3) dimk ExtiT (k, Hbj (T )) = 0 for i > dim SuppT Hbj (T ), where k denotes the residue field of T . To establish (5.1.3), our strategy will be to “replace" the T -modules in this statement with T /pT -modules, and then appeal to results in [Lyu93] (the key point being that T /pT is a regular local ring of characteristic p > 0). As a first step in this direction, observe that Corollary 4.4 allows us to restate (5.1.3) as dimk ExtiT /pT (k, Hbj−1 (T /pT )) = 0 for i > dim SuppT Hbj (T ). Assume, momentarily, that (5.1.4) SuppT Hbj (T ) = SuppT /pT Haj−1 (T /pT ), LYUBEZNIK NUMBERS AND INJECTIVE DIMENSION 15 where we regard the latter set as a subset of Spec(T /pT ) = V(pT ) ⊆ Spec T . Given this, we may rewrite (5.1.3) as the equivalent condition dimk ExtiT /pT (k, Hbj−1 (T /pT )) = 0 for i > dim SuppT /pT Haj−1 (T /pT ), and this condition holds by [Lyu93, Corollary 3.6]. Thus, to conclude the proof, it suffices to justify (5.1.4). Towards this, notice that localizing the sequence given by Lemma 4.3 at a prime ideal Q of T gives a short exact sequence p 0 → Haj−1 (T /pT )Q → Hbj−1 (Tp /T )Q → Hbj−1 (Tp /T )Q → 0. As p ∈ b, multiplication by p cannot be injective on Hbj−1 (Tp /T )Q unless it is the zero module, and so Hbj−1 (Tp /T )Q vanishes if and only if Haj−1 (T /pT )Q does. In other words, SuppT Hbj−1 (Tp /T ) = SuppT Haj−1 (T /pT ). To establish (5.1.4), simply note that the left-hand side above agrees with SuppT Hbj (T ) by Lemma 4.3, and that the right-hand side above agrees with SuppT /pT Haj−1 (T /pT ) by (5.1.1).  Theorem 5.4 (cf. Theorem D). Let S be a polynomial ring over Z, and I an ideal of S. Let W denote the set of prime integers that are zero divisors on HIj (S) for some j ≥ 0 (which is a finite set by Corollary 3.11). If Q ∈ SuppS HIj (S) does not lie over a prime in W , then inj. dimS HIj (S)Q ≤ dim SuppS HIj (S)Q . Proof. Suppose Q is a prime ideal in SuppS HIj (S) lying over p ∈ / W . Let T = SQ , and let k be its residue field. As in the proof of Proposition 5.3, we reduce to proving a statement about T /pT -modules, and then appeal to [Lyu93]. We stress, however, that the reduction step here is different from the one used in the proof of Proposition 5.3. j Since all Bass numbers of HIT (T ) with respect to non-maximal prime ideals of T vanish beyond the dimension of its support by Lemma 5.2, it is enough to show that j j dimK ExtiT (k, HIT (T )) = 0 if i > dim SuppT HIT (T ), which, by Corollary 4.2, may be restated as (5.1.5) j j dimK Exti−1 T /pT (k, HIT (T /pT )) = 0 if i > dim SuppT HIT (T ). By our choice of p, (5.1.2) shows that j j dim SuppT /pT HIT (T /pT ) + 1 ≤ dim SuppT HIT (T ), and therefore, the condition in (5.1.5) holds whenever j j dimK Exti−1 T /pT (k, HI (T /pT )) = 0 if i − 1 > dim SuppT HIT (T ). However, this last condition holds by [Lyu93, Corollary 3.6].  5.2. On the condition that local cohomology of Z-algebras descends from Q. Let S be a polynomial ring over Z, and I an ideal of S. In this subsection, we consider the condition that Q ⊗Z HIj (S) 6= 0, which appears in the statement of Theorem D, but not in the statement of Theorem 5.4. First, we consider what occurs when this condition fails: Fix an integer j ≥ 0, and set H = HIj (S). If Q ⊗Z H = 0, then no associated prime of H lies over 0 ∈ Z. Consequently, as any prime Q ∈ SuppS H contains an associated prime Q0 of H, the preceding observation shows that pZ = Q0 ∩ Z ⊆ Q ∩ Z for some prime integer p > 0, and it follows that Q ∩ Z = pZ as well. However, being an element of an associated prime, p is a zero divisor on H, and therefore, Q lies over a prime contained in the set W appearing in the statement of Theorem 5.4. It follows from 16 D. J. HERNÁNDEZ, L. NÚÑEZ-BETANCOURT, F. PÉREZ, AND E. E. WITT this that for such local cohomology modules, Theorem 5.4 is vacuous. We provide a concrete example of this situation below. Example 5.5. Let S be a polynomial ring in six indeterminates over Z, and let I denote the monomial ideal corresponding to Reisner’s variety [Rei76]. It is know that H = HI4 (S) is nonzero [Lyu84, Example 1]. On the other hand, Q ⊗Z S/I is Cohen-Macaulay [Rei76, Remark 3], and therefore, Q ⊗Z H = 0 by [ÀM00, Proposition 3.1]. By the preceding discussion, Theorem 5.4 is an empty statement in this case. See Section 6 for a closely-related example. Next, we consider the situation when local cohomology of S descends from Q. In Corollary 5.7 below, we show that this condition guarantees that Theorem 5.4 is not vacuous. We start with a lemma that is well known to experts. Lemma 5.6. If A is finitely generated as an algebra, and torsion-free, over Z, then there are only finitely many prime integers that are units of A. Proof. Suppose that there exists an infinite sequence of prime integers {pk }∞ k=1 that are units of A. By generic freeness [HR74, Lemma 8.1], there exists a non-zero integer d such that Ad is free over Zd . For k ≫ 0, the prime integer pk does not divide d, and therefore, ∩∞ k=1 pk Zd = 0; as Ad is p A = 0. However, because multiplication by d is injective free over Zd , we also have that ∩∞ k=1 k d ∞ p A = 0, which contradicts the p A ⊆ ∩ on A, we have that A ⊆ Ad , and therefore, ∩∞ k=1 k d k=1 k assumption that each pk A = A for every k ≥ 1.  Corollary 5.7. Let A be a domain that contains Z, and that is finitely generated as an algebra over Z, and let I be an ideal of A. If Q ⊗Z HIj (A) 6= 0, then for all but finitely many prime integers p > 0, there exists a prime Q ∈ SuppA HIj (A) such that Q ∩ Z = pZ. Proof. The hypothesis that Q ⊗Z HIj (A) 6= 0 implies that there exists an associated prime Q0 of HIj (A) such that Q0 ∩ Z = 0. Set B = A/Q0 . By hypothesis, Z ⊆ A, and as Q0 ∩ Z = 0, we also have that Z = Z/(Q0 ∩ Z) ⊆ B. As B is a domain, this shows that B is a finitely generated algebra and torsion-free over Z, and by Lemma 5.6, there exists N > 0 such that any prime p > N is not a unit of B. For such a prime, Q0 + pA 6= A (otherwise, q + pa = 1 for some q ∈ Q0 and a ∈ A, which would then imply that p was a unit of B). If Q is any prime ideal of A containing the proper ideal Q0 + pA, then Q contains p, by construction, and Q is in  SuppA HIj (A), since it contains Q0 . 5.3. Positive answers to the question on iterated local cohomology. In this subsection, we address Lyubeznik’s second question concerning the injectivity of certain iterated local cohomology modules. Before we start our discussion, we recall a result of the second and fourth authors (this result plays an important role in both the current and subsequent section). Remark 5.8 (A criterion for injectivity). Let A be a power series ring over a complete Noetherian discrete valuation ring (V, pV ) of mixed characteristic p > 0, and let D = D(A, V ) denote the ring of V -linear differential operators on A. If H is any (iterated) local cohomology module of A, then H is a D-module [Lyu93, Example 2.1(iv)] and has finite Bass numbers over A [NB13, Theorem 5.1]. Consequently, if we assume further that H is supported only at the maximal ideal of A, then [NBW13, Lemma 4.2] implies that H is an injective A-module if and only if multiplication by p on H is surjective. Remark 5.9 (A specialized criterion for injectivity). Fix a prime integer p > 0, and let S be either a polynomial ring over Z, or over the localization of Z at pZ. Let A be the completion of S at the maximal ideal m generated by p and the variables in S. Note that regardless of the choice of S, A is a power series ring over the p-adic integers. Fix an (iterated) local cohomology module H of S, and suppose that H is supported only at the maximal ideal m. This condition LYUBEZNIK NUMBERS AND INJECTIVE DIMENSION 17 implies that, when regarded as an A-module in the natural way, H is isomorphic to an (iterated) local cohomology module of A. Thus, by Remark 5.8, if multiplication by p on H is surjective, then H is injective over A. However, as the S-module H is supported only at m, this implies that H is injective over S as well. Proposition 5.10. Let S be a polynomial ring over Z, and let n be the ideal of S generated by the variables. Fix an ideal I of S, and let W denote the set of prime integers that are zero divisors on some (iterated) local cohomology module of the form HIj (S) or Hni HIj (S) (which is a finite set by / W and m = n + pS, then the iterated local cohomology modules Hmi HIj (S) Corollary 3.11). If p ∈ j j j−1 and Hmi HI+pS (S) are injective for all i, j (and, in fact, Hmi HI+pS (S) ∼ = Hmi+1 HI (S)). Proof. Consider the long exact sequence (5.3.1) · · · −→ Hmi HIj (S) −→ Hni HIj (S)−→Hni HIj (S)p −→Hmi+1 HIj (S) −→ · · · . By our choice of p, each localization map in this sequence is injective, and therefore, each map Hni HIj (S)p −→Hmi+1 HIj (S) is surjective. Therefore, as multiplication by p is surjective on Hni HIj (S)p , it is also surjective on the quotient Hmi HIj (S). As the iterated local cohomology module Hmi HIj (S) is supported only at m, it follows from Remark 5.9 that Hmi HIj (S) is an injective S-module. Next, consider the long exact sequence j · · · → HIj−1 (S) → HIj−1 (Sp ) → HI+pS (S) → · · · . As above, the localization maps in this sequence are injective, and so we obtain short exact j sequences 0 → HIj−1 (S) → HIj−1 (Sp ) → HI+pS (S) → 0, each of which induces j (S) → Hmi+1 HIj−1 (S) → Hmi+1 HIj−1 (Sp ) → · · · . · · · → Hmi HIj−1 (Sp ) → Hmi HI+pS j Since p is contained in m, all modules of the form Hmi HIj (Sp ) vanish. Thus, Hmi HI+pS (S) is isomorphic to Hmi+1 HIj−1 (S), the latter of which we have already established is injective.  6. A negative answer to Lyubeznik’s question on injective dimension In this section, we fix the following conventions. For clarity in notation, we set p = 2. We also let V be the localization of Z at pZ, and set S = V [x1 , . . . , x6 ]. Let m denote the maximal ideal of S generated by p and the indeterminates, and let I be the ideal generated by the following monomials (cf. [NBW13, Remark 6.11]): µ 1 = x1 x2 x3 , µ 2 = x1 x2 x4 , µ 3 = x1 x3 x5 , µ 4 = x1 x4 x6 , µ 5 = x1 x5 x6 , µ6 = x2 x3 x6 , µ7 = x2 x5 x6 , µ8 = x2 x4 x5 , µ9 = x3 x4 x5 , µ10 = x3 x4 x6 . The goal of this section is to show that HI4 (S) can be used to provide a negative answer to Lyubeznik’s question (i.e., Question C from the introduction). We begin with an important preliminary observation: Consider the long exact sequence 3 4 · · · → HI+pS (S) → HI3 (S) → HI3 (Sp ) → HI+pS (S) → HI4 (S) → HI4 (Sp ) → · · · . j (S) vanishes It was shown in [NBW13, Remark 6.3 and the proof of Proposition 6.10] that HI+pS unless j = 4. Furthermore, Sp /ISp is known to be Cohen-Macaulay [Rei76, Remark 3], and as Sp is a polynomial ring over Q, we may apply [ÀM00, Proposition 3.1] to conclude that (6.0.2) HIj (Sp ) = 0 if j 6= dim(Sp ) − dim (Sp /ISp ) = 3. Therefore, the long exact sequence above reduces to (6.0.3) 4 0 → HI3 (S) → HI3 (Sp ) → HI+pS (S) → HI4 (S) → 0. 18 D. J. HERNÁNDEZ, L. NÚÑEZ-BETANCOURT, F. PÉREZ, AND E. E. WITT Lemma 6.1. SuppS HI4 (S) ⊆ {m}. Proof. For 1 ≤ j ≤ 7, let aj denote the ideal of S generated by µj+1 , . . . , µ10 , and let bj = aj Sµj . A direct computation shows that b1 = (x4 , x5 , x6 )Sµ1 , and therefore,     Sx4 x5 x6 3 3 ∼ ∼ Hb1 (Sµ1 ) = H(x4 ,x5 ,x6 ) (S) = Sx4 x5 + Sx4 x6 + Sx5 x6 x1 x2 x3 x1 x2 x3 (6.0.4) Sx1 x2 x3 x4 x5 x6 ∼ = Sx1 x2 x3 x4 x5 + Sx1 x2 x3 x4 x6 + Sx1 x2 x3 x5 x6 . In particular, this module vanishes after localizing at any of x4 , x5 , or x6 , so that (6.0.5) SuppS Hb31 (Sµ1 ) ⊆ V(x4 , x5 , x6 ). The following computations imply that Hbi j (Sµj ) = 0 for i ≥ 3 and 2 ≤ j ≤ 7: b2 = (x5 , x6 )Sµ2 , b3 = (x4 , x6 )Sµ3 , b4 = (x3 , x5 )Sµ4 b5 = (x2 , x3 x4 )Sµ5 , b6 = (x4 , x5 )Sµ6 , b7 = (x4 )Sµ7 . Then, substituting this vanishing into the long exact sequence · · · → Hb3j (Sµj ) → Ha4j−1 S (S) → Ha4j S (S) → Hb4j (Sµj ) → · · · , shows Ha4j−1 S (S) ∼ = Ha4j S (S) for 2 ≤ j ≤ 6, and therefore, that Ha41 (S) = · · · = Ha47 (S) = 0, where the last equality holds since a7 is generated by the three monomials µ1 , µ2 , µ3 ∈ S. Substituting this vanishing into the long exact sequence · · · → Ha31 (S) → Hb31 (Sµ1 ) → HI4 (S) → Ha41 (S) → Hb41 (Sµ1 ) → · · · , shows that Hb31 (S) surjects onto HI3 (S). It follows from this and (6.0.5) that (6.0.6) SuppS HI4 (S) ⊆ Supp Hb31 (Sµ1 ) ⊆ V(x4 , x5 , x6 ). We next construct another closed set containing SuppS HI4 (S) using a similar argument. For 3 ≤ j ≤ 9, let cj be the ideal of S generated by µ1 , . . . , µj−1 . Let dj = cj Sµj , so that d4 = (x2 , x3 x5 )Sµ4 , d5 = (x3 , x4 )Sµ5 , d6 = (x1 )Sµ6 , d7 = (x1 , x3 )Sµ7 d8 = (x1 , x6 )Sµ8 , d9 = (x1 , x2 )Sµ9 , and d10 = (x1 , x2 , x5 )Sµ10 . As before, this shows that Hdij (Sµj ) = 0 for i ≥ 3 and 4 ≤ j ≤ 10, and combining this with the long exact sequence · · · → Hd3j (Sµj ) → Hc4j+1 S (S) → Hc4j S (S) → Hd4j (Sµj ) → · · · , shows that Hc410 (S) = · · · = Hc44 (S) = 0, where the last equality holds since c4 can be generated by the three monomials µ8 , µ9 , µ10 ∈ S. This and the long exact sequence · · · → Hc310 (S) → Hd310 (Sµ10 ) → HI4 (S) → Hc410 (S) → Hd410 (Sµ10 ) → · · · , imply that Hd310 (Sµ10 ) surjects onto HI4 (S), and by isomorphisms analogous to those in (6.0.4), we can conclude that SuppS Hd310 (Sµ10 ) is contained in V(x1 , x2 , x5 ), so that (6.0.7) SuppS HI4 (S) ⊆ SuppS Hd310 (Sµ10 ) ⊆ V(x1 , x2 , x5 ). Notice that after interchanging x1 with x3 and x4 with x6 , the set of generators of I remains the same. Thus, applying this symmetry to (6.0.7) implies that SuppS HI4 (S) ⊆ V(x2 , x3 , x5 ) as well, and these containments and (6.0.6) allow us to conclude that SuppS HI4 (S) ⊆ V(x4 , x5 , x6 ) ∩ V(x1 , x2 , x5 ) ∩ V(x2 , x3 , x5 ) = V(x1 , x2 , x3 , x4 , x5 , x6 ). LYUBEZNIK NUMBERS AND INJECTIVE DIMENSION 19 Finally, as HI4 (Sp ) = 0 by (6.0.2), we have that SuppS HI4 (S) ⊆ V(pS) ∩ V(x1 , x2 , x3 , x4 , x5 , x6 ) = {m}.  Lemma 6.2. Multiplication by p on HI4 (S) is not surjective. In particular, HI4 (S) 6= 0. 4 Proof. In [NBW13, Corollary 6.9], it was shown that multiplication by p on HI+pS (S) is not 4 surjective. With this in mind, suppose that multiplication by p on HI (S) is surjective. We aim 4 to derive a contradiction showing that this implies that p must also be surjective on HI+pS (S). Consider the following commutative diagram, whose rows are the sequence (6.0.3): 0 / H 3 (S) I α / H 3 (Sp ) I p  0 / H 3 (S) I β / H4 I+pS (S) p α  / H 3 (Sp ) I γ / H 4 (S) I p β  / H4 I+pS (S) /0 p γ  / H 4 (S) I / 0. Supposing that the last column is exact, we now show via a diagram chase that the third column is 4 also exact: Take v ∈ HI+pS (S). If w = γ(v), then there exists w′ ∈ HI4 (S) such that pw′ = w by 4 our assumption. In addition, there exists v ′ ∈ HI+pS (S) such that γ(v ′ ) = w′ . Since the diagram is commutative, we have that γ(pv ′ ) = w = γ(v), so that v − pv ′ ∈ Ker(γ) = Im(β). Thus, there exists u ∈ HI3 (Sp ) such that β(u) = v − pv ′ . Since multiplication by p is an automorphism on HI3 (Sp ), there exist u′ such that pu′ = u. Let z = β(u′ ) + v ′ . Then pz = pβ(u′ ) + pv ′ = β(pu′ ) + pv ′ = β(u) + pv ′ = v − pv ′ + pv ′ = v. 4 As v was arbitrary, this shows that multiplication by p on HI+pS (S) is surjective. By our earlier discussion, this is a contradiction.  With these results in hand, we are finally able to show that the answer to Lyubeznik’s question (Question C in the introduction) is not always positive. Theorem 6.3 (cf. Theorem E). Let T denote the completion of S at m. Then the injective 4 (T ) = H 0 H 4 (T ) is one, while its support is zero-dimensional. In particular, dimension of HIT mT IT this local cohomology module provides a negative answer to both parts of Lyubeznik’s question. Proof. As HI4 (S) is supported only at m by Lemma 6.1, it is naturally a module over the com4 (T ) ∼ H 4 (S) as T -modules, and as a T -module, it is pletion of S at m, T ; moreover, HIT = I supported only at the maximal ideal mT of T . As multiplication by p is not surjective on 4 (T ) = H 0 H 4 (T ) by Lemma 6.2, we know that this module is not injective by Remark 5.8. HIT mT IT We can then apply [Zho98, Theorem 5.1] to see that 4 4 1 ≤ inj. dimT HIT (T ) ≤ dim SuppT HIT (T ) + 1 = 0 + 1 = 1.  Acknowledgments The authors began this work in 2013 during the program in Commutative Algebra at the Mathematical Sciences Research Institute (MSRI). We thank MSRI for their hospitality and support, which includes Postdoctoral Fellowships for the first and fourth authors. The authors are also grateful to the National Science Foundation (NSF) and the National Council of Science and Technology of Mexico (CONACyT) for support: The first author is supported by a NSF Postdoctoral Research Fellowship (DMS-1304250), the second was partially supported by CONACyT grant 207063 and NSF grant DMS-1502282, the third by NSF grant DMS-1068190, and the last by NSF grant DMS-1501404. 20 D. J. HERNÁNDEZ, L. NÚÑEZ-BETANCOURT, F. PÉREZ, AND E. E. WITT References [ÀM00] Josep Àlvarez Montaner. Characteristic cycles of local cohomology modules of monomial ideals. J. Pure Appl. Algebra, 150(1):1–25, 2000. 16, 17 [ÀMBL05] Josep Àlvarez Montaner, Manuel Blickle, and Gennady Lyubeznik. Generators of D-modules in positive characteristic. Math. Res. Lett., 12(4):459–473, 2005. 8 [ÀMGLZA03] Josep Àlvarez Montaner, Ricardo García López, and Santiago Zarzuela Armengou. Local cohomology, arrangements of subspaces and monomial ideals. Adv. Math., 174(1):35–56, 2003. 1 [ÀMV14] J. Àlvarez Montaner and A. Vahidi. Lyubeznik numbers of monomial ideals. Trans. Amer. Math. Soc., 366(4):1829–1855, 2014. 1 [AMY14] Josep Àlvarez Montaner and Kohji Yanagawa. Lyubeznik numbers of local rings and linear strands of graded ideals. Preprint, arXiv: 1409.6486 [math.AC], 2014. 1 [BB05] Manuel Blickle and Raphael Bondu. Local cohomology multiplicities in terms of étale cohomology. Ann. Inst. Fourier (Grenoble), 55(7):2239–2256, 2005. 1 [BBL+ 14] Bhargav Bhatt, Manuel Blickle, Gennady Lyubeznik, Anurag K. Singh, and Wenliang Zhang. Local cohomology modules of a smooth Z-algebra have finitely many associated primes. Invent. Math., 197(3):509–519, 2014. 1, 3, 5, 9 [BS98] Markus P. Brodmann and Rodney Y. Sharp. Local cohomology: an algebraic introduction with geometric applications, volume 60 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 1998. 4 [Coh46] I. S. Cohen. On the structure and ideal theory of complete local rings. Trans. Amer. Math. Soc., 59:54–106, 1946. 2 [GLS98] R. García López and C. Sabbah. Topological computation of local cohomology multiplicities. Collect. Math., 49(2-3):317–324, 1998. Dedicated to the memory of Fernando Serrano. 1 [HH99] Melvin Hochster and Craig Huneke. Tight closure in equal characteristic zero. 1999. 3 [HR74] Melvin Hochster and Joel L. Roberts. Rings of invariants of reductive groups acting on regular rings are Cohen-Macaulay. Advances in Math., 13:115–175, 1974. 16 [HS93] Craig L. Huneke and Rodney Y. Sharp. Bass numbers of local cohomology modules. Trans. Amer. Math. Soc., 339(2):765–779, 1993. 1, 10 [ILL+ 07] Srikanth B. Iyengar, Graham J. Leuschke, Anton Leykin, Claudia Miller, Ezra Miller, Anurag K. Singh, and Uli Walther. Twenty-four hours of local cohomology, volume 87 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2007. 5 [Kaw00] Ken-ichiroh Kawasaki. On the Lyubeznik number of local cohomology modules. Bull. Nara Univ. Ed. Natur. Sci., 49(2):5–7, 2000. 1 [Kaw02] Ken-ichiroh Kawasaki. On the highest Lyubeznik number. Math. Proc. Cambridge Philos. Soc., 132(3):409–417, 2002. 1 [Lyu84] Gennady Lyubeznik. On the local cohomology modules Hai (R) for ideals a generated by monomials in an R-sequence. In Complete intersections (Acireale, 1983), volume 1092 of Lecture Notes in Math., pages 214–220. Springer, Berlin, 1984. 16 [Lyu93] Gennady Lyubeznik. Finiteness properties of local cohomology modules (an application of Dmodules to commutative algebra). Invent. Math., 113(1):41–55, 1993. 1, 2, 3, 8, 10, 14, 15, 16 [Lyu97] Gennady Lyubeznik. F -modules: applications to local cohomology and D-modules in characteristic p > 0. J. Reine Angew. Math., 491:65–130, 1997. 5, 6, 8 [Lyu00] Gennady Lyubeznik. Finiteness properties of local cohomology modules for regular local rings of mixed characteristic: the unramified case. Comm. Algebra, 28(12):5867–5882, 2000. Special issue in honor of Robin Hartshorne. 1, 3, 10 [Lyu06] Gennady Lyubeznik. On some local cohomology invariants of local rings. Math. Z., 254(3):627–640, 2006. 1 [NB13] Luis Núñez-Betancourt. Local cohomology modules of polynomial or power series rings over rings of small dimension. Illinois J. Math., 57(1):279–294, 2013. 10, 16 [NBW13] Luis Núñez-Betancourt and Emily E. Witt. Lyubeznik numbers in mixed characteristic. Math. Res. Lett., 20(6):1125–1143, 2013. 2, 3, 16, 17, 19 [NBWZ14] Luis Núñez-Betancourt, Emily E. Witt, and Wenliang Zhang. A survey on the Lyubeznik numbers. Preprint, arXiv:1406.7590 [math.AC], 2014. 2 [PS73] C. Peskine and L. Szpiro. Dimension projective finie et cohomologie locale. Applications à la démonstration de conjectures de M. Auslander, H. Bass et A. Grothendieck. Inst. Hautes Études Sci. Publ. Math., (42):47–119, 1973. 3 [Rei76] Gerald Allen Reisner. Cohen-Macaulay quotients of polynomial rings. Advances in Math., 21(1):30– 49, 1976. 16, 17 LYUBEZNIK NUMBERS AND INJECTIVE DIMENSION [Wal01] [Zha07] [Zho98] 21 Uli Walther. On the Lyubeznik numbers of a local ring. Proc. Amer. Math. Soc., 129(6):1631–1634 (electronic), 2001. 1 Wenliang Zhang. On the highest Lyubeznik number of a local ring. Compos. Math., 143(1):82–88, 2007. 1 Caijun Zhou. Higher derivations and local cohomology modules. J. Algebra, 201(2):363–372, 1998. 3, 14, 19 Department of Mathematics, University of Michigan, Ann Arbor, MI 48109 Email address: [email protected] Department of Mathematics, University of Virginia, Charlottesville, VA 22903 Email address: [email protected] Department of Mathematics & Statistics, Georgia State University, Atlanta, GA 30303 Email address: [email protected] Department of Mathematics, Kansas University, Lawrence, KS 66045 Email address: [email protected]
0math.AC
Tests for the weights of the global minimum variance portfolio in a high-dimensional setting Taras Bodnara, a arXiv:1710.09587v2 [q-fin.ST] 15 Nov 2017 b ∗ , Solomiia Dmytrivb , Nestor Parolyac, d and Wolfgang Schmidb Department of Mathematics, Stockholm University, Stockholm, Sweden Department of Statistics, European University Viadrina, Frankfurt(Oder), Germany c Institute of Statistics, Leibniz University Hannover, Hannover, Germany d Department of Economics, Heidelberg University, Heidelberg, Germany Abstract: In this study, we construct two tests for the weights of the global minimum variance portfolio (GMVP) in a high-dimensional setting, namely, when the number of assets p depends on the sample size n such that np → c ∈ (0, 1) as n tends to infinity. The considered tests are based on the sample estimator and on the shrinkage estimator of the GMVP weights. We derive the asymptotic distributions of both test statistics under the null and alternative hypotheses. Moreover, we provide a simulation study where the power functions of the proposed tests are compared with other existing approaches. We observe that the test performs well based on the shrinkage estimator even for values of c close to one. Keywords: Finance; Portfolio analysis; Global minimum variance portfolio; Statistical test; Shrinkage estimator; Random matrix theory. 1 Introduction Financial markets have developed rapidly in recent years, and the amount of money invested in risky assets has substantially increased. Due to this, an investor must have knowledge of optimal portfolio proportions in order to receive a large expected return and, at the same time, reduce the level of the risk associated with the investment decision. Since Markowitz (1952) presented his mean-variance analysis, many works about optimal portfolio selection have been published. However, investors are faced with some difficulties in the practical implementation of these investing theories since sampling error is present when unknown theoretical quantities are estimated. In classical asymptotic analysis, it is almost always assumed that the sample size increases while the size of the portfolio, namely the number of included assets p , remains constant (e.g., ∗ Corresponding Author: Taras Bodnar. E-Mail: [email protected]. Tel: +46 8 164562. Fax: +46 8 612 6717. 1 Jobson and Korkie (1981), Okhrin and Schmid (2006)). Nowadays, this case is often called standard asymptotics (see Cam and Yang (2000)). Here, the traditional plug-in estimator of the optimal portfolio, the so-called sample estimator, is consistent and asymptotically normally distributed. However, in many applications, the number of assets in a portfolio is large in comparison to the sample size (i.e. the portfolio dimension p and the sample size n tend to p tends to the concentration ratio c > 0 . In this case, infinity simultaneously) such that n we are faced with so-called high-dimensional asymptotics or ‘Kolmogorov’ asymptotics (see Bühlmann and Van De Geer (2011), Bai and Shi (2011), and Cai and Shen (2011)). Whenever the dimension of the data is large, the classical limit theorems are no longer suitable because the traditional estimators result in a serious departure from the optimal estimators under high-dimensional asymptotics (Bai and Silverstein (2010)). These methods fail to provide consistent estimators of the unknown parameters of the asset returns, that is, the mean vector and the covariance matrix. Generally, the greater the concentration ratio c , the worse the sample estimators are. In these cases, new test statistics must be developed, and completely new asymptotic techniques must be applied for their derivations. Several studies deal with high-dimensional asymptotics in portfolio theory using results from random matrix theory (see Frahm and Jaekel (2008) and Laloux et al. (2000)). Recently, Bodnar, Parolya and Schmid (2017) presented a shrinkage-type estimator for the global minimum variance portfolio (GMVP) weights, and Bodnar, Okhrin and Parolya (2017) derived the optimal shrinkage estimator of the mean-variance portfolio. Testing the efficiency of a portfolio is a classical problem in finance. The former literature focuses on the case of standard asymptotics or considers exact tests where both p and n are fixed. For example, Gibbons, Ross and Shanken (1989) provided an exact F -test for the efficiency of a given portfolio, and Britten-Jones (1999) derived inference procedures on the efficient portfolio weights based on the application of linear regression. More recently, Bodnar and Schmid (2008) presented a test for the general linear hypothesis of the portfolio weights in the case of elliptically contoured distributions. The contribution of this study is the derivation of statistical techniques for testing the efficiency of a portfolio under high-dimensional asymptotics. Two statistical tests are considered. Whereas the first approach is based on the asymptotic distribution of the test statistic suggested by Bodnar and Schmid (2008) in a highdimensional setting, the second test makes use of the shrinkage estimator of the GMVP weights and provides a powerful alternative to the existing methods. To the best of our knowledge, this analysis is the first time that the shrinkage approach has been applied to statistical test theory. The paper is structured as follows. In Section 2, we discuss the main results on distributional properties for optimal portfolio weights presented by Okhrin and Schmid (2006). In Section 3, a new test based on the shrinkage estimator for the GMVP weights is derived, and the highdimensional version of the test based on the test statistics given in Bodnar and Schmid (2008) is proposed. The asymptotic distributions of the test statistics under both the null hypothesis and the alternative hypothesis are obtained, and the corresponding power functions of both tests are presented. In Section 4, the power functions of the proposed tests are compared with 2 each other for different values of c ∈ (0, 1) . In our comparison study, a test of Glombeck (2014) is considered as well. We conclude in Section 5. All proofs are given in the Appendix. 2 Estimation of Optimal Portfolio Weights We consider a financial market consisting of p risky assets. Let Xt denote the p -dimensional vector of the returns on risky assets at time t . Suppose that E(Xt ) = µ and Cov(Xt ) = Σ . The covariance matrix Σ is assumed to be positive definite. Let us consider a single period investor who invests in the GMVP, one of the most commonly used portfolios (see, for example, Memmel and Kempf (2006), Frahm and Memmel (2010), Okhrin and Schmid (2006), Bodnar and Schmid (2008), Glombeck (2014), and others). This portfolio exhibits the smallest attainable portfolio variance w0 Σw under the constraint w0 1 = 1 , where 1 = (1, . . . , 1)0 denotes the p -dimensional vector of ones and w stands for the vector of portfolio weights. The weights of GMVP are given by wGM V P = Σ−1 1 . 10 Σ−1 1 (1) The practical implementation of the mean-variance framework in the spirit of Markowitz (1952) relies on estimating the first two moments of the asset returns. Because we do not know the true covariance matrix, it is usually replaced by its sample estimator, which is based on a sample of n > p historical asset returns X1 , . . . , Xn given by n Σ̂n = n  0 1 X 1X Xv . Xj − X̄n Xj − X̄n with X̄n = n − 1 j=1 n v=1 (2) Replacing Σ in (1) by the sample estimator Σ̂n , we obtain an estimator of the GMVP weights expressed as Σ̂−1 1 ŵn = 0 n . (3) 1 Σ̂−1 n 1 Note that the estimator of the GMVP weights is exclusively a function of the estimator Σ̂n of the covariance matrix. Assuming that the asset returns {Xt } follow a stationary Gaussian process with mean µ and covariance matrix Σ , Okhrin and Schmid (2006) proved that the vector of estimated optimal portfolio weights is asymptotically normal. Under the additional assumption of independence, they derived the exact distribution of ŵn . Okhrin and Schmid (2006) showed that the distribution of arbitrary p − 1 components of ŵn is a (p − 1) - dimensional t -distribution with n − p + 1 degrees of freedom and E(ŵn ) = wGM V P , Cov(ŵn ) = Ω = Q Σ−1 110 Σ−1 1 −1 , Q = Σ − . (4) n − p − 1 10 Σ−1 1 10 Σ−1 1 3 ∗ Consequently, if ŵn∗ and wGM V P are obtained by deleting the last element of ŵn and ∗ ∗ wGM V P and if Ω and Q consist of the first (p − 1) × (p − 1) elements of Ω and Q , then ŵn∗ has a (p − 1) -variate t-distribution with n − p + 1 degrees of freedom and param1 Q∗ ∗ eters wGM and . This distribution is denoted by ŵn∗ ∼ tp−1 (n − p + VP n − p + 1 10 Σ−1 1 n−p−1 ∗ n−p−1 ∗ 1 Q∗ ∗ 1, wGM , Ω ) , since Ω = . VP n−p+1 n−p+1 n − p + 1 10 Σ−1 1 3 Test Theory for the GMVP in High Dimensions At each time point, an investor is interested to know whether the portfolio he is holding coincides with the true GMVP or has to be reconstructed. For that reason, we consider the following testing problem: H0 : wGM V P = r against H1 : wGM V P 6= r, (5) where r with r0 1 = 1 is a known vector of, for example, the weights of the holding portfolio. Thus, this problem analyses whether the true GMVP weights are equal to some given values. Bodnar and Schmid (2008) analysed a general linear hypothesis for the GMVP portfolio weights and introduced an exact test assuming that the asset returns are independent and elliptically contoured distributed. Moreover, they derived the exact distribution of the test statistic under the null hypothesis and the alternative hypothesis. The main focus of this study is high-dimensional portfolios. We want to consider the p testing problem (5) in a high-dimensional environment, that is, assuming that → c ∈ (0, 1) n as n → ∞ . Note that, in this case, H0 and H1 depend on n as well. Thus, it would be more precise to write H0,n : w∗GM V P,n = r∗n and H1,n : w∗GM V P,n 6= r∗n . In the following, we will ignore this fact in order to simplify our notation. Moreover, it turns out that the sample covariance matrix is no longer a good estimator of the covariance matrix (see Bai and Silverstein (2010) and Bai and Shi (2011)). For that reason, it is unclear how well the test of Bodnar and Schmid (2008) behaves in that context. First, we study its behaviour under the high-dimensional asymptotics, and, after that, we propose an alternative test that makes use of the shrinkage estimator for the portfolio weights (cf. Bodnar, Parolya and Schmid (2017)). In recent years, several studies have dealt with estimators of unknown portfolio parameters under high-dimensional asymptotics with applications to portfolio theory. Glombeck (2014) formulated tests for the portfolio weights, variances of the excess returns, and Sharpe ratios of the GMVP for c ∈ (0, 1) . Bodnar, Parolya and Schmid (2017) and Bodnar, Okhrin and Parolya (2017) derived the shrinkage estimators for the GMVP and for the mean-variance portfolio, respectively, under the Kolmogorov asymptotics for c ∈ (0, ∞) . 3.1 A Test Based on the Mahalanobis Distance Bodnar and Schmid (2008) proposed a test for a general linear hypothesis of the weights of the 4 global minimum variance portfolio. Here, we are interested in the special case (5). For this case, the test statistic is given by Tn = n − p 0 −1 (1 Σ̂ 1)(ŵn∗ − r∗ )0 (Q̂∗n )−1 (ŵn∗ − r∗ ), p−1 (6) where Q̂∗n consists of the first (p − 1) × (p − 1) elements of Q̂n = Σ̂−1 n − 0 −1 Σ̂−1 n 11 Σ̂n and 10 Σ̂−1 n 1 the number of assets p in the portfolio is fixed. It was shown that, under the null hypothesis, Tn ∼ Fp−1,n−p . Moreover, the density of Tn under the alternative hypothesis H1 is equal to fTn (x) = fp−1,n−p (x) (1 + λ)−(n−1)/2   n−1 n−1 p−1 (p − 1)x λ × , , ; , 2 F1 2 2 2 n − p + (p − 1)x 1 + λ (7) where ∗ ∗ 0 ∗ −1 ∗ ∗ λ = 10 Σ−1 1(wGM V P − r ) (Q ) (wGM V P − r ) (8) and 2 F1 stands for the hypergeometric function (see Abramowitz and Stegun (1964), chap. 15), that is, ∞ Γ(c) X Γ(a + i)Γ(b + i) z i . 2 F1 (a, b, c; x) = Γ(a)Γ(b) i=0 Γ(c + i) i! Thus, the exact power function of the test is known. Note that this result is also valid for matrix-variate elliptically contoured distributions (see Bodnar and Schmid (2008)). On the other hand, several computational difficulties appear when the power function of the test is calculated for large values of p and n , since doing so involves a hypergeometric function whose computation is very challenging for large values of p and n . In order to deal with this problem, we derive the asymptotic distribution of Tn in a high-dimensional setting. This result is given in Theorem 1. The proof is in the Appendix. Since λ depends on p (i.e. on n ) through Σ , we write λn in the rest of the paper. Theorem 1 Let p ≡ p(n) and cn = np → c ∈ (0, 1) . Assume that {Xt } is a sequence of independent and normally distributed p -dimensional random vectors with mean µ and covariance matrix Σ , which is assumed to be positive definite. Let Cn2 λ2 λn c =2+2 n +4 +2 c c 1−c  2 λn 1+ . c Then, it holds that p p−1 Tn − 1 − λn n−1 p−1 Cn ! d → N (0, 1) for p/n → c ∈ (0, 1) as n → ∞ . Under the null hypothesis, for p/n → c ∈ (0, 1) as n → ∞ . 5 √ d p − 1 (Tn −1) → N (0, 2/(1 − c)) The results of Theorem 1 lead to an asymptotic expression of the power function given by √ ! p − 1 (Tn − 1) p P > z1−α 2/(1 − c)   √  p √ n−1 p − 1 Tn − 1 − λn n−1 2/(1 − c)z − p − 1λ 1−α n p−1 p−1  = 1−P  ≤ Cn Cn ! p √ 2/(1 − c)z1−α − p − 1 λcn ≈ 1−Φ , Cn (9) where z1−α is the (1 − α) -quantile of the standard normal distribution. In Figure 1, we plot the power function (9) as a function of λn for several values of c and n (solid line). In addition, the empirical power of the test is shown for the same values of c and n (dashed line) and is equal to the relative number of rejections of the null hypothesis obtained via a simulation study. It is remarkable that, following the proof of Theorem 1, the considered simulation study can be considerably simplified. Instead of generating a p × n random matrix of asset returns in each simulation run, we simulate four independent random variables from standard univariate distributions and then compute the statistic Tn for the given value of λn following the stochastic representation (25) in the Appendix. Namely, the simulation study is performed in the following way: (b) (b) (b) (i) Generate four independent random variables ω1 ∼ N (0, 1) , ξ2 ∼ χ2n−p , ξ3 ∼ χ2n−1 , (b) and ξ4 ∼ χ2p−2 (ii) For fixed λn , compute d Tn(b) = q (b) (b) 2 (b) n − p ( λn ξ3 + ω1 ) + ξ4 (b) p−1 ξ2 (iii) Repeat steps (a) and (b) for b = 1, ..., B and calculate the empirical power by √  p−1 −1 1 , p 1(z1−α ,+∞)  P̂ = B b=1 2/(1 − c) B X  (b) Tn (10) where 1A (.) is the indicator function of the set A . In Figure 1, we observe a good performance of the asymptotic approximation of the power function. This approximation works perfectly for both small and large values of c . 3.2 Test Based on a Shrinkage Estimator In most cases, the unknown parameters of the asset return distribution are replaced by their sample counterparts when an optimal portfolio is constructed. In recent years, however, other 6 0.9 0.8 0.7 0.6 0.5 Power 0.1 0.2 0.3 0.4 0.5 0.1 0.2 0.3 0.4 Power 0.6 0.7 0.8 0.9 1.0 c = 0.5 1.0 c = 0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Asymptotic Empirical 0.0 0.0 Asymptotic Empirical 1.0 0.0 0.1 0.2 Noncentrality parameter, λn 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Noncentrality parameter, λn 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.1 0.2 0.3 0.4 0.5 Power 0.6 0.7 0.8 0.9 1.0 c = 0.9 1.0 c = 0.7 Power 0.3 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Asymptotic Empirical 0.0 0.0 Asymptotic Empirical 1.0 0.0 Noncentrality parameter, λn 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Noncentrality parameter, λn Figure 1: Asymptotic power function vs. empirical power function for different values as a function of λn in (7) for various values of c ∈ {0.1, 0.5, 0.7, 0.9} and a 5% significance level. types of estimators, such as shrinkage estimators, have been discussed as well (see Okhrin and Schmid (2007) and Bodnar, Parolya and Schmid (2017)). The shrinkage methodology was introduced by Stein (1956). His results were extended by Efron and Morris (1976) to the case in which the covariance matrix is unknown. The shrinkage methodology can be applied to the expected asset returns (e.g. Jorion (1986)) and the covariance matrix (Bodnar, Gupta and Parolya (2014, 2016)). Both of these applications appear to be very successful in reducing damaging influences on the portfolio selection. A shrinkage estimator was applied directly to the portfolio weights by Golosnoy and Okhrin (2007) and Okhrin and Schmid (2008). They showed that the shrinkage estimators of the portfolio weights lead to a decrease in the variance of the portfolio weights and to an increase in utility. Bodnar, Parolya and Schmid (2017) proposed a new shrinkage estimator for the weights of the GMVP that turns out to provide better results in the high-dimensional case than the existing estimators do. This estimator is based on a convex combination of the sample estimator of the GMVP weights and an arbitrary constant vector expressed as 7 ŵn;GSE = αn Σ̂−1 n 1 0 1 Σ̂−1 n 1 + (1 − αn )bn with b0n 1 = 1. (11) Here, the index GSE stands for ‘general shrinkage estimator’. It is assumed that bn ∈ Rp is a vector of constants such that b0n Σbn is uniformly bounded. Bodnar, Parolya and Schmid (2017) proposed determining the optimal shrinkage intensity αn for a given target portfolio bn such that the out-of-sample risk is minimal, that is, L = (ŵn;GSE − wGM V P )0 Σ(ŵn;GSE − wGM V P ) (12) is minimized with respect to αn . This result leads to (bn − ŵn )0 Σ bn α̂n = . (bn − ŵn )0 Σ (bn − ŵn ) (13) The authors showed that the optimal shrinkage intensity α̂n is almost surely asymptotically equivalent to a non-random quantity α̃n ∈ [0, 1] when np → c ∈ (0, 1) as n → ∞ , which is given by α̃n = where Rbn = (1 − c)Rbn , c + (1 − c)Rbn (14) σb2 n − σn2 = 10 Σ−1 1b0n Σbn − 1 σn2 is the relative loss of the target portfolio bn , σb2 n = b0n Σbn is the variance of the target portfolio, and σn2 = 1/10 Σ−1 1 is the variance of the GMVP. This result provides an estimator of the optimal shrinkage intensity given by ˆn = α̃ (1 − np )R̂bn p n + (1 − p )R̂bn n , R̂bn = (1 − p 0 )bn Σ̂n bn 10 Σ̂−1 n 1 − 1. n (15) ˆ n , the corresponding portfolio weights are given by Using the estimated shrinkage intensity α̃ ˆ n ŵn + (1 − α̃ ˆ n )bn . ŵn;ESI = α̃ (16) ˆn α̃ → 1 if np → c ∈ (0, 1) α̃n as n → ∞ . In Theorem 2, we show that the estimated intensity is asymptotically normally distributed. The proof of Theorem 2 is given in the Appendix. Bodnar, Parolya and Schmid (2017) proved that the ratio Theorem 2 Let p ≡ p(n) and cn = np → c ∈ (0, 1) . Assume that {Xt } is a sequence of independent and normally distributed p -dimensional random vectors with mean µ and covariance matrix Σ , which is assumed to be positive definite. Then 8 ˆ n − An d √ α̃ n → N (0, 1) for p/n → c ∈ (0, 1) as n → ∞ , Bn (17) where (1 − cn )Rbn , cn + (1 − cn )Rbn c3 (1 − cn )(1 + Rbn )2 = 2 n . (cn + (1 − cn )Rbn )4 An = Bn2 Next, we introduce a test based on the estimated shrinkage intensity. The motivation is based on the following equivalences: α̃n = 0 ⇐⇒ Rbn = 0 ⇐⇒ σb2 n = σn2 . This result means that α̃n = 0 if and only if the variance of the portfolio based on bn is equal = to the variance of the GMVP. This finding in turn means that b0n Σbn = 1/10 Σ−1 1 = min 0 w: w 1=1 0 w0 Σw = wGM V P Σ wGM V P . Since the GMVP weights are uniquely determined, this result is valid if and only if bn = wGM V P . Choosing bn = r , it holds that wGM V P = r ⇐⇒ α̃n = 0. Thus, it is possible to obtain a test for the structure of the GMVP using the shrinkage intensity with the hypothesis given by H0 : α̃n = 0 against H1 : α̃n > 0. (18) ˆ = α̃(b ˆ n ) . Let Sn = √n α̃(b ˆ n = r) . For testing (18), we use the test statistic Sn . Note that α̃ Theorem 3 Suppose that the conditions of Theorem 2 are satisfied. Then √ Sn − nAn d → N (0, 1) for p/n → c ∈ (0, 1) as n → ∞ , Bn d where An and Bn are given in the statement of Theorem 2. Under the null hypothesis, Sn → N (0, 2(1 − c)/c) for p/n → c ∈ (0, 1) as n → ∞ . The proof of Theorem 3 follows directly from Theorem 2. This result gives us a promising new approach for detecting deviations of the true portfolio weights from the given quantities. Using Theorem 3, we are able to make a statement about the power function of this test. Since 9 An and Bn depend on bn , we only have to replace this quantity with r . It holds that   q  2 1−c z1−α − An (bn = r) c Sn Sn − An (bn = r) P q ≤ > z1−α  = 1 − P  Bn (bn = r) Bn (bn = r) 2 1−c c q  2 1−c z − A (b = r) 1−α n n c . ≈ 1 − Φ Bn (bn = r)   (19) Note that the approximation given in (19) is purely a function of Rbn =r . This property is a main difference from the test discussed in Section 3.1, where the power function is a function of λn . These properties are very useful to analyse the performances of both tests and simplify the power analysis. 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.6 0.8 0.1 0.2 c=0.1 c=0.3 c=0.5 c=0.7 c=0.9 0.0 0.0 0.0 0.4 Power 0.6 0.4 c=0.1 c=0.3 c=0.5 c=0.7 c=0.9 0.2 Power 0.8 1.0 n=1500 1.0 n=500 1.0 0.0 Relative loss of the portfolio, Rbn 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Relative loss of the portfolio, Rbn Figure 2: The asymptotic power function of the test in (18) as a function of Rbn for various values of c ∈ {0.1, 0.3, 0.5, 0.7, 0.9} and a 5% significance level. The number of observations is n = 500 . In Figure 3, the power of the test is shown as a function of Rbn and n . It can be seen that the test performs better for smaller values of c . With increasing values of c , the power of the test decreases. We determine the power function for two different sample numbers, n = 500 and n = 1500 . As expected, the test shows a better performance for larger values of n , since An (bn = r) increases, the numerator of the expression in the cumulative distribution function in (19) becomes negative, and the whole expression tends to one. 4 Comparison Study The aim of this section is to compare several tests for the weights of the GMVP. In the preceding two subsections, we considered two tests for the weights of the GMVP. For the test based on the empirical portfolio weights, the exact distribution of the test statistic is known. In Section 3.1, the asymptotic power function of the test proposed by Bodnar and Schmid (2008) is derived in a high-dimensional setting. In Section 3.2, a new test is proposed, and its asymptotic power function, which purely depends on Rbn =r , is determined. The fact 10 that both tests depend on different quantities complicates the comparison of both tests. Note that r0 Σr Rbn =r = 10 Σ−1 1 r0 Σr − 1 = λn − 1. ∗ ∗ ∗ 0 ∗ −1 ∗ (wGM V P − r ) (Q ) (wGM V P − r ) In Section 4.1, both tests are compared with each other. Additionally, we include the test presented by Glombeck (2014, Theorem 10) in our comparison study. 4.1 Design of the Comparison Study Let Σ be a p×p positive definite covariance matrix of asset returns, n the number of samples, and p ≡ p(n) . The structure of the covariance matrix is chosen in the following way: oneninth of the eigenvalues are set equal to 2 , four-ninths are set equal to 5 , and the rest are set equal to 10 . In doing so, we can ensure that the eigenvalues are not very dispersed, and if p increases, then the spectrum of the covariance matrix does not change its behaviour. Then, the covariance matrix is determined as follows 0 Σ = ΘΛΘ , where Λ is the diagonal matrix whose diagonal elements are the predefined eigenvalues and Θ is the p × p matrix of eigenvectors obtained from the spectral decomposition of a standard Wishart-distributed random matrix. We consider the following scenario for modelling the changes. Under the alternative hypothesis, the covariance matrix is defined by Σ1 = ∆Σ∆, (20) where ∆ denotes the change in the standard deviation and is given by ∆= 0 Dm 0 Ip−m ! , (21) with Dm = diag(a) and a = 1 + 0.1k , k ∈ {1, 2, . . . , 15} , and m ∈ {0.1p, 0.2p, 0.5p, 0.8p} . In order to demonstrate the influence of ∆ , we build the `1 norm of the difference between wGM V P (Σ1 ) and wGM V P (Σ) as a function of a. This difference relates to the proportional transaction costs for moving to the new optimal portfolio weights wGM V P (Σ1 ) . In Figure 3, we can see that the largest influence on the portfolio composition is observed if m = 0.5p . For m larger than 0.5p , this influence decreases. The results obtained for this scenario present an almost linear relationship between the `1 norm of the difference vector and the size of the change. It is worth mentioning that the differences are all zero when no changes occur (i.e. under H0 ). 11 12 l 1 − difference l 1 − difference 0.0 0.1 0.2 0.3 0.4 0.5 0.00 0.05 0.10 0.15 0.20 0.25 0.30 1.1 1.2 1.0 1.1 1.2 1.20 1.22 1.24 1.26 1.28 1.30 0.18 0.20 0.22 0.24 0.26 1.3 Change, a 1.4 1.4 Figure 3: `1 - differences Change, a 1.3 50 % changes on the main diagonal 1.0 0.10 0.09 0.08 0.07 0.06 0.05 1.20 1.22 1.24 1.26 1.28 1.30 10 % changes on the main diagonal i=1 Pp 1.6 c = 0.1 c = 0.3 c = 0.5 c = 0.7 c = 0.9 1.6 0.0 0.1 0.2 0.3 0.4 0.5 0.00 0.05 0.10 0.15 0.20 0.25 0.30 1.1 1.2 1.0 1.1 1.2 0.20 0.18 0.16 0.14 0.12 0.10 1.20 1.22 1.24 1.26 1.28 1.30 1.3 Change, a Change, a 1.3 80 % changes on the main diagonal 1.0 0.16 0.15 0.14 0.13 0.12 0.11 0.10 1.20 1.22 1.24 1.26 1.28 1.30 20 % changes on the main diagonal wi,GM V P (Σ1 ) − wi,GM V P (Σ) as a function of change, a . Here, n = 500 . 1.5 1.5 c = 0.1 c = 0.3 c = 0.5 c = 0.7 c = 0.9 l 1 − difference l 1 − difference 1.4 1.4 1.5 1.5 1.6 c = 0.1 c = 0.3 c = 0.5 c = 0.7 c = 0.9 1.6 c = 0.1 c = 0.3 c = 0.5 c = 0.7 c = 0.9 4.2 Comparison of the Tests In this section, we present the results of a simulation study to compare the powers of the three tests. The simulation is based on 105 independent realizations of ∆ . The significance level α is chosen to be 5% , and the concentration ratio c takes a value within the set {0.1, 0.5, 0.7, 0.9} . In order to illustrate the performance of the tests based on the shrinkage approach, the test based on the statistic of Bodnar and Schmid (2008), and the test proposed by Glombeck (2014), the empirical power functions for the general hypothesis are evaluated for m = 0.2p (Figure 4a) and m = 0.5p (Figure 4b). 0.9 0.8 0.7 0.6 0.2 0.3 0.4 0.5 Power 0.5 0.2 0.3 0.4 Power 0.6 0.7 0.8 0.9 1.0 c = 0.5 1.0 c = 0.1 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0 2.1 2.2 2.3 2.4 0.1 Shrinkage Glombeck Bodnar and Schmid 0.0 0.0 0.1 Shrinkage Glombeck Bodnar and Schmid 2.5 1.0 1.1 1.2 1.3 1.4 1.5 1.6 Change, a 1.8 1.9 2.0 2.1 2.2 2.3 2.4 2.5 Change, a 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.2 0.3 0.4 0.5 Power 0.6 0.7 0.8 0.9 1.0 c = 0.9 1.0 c = 0.7 Power 1.7 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0 2.1 2.2 2.3 2.4 0.1 Shrinkage Glombeck Bodnar and Schmid 0.0 0.0 0.1 Shrinkage Glombeck Bodnar and Schmid 2.5 1.0 Change, a 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0 2.1 2.2 2.3 2.4 2.5 Change, a Figure 4a: Empirical power functions of the three tests for different values of c and 20% changes on the main diagonal according to scenario given in (20) and n = 500 . In Figure 4a, where 20% of the elements in the main diagonal of the covariance matrix are contaminated, we observe a slow increase of the power functions for c = 0.9 and better behaviour for smaller values of c . In the case c = 0.1 , there is no significant difference in the performance of the tests. For c = 0.5 and c = 0.7 , the power curves of Glombeck’s test and the test of Bodnar and Schmid (2008) are very close to each other, whereas the test presented 13 by Glombeck (2014) shows a slightly better performance than the one presented by Bodnar and Schmid (2008) when c = 0.9 . For c = {0.5, 0.7, 0.9} , the test based on the shrinkage approach outperforms its competitors. Figure 4b illustrates the behaviour of the tests for m = 0.5p . We detect an improvement in the performances of the tests for all values of c . In the case c = 0.1 , the test of Bodnar and Schmid (2008) outperforms both the shrinkage approach and Glombeck’s test, whereas, for c = 0.9 , this test appears to be the worst one. For c = 0.5 and c = 0.7 , the same situation occurs as when m = 0.2p , where the powers of Glombeck’s test and the test of Bodnar and Schmid (2008) almost coincide with each other. The test based on the estimated shrinkage intensity lies above the rest of the competitors for c = 0.5 , c = 0.7 , and c = 0.9 . 0.9 0.8 0.7 0.6 0.2 0.3 0.4 0.5 Power 0.5 0.2 0.3 0.4 Power 0.6 0.7 0.8 0.9 1.0 c = 0.5 1.0 c = 0.1 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0 2.1 2.2 2.3 2.4 0.1 Shrinkage Glombeck Bodnar and Schmid 0.0 0.0 0.1 Shrinkage Glombeck Bodnar and Schmid 2.5 1.0 1.1 1.2 1.3 1.4 1.5 1.6 Change, a 1.8 1.9 2.0 2.1 2.2 2.3 2.4 2.5 Change, a 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.2 0.3 0.4 0.5 Power 0.6 0.7 0.8 0.9 1.0 c = 0.9 1.0 c = 0.7 Power 1.7 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0 2.1 2.2 2.3 2.4 0.1 Shrinkage Glombeck Bodnar and Schmid 0.0 0.0 0.1 Shrinkage Glombeck Bodnar and Schmid 2.5 1.0 Change, a 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0 2.1 2.2 2.3 2.4 2.5 Change, a Figure 4b: Empirical power functions of the three tests for different values of c and 50% changes on the main diagonal according to the scenario given in (20) and n = 500 . 5 Summary The main focus of this study is the inference of the GMVP weights. After constructing an optimal portfolio, an investor is interested to know whether or not the weights of the portfolio he 14 is holding are still optimal at a fixed time point. For that reason, we investigate several asymptotic and exact statistical procedures for detecting deviations in the weights of the GMVP. One test is based on the sample estimator of the GMVP weights, whereas another uses its shrinkage estimator. To the best of our knowledge, the shrinkage approach, which is very popular in point estimation, is applied in test theory for the first time. The asymptotic distributions of both test statistics are obtained under the null and alternative hypotheses in a high-dimensional setting. This finding is a great advantage with respect to other approaches that appear in the literature which do not elaborate on the distribution under the alternative hypothesis (e.g. Glombeck (2014)). In order to compare the performances of the proposed procedures, the empirical power functions of the derived tests are determined. It is shown that the test based on the shrinkage approach performs uniformly better than the other tests considered in the analysis for moderate and large values of the concentration ratio c . This approach seems to be very promising for testing portfolio weights in a high-dimensional situation. Acknowledgement This research was partly supported by the German Science Foundation (DFG) via the projects BO 3521/3-1 and SCHM 859/13-1 ‘Bayesian Estimation of the Multi-Period Optimal Portfolio Weights and Risk Measures’. References Abramowitz, M. and Stegun, I. A. (1964). Pocketbook of Mathematical Functions, Verlag Harri Deutsch, Frankfurt(Main). Bai, J. and Shi, S. (2011). Estimating high dimensional covariance matrices and its applications, Annals of Economics and Finance 12: 199–215. Bai, Z. and Silverstein, J. W. (2010). Spectral Analysis of Large Dimensional Random Matrices, Springer, New York. Bodnar, T., Gupta, A. K. and Parolya, N. (2014). On the strong convergence of the optimal linear shrinkage estimator for large dimensional covariance matrix, Journal of Multivariate Analysis 132: 215–228. Bodnar, T., Gupta, A. K. and Parolya, N. (2016). Direct shrinkage estimation of large dimensional precision matrix, Journal of Multivariate Analysis 146: 223–236. Bodnar, T., Okhrin, Y. and Parolya, N. (2017). Optimal shrinkage-based portfolio selection in high dimensions, arXiv:1611.01958 . URL: http://adsabs.harvard.edu/abs/2016arXiv161101958B 15 Bodnar, T., Parolya, N. and Schmid, W. (2017). Estimation of the global minimum variance portfolio in high dimensions, European Journal of Operational Research, in press . URL: https://doi.org/10.1016/j.ejor.2017.09.028 Bodnar, T. and Schmid, W. (2008). A test for the weights of the global minimum variance portfolio in an elliptical model, Metrika 67: 127–143. Britten-Jones, M. (1999). The sampling error in estimates of mean-variance efficient portfolio weights, The Journal of Finance 54: 655–671. Bühlmann, P. and Van De Geer, S. (2011). Statistics for High-Dimensional Data: Methods, Theory and Applications, Springer, Berlin, Heidelberg. Cai, T. and Shen, X. (2011). High-Dimensional Data Analysis, World Scientific, Singapore. Cam, L. and Yang, G. (2000). Asymptotics in Statistics: Some Basic Concepts, Springer, New York. DasGupta, A. (2008). Asymptotic Theory of Statistics and Probability, Springer, New York. Efron, B. and Morris, C. (1976). Families of minimax estimators of the mean of a multivariate normal distribution, Annals of Statistics 4: 11–21. Frahm, G. and Jaekel, U. (2008). Tyler’s m-estimator, random matrix theory, and generalized elliptical distributions with applications to finance, SSRN Electronic Journal . URL: https://ssrn.com/abstract=1287683 Frahm, G. and Memmel, C. (2010). Dominating estimators for minimum-variance portfolios, Journal of Econometrics 159: 289–302. Gibbons, M. R., Ross, S. A. and Shanken, J. (1989). A test of the efficiency of a given portfolio, Econometrica 57: 1121–1152. Glombeck, K. (2014). Statistical inference for high-dimensional global minimum variance portfolios, Scandinavian Journal of Statistics 41: 845–865. Golosnoy, V. and Okhrin, Y. (2007). Multivariate shrinkage for optimal portfolio weights, The European Journal of Finance 13: 441–458. Jobson, J. and Korkie, B. M. (1981). Performance hypothesis testing with the Sharpe and Treynor measures, The Journal of Finance 36: 889–908. Jorion, P. (1986). Bayes-Stein estimation for portfolio analysis, Journal of Financial and Quantitative Analysis 21: 279–292. Laloux, L., Cizeau, P., Potters, M. and Bouchaud, J.-P. (2000). Random matrix theory and financial correlations, International Journal of Theoretical and Applied Finance 3: 391–397. 16 Markowitz, H. (1952). Portfolio selection, The Journal of Finance 7: 77–91. Memmel, C. and Kempf, A. (2006). Estimating the global minimum variance portfolio, Schmalenbach Business Review 58: 332–348. Muirhead, R. J. (1982). Aspects of Multivariate Statistical Theory, Wiley, New York. Okhrin, Y. and Schmid, W. (2006). Distributional properties of portfolio weights, Journal of Econometrics 134: 235–256. Okhrin, Y. and Schmid, W. (2007). Comparison of different estimation techniques for portfolio selection, AStA Advances in Statistical Analysis 91: 109–127. Okhrin, Y. and Schmid, W. (2008). Estimation of optimal portfolio weights, International Journal of Theoretical and Applied Finance 11: 249–276. Stein, C. (1956). Inadmissibility of the usual estimator for the mean of a multivariate normal distribution, Proceedings of the Third Berkeley Symposium on Mathematical Statistics and Probability, Volume 1: Contributions to the Theory of Statistics, University of California Press, Berkeley, California: 197–206. Appendix In this section, the proof of Theorems 1 and 2 are given. Proof of Theorem 1: d We first derive a stochastic representation for Tn . Let the symbol = denote equality in distribution. Then, it holds that (see, the proof of Theorem 2 in Bodnar and Schmid (2008)) d Tn = n − p ξ1 , p − 1 ξ2 (22) where ξ1 |Q̂∗n 10 Σ−1 1 ∼ χ2n−p , 10 Σ̂−1 1  = (n − 1) 10 Σ−1 1 (ŵn∗ − r∗ )0 (Q̂∗n )−1 (ŵn∗ − r∗ ) ∼ χ2p−1,λn (Q̂∗ ) , ξ2 = (n − 1) (23) n with  ∗ ∗ 0 ∗ −1 ∗ ∗ λn (Q̂∗n ) = (n − 1) 10 Σ−1 1 (wGM (wGM V P − r ) (Q̂n ) VP − r ), and ξ2 is independent of both Q̂∗n and (Q̂∗n )−1 (ŵn∗ − r∗ ) . Moreover, in using (n − 1)(Q̂∗n )−1 ∼ Wp (n − 1, (Q∗ )−1 ) (cf. Muirhead (1982, Theorems 3.2.10 and 3.2.11)), we obtain λn (Q̂∗n ) ∗ ∗ 0 ∗ −1 ∗ ∗ (n − 1) (wGM (wGM V P − r ) (Q̂n ) VP − r ) = λn ∼ λn ξ3 , 0 ∗ ∗ ∗ −1 (w∗ ∗ (wGM V P − r ) (Q ) GM V P − r ) 17 (24) where ξ3 ∼ χ2n−1 . The last equality shows that the conditional distribution of ξ1 given Q̂∗n depends only on Q̂∗n over ξ3 , and, consequently, the conditional distribution ξ1 |Q̂∗n coincides with ξ1 |ξ3 . Using the distributional properties of the non-central F -distribution, we obtain the following stochastic representation for ξ1 given by p d ξ1 = ( λn ξ3 + ω1 )2 + ξ4 , and, hence, the stochastic representation of Tn is expressed as √ n − p ( λn ξ3 + ω1 )2 + ξ4 , Tn = p−1 ξ2 d (25) where ω1 ∼ N (0, 1) , ξ2 ∼ χ2n−p , ξ3 ∼ χ2n−1 , and ξ4 ∼ χ2p−2 ; ω1 , ξ2 , ξ3 , and ξ4 are independent. Applying (25), we obtain √    λn ξ3 + 2 λn ξ3 ω1 + ω12 + ξ4 n−1 ξ2 − 1 + λn p−1 p−1 n−p     p n−p n − 1p ξ3 ξ4 = −1 + p−1 −1 λn p−1 ξ2 p−1 n−1 p−1 s !     p n−1 p ξ3 ω12 ξ2 − 1 + λn − 1 + 2 λn ω1 + √ . p−1 p−1 n−p p−1 p−1 n − pp p−1 ξ2  Using the asymptotic properties of a χ2 -distribution with infinite degrees of freedom and the independence of ω1 , ξ2 , ξ3 , ξ4 , the application of Slutsky’s lemma (see, for example, Theorem 1.5 in DasGupta (2008)) leads to p p−1 Tn − 1 − λn n−1 p−1 Cn ! where Cn2 d → N (0, 1) , λ2 λn c =2+2 n +4 +2 c c 1−c  2 λn 1+ . c 2 Proof of Theorem 2: In order to stress the dependence on n , we use the notation Σn in the following. a) Using Proposition 3 of Glombeck (2014), we have  b0n Σ̂n bn " ! !# 1 √  b0 Σn bn − 1  d 1 − 0 1−c n →N n , ,2 1 1  10 Σ̂−1 1  1 − 0 3 n 1−c (1−c) − 10 Σ−1 1 − cn n 1  18 (26) if n → ∞ . We can rewrite R̂bn as b0n Σ̂n bn 10 Σ̂−1 n 1 0 b Σn bn 10 Σ−1 n 1−1 0 0 −1 bn Σn bn 1 Σn 1 n Dn = ∆n (1 − cn )(Dn En + + En ) + ∆ n − 1 1 − cn R̂bn = (1 − cn ) with ∆n = b0n Σn bn 10 Σ−1 n 1, Dn = b0n Σ̂n bn − 1, b0n Σn bn En = 10 Σ̂−1 1 n 1 − . 10 Σ−1 1 − cn n 1 Using (26), it follows that √ R̂bn − ∆n + 1 √ n = n(1 − cn )(Dn En + Dn /(1 − cn ) + En ) ∆n √ = (1 − cn ) n(Dn /(1 − cn ) + En ) + op (1) !   √ Dn c d = (1 1 − cn ) n + op (1) → N 0, 2 1−c En √ √ d d 2 since nDn → N (0, 2) and nEn → N (0, (1−c) 3). b) Since ˆn = α̃ (1 − cn ) cn ∆n  R̂bn −∆n +1 ∆n + (1 − cn )  + R̂bn −∆n +1 ∆n ∆n −1 ∆n +  ∆n −1 ∆n , it follows that  √ ˆn − n α̃ (1 − cn )(∆n − 1) cn + (1 − cn )(∆n − 1)  = In + IIn with In = √ n = √ n = √ n   R̂ −∆ +1 (1 − cn ) bn ∆n n   R̂bn −∆n +1 cn ∆n −1 + (1 − c ) + n ∆n ∆n ∆n   R̂ −∆ +1 cn (1 − cn ) bn ∆n n + (1 − cn ) ∆∆n −1 ∆n n   cn ∆n −1 R̂bn −∆n +1 c ∆n −1 n + (1 − c ) n ∆n + (1 − c ) + ∆n n ∆n ∆n ∆n   R̂ −∆ +1 (1 − cn ) bn ∆n n 1 √ R̂bn −∆n +1 . 1−2cn 1−cn 1 − cn − ∆ n √ 1+ n 1−2c n(1−cn − 19 ∆n n) ∆n Furthermore, IIn  √ 1  = ) n(1 − cn )(1 − ∆n = √ n  1 cn ∆n + (1 − cn )  1 − cn 1  )  1−2cn (1 − ∆n 1 − cn − ∆ n 1+  R̂bn −∆n +1 ∆n + ∆n −1 ∆n − cn ∆n 1  + (1 − cn ) ∆∆n −1 n  1 1−cn n 1−cn − 1−2c ∆ n R̂bn −∆n +1 ∆n  − 1 √ R̂bn −∆n +1 n (1 − cn )2 1 ∆n = − ) . 1−2cn 2 (1 − ∆n 1 + √ 1−cn1−2c √n R̂bn −∆n +1 (1 − cn − ∆n ) n) ∆n n(1−c − n ∆n Consequently, In + IIn = √ R̂bn − ∆n + 1 cn (1 − cn ) 1 n 1−2cn 2 ∆n (1 − cn − ∆n ) ∆n 1 + 1 √ √ R̂bn −∆n +1 1−cn n n) ∆n n(1−cn − 1−2c ∆ n √ R̂bn − ∆n + 1 cn (1 − cn )∆n n (1 + op (1)) ∆n (cn + (∆n − 1)(1 − cn ))2   d c3n (1 − cn )∆2n ≈ N 0, 2 (cn + (∆n − 1)(1 − cn ))4 = if √ n(1 − cn − 1−2cn ) ∆n → ∞ as n → ∞ . Since b0n Σn bn ≥ min w0 Σn w = w that ∆n ≥ 1 , and, thus, this condition is fulfilled. 20 1 10n Σ−1 n 1n , it holds 2
10math.ST
EMBEDDINGS OF SPHERICAL HOMOGENEOUS SPACES IN CHARACTERISTIC p arXiv:1603.04705v2 [math.AG] 17 Feb 2017 RUDOLF TANGE Summary. Let G be a reductive group over an algebraically closed field of characteristic p > 0. We study properties of embeddings of spherical homogeneous G-spaces. We look at Frobenius splittings, canonical or by a (p −1)th power, compatible with certain subvarieties. We show the existence of rational G-equivariant resolutions by toroidal embeddings, and give results about cohomology vanishing and surjectivity of restriction maps of global sections of line bundles. We show that the class of homogeneous spaces for which our results hold contains the symmetric homogeneous spaces in characteristic 6= 2 and is closed under parabolic induction. Introduction Let k be an algebraically closed field of prime characteristic p and let G be a connected reductive group over k. Let H be a closed k-subgroup scheme of G such that the homogeneous space G/H (see [10, III.3.5.4]) is spherical. We are interested in properties of embeddings (always normal and equivariant) of G/H: existence of equivariant “rational” resolutions by toroidal embeddings, (canonical) Frobenius splittings, compatible splittings, cohomology vanishing, surjectivity of restriction maps of global sections of line bundles. The idea is to show that there is a class of spherical homogeneous spaces whose embeddings have nice properties, which contains symmetric homogeneous spaces in characteristic 6= 2 and the reductive group embeddings in any characteristic, and is closed under parabolic induction. Previously, special cases were studied in [27] (the wonderful compactification of an adjoint group), [5] (p large), [9] (the wonderful compactification of an “adjoint” symmetric homogeneous space), [25] (arbitrary embeddings of reductive groups), [30] (embeddings of homogeneous spaces induced from reductive groups). We describe some of the issues that arise in this general approach. Let x ∈ G/H be the image of the neutral element e under the canonical map G → G/H, let B be a Borel subgroup of G such that the orbit B ·x is open in G/H and let T be a maximal torus of B. In [30] it was shown that, after reducing to the toroidal case, one has a B-canonical splitting by a (p − 1)-th power which compatibly splits the B-stable prime divisors, and this implies certain cohomology vanishing results. In more general situations, e.g. symmetric homogeneous spaces, this need no longer be true. Firstly, one cannot construct a B-canonical splitting for the wonderful compactification which is a (p − 1)-th power, since the closed Key words and phrases. equivariant embedding, spherical homogeneous space, Frobenius splitting. 2010 Mathematics Subject Classification. 14L30, 14M27, 57S25, 20M32. 1 2 R. H. TANGE orbit is G/P for some P which is not necessarily a Borel subgroup, and therefore its B-canonical splitting is in general not a (p − 1)-th power. Since it seems necessary to have a splitting by a (p − 1)-th power to get cohomology vanishing results using the Mehta-Van der Kallen theorem [6, Thm. 1.2.12, 1.3.14], and B-canonical splitting are of interest for results about good filtrations we use both types of splittings. Furthermore, one cannot expect a splitting which compatibly splits all B-stable prime divisors, since their schematic intersections need not be reduced, see [3, p266]. Finally, one cannot expect the “usual” cohomology vanishing results for arbitrary spherical varieties in characteristic p, see [14, Example 3] for an example of an ample line bundle on a projective spherical variety with nonzero H 1 . So we need to make some assumptions on the homogeneous space. The paper is organised as follows. In Section 1 we introduce the notation, state the basic assumptions (A1)-(A4), and prove some technical lemmas. In Section 2 we prove our three main results, each result requires one of two extra assumptions (B1) and (B2). As it turns out, (B2) implies (B1), see Remark 2.1.2. In Proposition 2.1 we show that, assuming (B1), every smooth toroidal embedding is split by a (p − 1)-th power which compatibly splits the G-stable closed subvarieties. Theorem 2.1 states that, assuming (B1), every G/H-embedding has an equivariant rational resolution by a toroidal embedding. This then leads to results about Frobenius splitting, cohomology vanishing, and surjectivity of restriction maps of global sections of line bundles. The proofs of Proposition 2.1 and Theorem 2.1 use ideas from [27], [25], [6] together with the fact that G/P is split by a (p − 1)-th power, Lemma 1.1, and the fact that we have a certain expression for the canonical divisor which contains every B-stable prime divisor with strictly negative coefficient, Lemma 1.4. Theorem 2.2 states that, assuming (B2), every G/H-embedding is B-canonically split compatible with all G-stable closed subvarieties. In Section 3 we show that the class of homogeneous spaces which satisfy (A1)-(A4) and (B2) contains the symmetric spaces. In Section 4 we show that this class is stable under parabolic induction. 1. Preliminaries We retain the notation k, p, G, H, x, B, T from the introduction. Recall that the canonical morphism G/Hred → G/H, Hred = H(k) the reduced scheme associated to H, is a homeomorphism. See [10, III §3]. We will usually deal with varieties, unless we explicitly use the term “scheme”. Schemes are always supposed to be algebraic over k. We will use ordinary notation for notions like inverse image and intersection, and special notation for their schematic analogues. 1.1. Frobenius splitting. We give a very brief sketch of the basics of Frobenius splitting, introduced by Mehta and Ramanathan in [23]. For precise statements we refer to [6]. We start with the affine case. For a commutative ring A of characteristic p put Ap = {ap | a ∈ A}, a subring of A. First assume A is reduced. Then one can define a Frobenius splitting of A as an Ap -module direct complement to Ap in A. For example for a polynomial ring A = k[x1 , . . . , xn ], a Frobenius splitting is given by the Ap -span of the restricted monomials (i.e. EMBEDDINGS OF SPHERICAL HOMOGENEOUS SPACES IN CHARACTERISTIC p 3 exponents < p) different from 1. Of course a Frobenius splitting of A can also be seen as an Ap -linear projection from A onto Ap , i.e. as an Ap -linear map σ : A → Ap with σ(1) = 1. This is the point of view of [22]. If we “take the p-th root from this definition” we obtain the usual definition: a Frobenius splitting of A is a homomorphism of abelian groups σ : A → A with σ(ap b) = aσ(b) for all a, b ∈ A and σ(1) = 1. This definition also works for non-reduced rings, and then it follows immediately that a Frobenius split ring is always reduced. Finally one can “sheafify” this definition to obtain the definition of Frobenius splitting for any scheme X (separated and of finite type) over k: Let F : X → X be the absolute Frobenius morphism, i.e. the morphism of schemes (not over k) which is the identity on the underlying topological space and for which F # : OX → OX is the p-th power map. Then a Frobenius splitting of X is a morphism of OX -modules σ : F∗ (OX ) → OX which maps 1 to 1 (in OX (X) and therefore in any OX (U )). We can also consider F # as a morphism of OX modules : OX → F∗ (OX ) and then a Frobenius splitting of X is a left-inverse of the OX -module morphism F # . If σ is a Frobenius splitting of X, then a closed subscheme of X is called compatibly split (with σ) if its ideal sheaf is stable under σ. A compatibly split subscheme is itself Frobenius split. From the existence of a Frobenius splitting one can deduce cohomology vanishing results. For example, if X is a Frobenius split variety which is proper over an affine, and L is an ample line bundle on X, then H i (X, L) = 0 for all i > 0. See [6, Thm. 1.2.8]. This is based on the fact that, in the presence of a r Frobenius splitting, one can embed H i (X, L) in H i (X, Lp ) for all r ≥ 1. Using compatible splittings or splitting relative to a divisor one can obtain more refined results, see [6, Lem. 1.4.7, 1.4.11]. An important result about Frobenius splittings of a smooth variety X is that they correspond to certain nonzero global sections of the (1 − p)-th power of the canonical line bundle ωX , see [6, Thm. 1.3.8]. This result was first proved 1−p for X smooth and projective in [23, Sect. 2]. Since H 0 (X, ωX ) may be zero, this also shows that, not every smooth variety is Frobenius split. If G acts on a variety X, then it also acts on HomOX (F∗ (OX ), OX ) and turns this vector space into a rational G-module, see [6, 4.1.5]. If X is smooth then 1−p the same holds for H 0 (X, ωX ). Now one can define a Frobenius splitting to (n) be B-canonical if it is T -fixed and is killed by all divided powers eα ∈ Dist(G), α simple relative to B, n ≥ p. Here Dist(G) is the distribution algebra or hyper algebra of G. The notion of canonical splitting was introduced by Mathieu who proved the following result. If L is a G-linearised line bundle on a B-canonically split G-variety X, then the G-module H 0 (X, L) has a good filtration. See [22, Sect. 5,6] and [6, Ch. 4] for more detail. The notion of good filtration can be found in [16], [22] or [6]. We finish our discussion of Frobenius splittings, with a lemma about splittings of G/P . The lemma seems to be known, but we provide a proof for lack of reference. It will be needed in the proof of Proposition 2.1. For a parabolic subgroup P containing B we denote by ρP the half sum of the roots of T in the unipotent radical Ru P and we put ρ = ρB . We denote the opposite parabolic by P − . For λ a character of P − (viewed as a character of T orthogonal to 4 R. H. TANGE the simple coroots of the Levi of P containing T ) we denote the line bundle − G ×P kλ on G/P − by L(λ). Lemma 1.1. Let P be a parabolic subgroup of G. Then G/P is split by a (p − 1)-th power. Proof. We prove the claim for G/P − . It is well known and easy to see that ρP is dominant, see e.g. [6, Sect 3.1] (ρP is there denoted δP ). So by [6, Thm 3.1.2(c)] the multiplication map H 0 (G/P − , L(2ρP ))⊗(p−1) → H 0 (G/P − , L(2(p − 1)ρP )) is surjective. Now L(−2ρP ) is the canonical bundle of G/P − , so the assertion follows from [6, Ex. 1.3.E(2)].  1.2. Spherical embeddings. We recall some facts from the theory of spherical embeddings. For more detail we refer to [4], [19], and [31]. A G/H-embedding is a normal G-variety X with a G-equivariant embedding G/H ֒→ X of which the image is open and dense. The set of T -weights of the nonzero B-semi-invariant rational functions on G/H is denoted by Λ = ΛG/H . Since B · x is open in G/H every B-semi-invariant rational function on G/H is defined at x, and completely determined by its T -weight and its value at x. So k(G/H)(B) /k× ∼ = Λ, where k(G/H)(B) is the group of B-semi-invariant rational functions and k× is the multiplicative group of k. Put ΛQ = Q ⊗Z Λ. Then HomZ (Λ, Q) = Λ∗Q . Any (discrete) valuation of the function field k(G/H) determines a Z-linear function k(G/H)(B) /k× → Q, i.e. an element of Λ∗Q . In particular, any prime divisor (irreducible closed subvariety of codimension 1) of any G/H-embedding determines an element of Λ∗Q . A G-invariant valuation is completely determined by its image in Λ∗Q , so we can identify the set of G-invariant valuations of G/H with a subset V = V(G/H) of Λ∗Q . It is well-known that V is a finitely generated convex cone which spans Λ∗Q . If X is a G/H-embedding, then the complement of the open B-orbit has the B-stable prime divisors as its irreducible components. These we divide in two classes: The ones that are G-stable are called boundary divisors and will be denoted by Xi . The others are the B-stable prime divisors that intersect G/H. These will be denoted by D, possibly with a subscript. A G/H-embedding is called simple if it has a unique closed G-orbit. The coloured cone of a simple embedding X is the pair (C, F) where F, the set of colours of X, is the set of B-stable prime divisors of G/H whose closure in X contains the closed orbit of X and C ⊆ Λ∗Q is the cone generated by the images of all boundary divisors of X and the image of F. A simple embedding is completely determined by its coloured cone. For an arbitrary embedding each orbit is the closed orbit of a unique simple open sub-embedding. The coloured cones of these sub-embeddings form the coloured fan of the embedding. An embedding is completely determined by its coloured fan. More generally, coloured fans can be used to describe morphisms between embeddings of different homogeneous spaces, see [19, Sect. 4]. A G/H-embedding is called toroidal if no B-stable prime divisor of X which intersects G/H (i.e. which is not G-stable) contains a G-orbit. The toroidal embeddings are the embeddings corresponding to the coloured fans without colours, that is, the fans (in the sense of toric geometry) whose support is EMBEDDINGS OF SPHERICAL HOMOGENEOUS SPACES IN CHARACTERISTIC p 5 contained in V(G/H). For a toroidal G/H-embedding X we denote by X0 the complement in X of the union of the B-stable prime divisors that intersect G/H. Note that every G-orbit in X intersects X0 . We denote the closure of X T · x in X0 by T · x 0 . We will say that the local structure theorem holds for X (relative to x, T and B) if there exists a parabolic subgroup P of G containing B such that, with M the X Levi subgroup of P containing T , the variety T · x 0 is M -stable, the derived group DM acts trivially on it, and the action of G induces an isomorphism X ∼ Ru P × T · x 0 → X0 . The local structure theorem for X implies the local structure theorem for G/H itself. Furthermore, the parabolic subgroup P only depends on G/H: it can be characterised as the stabiliser of the open B-orbit B ·x or as the stabiliser of the open subset BHred of G (for the action by left multiplication). Whenever M is a closed k-subgroup scheme of G and L a Levi subgroup of G containing T we denote by ML the schematic intersection of L and M . When the local structure theorem holds for G/H, then every rational function on T · x = T /HT has a unique U = Ru B-invariant extension to G/H. So Λ = X (T /HT ), the character group of T /HT , and Λ∗Q ∼ = Q ⊗Z Y(T /HT ), where Y(T /HT ) is the cocharacter group of T /HT .1 Lemma 1.2. Let X be a toroidal G/H-embedding for which the local structure theorem holds. X X (i) Every G-orbit intersects T · x 0 in a T -orbit, and T · x 0 is a T /HT embedding whose fan is the same as that of X. Furthermore, all G-orbit closures of X are normal and the local structure theorem also holds for them (with the same T, B, P ). (ii) If X is complete, then every closed G-orbit is isomorphic to G/P − . Proof. (i). The first assertion follows by the same arguments as in the proof of [31, Thm. 29.1] or in [4, Sect. 2.4]. The normality of the G-orbit closures follows from the toric case [18, Prop. I.2] and the final assertion is obvious. (ii). Let Y be a closed G-orbit in X and let X ′ be the corresponding simple open sub-embedding of X. Then Y is complete. So the cone in Λ∗Q corresponding to X ′ has dimension dim Λ∗Q = rkX (T /HT ) = dim T /HT , by [19, Thm. 6.3]. X′ So T · x 0 has a unique T -fixed point y which must be the intersection of Y X′ with T · x 0 . Let P0 ≤ G be the stabiliser of y, then we have a bijective Gequivariant morphism ϕ : G/P0 → G · y = Y . By [1, Prop. AG.18.3(1)], G/P0 is complete, i.e. P0 is a parabolic. Since P0 contains T , we have that for every root of T in G, at least one of the root subgroups Uα , U−α is contained in P0 . It is then obvious from the local structure theorem for Y , that P0 = P − and that ϕ is an isomorphism.  A wonderful compactification of G/H is a smooth simple complete toroidal G/H-embedding. Since it is is simple and toroidal it is determined by a cone 1 In [30, Rem. 1.2], where this is also stated, it should have been assumed that the local structure theorem holds for G/H. 6 R. H. TANGE in V. Since it is also complete the cone has to be all of V, see [19, Thm. 4.2]. So a wonderful compactification is unique if it exists. Furthermore, since cones in coloured fans of G/H-embeddings are always strictly convex, V contains no line. If X is a wonderful compactification of G/H for which the local structure X theorem holds, then T · x 0 (and therefore X0 ) is an affine space. We will also need the following property of a smooth G/H-embedding X: −1 has a nonzero B-semi-invariant global section The anti-canonical bundle ωX of weight 2ρP with divisor X X aD D + Xi , (∗) D i where the first sum is over the B-stable prime divisors of X that are not Gstable with the aD strictly positive and the second sum is over the boundary divisors of X. If X is a toroidal smooth G/H-embedding which satisfies the local structure −1 has a nonzero B-semi-invariant global theorem, then it is easy to see that ωX section of weight 2ρP , unique up to scalar multiples, and that its divisor is given by the formula above with the aD integers. So the crucial point of (*) is that the coefficients aD are strictly positive. Lemma 1.3 (cf. proof of [2, Prop. 4.1]). Let X be a smooth G/H-embedding. Assume the local structure theorem holds for the toroidal embedding obtained by removing all orbits of codimension > 1 from X, and assume the T -orbit map of x is separable, then X has property (*). Proof. Since the proof in [2] is very brief we give a bit more detail. We reduce as in [loc. cit.] to the case that X is smooth and toroidal by removing the orbits of X ∼ codimension > 1. Then we have the local structure theorem Ru P × T · x 0 → X0 . If an algebraic group K acts on X by automorphisms, then Lie(K) acts on OX by derivations, put differently, we have a Lie algebra homomorphism ϕ : Lie(K) → Vect(X). Furthermore, ϕ(u)y ∈ Ty (K · y) (†) for all u ∈ Lie(K) and y ∈ X and Ty (−) is the Zariski tangent space at y. We have dim(Ty (B · y)) = dim(B · y) < dim(B · x) = dim(Ru P ) + dim(T · x) for all y ∈ X \ B · x. Let θ be the wedge product of the ϕ-images of a basis of Lie(Ru P ) together with a lift to Lie(T ) of a basis of Lie(T ·x). Here the lift of the basis of Lie(T · x) to elements in Lie(T ) exists by the separability assumption. Clearly θ is B-semi-invariant of weight 2ρP . By (†) applied to K = B we have that θ vanishes on the complement of the open B-orbit B · x = Ru (P )T · x, i.e. on the union of B-stable prime divisors. Furthermore, it is well known from toric geometry, using the local structure theorem, that it vanishes to the order one on the boundary divisors and it is easy to see that it is nonzero on Ru (P )T P · x. So θ is Pa global section of the anti-canonical bundle with divisor of zeros D aD D + i Xi , aD > 0.  For anyPsmooth G/H-embedding X with boundary divisors Xi we put ω̃X = ωX ⊗OX ( i Xi ). Like ωX , the sheaf ω̃X comes with a canonical G-linearisation. EMBEDDINGS OF SPHERICAL HOMOGENEOUS SPACES IN CHARACTERISTIC p 7 When X is also toroidal ω̃X is the log P canonical sheaf. Note that if X has property (*), then we have ω̃X = OX (− D aD D). Let Ha (“a” for adjoint, see Section 3) be a closed subgroup scheme of G containing H and let xa ∈ G/Ha be the image of e under the canonical map G → G/Ha . From the next section on we will make the following assumptions; see Remark 1.1.1 for further comments. (A1) Ha is generated by H and a closed subgroup scheme of T that normalises H; (A2) the homogeneous space G/Ha has a wonderful compactification X for which the local structure theorem (see Section 1) holds; (A3) the wonderful compactification X of G/Ha has property (*) above; (A4) the characters through which H acts on the top exterior powers of Tx (G/H) and Txa (G/Ha ) are the same. From (A1) and (A2) it follows easily that the local structure theorem holds for all toroidal G/H-embeddings. One just has to use the fact that for every toroidal G/H-embedding X there is a unique G-equivariant morphism X → X which maps x to xa , see [19, Thm. 4.1]. This was also observed (in more specific situations) in [6, Prop. 6.2.3(i)] and [30, Prop. 3.2(i)]. Concerning (A4), to see that H acts on Tx (G/H) it is easiest to use the “functor of points” approach and dual numbers. For a commutative k-algebra R put R[δ] = R[t]/(t2 ), where t is an indeterminate and δ = t + (t2 ). Let X be a smooth scheme over k. Then we can define the “tangent bundle functor” T (X) : {commutative k-algebras} → {sets} as follows: T (X)(R) = X(R[δ]). Furthermore, we obtain a morphism : T (X) → X of functors given by the homomorphism R[δ] → R which maps δ to 0. Let x ∈ X(k). Then the fiber over x is Tx (X), see [10, Cor. II.4.3.3] or [1, AG.16.2]. If H acts on X, see [10, Sect. II.1.3] or [16, I.2.6], then it also acts on T (X) and the above morphism is H-equivariant. So if H fixes x, then H acts on Tx (X). This action is clearly linear, so H will then also act on the top exterior power of Tx (X), see [16, I.2.7(3)]. Note that if f : G → G′ is a central (cf. [1, V.2.2]) surjective morphism of reductive groups, H ′ ⊆ Ha′ closed subgroup schemes of G′ and H, Ha the schematic inverse image of H ′ , Ha′ under f , then G, H, Ha satisfy (A1)-(A4), if G′ , H ′ , Ha′ do (relative to B ′ = f (B), T ′ = f (T ) etc.). We denote the boundary divisors of X by X1 , . . . , Xr . Note that by the adjunction formula (see e.g. [15, Prop. II.8.20]) the restriction of ω̃X to G/P − is isomorphic to ωG/P − . Lemma 1.4. Assume (A1)-(A4) and let X be a smooth toroidal G/H-embedding. Consider the morphism π : X → X which extends the canonical map G/H → G/Ha Then ω̃X is G-equivariantly isomorphic to the pull-back π ∗ ω̃X . In particular, all smooth G/H-embeddings have property (*). Proof. Let s1 and s2 be the (unique up to scalar multiples) nonzero B-semiinvariant rational section of ω̃X resp. ω̃X of weight −2ρP . They are the nonzero B-semi-invariant rational sections of ωX and ωX multiplied with canonical sections of the boundary divisors of X resp. X. Let s′2 be the pull-back of s2 to 8 R. H. TANGE π ∗ ω̃X . First we show that the divisor (s1 ) equals (s′2 ). From the local structure theorems for X and X it is clear that (s1 ) and (s′2 ) do not involve the boundary divisors of X, so it is enough to show that (s1 ) and (s′2 ) are equal on G/H. Since the equivariant Picard group PicG (G) is trivial s1 and s′2 pull back to (B × H)-semi-invariant functions on G. Since they have the same B-weight −2ρP it is enough to show they have the same H-weight and this follows from (A4). Since (s1 ) equals (s′2 ), there is an isomorphism between ω̃X and π ∗ ω̃X such that s1 corresponds to s′2 . Since any two G-linearisations of a line bundle on X differ by a character of G and s1 and s′2 have the same weight it is clear that this isomorphism is G-equivariant. The final assertion follows by reducing to the case that X is (smooth and) toroidal by removing the orbits of codimension > 1.  Note that if we define a canonical divisor of a G/H-embedding X to be a divisor whose restriction to the smooth locus X ′ of X is a canonical divisor of X ′ , then, assuming (A1)-(A4), a canonical divisor of any G/H-embedding X is given by the formula in (*). Remarks 1.1. 1. Of the assumptions (A1)-(A4), (A2) is the most important one; it is for example not satisfied by the varieties of unseparated flags from [14]. It is essential that in the local structure theorem the isomorphism is given by the action of G. If H and Ha are reduced we can deduce (A4) as follows. We have canonical isomorphisms Tx (G/H) ∼ = Lie(G)/Lie(H) and Txa (G/H) ∼ = Lie(G)/Lie(Ha ) of H-modules and we obtain an exact sequence of H-modules 0 → Lie(Ha /H) → Tx (G/H) → Txa (G/Ha ) → 0 , where we used that, by (A1), H is normal in Ha . So the top exterior power of Tx (G/H) is the tensor product of the top exterior powers of Txa (G/Ha ) and Lie(Ha /H). The action of H on Lie(Ha /H) is trivial, since the H-action on Ha /H induced by the conjugation action of H on Ha is trivial. Thus (A4) follows. Now (A4) holds in characteristic 0 by Cartier’s Theorem [10, II.6.1.1]. I do not know of any examples where (A1) and (A2) are satisfied, but not (A3) or (A4). 2. In [2] (see also [21, Prop. 3.6]) the coefficients aD from (*) are determined and, at least in many special cases, this can also be done in prime characteristic. For example, when p > 2 and G/H is a symmetric space G/Gθ , G adjoint semisimple (see also Section 3), then this amounts to writing the canonical divisor plus the sum of the boundary divisors as a linear combination of the images in Pic(X) of the B-stable prime divisors of X that are not G-stable, where X is the wonderful compactification of G/H (these form basis of Pic(X)). Since Pic(X) embeds naturally in Pic(G/P − ), G/P − the closed orbit, via restriction (see [9]), and the line bundle corresponding to the above divisor restricts to the canonical bundle of Pic(G/P − ) this amounts to writing −2ρP as a linear combination of the above images. In the notation of [31, Prop 26.22] these images are: the ̟ ˆ i , i a ι-fixed node of the Satake diagram, and the ̟j + ̟ι(j) , j a ι-unstable node of the Satake diagram, except that we split the weights in the second list that correspond to the exceptional simple roots (see [9, Sect. 4]) EMBEDDINGS OF SPHERICAL HOMOGENEOUS SPACES IN CHARACTERISTIC p 9 into their two summands. 3. In characteristic 0 all orbit closures in G/H-embeddings are normal, see [7, Prop. 3.5]. The following example in characteristic p was mentioned to me by M. Brion. Put G = SL2 ×k× ×k× , V = k2 , and W = V ⊕V . Let G act on W by (g, λ, µ) · (v, w) = (λ g · v, µFr(g) · w), where Fr is the Frobenius morphism which raises all matrix entries to the power p. Now let h−, −i be the symplectic form on V given by hv, wi = v1 w2 − v2 w1 . This form is preserved by SL2 . Now define f ∈ k[W ] by f (v, w) = hFr(v), wi, where Fr is the Frobenius morphism of k 2 . Then W is spherical and the zero locus of f is a non-normal orbit closure, since its singular locus has codimension 1. See [30, Cor to Prop. 2.1] for a result that guarantees the normality of the G-orbit closures under a certain assumption. 2. Frobenius splittings and rational resolutions We retain the notation from the previous sections and we assume that properties (A1)-(A4) hold. We will consider the following further assumptions −1 −1 ) → H 0 (G/P − , ωG/P (B1) the restriction map H 0 (X, ω̃X − ) is surjective; 1−p 1−p ) → H 0 (G/P − , ωG/P (B2) the restriction map H 0 (X, ω̃X − ) is surjective and its kernel has a good filtration. Recall that a line bundle is called semi-ample if some positive power of it is generated by its global sections. Proposition 2.1. Assume (B1) and let X be a smooth toroidal G/H-embedding. Then (i) X is Frobenius split by a (p − 1)-th power which compatibly splits all Gstable closed subvarieties. (ii) Assume X is projective over an affine, let L be a semi-ample line bundle on X and let Y be a G-stable closed subvariety of X. (a) H i (Y, L) = 0 for all i ≥ 1. (b) If X is simple or Y is irreducible, then the restriction map H 0 (X, L) → H 0 (Y, L) is surjective. Proof. (i). The local structure theorem holds for all toroidal G/H-embeddings, so, by arguing as at the beginning of [31, Sect. 29.2], X has a toroidal smooth completion. So we may assume X complete as well. Let π : X → X be as in Lemma 1.4. By Lemma 1.2 its restriction to any closed orbit is an isomorphism. −1 p−1 a splitting of By Lemma 1.1 there exists s ∈ H 0 (G/P − , ωG/P − ) with s −1 ) of s. Note that by G/P − . By assumption we can find a lift s ∈ H 0 (X, ω̃X ∗ Lemma 1.4 we have π (ω̃X ) = ω̃X . Let σi be the canonical section of OX (Xi ) Q 1−p −1 ) is a splitting and put τ = π ∗ (s) i σi ∈ H 0 (X, ωX ). Then τ p−1 ∈ H 0 (X, ωX of X which compatibly splits the boundary divisors Xi and therefore all Gstable closed subvarieties. This follows by considering its local expansion at a point in a closed orbit. See [25, Thm. 2], [27, Thm. 3.1] or [6, Thm. 6.1.12, Ex. 1.3.E.4] for more details. (ii)(a). Let s, the σi and τ be as in (i) (restricted to the original X) and put E = (π ∗ (s)), the divisor of π ∗ (s). By [6, Prop. 1.3.11] E is reduced and effective 10 R. H. TANGE and contains Pnone of the Xi , and the splitting given by τ is compatible with (τ ) = E + i Xi . By [6, Lem. 1.4.11] there is for all j, 0 ≤ nj , m < pr and r ≥ 0 a split injection X r H i (X, L) → H i (X, Lp ( nj Xj + mE)) . (∗∗ ) j P Now E is linearly equivalent to the divisor D aD D, the aD as in (∗), see Lemma 1.4. Since X is quasiprojective it has a B-stable effective divisor P ample P E ′ (cf. proof of [4, Cor. 6.2.8]). We can write E ′ = j n′j Xj + D bD D. If we P choose n = n′ and m such that maD ≥ bD for all D, then j nj Xj + mE ∼ P ′ j Pj j nj Xj + D maD D will be ample, since each OX (D) is generated by its global sections. The latter follows from the fact that no D contains a G-orbit, since X is toroidal. See the argument in the proof of [4, Prop. 2.2]. Finally, we choose r such that nj , m < pr . Now the assertion for Y = X follows from (∗∗ ) and [6, Thm. 1.2.8]. The case of arbitrary Y follows by observing that τ restricts to a splitting of Y and that Y is a smooth toroidal spherical variety which is projective over an affine. Furthermore, the local structure theorem clearly holds for Y . So we can apply the same P arguments as above. (b). First assume X is simple. Then E ∼ D aD D is ample by [31, Cor. 17.20] and it is compatibly split by τ p−1 . So τ p−1 is an (p − 1)E-splitting and therefore also an E-splitting by [6, Thm. 1.4.10, Rem. 1.4.2(ii)]. Furthermore, E contains no G-orbit, since π ∗ (s) is nonzero on the closed orbit. So all closed G-stable subvarieties are compatibly E-split and the result follows from [6, Thm. 1.4.8(ii)]. Now assume Y is irreducible. Then X has an ample G-equivariant line bundle which, by [20, Cor. 2.3], has a B-semi-invariant global section which is nonzero on Y . Take E ′ to be the divisor of this section. Then E ′ is B-invariant, ample and doesn’t contain Y . Now we proceed as in the proof of (a) where P we have that  ′ nj = nj = 0 whenever Y ⊆ Xj . Since Y is compatibly (p − 1) E + Xj +Y Xj split we can combine the arguments from Lemmas 1.4.7 and 1.4.11 in [6]. We form a commutative square  P / H i X, Lpr ( H i (X, L) j nj Xj + mE)  H i (Y, L) / H i Y, Lpr ( P   n j j Xj + mE) , where the top horizontal row is (∗∗ ), the bottom horizontal row is (∗∗ ) with X replaced by Y , and the vertical arrows are restriction maps. Then the splittings of the horizontal arrows are compatible, i.e. they form a commutative square with the vertical arrows. Since the right vertical arrow is surjective, it follows that the same holds for the left vertical arrow. This is just a consequence of the fact that if a homomorphism of abelian groups f : M = M1 ⊕ M2 → N = N1 ⊕ N2 with f (Mi ) ⊆ Ni , i ∈ {1, 2}, is surjective, then the restrictions f |M1 : M1 → N1 and f |M2 : M2 → N2 are surjective.  EMBEDDINGS OF SPHERICAL HOMOGENEOUS SPACES IN CHARACTERISTIC p 11 As in [6, Sect. 3.3] (following Kempf) we define a morphism f : X → Y of varieties to be rational if the direct image under f of the structure sheaf of X is that of Y and if the higher direct images are zero, that is, if f∗ (OX ) = OY and Ri f∗ (OX ) = 0 for i > 0. Recall that a resolution (of singularities) of an irreducible variety X is a smooth irreducible variety X̃ together with a proper birational morphism ϕ : X̃ → X. Note that if X is normal we have ϕ∗ (OX̃ ) = OX . A rational resolution is a resolution ϕ : X̃ → X which is a rational morphism and satisfies Ri ϕ∗ ωX̃ = 0 for all i > 0. The following lemma was implicit in the proof of [6, Cor. 6.2.8]. Lemma 2.1. Assume that every projective G/H-embedding X has a G-equivariant resolution ψ : X̃ → X with ψ projective and X̃ toroidal and assume that every such resolution is a rational morphism. Then every projective birational G-equivariant morphism ϕ : X̃ → X is rational for every G/H-embedding X. Proof. Let X be a G/H-embedding and let ϕ : X̃ → X be G-equivariant, projective and birational. By the Sumihiro Theorem [28],[29] we may assume that X is quasi-projective. Then X and X̃ embed as open subsets in projective G/H-embeddings Y and Ỹ . Now let Z be the normalisation of the closure of the graph of ϕ in Ỹ × Y . Then Z is projective and X̃ identifies with a G-stable open subset of Z. Furthermore, the natural morphism Z → Y extends ϕ and we denote it again by ϕ. Note that ϕ−1 (X) = X̃. Now let ψ : Z̃ → Z be a resolution as above. Then Z̃ is projective. Since ψ is a rational morphism, the Grothendieck spectral sequence for ϕ∗ , ψ∗ and OZ̃ collapses and we obtain (Ri ϕ∗ )(OZ ) = (Ri ϕ∗ )(ψ∗ OZ̃ ) = Ri (ϕ ◦ ψ)∗ (OZ̃ ). So the rationality of the morphism ϕ follows from that of ϕ ◦ ψ.  Theorem 2.1. Assume (B1) and let X be a G/H-embedding. Then (i) X has a G-equivariant rational resolution ϕ : X̃ → X with ϕ projective and X̃ quasiprojective toroidal. (ii) X is Frobenius split compatible with all G-stable closed subvarieties. (iii) Assume X is proper over an affine, let L be a semi-ample line bundle on X and let Y be an irreducible G-stable closed subvariety of X. (a) H i (X, L) = 0 for all i ≥ 1. (b) The restriction map H 0 (X, L) → H 0 (Y, L) is surjective. (iv) If Y is a scheme which is projective over an affine and ϕ : X → Y is a proper morphism, then Ri ϕ∗ (OX ) = 0 for all i > 0. Proof. (i). We construct a resolution ϕ : X̃ → X with ϕ projective and X̃ smooth, quasiprojective and toroidal as in [6, Prop. 6.2.5], [25, Prop. 3]. First we consider the normalisation X ′ of the closure in X × X of the graph of G/H → G/Ha . In the language of coloured fans this means that we form a toroidal covering of X by “removing the colours” from the fan of X. Then we construct the desingularisation X̃ → X ′ by simplicial subdivision of the fan of X ′ . The “toric slice” in X̃0 has the same fan as X̃. So, by the local structure theorem, we deduce the smoothness of X̃ from that of the toric slice. By the local structure theorem we can also deduce the quasi-projectivity of X̃ from that of the toric slice as in [6, Prop. 6.2.3(iv)]: Since the toric slice is quasiprojective, 12 R. H. TANGE a nonnegative combination of the boundary divisors of the toric slice is ample. Now we form the same combination of the corresponding boundary divisors of X̃. This is ample on X̃0 which is the inverse image of X0 under π : X̃ → X. By the G-equivariance of π we obtain the this divisor is ample relative to π. So, since X is projective, it follows that X̃ is quasi-projective, see [13, Prop. 4.6.13(ii)]. Alternatively one can deduce the quasi-projectivity of X̃ from the description of ample divisors on a spherical variety, see [31, Sect. 17.5], [4, Sect. 5]. We show that ϕ is rational. The arguments are very similar to the proof of [6, Cor. 6.2.8]. Only for the vanishing of the higher direct images of OX̃ we have to use Proposition 2.1(ii)(a) rather than the arguments in [loc. cit.]. Since ϕ is proper and birational and X is normal, we have ϕ∗ (OX̃ ) = OX . Let τ p−1 be the Frobenius splitting of X̃ given by Proposition 2.1(i). Then τ vanishes on the exceptional locus of ϕ, since the latter is contained in the complement of the open G-orbit, i.e. the union of the boundary divisors. So, by [6, Thm. 1.3.14] we have Ri ϕ∗ (ωX̃ ) = 0 for all i ≥ 1. It remains to show that Ri ϕ∗ (OX̃ ) = 0 for all i ≥ 1. By Lemma 2.1 we may assume that X is projective. By Kempf’s Lemma, [17, Lem. 1] or [6, Lem. 3.3.3(a)], it suffices to show that there exists an ample line bundle L on X such that H i (X̃, ϕ∗ (L)) = 0 for all i ≥ 1. Here we note that for Kempf’s Lemma it is enough that ϕ is proper and X proper over an affine. Note furthermore that ϕ∗ (L) is semi-ample if L is, so by Proposition 2.1(ii)(a) any ample line bundle L on X will do, and there exists one, since X is projective. (ii). We take a resolution ϕ : X̃ → X as in (i). Then, by [6, Lem. 1.1.8], we can push a splitting of X̃ as in Proposition 2.1(i) forward to X, since ϕ∗ (OX̃ ) = OX . It will compatibly split all G-stable closed subvarieties, since this holds for the splitting of X̃. (iii). Again we take a resolution ϕ : X̃ → X as in (i). Note that X̃ is proper over an affine, since X is. Since X̃ is also quasi-projective, it is projective over ∼ an affine. Since ϕ is rational H i (X, L) → H i (X̃, ϕ∗ (L)) for all i ≥ 0, see e.g. [6, Lemma 3.3.2]. Choose Ỹ ⊆ X̃ irreducible G-stable with ϕ(Ỹ ) = Y . Then the pull-back map H 0 (Y, L) → H 0 (Ỹ , ϕ∗ (L)) is injective. Now ϕ∗ (L) is semi-ample since L is, so the assertions follow from Proposition 2.1(ii). (iv). This follows from (iii) and Kempf’s Lemma, see the proof of (i).  Theorem 2.2. Assume (B2). Let X be a G/H-embedding. Then X has a B-canonical Frobenius splitting which compatibly splits all G-stable closed subvarieties. Proof. After replacing G by the connected centre times the simply connected cover of the derived group DG, we may assume that ρ ∈ X(T ). Let ϕ : X̃ → X be a resolution as in Theorem 2.1(i). Recall that the (first) Steinberg module St is irreducible and also isomorphic to the induced module ∇((p − 1)ρ) and to the Weyl module ∆((p − 1)ρ), see [16, II.3.18,19]. Let v− and v+ be a (nonzero) lowest and highest weight vector of St. EMBEDDINGS OF SPHERICAL HOMOGENEOUS SPACES IN CHARACTERISTIC p 13 1−p If X ′ is a smooth G/H-embedding, then, by [6, Lem. 4.1.6], σ ∈ H 0 (X ′ , ωX ′ ) is B-canonical if an only if it is the image of v− ⊗v+ under a G-module homomor1−p 1−p 0 − phism St ⊗ St → H 0 (X ′ , ωX ′ ). Let s ∈ H (G/P , ωG/P − ) be the B-canonical splitting of G/P − , see [6, Thm. 4.1.15]. Since St is a Weyl module, St⊗St has a Weyl filtration by [16, Prop. II.4.21] or [6, Cor. 4.2.14]. From [16, Prop. II.4.13] one now easily deduces that the functor HomG (St ⊗ St, −) is exact on short exact sequences of G-modules with a good filtration. So from (B2) it follows that the map  1−p 1−p  ) → HomG St ⊗ St, H 0 (G/P − , ωG/P HomG St ⊗ St, H 0 (X, ω̃X −) 1−p is surjective. Now let f ∈ HomG (St ⊗ St, H 0 (G/P − , ωG/P − )) be such that 1−p f (v− ⊗ v+ ) = s. We lift f to f ∈ HomG (St ⊗ St, H 0 (X, ω̃X )) and we put s = f (v− ⊗ v+ ). Clearly s restricts to s. Let π : X̃ → X be as in Lemma 1.4. As in the proof of Proposition 2.1(i) one can a toroidal smooth completion of X̃, that Q show now, by considering 1−p π ∗ (s) i σip−1 ∈ H 0 (X̃, ωX̃ ) is a splitting of X̃ which compatibly splits all Gstable closed subvarieties. Clearly this splitting is B-canonical. Finally we push the splitting down to X by means of ϕ and apply [6, Lem. 1.1.8, Ex. 4.1.E(3)].  Remarks 2.1. 1. The existence of a B-canonical splitting as in Theorem 2.2 implies existence of good filtrations in several situations, see [6, Thm. 4.2.13, Ex. 4.2.E(2)]. It also implies normality of the G-orbit closures, see [30, Cor to Prop. 2.1]. 2. Assumption (B2) is in fact equivalent to the existence of a B-canonical splitting of X which compatibly splits the closed orbit G/P − . Indeed, assum1−p 1−p ing the latter, we have that H 0 (X, ω̃X ) → H 0 (G/P − , ωG/P − ) is surjective, 1−p since ω̃X is ample by [31, Cor. 17.20] and G/P − is compatibly split, see [6, Thm. 1.4.8]. Furthermore, the kernel of the above map has a good filtration by [6, Ex. 4.2.E(2)]. Similarly, (B1) is equivalent to the existence of a splitting of X which compatibly splits G/P − . In particular, (B2) implies (B1). 3. Theorem 2.1 does not hold for all equivariant embeddings of spherical homogeneous spaces. The following example was mentioned to me by one of the referees. Let Y be a G-variety of unseparated flags admitting an ample line bundle L such that H 1 (Y, L) 6= 0 (examples of such varieties are given in [14], Examples 3 and 4 in Section 6). Let π : X̃ → Y be the structure map of the dual line bundle, put X := Spec(k[X̃]), the “affine cone over Y ”, and let ϕ : X̃ → X be the canonical map. Then X is normal and X̃ and X are spherical G × k× varieties. Furthermore, ϕ is an equivariant resolution, see [6, Lem. 1.1.13]. But ϕ is not rational, since M H 0 (X, R1 ϕ∗ OX̃ ) = H 1 (X̃, OX̃ ) = H 1 (Y, π∗ OX̃ ) = H 1 (Y, Ln ) 6= 0 . n≥0 Here the first equality holds since X is affine, and the second one since π is affine. See [15, Prop. III.8.5, Ex. III.8.2]. 14 R. H. TANGE 3. Symmetric spaces We notice that the case of group embeddings in any characteristic is already treated in [25, 30]. In the remainder of this section we assume that char(k) 6= 2. For background on symmetric spaces we refer to [24, 32, 9, 31]. Let θad 6= id be an involution of the adjoint group Gad of G, let π : G → Gad be the canonical homomorphism, let Z(G) be the ordinary centre of G and let Zsch (G) be the schematic centre of G which is also the schematic kernel of π. The fixed point subgroup Gθadad is a (smooth) reductive subgroup of G. Let π −1 (Gθadad ) be the (ordinary) inverse image of Gθadad under π. Then Ha = Zsch (G)π −1 (Gθadad ) (see [16, I.6.2] for this notation) is the schematic inverse image of Gθadad under π. Let Tad be a maximal torus of Gad which contains a maximal θad -split torus, let Pad be a minimal θad -split parabolic subgroup containing Tad and let Bad be a Borel subgroup of Pad containing Tad as in e.g. [31, Sect. 26.4] or [9, Sect. 1]. Let T, P, B be the corresponding maximal torus, parabolic and Borel subgroup of G. Now let H be a closed subgroup scheme of Ha such that (A1) and (A4) are satisfied. Then B · x is open in G/H. For example, assume H is a closed subgroup scheme of Ha with (Gθ )0 ⊆ H for an involution θ of G with π ◦ θ = θad ◦ π.2 Then we have π −1 (Gθadad ) = {g ∈ G | gθ(g)−1 ∈ Z(G)} . Since T contains a maximal θ-split torus, it is θ-stable. By the arguments from [24, Lem. 8.1] (or [32, Lemme 2]) Ha (k) is generated by (Gθ )0 and Ha (k) ∩ T . Adding Zsch (G) ⊆ T to Ha (k) gives us Ha , so (A1) is satisfied in this case. Now assume in addition H is reduced. We check that (A4) is satisfied. Since H = (Gθ )0 (H ∩ T ) and, by property (*), the character of H ∩ T on Tx (G/H) is given by −2ρP it is enough to look at the H 0 -character. The G-action on the top exterior power of Lie(G) is trivial and the H 0 -action on the top exterior power of Lie(H) is trivial, so the H 0 -action on the top exterior power of Lie(G)/Lie(H) is trivial. Similarly, the (Gθadad )0 -action on the top exterior power of Lie(Gad )/Lie(Gθadad ) is trivial. Since H 0 maps to (Gθadad )0 , (A4) follows. Now we check that the assumptions (A2) and (A3) from Section 3, which only depend on Ha , are satisfied. By [9, Prop. 3.8, Thm. 3.9], G/Ha = Gad /Gθadad has a wonderful compactification for which the local structure theorem, with T and P as above, holds, so (A2) is satisfied. Furthermore, the Tad = π(T )-orbit map of x is separable by [9, 1.8, 3.5], so (A3) is satisfied by Lemma 1.3. Finally, we check the assumptions in the results from Section 2. It follows −1 is ample by [31, Cor. 17.20, from [9, Thm. 5.9] that (B1) is satisfied. Indeed ω̃X Rem. 17.21], so [9, Thm. 5.9] and [6, Thm. 1.2.8] imply that the restriction map −1 −1 ) → H 0 (G/P − , ωG/P H 0 (X, ω̃X − ) is surjective. Similarly, it is clear from [9, Prop. 5.7, Thm. 5.10] that (B2) is satisfied in this situation. Here one has to look at the construction of the good filtration of [9, Thm. 5.10] to see that it contains a good filtration for the kernel of the restriction map. 2A θ with this last property exists when G is semi-simple simply connected, see [26, 9.16]. EMBEDDINGS OF SPHERICAL HOMOGENEOUS SPACES IN CHARACTERISTIC p 15 4. Parabolic induction In this final section we show that most of the assumptions that we used are preserved if we apply parabolic induction to G/H. We retain the notation and assumptions (A1)-(A4) of Section 1 and 2. Let G be a connected reductive group and let Q be a parabolic subgroup of G with a surjective homomorphism π : Q → G. We assume that (P) the isogeny Q/Ker(π) → G induced by π is central (cf. [1, V.2.2]). Let T be a maximal torus of G whose image in G is T , let B be a Borel subgroup of Q containing T with π(B) = B. Let P ⊆ Q be the inverse image of P under π and let Q− , P − , B − be the opposites of Q, P, B relative to T .3 Let K be the Levi subgroup of Q containing T and denote the natural map Q− → K → G by π − . Note that (P) also holds when we replace Q and π by Q− and π − . Now let H, Ha ⊆ Q− be the schematic inverse images under π − of H and Ha . Let K′ be the subgroup of K generated by T and the simple factors of K that do not lie in the kernel of π. Then the isogeny K′ → G is central and Q− /H = G/H = K′ /HK′ . Note that Q− /P − = G/P − , so − − G ×Q G/H = G/H and G ×Q G/P − = G/P − . − Recall also that for any Q− -variety X, the induced variety G ×Q X comes with − a fibration pr : G ×Q X → G/Q− for which the fiber over the base point Q− of G/Q− is X and with pr −1 Ru (Q)Q− /Q− ∼ = Ru Q × X. By [30, Lem. 3.1] G/H is spherical, its open B-orbit is B · x. Moreover, if Di , i ∈ I are the B-stable prime divisors of G/H, then we have unique B-stable prime divisors Di = BDi , i ∈ I of G/H which intersect G/H in the Di . The other B-stable prime divisors of G/H are the pull-backs Dα = Bsα · x to G/H of the B-stable prime divisors BsαQ− /Q− of G/Q− , α a simple root in Ru Q and sα the corresponding reflection in the Weyl group of G. We will denote the closures of the B-stable prime divisors of G/H in any G/H-embedding by the same letters. Recall that the valuation cone of G/H is the same as that of G/H, see [31, Sect. 20.6] or [30, Prop. 3.1], and it is easy to determine the images of the Dα in the valuation cone, cf. [30, Prop. 4.1]. Furthermore, the toroidal embeddings of G/H are obtained by parabolically inducing the toroidal embeddings of G/H, see [30, Prop. 3.2]. Now we will show that assumptions (A1)-(A4) hold for G, H, B etc. If T1 is a closed subgroup scheme of HT which normalises H such that T1 and H generate Ha , then the schematic inverse image of T1 under π : HT → HT normalises H and together with H it generates Ha . Thus (A1) holds for G, H and Ha . By (A2) Q− /Ha = G/Ha has a wonderful compactification X for which the local structure theorem holds. So, by [30, Prop. 3.3], G/Ha has the wonderful − compactification X = G ×Q X for which the local structure theorem (relative to x, T , B and with P as above) holds. Thus (A2) holds for G/Ha . 3In [30] P, P and Q were denoted by Q, Q and P. 16 R. H. TANGE Lemma 4.1. Let X be an irreducible K-variety and let L be a G-linearised line − bundle on X = G ×Q X (Ru Q− acting trivially on X). Let s̃ be the B-semiinvariant lift of a nonzero BK -semi-invariant rational section s of L = L|X of weight λ. Then s̃ is a global section if and only if s is global and λ is dominant relative to B. If s̃ is global, then the divisor (s̃) of s̃ contains Dα = Bsα · X with coefficient hλ, α∨ i for every simple root α of Ru Q. Proof. Since the global sections of L form a rational G-module it is clear that for s̃ to be global s must be global and λ dominant. Now assume the latter. Let ∆K (λ) be the K-Weyl module of highest weight λ. By [16, II.2.11, 5.21], [11, Sect. 1] we can consider ∆K (λ) as the sub K-module of the G-Weyl module ∆(λ) generated by the highest weight vector vλ . In fact ∆K (λ) is the sum of the weight spaces of ∆(λ) corresponding to the weights which are congruent to λ modulo the root lattice of K. The sum of the other weight spaces is Q− -stable and the quotient by it is isomorphic to ∆K (λ) with Ru Q− acting trivially. Using the projection ∆(λ) → ∆K (λ) and the universal property of the Weyl module ∆K (λ), we obtain a homomorphism of Q− -modules f : ∆(λ) → H 0 (X, L) with − f (vλ ) = s. Since L = G ×Q L we have H 0 (X , L) = indGQ− H 0 (X, L) = MorQ− (G, H 0 (X, L)) , Q− acting by right multiplication on G, and Frobenius reciprocity gives us the homomorphism of G-modules f˜ : v 7→ (g 7→ f (g−1 v)) : ∆(λ) → H 0 (X , L). Now we have s̃ = f˜(vλ ) = g 7→ f (g−1 vλ ). A direct description of s̃ without the above identification is s̃(g · y) = g · f (g −1 vλ )y for all g ∈ G and y ∈ X. Let Xs be the nonzero locus of s and let α be a simple root in Ru Q. The pull-back prime divisor Dα = Bsα · Xs intersects the open set sα B · Xs = Vα sα Uα · Xs ∼ = Vα × Uα × Xs , where Vα is the product of the root subgroups Uβ with β > 0, 6= α ∼ and sα (β) a root of Ru Q. Here the isomorphism Vα × Uα × Xs → Vα sα Uα · Xs is given by (u, u′ , y) 7→ unα u′ · y, where nα ∈ NG (T ) is a representative of sα . If we choose an isomorphism θα : ka → Uα , ka the additive group of k, and denote the corresponding ka -coordinate by a then Dα ∩ sα B · Xs is defined by a = 0. Let u ∈ Vα , y ∈ X and a ∈ k. Then s̃(unα θ(a) · y) = unα θ(a) · f (θ(−a)n−1 α vλ )y . −1 Now w = nα vλ is a nonzeroP weight vector of weight sα (λ) = λ − mα, with i m = hλ, α∨ i, and θ(−a)w = m i=0 (−a) Xα,i w in the notation of [16, II.1.12]. Furthermore, we have f (Xα,i w) = 0 for i < m and Xα,m w = cvλ for some c ∈ k \ {0}. So s̃(unα θ(a) · y) = c(−a)m unα θ(a) · sy . Since unα θ(a) · y 7→ unα θ(a) · sy is nowhere zero on sα B · Xs it follows that (s̃) contains Dα with coefficient m.  Corollary. Assumption (A3) holds for G/Ha . Proof. We have ωX = ωpr ⊗ pr ∗ ωG/Q− , where pr : X → G/Q− is the canonical − projection and ωpr is the relative canonical bundle. Now pr ∗ ωG/Q− = G ×Q − − −1 −1 ⊗ k2ρQ ). Let s′ be = G ×Q (ωX (X × k−2ρQ ) and ωpr = G ×Q ωX , so ωX −1 and let s the B- (or BK -) semi-invariant global section of weight 2ρP of ωX −1 ′ be s considered as a (BK -semi-invariant) global section of ωX ⊗ k2ρQ . Note that we can consider 2ρP as a character of T and then it is the sum of the EMBEDDINGS OF SPHERICAL HOMOGENEOUS SPACES IN CHARACTERISTIC p 17 roots in Ru PK = Ru PK′ , so ρP = ρP + ρQ . Since pr −1 (BQ− /Q− ) ∼ = Ru Q × X and since (A3) holds for G/Ha , it is enough to show that s lifts to a B-semi−1 of weight 2ρP which vanishes along the prime invariant global section s̃ of ωX divisors Dα , α a simple root in Ru Q. So the result follows from Lemma 4.1 with λ = 2ρP , noting that hλ, α∨ i > 0 for α a simple root in Ru Q ⊆ Ru P.  Clearly (A4) is equivalent to ωG/H is G-equivariantly isomorphic to the pull-back of ωG/Ha along G/H → G/Ha . − But, as in the proof of the above corollary, we have ωG/H = (G ×Q ωG/H ) ⊗ − pr ∗ ωG/Q− and ωG/Ha = (G ×Q ωG/Ha ) ⊗ pra∗ ωG/Q− , where pr : G/H → G/Q− and pra : G/Ha → G/Q− are the canonical projections. From this it follows that (A4) holds for G, H and Ha . We now show that assumption (B2) is preserved under parabolic induction. Proposition 4.1. If X satisfies (B2), then so does X . Proof. By Remark 2.1.2 the assumption is equivalent to the existence of a Bcanonical splitting of X which compatibly splits the closed orbit G/P − . Assume the latter holds. Then this splitting is also B − -canonical, by [6, Prop. 4.1.10], and therefore also B − -canonical, when we consider X as a Q− -variety via π : Q− → G. So, by a result of Mathieu, see [22, Prop. 5.5] or [6, Thm. 4.1.17, − Ex. 4.1.E(4)], G×B X is B − -canonically and therefore also B-canonically Frobe− nius split, compatible with G ×B Q− /P − . From the fact that the morphism − − G ×B X → X = G ×Q X is a locally trivial fibration with fiber Q− /B − we − deduce easily that the push-forward of the structure sheaf of G ×B X is that − of X . Now we push the splitting down to X = G ×Q X and obtain that X is B-canonically Frobenius split compatible with the closed G-orbit.  Remark 4.1. It is not clear to me whether (B1) is preserved under parabolic induction, but recall from Remark 2.1.2 that it is implied by (B2). We end with a description of the Picard group of X in terms of that of X similar to [30, Prop. 4.2]4 in the case of induction from a reductive group. Since we are only interested in X and X we may assume that Ha = H and Ha = H, and since Ha and Ha contain the connected centres of G and G we may assume that G and G are semi-simple and G simply connected. Then Pic(X ) = PicG (X ) and the restriction of π to the derived group DK′ is the simply connected cover G̃ → G of G. It is now easy to see that we obtain a diagram with exact rows as in [8, p26] 0 / Pic(G/Q− ) p∗ / Pic(X ) 0 / Pic(G/Q− ) / Pic(X) j∗ id  h̃∗ (pj)∗  / Pic(G/P − ) /0 i∗ h∗  / Pic(G/P − ) /0 4 In the proof of [30, Prop. 4.2(i)] the first occurrence of (−̟ , ̟ ) should be replaced by i i ̟i1 + ̟i2 and the second occurrence should be omitted. 18 R. H. TANGE where the arrows are pull-backs associated to the maps i : G/P − ֒→ X, h̃ : X ֒→ X , h : G/P − ֒→ G/P − , j : G/P − ֒→ X , and p : X → G/Q− . Here one uses that Pic(Ru Q × X) = Pic(X) and similar for G/P − = Q− /P − . Furthermore, j ∗ is injective if i∗ is, see [8, Prop. 4.1]. As is well-known, Pic(G/P − ) and Pic(G/P − ) are isomorphic to certain subgroups Ω and Σ of the weight lattices of G and G̃. The first is freely generated by the fundamental weights ωβ , β a simple root in Ru P (or in Ru PK′ = Ru PK ), and ωα , α a simple root in Ru Q; the second is freely generated by the fundamental weights ̟β , β a simple root in Ru P . Furthermore, it is well-known that Pic(X ) is freely generated by the line bundles corresponding to the B-stable prime divisors that are not G-stable, and similar for Pic(X). Using the fact that Pic(X0 ) = 0 one can describe j ∗ (and similarly i∗ ) as follows: a line bundle L on X is mapped to the weight of the (unique up to scalar multiples) B-semi-invariant rational section of L which is nonzero on G/P − . In the case of the line bundle corresponding to a B-stable prime divisor which is not G-stable this is the weight of the canonical section. Similar for X, where we use the Borel subgroup B̃ of G̃ corresponding to B. We describe a right inverse to h∗ . The natural restriction map Ω → Σ sends ωβ to ̟β and the ωα to 0. This map has a natural right inverse: for each λ ∈ Σ there is a unique λ̃ ∈ Ω which restricts to λ and with hλ̃, α∨ i = 0, for all simple roots α in Ru Q. Under this map ̟β goes to ωβ . Assume from now on that i∗ (and j ∗ ) is injective.5 We describe the right inverse to h̃∗ corresponding to that of h∗ . If for the B-stable prime divisor Di , the line bundle L(Di ) on X has image µi in Σ, then the line bundle L(Di ) has image µ̃i in Ω. Indeed, if s is the canonical section of L(Di ), then (s) = Di and s|X is the canonical section of L(Di ). So, by Lemma 4.1, s must have weight µ̃i . Thus we can conclude that L(Di ) is mapped to L(Di ). We note that, under this right inverse, the line bundles L(Xj ) corresponding to the boundary divisors Xj of X will in general not be mapped to the line bundles L(Xj ) of X . − The L(Xj ) have a natural G-linearisation and we have L(Xj ) = G ×Q L(Xj ). So the images of the L(Xj ) in Σ actually lie in the character group of T and the images of the L(Xj ) in Ω are obtained by pulling these back along π : T → T . So in the case X is a symmetric space, one simply obtains Pic(X ) by replacing in the formulas at the end of Remark 1.1.2 the fundamental weights of G̃ by the corresponding fundamental weights of G and then adding to these the fundamental weights ωα , α a simple root in Ru Q. The images of the L(Xj ) in Ω are obtained by pulling the weights αj − θ(αj ) back along π : T → T . Acknowledgement. I would like to thank D. Timashev and M. Brion for helpful discussions, and the referee for helpful comments. This research was funded by the EPSRC grant EP/L013037/1. References [1] A. Borel, Linear algebraic groups, Second edition, Graduate Texts in Mathematics 126, Springer-Verlag, New York, 1991. 5This is the case for symmetric spaces, see [9, Thm. 4.2(ii)]. EMBEDDINGS OF SPHERICAL HOMOGENEOUS SPACES IN CHARACTERISTIC p 19 [2] M. Brion, Curves and divisors in spherical varieties, Algebraic groups and Lie groups, a volume of papers in honor of the late R. W. Richardson (G. Lehrer et al., ed.), 21-34, Austral. Math. Soc. Lect. Ser. 9, Cambridge University Press, Cambridge, 1997. [3] M. Brion, On orbit closures of spherical subgroups in flag varieties, Comment. Math. Helv. 76 (2001), no. 2, 263-299. [4] M. Brion, Variétés sphériques, Notes de la session de la S. M. F. “Opérations hamiltoniennes et opérations de groupes algébriques”, Grenoble, 1997, http://www-fourier.ujf-grenoble.fr/∼mbrion/spheriques.ps. [5] M. Brion, S. P. Inamdar, Frobenius splitting of spherical varieties, Algebraic groups and their generalizations: classical methods, 207-218, Proc. Sympos. Pure Math., 56, Part 1, Amer. Math. Soc., Providence, RI, 1994. [6] M. Brion, S. Kumar, Frobenius splitting methods in geometry and representation theory, Progress in Mathematics, 231, Birkhäuser Boston, Inc., Boston, MA, 2005. [7] M. Brion, F. Pauer, Valuations des espaces homogènes sphériques, Comment. Math. Helv. 62 (1987), no. 2, 265-285. [8] De Concini, Normality and non normality of certain semigroups and orbit closures, Algebraic transformation groups and algebraic varieties, 15-35, Encyclopaedia Math. Sci., 132, Springer, Berlin, 2004. [9] C. De Concini, T. A. Springer, Compactification of symmetric varieties, Transform. Groups 4 (1999), no. 2-3, 273-300. [10] M. Demazure, P. Gabriel, Groupes algébriques. Tome I: Géométrie algébrique, généralités, groupes commutatifs, Masson & Cie, Éditeur, Paris, North-Holland Publishing Co., Amsterdam, 1970. [11] S. Donkin, On tilting modules for algebraic groups, Math. Z. 212 (1993), no. 1, 39-60. [12] W. Fulton, Introduction to toric varieties, Annals of Mathematics Studies 131, Princeton University Press, 1993. [13] A. Grothendieck, Eléments de géométrie algébrique II, Étude globale élémentaire de quelques classes de morphismes, Inst. Hautes Études Sci. Publ. Math. No. 8, 1961. [14] Haboush, W. J. and Lauritzen, N.Varieties of unseparated flags, Linear Algebraic Groups and Their Representations (Richard S. Elman, ed.), Contemp. Math. vol. 153, Amer. Math. Soc., 1993 pp. 35-57. [15] R. Hartshorne, Algebraic geometry, Graduate Texts in Mathematics, No. 52. SpringerVerlag, New York-Heidelberg, 1977. [16] J. C. Jantzen, Representations of algebraic groups, Second edition, American Mathematical Society, Providence, RI, 2003. [17] G. Kempf, Schubert methods with an application to algebraic curves, Publications of Mathematisch Centrum, Afdeling Zuivere Wiskunde, Amsterdam, 1973. [18] G. Kempf, F. F. Knudsen, D. Mumford, B. Saint-Donat, Toroidal embeddings I, Lecture Notes in Mathematics, Vol. 339, Springer-Verlag, Berlin-New York, 1973. [19] F. Knop, The Luna-Vust theory of spherical embeddings, Proceedings of the Hyderabad Conference on Algebraic Groups (Hyderabad, 1989), 225-249, 1991. [20] F. Knop, Über Bewertungen, welche unter einer reduktiven Gruppe invariant sind, Math. Ann. 295 (1993), no. 2, 333-363. [21] D. Luna, Grosses cellules pour les variétés sphériques, Algebraic groups and Lie groups, a volume of papers in honor of the late R. W. Richardson (G. Lehrer et al., ed.), 267-280, Austral. Math. Soc. Lect. Ser. 9, Cambridge University Press, Cambridge, 1997. [22] O. Mathieu, Tilting modules and their applications, Analysis on homogeneous spaces and representation theory of Lie groups, Okayama-Kyoto (1997), 145-212, Adv. Stud. Pure Math., 26, Math. Soc. Japan, Tokyo, 2000. [23] V. B. Mehta, A. Ramanathan, Frobenius splitting and cohomology vanishing for Schubert varieties, Ann. of Math. (2) 122 (1985), no. 1, 27-40. [24] R. W. Richardson, Orbits, invariants, and representations associated to involutions of reductive groups, Invent. Math. 66 (1982), no. 2, 287-312. [25] A. Rittatore, Reductive embeddings are Cohen-Macaulay, Proc. Amer. Math. Soc. 131 (2003), no. 3, 675-684. 20 R. H. TANGE [26] R. Steinberg, Endomorphisms of linear algebraic groups, Memoirs of the American Mathematical Society, No. 80, Providence, R.I. 1968. [27] E. Strickland, A vanishing theorem for group compactifications, Math. Annalen, 277 (1987), 165-171. [28] H. Sumihiro, Equivariant completion, J. Math. Kyoto Univ. 14 (1974), 1-28. [29] H. Sumihiro, Equivariant completion II, J. Math. Kyoto Univ. 15 (1975), no. 3, 573-605. [30] R. H. Tange, On embeddings of certain spherical homogeneous spaces in prime characteristic, Transform. Groups. 17 (2012), no. 3, 861-888. [31] D. Timashev, Homogeneous spaces and equivariant embeddings, Encyclopaedia of Mathematical Sciences 138, Invariant Theory and Algebraic Transformation Groups 8, Springer, Heidelberg, 2011. [32] T. Vust, Plongements d’espaces symétriques algébriques: une classification, Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4) 17 (1990), no. 2, 165-195. School of Mathematics, University of Leeds, Leeds, LS2 9JT , UK. E-mail address : [email protected]
4math.GR
arXiv:1501.00224v2 [math.CO] 1 Oct 2017 Coloring Games and Algebraic Problems on Matroids Michał Lasoń Ph.D. thesis Theoretical Computer Science Department Faculty of Mathematics and Computer Science Jagiellonian University advised by Jarosław Grytczuk Kraków 2013 Contents Introduction 5 Chapter 0. Introduction to Matroids 1. Definitions and Terminology 2. Examples and Operations 3. Matroid Union Theorem 9 9 11 13 Part 1. 17 Coloring Games on Matroids Chapter 1. Colorings of a Matroid 1. Chromatic Number 2. List Chromatic Number 3. Basis Exchange Properties 19 19 21 23 Chapter 2. Game Colorings of a Matroid 1. Game Coloring 2. On-line List Coloring 3. Indicated Coloring 25 26 30 32 Part 2. 35 Algebraic Problems on Matroids Chapter 3. f -Vectors of Simplicial Complexes 1. Kruskal-Katona Inequality 2. Vertex Decomposable Simplicial Complexes 3. Extremal Simplicial Complexes 37 37 39 40 Chapter 4. The Conjecture of White 1. Introduction 2. White’s Conjecture for Strongly Base Orderable Matroids 3. White’s Conjecture up to Saturation 4. Remarks 43 43 44 46 50 Part 3. 53 Applications of Matroids Chapter 5. Obstacles for Splitting Necklaces 1. Introduction 2. The Topological Setting 3. Axis-aligned Hyperplane Cuts 4. Arbitrary Hyperplane Cuts 5. Colorings Distinguishing Cubes 55 56 59 60 64 66 Bibliography 67 3 Introduction This thesis is basically devoted to matroids — fundamental structure of combinatorial optimization — though some of our results concern simplicial complexes, or Euclidean spaces. We study old and new problems for these structures, with combinatorial, algebraic, and topological flavor. Therefore the thesis splits naturally into three parts, accordingly to these three aspects. In the first part we study several coloring games on matroids. Proper colorings of a ground set of a matroid (colorings in which every color class forms an independent set) were already studied by Edmonds [33]. He showed an explicit formula for the chromatic number of a matroid (the least number of colors in a proper coloring) in terms of rank function. We generalize (Theorem 1.7) a theorem of Seymour [98] on the list chromatic number of a matroid (another important result concerning colorings). Connections between matroids and games go back to the famous Shannon switching game (cf. [83]), invented independently by Gale (cf. [110]). Its matroidal version, introduced by Lehman [73], was solved by Edmonds [34] already in 1960’s. However our games are of a slightly different nature. For instance, in one variant two players are coloring a matroid properly, but only the first player is interested in finishing the job (with a given set of colors). The second, ‘bad’ one, tries to prevent it from happening. A natural question arises: how many colors are needed for the ‘good guy’ to win, compared to the chromatic number of a matroid. We show that this parameter is at most twice as big as the chromatic number (Theorem 2.1). This improves and extends to arbitrary matroids the result of Bartnicki, Grytczuk and Kierstead [12], which concerns only graphic matroids. Moreover, we show that this bound is in general almost optimal (up to at most one). In another game, we consider a coloring of a matroid from lists, but in a situation when only part of information about colors in the lists is known. The ‘bad player’ assigns consecutive colors to lists of some elements arbitrarily, while the ‘good’ one colors properly elements, each with a color from its list immediately after the color is assigned. It is called the on-line list coloring game. We prove, with Wojciech Lubawski, that the chromatic parameter of this game is the same as its off-line version (Theorem 2.7). This generalizes a theorem of Seymour [98] to the on-line setting. Games of this type were initially investigated for graphs, leading to lots of interesting results and fascinating conjectures (cf. [40, 17, 11] and [92, 116] respectively). We hope that our results add a new aspect of structural type, by showing that most of pathological phenomena appearing for graph coloring games are no longer present in the realm of matroids. 5 6 INTRODUCTION The second part of the thesis studies two problems of algebraic nature. The first one concerns f -vectors of simplicial complexes. These vectors encode the number of faces of a given size in the complex, and are characterized in the celebrated Kruskal-Katona inequality [62, 60]. A similar characterization for matroids or, more generally, for pure complexes (those with all maximal faces of the same size) remains elusive. Our main result (Theorem 3.14) asserts that every extremal simplicial complex (pure, with equality in the Kruskal-Katona inequality) is vertex decomposable. This extends a theorem of Herzog and Hibi [50] asserting that extremal complexes have the Cohen-Macaulay property (it is a simple fact that the last property is implied by vertex decomposability). Our argument is purely combinatorial with the main inspiration coming from the proof of the Kruskal-Katona inequality. The second problem we consider in this part of the thesis is a long standing conjecture of White [112] from 1980 (Conjecture 4.2). The problem concerns the symmetric exchange phenomenon in matroids, and it is solved only for some very special classes of matroids (cf. [16, 19, 59, 96, 22]). In a simplest form White’s conjecture says that if two families of bases of a matroid have equal unions (as multisets), then one can pass between them by a sequence of single element symmetric exchanges. In the algebraic language this means that the toric ideal of a matroid is equal to the ideal generated by quadratic binomials corresponding to symmetric exchanges. In a joint work with Mateusz Michałek we prove White’s conjecture up to saturation (Theorem 4.4), that is that the saturations of both ideals are equal. We believe that it is the first result in this direction valid for arbitrary matroids. Additionally we prove the full conjecture for a new large class – strongly base orderable matroids (Theorem 4.3). The last part of the thesis concerns the famous necklace splitting problem. Suppose we are given a colored necklace (a segment of integers, or a line segment), and we want to cut it so that the resulting pieces can be fairly split into two (or more) parts, which means that each part captures the same amount of every color. A theorem of Goldberg and West [42] asserts that if the number of parts is two, then every k-colored necklace has a fair splitting using at most k cuts. This fact has a simple proof [2] using the celebrated Borsuk-Ulam theorem (cf. [79]). Alon [3] extended this result to arbitrary number of parts by showing that k(q − 1) cuts suffice in the case of q parts. In a joint paper with Alon, Grytczuk, and Michałek [7] we studied a kind of opposite question, which was motivated by the problem of Erdős on strongly nonrepetitive sequences [35] (cf. [25, 30, 45]). We proved that for every k there is a (k + 3)-coloring of the real line such that no segment has a fair splitting into two parts with at most k cuts. The main results of this part of the thesis generalize this theorem to arbitrary dimension and arbitrary number of parts. We consider two versions – with cuts made by axis-aligned hyperplanes (Theorem 5.7), and by arbitrary hyperplanes (Theorem 5.13). In the first case the upper bound we achieve almost matches the lower bound implied by a theorem of Alon [3] (or by a more general theorem of de Longueville and Živaljevič [75]). The methods we use relay on the topological Baire category theorem and several applications of algebraic matroids. Acknowledgements Firstly, I would like to thank my advisor Jarosław Grytczuk for his scientific guidance during my entire Ph.D. studies. It was a great pleasure and opportunity to learn from him the taste of combinatorics. Throughout the thesis he provided me with encouragement and a lot of interesting problems. In particular he introduced me to the game chromatic number of a matroid. He also was the first person who asked if Seymour’s theorem generalizes to the on-line setting. Finally, Jarek is a coauthor of my first scientific publication. I am sincerely grateful for his effort. I would like to express my gratitude to professor Jürgen Herzog for pointing a direction in which I can generalize one of his results, and also for showing me the conjecture of White. I am especially grateful to Wojciech Lubawski and Mateusz Michałek, who are coauthors of some of the results presented in this thesis. I thank them for inspiration, motivation and hours of fruitful discussions. I thank my friends: Łucja Farnik, Michał Farnik, Adam Gągol, Andrzej Grzesik, Grzegorz Gutowski, Jakub Kozik, Tomasz Krawczyk, Piotr Micek, Arkadiusz Pawlik, Małgorzata Sulkowska and Bartosz Walczak for interesting talks, questions, answers and a great working atmosphere. Writing this thesis would not be possible without the constant support of my family. I thank a lot my mum for her love and motivation. The author has been granted scholarship in the project Doctus Małopolski fundusz stypendialny dla doktorantów co-funded by the European Union through the European Social Fund, and also supported by the grant of the Polish Ministry of Science and Higher Education On-line tasks selection N N206 568240. 7 CHAPTER 0 Introduction to Matroids Matroids are one of basic mathematical structures. They abstract the idea of independence from various areas of mathematics, ranging from algebra to combinatorics. The origins of matroid theory turn back to 1930’s, when they were introduced by Whitney [113]. For an adequate introduction we refer the reader to the book of Oxley [85] (and to [86] for a sketchy one). In this Chapter we give a brief sketch of basic concepts of matroid theory. We present definitions, examples and several classes of matroids. We also prove Matroid Union Theorem 0.20, a basic tool, which already reveals regularity of the structure of matroid. Therefore, this Chapter can be treated as preliminaries, in particular it does not contain any our result. 1. Definitions and Terminology Matroids are cryptomorphic structures, that is they can be defined in many different but nontrivially equivalent ways. We will introduce five of this axiom systems which are most important for us. We show correspondence between the first one, which can be considered as basic, and the others (for proofs see [85, Chapter 1]). This will also provide a necessary notation. Usually by a matroid M on a finite ground set E we mean a collection I of subsets of E satisfying the following conditions. Definition 0.1. (Independent Sets) A matroid M consists of a finite set E, called the ground set, and a set I of independent subsets of E satisfying the following conditions: (1) empty set is an independent set, (2) subset of an independent set is independent, (3) if I, J ∈ I and |I| < |J|, then we can extend I with an element of J, that is I ∪ e ∈ I for some e ∈ J \ I (augmentation axiom). Actually, as we will see in Definition 0.7, any finite subset E of a vector space together with a family of linearly independent sets contained in E constitutes a basic example of a matroid. This connection justifies a large part of terminology used in matroid theory. For example, one may define bases of a matroid, the rank of a subset, or the closure operation just the same way as for vector spaces. Definition 0.2. (Bases) A matroid M consists of a finite set E, called the ground set, and a set B of subsets of E, called bases, satisfying the following conditions: (1) B is non-empty, (2) if B1 , B2 are bases and e ∈ B1 \ B2 , then there is f ∈ B2 \ B1 , such that B1 ∪ f \ e is a basis. 9 10 0. INTRODUCTION TO MATROIDS Bases are just maximal independent sets. Due to augmentation axiom all of them have equal size, so bases are in fact maximum independent sets. The other direction is also clear, having the set of bases, independent sets are exactly those which are contained in some basis. Definition 0.3. (Rank Function) A matroid M consists of a finite set E, called the ground set, and a function r : P(E) → N, called the rank function, satisfying the following conditions: (1) if A ⊂ E, then 0 ≤ r(A) ≤ |A|, (2) rank is monotone, that is if A ⊂ B ⊂ E, then r(A) ≤ r(B), (3) if A, B ⊂ E, then r(A∪B)+r(A∩B) ≤ r(A)+r(B) (submodularity axiom). The rank of a set A ⊂ E is the size of the maximal independent set contained in A. Knowing the rank function, independent sets are exactly whose sets A ⊂ E for which r(A) = |A|. So as independent sets abstract among others linear independence or algebraic independence, the rank function abstracts dimension or transcendence degree. Sometimes the rank of the whole matroid M , which is r(E), will be denoted by r(M ) in order not to refer to the ground set. Definition 0.4. (Closure) A matroid M consists of a finite set E, called the ground set, and a closure operator cl : P(E) → P(E) satisfying the following conditions: (1) (2) (3) (4) if if if if A ⊂ E, then A ⊂ cl(A), A ⊂ B ⊂ E, then cl(A) ⊂ cl(B), A ⊂ E, then cl(cl(A)) = cl(A), A ⊂ E, e ∈ E, and f ∈ cl(A ∪ e) \ cl(A), then e ∈ cl(A ∪ f ). The closure of a set A ⊂ E is the set of all elements e ∈ E satisfying r(A ∪ e) = r(A). Or, in other words, cl(A) is the maximum set B containing A for which r(B) = r(A). Independent sets are exactly those sets which do not have a proper subset with the same closure. Another source of matroid terminology comes from graph theory. This is because, as we will see in Definition 0.9, every graph G can be turned into a graphic matroid M (G) on the set of edges by defining I as the family of sets not containing a cycle. Therefore circuits in a matroid, which generalize cycles of graphic matroid, are minimal sets which are not independent. Thus, a set is independent if it does not contain any circuit. Definition 0.5. (Circuits) A matroid M consists of a finite set E, called the ground set, and a set C of subsets of E, called circuits, satisfying the following conditions: (1) empty set is not a circuit, (2) there is no inclusion between any two circuits, (3) if C1 , C2 are two circuits and e ∈ C1 ∩ C2 , then there is a circuit C3 such that C3 ⊂ C1 ∪ C2 \ e. There are also several other ways to define a matroid, for example by flats or hyperplanes. 2. EXAMPLES AND OPERATIONS 11 2. Examples and Operations We show basic examples and classes of matroids (for proofs that they satisfy axioms of matroid see [85]). More examples can be produced using operations on matroids described in the second part of this Section. We begin with an almost trivial example. Definition 0.6. (Uniform Matroid) Let E be a finite set, and let b be an integer. Then, (E, I) = (E, {A ⊂ E : |A| ≤ b}) is a matroid denoted by Ur(E),|E| , with the rank function r(A) = min{|A| , b}. The next two examples come from algebra. The second will play a big role in the last part of the thesis. Definition 0.7. (Representable Matroid) Let V be a vector space over a field K, and let E ⊂ V be a finite set. Then, (E, I) = (E, {A ⊂ E : A is linearly independent over K}) is a matroid with the rank function r(A) = dimK (span(A)) and closure operator cl(A) = span(A) ∩ E. Such a matroid is called representable over the field K. If a matroid is representable over some field it is called representable. Definition 0.8. (Algebraic Matroid) Let K ⊂ L be a field extension, and let E ⊂ L be a finite set. Then, (E, I) = (E, {A ⊂ E : A is algebraically independent over K}) is a matroid with the rank function r(A) = trdegK (K(A)). The following example is probably the most natural for combinatorics. Its big advantage is that it can be easily visualized. We will often refer to graphic matroids in examples and some theorems. Definition 0.9. (Graphic Matroid) Let G = (V, E) be a multigraph. Then, (E, I) = (E, {A ⊂ E : A does not contain a cycle}) is a matroid with the set of bases equal to the set of spanning forests of G, and the set of circuits equal to the set of cycles of G. We denote it by M (G). Graphic matroids are also interesting because they are very different from previous two families of matroids coming from algebra. Some classes of matroids are defined by a certain property, instead of by a specific presentation (as it was in the previous examples). Definition 0.10. (Base Orderable Matroid) A matroid is base orderable if for any two bases B1 and B2 there is a bijection π : B1 → B2 such that B1 ∪ π(e) \ e and B2 ∪ e \ π(e) are bases for any element e ∈ B1 . Definition 0.11. (Strongly Base Orderable Matroid) A matroid is strongly base orderable if for any two bases B1 and B2 there is a bijection π : B1 → B2 such that B1 ∪ π(A) \ A is a basis for any subset A ⊂ B1 . 12 0. INTRODUCTION TO MATROIDS The condition of strongly base orderable matroid implies also that B2 ∪ A \ π(A) is a basis for any A ⊂ B1 . Moreover, we can assume that π is the identity on B1 ∩ B2 . The class of strongly base orderable matroids is closed under taking minors. As we will see in Section 3 of Chapter 4, one of our results holds only for strongly base orderable matroids. Definition 0.12. (Transversal Matroid) Let E be a finite set, and let A = {Aj : j ∈ J} be a multiset of subsets of E. The following is a matroid (E, I) = (E, {A ⊂ E : there exists an injection φ : A → J such that a ∈ Aφ(a) for every a ∈ A}. In fact for any transversal matroid one can find a corresponding multiset A of cardinality equal to its rank. Another similar, and not so well-known class of matroids, is the class of laminar matroids, see [41]. Definition 0.13. (Laminar Matroid) A family of sets is laminar if any two sets are either disjoint or one is contained in the other. Let F be a laminar family of subsets of a finite set E, and let k : F → N be a function. The following is a matroid (E, I) = (E, {A ⊂ E : |A ∩ F | ≤ k(F ) for every F ∈ F}). As one can prove both transversal matroids and laminar matroids are strongly base orderable. Moreover, none of these classes is contained in the other. Definition 0.14. (Regular Matroid) A matroid is regular if it is representable over every field. One can easily show that graphic matroids are regular. For characterizations of regular matroids see the paper of Tutte [105] (cf. [85]). Seymour [97] proved a very interesting structural result. It asserts that every regular matroid may be constructed by piecing together graphic and cographic matroids and copies of a certain 10-element matroid. Now we pass to operations on matroids. Firstly we are going to define two basic, restriction and contraction, which produce from a given matroid a matroid on a smaller ground set with useful properties. Let M be a matroid on a ground set E with the set of independent sets I, the set of bases B, rank function r, and let F ⊂ E. Definition 0.15. (Restriction) Restriction of a matroid M to the set F , denoted by M |F or M \ (E \ F ), is a matroid on the ground set F satisfying: (1) I ⊂ F is independent in M |F if I is independent in M , (2) rank function of M |F equals to r|F – restriction of rank function of M to the set F . Definition 0.16. (Contraction) Contraction of the set F in a matroid M , denoted by M/F , is a matroid on the ground set E \ F satisfying: (1) I ⊂ E \ F is independent in M/F if for every J ⊂ F independent in M the set I ∪ J is independent in M , (2) rank function r ′ of M/F satisfies r ′ (A) = r(A ∪ F ) − r(F ) for every A ⊂ E \ F. 3. MATROID UNION THEOREM 13 A matroid is called a minor of another matroid if it is obtained from it by a sequence of restrictions and contractions. One can show that in fact any minor is obtained by at most two such operations. Matroid theory has a concept of duality. Definition 0.17. (Dual Matroid) Dual matroid of a matroid M , denoted by M ∗ , is a matroid on the ground set E, such that: (1) B ⊂ E is a basis in M ∗ if E \ B is a basis in M , (2) rank function r ′ of M ∗ satisfies r ′ (A) = |A| − r(E) + r(E \ A) for every A ⊂ E. Of course (M ∗ )∗ = M . Now we can describe the duality between operations of restriction and contraction. Namely, we have M/F = (M ∗ \ F )∗ and M \ F = (M ∗ /F )∗ . Bases, circuits, etc. of the dual matroid are called cobases, cocircuits, etc. of the original matroid. It is also easy to define direct sum of matroids. Definition 0.18. (Direct Sum) Suppose that M1 = (E1 , I1 , r1 ), . . . , Mk = (Ek , Ik , rk ) are matroids on disjoint ground sets. Then, there exists a matroid M1 ⊔ · · · ⊔ Mk = (E1 ∪ · · · ∪ Ek , I, r) satisfying: (1) I = {I1 ∪ · · · ∪ Ik : I1 ∈ I1 , . . . , Ik ∈ Ik }, (2) r(A) = r1 (A ∩ E1 ) + · · · + rk (A ∩ Ek ) for every A ⊂ E1 ∪ · · · ∪ Ek . As we will see in Theorem 0.22 it is possible to extend this definition to matroids with not necessary disjoint ground sets. One can also blow up an element of a matroid by adding ‘parallel’ elements. This operation extends the idea of adding parallel edges in graphic matroids. Definition 0.19. (Blow Up) Let M = (E, I, C) be a matroid and let e ∈ E, f ∈ / E. Then, there exists a matroid M ′ = (E ∪ f, I ′ , C ′ ) satisfying: (1) I ′ = I ∪ {I ∪ f \ e : e ∈ I ∈ I}, (2) C ′ = C ∪ {C ∪ f \ e : e ∈ C ∈ C} ∪ {e, f }. The crucial fact about above definitions is that in each case the matroid with described set of independent sets, bases, circuits or rank function exists (for the proofs see [85]). 3. Matroid Union Theorem We will describe one of possible formulations of the Matroid Union Theorem, which is best for our applications. Theorem 0.20 (Nash-Williams, cf. [85]). (Matroid Union Theorem) Let M1 , . . . , Mk be matroids on the same ground set E with rank functions r1 , . . . , rk respectively. The following conditions are equivalent: (1) there exist sets Vi with V1 ∪ · · · ∪ Vk = E, such that the set Vi is independent in Mi for each i, (2) for every A ⊂ E we have r1 (A) + · · · + rk (A) ≥ |A|. 14 0. INTRODUCTION TO MATROIDS The theorem follows from its weighted version, which we prove below. Let w : E → N be a weight assignment of elements of E. A collection of (not necessarily different) subsets V1 , . . . , Vd ⊂ E is said to be a w-covering of E if for every element e ∈ E there are exactly w(e) members of the collection containing e. If w ≡ k for some integer k, then we write shortly k-covering. By a covering we mean a 1-covering. Theorem 0.21 (Nash-Williams, cf. [85]). (Weighted Matroid Union Theorem) Let M1 , . . . , Mk be matroids on the same ground set E with rank functions r1 , . . . , rk respectively. The following conditions are equivalent: (1) there exists a w-covering of E by sets V1 , . . . , Vk with Vi independent in Mi for each i, P (2) for every A ⊂ E we have r1 (A) + · · · + rk (A) ≥ e∈A w(e). Proof. Implication (1) ⇒ (2) is easy. For every A ⊂ E we have X r1 (A) + · · · + rk (A) ≥ |A ∩ V1 | + · · · + |A ∩ Vk | = w(e). e∈A We focus on the opposite implication (2) ⇒ (1), which we will prove by double induction. Firstly on the size of the ground set E, and secondly on P the sum e∈E w(e). The basis of induction, that is when |E| = 1 or w ≡ 0, is clearly true. Consider a function f : P(E) → N defined by X f (A) = r1 (A) + · · · + rd (A) − w(e). e∈A Notice that the condition (2) guarantees that f is nonnegative for all A ⊂ E. We distinguish two cases. Case 1. For some non-empty proper subset A ( E there is f (A) = 0. Consider matroids M1 , . . . , Mk restricted to A and weight function w|A . Since for them the condition (2) of the theorem is satisfied, then from the inductive assumption there is a w|A -covering by U1 , . . . , Uk with Ui is independent in Mi |A . Consider also weight function w|E\A and matroids M1′ , . . . , Md′ on the set E\A obtained by contracting set A in M1 , . . . , Md , that is Mi′ = Mi /A. Then ri′ (X) = ri (X ∪ A) − ri (A), so they satisfy the condition (2) of the theorem. From the inductive assumption there is a w|E\A -covering by U1′ , . . . , Uk′ with Ui′ is independent in Mi /A. Now unions Vi = Ui ∪ Ui′ form a w-covering of E, and Vi is independent in Mi . Case 2. For every non-empty proper subset A ⊂ E there is f (A) > 0. Pick an arbitrary element e ∈ E with w(e) > 0, and pick an arbitrary i such that ri (e) > 0, which clearly exists because f (e) > 0. Consider matroids M1 , . . . , Mi′ , . . . , Mk , where Mi′ = Mi /e, and weight function w′ such that w′ |E\e ≡ w|E\e and w′ (e) = w(e) − 1. For them condition (2) of the theorem still holds, so from the inductive assumption there is a w′ -covering by sets V1 , . . . , Vi′ , . . . , Vk independent in corresponding matroids. Observe that if we define Vi = Vi′ ∪ e, then V1 , . . . , Vi , . . . , Vk will be an eligible w-covering for matroids M1 , . . . , Mi , . . . , Mk .  3. MATROID UNION THEOREM 15 Another formulation of the Matroid Union Theorem describes the rank function of the join of matroids. We derive it as a corollary of the previous formulation. Theorem 0.22 (Nash-Williams, cf. [85]). (Join of Matroids) Let M1 = (E, I1 , r1 ), . . . , Mk = (E, Ik , rk ) be matroids on the same ground set E. Then, there exists a matroid M1 ∨ · · · ∨ Mk = (E, I, r) satisfying: (1) I = {I1 ∪ · · · ∪ Ik : I1 ∈ I1 , . . . , Ik ∈ Ik }, (2) r(A) = min{r1 (B) + · · · + rk (B) + |A \ B| : B ⊂ A}. Proof. We present a proof in the case k = 2, a generalization to arbitrary number of matroids is straightforward. According to Definition 0.3, matroid M1 ∨M2 = (E, r) exists if conditions (1) − (3) of rank function hold. It is easy to check that first two are satisfied. To check the submodularity axiom r(A∪B)+r(A∩B) ≤ r(A)+r(B) assume that minima from the definition of r(A) and r(B) are realized for sets C, D respectively. We have r(A ∪ B) + r(A ∩ B) ≤ r1 (C ∪ D) + r2 (C ∪ D) + |A ∪ B \ (C ∪ D)| + r1 (C ∩ D) + r2 (C ∩ D) + |A ∩ B \ (C ∩ D)| ≤ r1 (C) + r2 (C) + |A \ C| + r1 (D) + r2 (D) + |B \ D| = r(A) + r(B), where the second inequality follows from submodularity of r1 and r2 . Suppose now that r(A) = |A|. It means that for each B ⊂ A there is r1 (B) + r2 (B) + |A \ B| ≥ |A| , so r1 (B) + r2 (B) ≥ |B| . From Matroid Union Theorem 0.20 it follows that A = I1 ∪ I2 for some I1 ∈ I1 , I2 ∈ I2 . Contrary, if A = I1 ∪ I2 is the sum of two independent sets, then for each B ⊂ A we have r1 (B) + r2 (B) + |A \ B| ≥ r1 (B ∩ I1 ) + r2 (B ∩ I2 ) + |A \ B| ≥ |B ∩ I1 | + |B ∩ I2 | + |A \ B| ≥ |A| , so A is independent.  Notice that a special case of join of matroids is the direct sum of matroids (see Definition 0.18). Direct sum of matroids on disjoint ground sets is just the join of their extensions to the sum of all ground sets. Corollary 0.23. A matroid M = (E, r) can be covered with k independent sets if and only if for every A ⊂ E there is kr(A) ≥ |A|. Proof. It is exactly Matroid Union Theorem 0.20 applied to M1 = · · · = Mk = M . We can also easily get it from Theorem 0.22. Namely, matroid M can be covered with k independent sets if and only if E is independent in set (rank of E is |E|) in the matroid M ′ = M1 ∨ · · · ∨ Mk , where all Mi = M . Rank of E in M ′ equals to min{kr(A) + |E \ A| : A ⊂ E}.  Corollary 0.24. A matroid M = (E, r) has k disjoint bases if and only if for every A ⊂ E there is kr(A) + |E \ A| ≥ kr(E). Proof. Matroid M has k disjoint bases if and only if E has rank at least kr(E) in the matroid M ′ = M1 ∨ · · · ∨ Mk , where all Mi = M . Due to Theorem 0.22 rank of E in M ′ equals to min{kr(A) + |E \ A| : A ⊂ E}.  16 0. INTRODUCTION TO MATROIDS The last theorem of this Section gives an explicit formula, in terms of rank functions, for the largest common independent set in two matroids. Unfortunately, it does not have an easy generalization to larger number of matroids. Theorem 0.25 (Edmonds, cf. [85]). (Intersection of Two Matroids) Let M1 = (E, I1 , r1 ) and M2 = (E, I2 , r2 ) be two matroids on the same ground set E. Then max{|I| : I ∈ I1 ∩ I2 } = min{r1 (A) + r2 (E \ A) : A ⊂ E}. Proof. Inequality ≤ is obvious. To prove ≥ suppose that for each A ⊂ E there is r1 (A)+r2 (E\A) ≥ k. Observe that rM1 ∨M2∗ (E) ≥ k+rM2∗ (E), because for each A ⊂ E there is r1 (A) + |A| − r2 (E) + r2 (E \ A) + |E \ A| ≥ k + |E| − r2 (E) = k + rM2∗ (E). Thus there are two disjoint sets I1 independent in M1 and I2 independent in M2∗ , such that |I1 | + |I2 | ≥ k + rM2∗ (E). Suppose c = rM2∗ (E) − |I2 |. Extend I2 with c elements to a basis B ∗ of M2∗ . Now I1 ∩ (E \ B ∗ ) is independent in both M1 and M2 . Moreover, it has at least |I1 | − c ≥ k elements.  Part 1 Coloring Games on Matroids CHAPTER 1 Colorings of a Matroid This Chapter extends the concept of a proper graph coloring to matroids. Namely, a coloring of a matroid is proper if elements of the same color form an independent set. In that way we get a notion of chromatic number and list chromatic number of a matroid (several game versions of this parameters will be studied in the following Chapter 2). We present explicit formula of Edmonds [33] for chromatic number in terms of rank function, and a theorem of Seymour [98] asserting that the list chromatic number of a matroid equals to the chromatic number. The main result of this Chapter, Theorem 1.7, generalizes Seymour’s theorem from lists of constant size to the case when lists have arbitrary fixed size. We prove that a matroid is colorable from any lists of a fixed size if and only if it is colorable from particular lists of that size (this theorem will be later even further generalized, see Theorem 2.7). The last Section shows applications of Theorem 1.7 to several basis exchange properties. The results of this Chapter come from our paper [64]. 1. Chromatic Number Let M be a matroid on a ground set E. A coloring of M is an assignment of colors (usually natural numbers) to elements of E. A coloring is proper if elements of the same color form an independent set. For the rest of this thesis whenever we write the term coloring, we always mean proper coloring. An element e ∈ E in a matroid M = (E, r) is called a loop if r(e) = 0 (in analogy to a loop in a graph, which is a loop in the corresponding graphic matroid). Notice that if a matroid contains a loop, then it does not admit any proper coloring. On the other hand, if a matroid does not have any loop, it is loopless, then it has at least one proper coloring. Thus, when considering colorings we restrict to loopless matroids. Definition 1.1. (Chromatic Number) The chromatic number of a loopless matroid M , denoted by χ(M ), is the minimum number of colors in a proper coloring of M . For instance, if M = M (G) is a graphic matroid, then χ(M ) is the least number of colors needed to color edges of G so that no cycle is monochromatic. This number is known as the arboricity of a graph G and should not be confused with the usual chromatic number χ(G) of a graph G. In further analogy to graph theory we define fractional chromatic number. For a fixed integer a an a-coloring of a matroid M on a ground set E is an assignment of a colors to each element of E. Still a coloring is proper if for each color elements to which it is assigned form an independent set in 19 20 1. COLORINGS OF A MATROID M . Notice, that there is no difference between proper a-coloring of a matroid and a-covering with independent sets. Definition 1.2. (Fractional Chromatic Number) The fractional chromatic number of a loopless matroid M , denoted by χf (M ), is the infimum of fractions ab such that M has a proper a-coloring with b colors. For a matroid the chromatic number as well as the fractional chromatic number can be easily expressed in terms of a rank function. Extending a theorem of Nash-Williams [84] for graph arboricity, Edmonds [33] proved the following explicit formula for the chromatic number of a matroid. It was observed that his proof gives also a formula for fractional chromatic number. Theorem 1.3 (Edmonds, [33]). (Formula for Chromatic Number) For every looples matroid M = (E, r) we have   |A| |A| and χ(M ) = max χf (M ) = max . ∅6=A⊂E r(A) ∅6=A⊂E r(A) Proof. Denote |A| . ∅6=A⊂E r(A) Suppose M has an a-covering with b independent sets V1 , . . . , Vb , then for each ∅ = 6 A ⊂ E we have ∆(M ) = max br(A) ≥ r(A ∩ V1 ) + · · · + r(A ∩ Vb ) = |A ∩ V1 | + · · · + |A ∩ Vb | = a |A| . |A| Hence ab ≥ r(A) , and as a consequence χf (M ) ≥ ∆. Since χ(M ) is an integer, and obviously χ(M ) ≥ χf (M ), we get immediately that χ(M ) ≥ ⌈∆⌉. To prove opposite inequalities suppose that the maximum in the defini|A| tion of ∆ is reached for some non-empty subset A ⊂ E, that is ∆(M ) = r(A) . For fractional chromatic number consider |A| matroids M1 , . . . , M|A| on the ground set E, all equal to M . Apply Weighted Matroid Union Theorem 0.21 with weight function w ≡ r(A). For each ∅ = 6 B ⊂ E we have r1 (B) + · · · + r|A| (B) = |A| r(B) ≥ |B| r(A), so the second condition of the theorem is satisfied. Hence, there exists an r(A)-covering of M with |A| independent sets. This means that χf (M ) ≤ ∆. For chromatic number consider ⌈∆⌉ matroids M1 , . . . , M⌈∆⌉ on the ground set E, all equal to M . We will apply Matroid Union Theorem 0.20. For each ∅= 6 B ⊂ E we have   |A| |A| r(B) ≥ |B| . r(B) ≥ r1 (B) + · · · + r⌈∆⌉ (B) = ⌈∆⌉ r(B) = r(A) r(A) The second condition of the theorem holds, hence there exists a proper coloring of M with ⌈∆⌉ independent sets. This means that χf (M ) ≤ ⌈∆⌉.  In particular, the chromatic number of a matroid is less than one bigger than its fractional chromatic number. In contrary for graphs chromatic number is not bounded from above by any function of fractional chromatic number (see for example Kneser’s graphs [43]). Obviously in graph theory there is no chance for an analogical formula, however there is one example of an almost exact formula for the chromatic 2. LIST CHROMATIC NUMBER 21 number. Vizing [106] proved that the chromatic index (chromatic number of line graph – incidence graph of edges of a graph) is always equal to the maximum degree or to the maximum degree plus one. 2. List Chromatic Number Suppose each element e of the ground set E of a matroid M is assigned with a set (a list) of colors L(e). In analogy to the list coloring of graphs (initiated by Vizing [107], and independently by Erdős, Rubin and Taylor [36]), we want to find a proper coloring of M with additional restriction that each element gets a color from its list. By a list assignment (or simply lists) we mean a function L : E → P(N), and by the size of L we mean the function ℓ : E → N with |L(e)| = ℓ(e) for each e ∈ E. A matroid M is said to be colorable from lists L if there exists a proper coloring c : E → N of M , such that c(e) ∈ L(e) for every e ∈ E. Definition 1.4. (List Chromatic Number) The list chromatic number of a loopless matroid M , denoted by ch(M ), is the minimum number k such that M is colorable from any lists of size at least k. Clearly, we have ch(M ) ≥ χ(M ). The theorem of Seymour [98] asserts that actually we have equality here. We provide a proof of this result for completeness. It is a simple application of Matroid Union Theorem 0.20. Theorem 1.5 (Seymour, [98]). For every loopless matroid M we have equality ch(M ) = χ(M ). Proof. Suppose χ(M ) = k, and take any list assignment L : E → P({1, . . . , d}) of size at least k. Let Qi = {e ∈ E : i ∈ L(e)}, and let Mi = (E, ri ) be the restriction of M to Qi with the ground set extended to E. From the assumption there exists a 1-covering with k independent sets V1 , . . . , Vk . Then, for every A ⊂ E we have X X 1 X |Vj ∩ A ∩ Qi | ≥ |Vj ∩ A| = |A| , k 1≤j≤k 1≤i≤d 1≤j≤k since each element e ∈ Vj ∩ A belongs to Qi for at least k values of i. But ri (A) ≥ |Vj ∩ A ∩ Qi |, so X X 1 X ri (A) = ri (A) ≥ |A| . k 1≤i≤d 1≤j≤k 1≤i≤d The result follows from Matroid Union Theorem 0.20.  For a weight assignment w and a list assignment L we say that M is w-colorable from lists L, if it is possible to choose for each e ∈ E a set of w(e) colors from its list L(e), such that for each color elements to which it is assigned form an independent set in M . In other words each list L(e) contains a subset L′ (e) of size w(e) such that choosing any color from L′ (e) for e results in a proper coloring of a matroid M . Definition 1.6. (Fractional List Chromatic Number) The fractional list chromatic number of a loopless matroid M , denoted by chf (M ), is the infimum of fractions ab for which M is a-colorable from any lists of size at least b. 22 1. COLORINGS OF A MATROID For graphs list chromatic number is not bounded from above by any function of chromatic number, as it can already be arbitrarily large for graphs with chromatic number two. Dinitz stated a conjecture that chromatic number and list chromatic number coincide for line graphs. It was confirmed by Galvin [39] for bipartite graphs, and for arbitrary graphs in the asymptotic sense by Kahn [58]. Surprisingly, Alon, Tuza and Voigt [4] proved that fractional versions of chromatic number and list chromatic number coincide for all graphs. We generalize Seymour’s Theorem 1.5 to the setting, where sizes of lists are still fixed, but not necessarily equal. Theorem 1.7 (Lasoń, [64]). (Generalization of Seymour’s Theorem) Let M = (E, r) be a matroid, and let ℓ : E → N, w : E → N be list size and weight functions. Then the following conditions are equivalent: P P (1) for each A ⊂ E there is i∈N+ r({e ∈ A : ℓ(e) ≥ i}) ≥ e∈A w(e), (2) matroid M is w-colorable from lists Lℓ (e) = {1, . . . , ℓ(e)}, (3) matroid M is w-colorable from any lists of size ℓ. Proof. Let L be a fixed list assignment S of size ℓ, and let us denote by Qi = {e ∈ E : i ∈ L(e)} and d = max e∈E L(e) (and respectively Qℓi for lists Lℓ ). Consider matroids M1 , . . . , Md , such that Mi equals to M |Qi with the ground set extended to E. Observe that matroid M is properly w-colorable from lists L if and only if its ground set E can be w-covered by sets V1 , . . . , Vd , such that Vi is independent in Mi . Namely two expressions to cover and to color have the same meaning. From Weighted Matroid Union Theorem 0.21 we get that this condition is equivalent to the fact, that for each A ⊂ E there is an inequality X w(e). r(A ∩ Q1 ) + · · · + r(A ∩ Qd ) ≥ e∈A Now equivalence between conditions (1) and (2) follows from the fact that {e ∈ A : ℓ(e) ≥ i} = A ∩ Qℓi . Implication (3) ⇒ (2) is obvious. To prove (2) ⇒ (3) it is enough to show that for each subset A ⊂ E there is an inequality r(A ∩ Q1 ) + · · · + r(A ∩ Qd ) ≥ r(A ∩ Qℓ1 ) + · · · + r(A ∩ Qℓd ). P We will use the notion or + for the sum P of multisets. P Since Pboth list asℓ signments Lℓ and L are of size ℓ, we get that i Qi = i Qi = e∈E ℓ(e){e} as multisets. Observe that to prove the inequality, whenever two sets Qi and Qj are incomparable in the inclusion order, we may replace them by sets Qi ∪ Qj , and Qi ∩ Qj . It is because: (1) r(A ∩ Qi ) + r(A ∩ Qj ) ≥ r(A ∩ (Qi ∪ Qj )) + r(A ∩ (Qi ∩ Qj )), by submodularity of r (see Definition 0.3), (2) Qi + Qj = (Qi ∪ Qj ) + (Qi ∩ Qj ) as multisets, (3) |Qi |2 + |Qj |2 < |Qi ∪ Qj |2 + |Qi ∩ Qj |2 . The last parameter grows, and it is bounded (because the number of non-empty sets, and their sizes are bounded). Hence, after a finite number of steps the replacement procedure stops. Then sets Qi are linearly ordered by inclusion. Let us reorder them in such a way that Q1 ⊃ Q2 ⊃ · · · ⊃ Qd . 3. BASIS EXCHANGE PROPERTIES 23 Then, it is easy to see that Q1 = Qℓ1 , Q2 = Qℓ2 , . . . , Qd = Qℓd , so the inequality r(A ∩ Q1 ) + · · · + r(A ∩ Qd ) ≥ r(A ∩ Qℓ1 ) + · · · + r(A ∩ Qℓd ) is satisfied. This proves the assertion.  Seymour’s Theorem 1.5 follows from the above result by taking a constant size function ℓ ≡ χ(M ) and a weight function w ≡ 1. We get also that chf (M ) = χf (M ). Our theorem implies as well the following stronger statement. Corollary 1.8. Let I1 , . . . , Ik be a partition of the ground set of a matroid M into independent sets, and let ℓ(e) = i whenever e ∈ Ii . Then matroid M is colorable from any lists of size ℓ. In the next Section we show applications of our Generalization of Seymour’s Theorem 1.7 to basis exchange properties. 3. Basis Exchange Properties Let E be a finite set. For a family B ⊂ P(E) of subsets of E consider the following conditions: (1) B is non-empty, (2) for every B1 , B2 ∈ B and e ∈ B1 \ B2 there exists f ∈ B2 \ B1 , such that B1 ∪ f \ e ∈ B, (3) for every B1 , B2 ∈ B and e ∈ B1 \ B2 there exists f ∈ B2 \ B1 , such that B1 ∪ f \ e ∈ B and B2 ∪ e \ f ∈ B, (4) for every B1 , B2 ∈ B and A1 ⊂ B1 there exists A2 ⊂ B2 , such that B1 ∪ A2 \ A1 ∈ B and B2 ∪ A1 \ A2 ∈ B. According to Definition 0.2, family B ⊂ P(E) is a set of bases of a matroid if conditions (1) and (2) are satisfied. It is a nice exercise to show that in this case condition (3) also holds. It is called symmetric exchange property, and was discovered by Brualdi [20]. We will often refer to it in Chapter 4. Surprisingly even condition (4), known as multiple symmetric exchange property, is true (for simple proofs see [66, 114], and for more exchange properties [64, 63]). We will demonstrate usefulness of our Generalization of Seymour’s Theorem 1.7 by giving easy proofs of several basis exchange properties, in particular of multiple symmetric exchange property. The idea of the proofs is to choose a suitable list assignment which guarantees existence of a required coloring. The crucial point in our Generalization of Seymour’s Theorem is that lists may have distinct sizes. In particular, if a list of an element has size one, then the color of that element is already determined by a list assignment. Proposition 1.9. (Multiple Symmetric Exchange Property) Let B1 and B2 be bases of a matroid M . Then for every A1 ⊂ B1 there exists A2 ⊂ B2 , such that (B1 \ A1 ) ∪ A2 and (B2 \ A2 ) ∪ A1 are both bases. Proof. Observe that we can restrict our attention to the case when B1 and B2 are disjoint. Indeed, otherwise consider matroid M/(B1 ∩ B2 ), its disjoint bases B1 \ B2 , B2 \ B1 and a set A1 \ B2 ⊂ B1 \ B2 . If for them there exists an appropriate A2 , then A2 ∪ (A1 ∩ B2 ) solves the original problem. 24 1. COLORINGS OF A MATROID When bases B1 , B2 are disjoint, then restrict matroid M to their union. Let L assign list {1} to elements of A1 , list {2} to elements of B1 \ A1 and list {1, 2} to elements of B2 . Observe that for w ≡ 1 the condition (1) of Theorem 1.7 is satisfied, so we get that there is a 1-coloring from lists L. Denote by C1 elements colored with 1, and by C2 those colored with 2. Now A2 = C2 ∩ B2 is a good choice, since sets (B1 \ A1 ) ∪ A2 = C2 and (B2 \ A2 ) ∪ A1 = C1 are independent.  We formulate also a more general multiple symmetric exchange property for independent sets. We will use it in Section 2 of Chapter 2. Exactly the same proof applies. Proposition 1.10. Let I1 and I2 be independent sets in a matroid M . Then for every A1 ⊂ I1 there exists A2 ⊂ I2 , such that (I1 \ A1 ) ∪ A2 and (I2 \ A2 ) ∪ A1 are both independent sets. Multiple Symmetric Exchange Property 1.9 can be slightly generalized. Instead of having a partition of one of bases into two pieces we can have an arbitrary partition of it. We prove that for any such partition there exists a partition of the second basis which is consistent in two different ways. Proposition 1.11. Let A and B be bases of a matroid M . Then for every partition B1 ⊔ · · · ⊔ Bk = B there exists a partition A1 ⊔ · · · ⊔ Ak = A, such that (B \ Bi ) ∪ Ai are all bases for 1 ≤ i ≤ k. Proof. Analogously to the proof of Proposition 1.9 we can assume that bases A, B are disjoint (on their intersection sets Ai and Bi should coincide). When bases A and B are disjoint, then restrict matroid M to their sum. Let L assign list {1, . . . , k} \ {i} to elements of Bi , and list {1, . . . , k} to elements of A. Define weight function by w ≡ k − 1 on B and w ≡ 1 on A. Observe that the condition (1) of Theorem 1.7 is satisfied, so there exists a wcoloring from lists L. Denote by Ci elements colored with i. Now Ai = Ci ∩A is a good partition, since sets (B \ Bi ) ∪ Ai = Ci are independent.  Proposition 1.12. Let A and B be bases of a matroid M . Then for every partition B1 ⊔ · · · ⊔ Bk = B there exists a partition A1 ⊔ · · · ⊔ Ak = A, such that (A \ Ai ) ∪ Bi are all bases for 1 ≤ i ≤ k. Proof. Analogously to the proofs of previous propositions we can assume that bases A and B are disjoint. Then we restrict matroid M to A ∪ B. Let L assign list {i} to elements of Bi , and list {1, . . . , k} to elements of A. Define weight function by w ≡ 1 on B and w ≡ k − 1 on A. Observe that the condition (1) of Theorem 1.7 holds, so there is a w-coloring from lists L. Denote by Ci elements colored with i. Now Ai = A \ Ci is a good partition, since sets (A \ Ai ) ∪ Bi = Ci are independent.  CHAPTER 2 Game Colorings of a Matroid The typical problem in game coloring is to get optimal coloring of a given combinatorial structure (graph, poset, matroid, etc.) even if one of the players craftily aims to fail this task. The scheme of our games is the following. There are two players – Alice and Bob. Both know the structure of a fixed loopless matroid M . During the play, they (or one of them) are coloring properly elements of the ground set of M obeying some rules. The game ends when the whole matroid has been colored, or if they arrive to a partial coloring that cannot be further extended by making a legal move. Alice wins in the first case, so her goal is to color the whole matroid. Therefore she is called a ‘good player’. Bob’s goal is the opposite, and he is called a ‘bad guy’. The chromatic parameter of this kind of game, denoted by χgame (M ), is the minimum number of colors available for an element (usually the size of the set of colors) for which Alice has a winning strategy. We always have χgame (M ) ≥ χ(M ), since Alice’s win results in a proper coloring of M . Our goal is to bound χgame form above by the best possible function of χ. That is, to find the pointwise smallest function f : N → N, such that every loopless matroid M satisfies χgame (M ) ≤ f (χ(M )). We consider three coloring games on matroids, all of which were initially investigated for graphs (for graphs function f usually does not exist). In two cases we get optimal functions, and in the third case an almost optimal. In Section 1 we study game coloring, introduced for planar graphs by Brams (cf. [40]), and independently by Bodlaender [17]. In Theorem 2.1 we prove that every matroid M satisfies χg (M ) ≤ 2χ(M ). This improves and extends a theorem of Bartnicki, Grytczuk, and Kierstead [12], who proved that χg (M ) ≤ 3χ(M ) holds for every graphic matroid M . Moreover, in Theorem 2.5 we show that our bound is in general almost tight (up to at most one). In Section 2 we analyse on-line list coloring, introduced for graphs by Schauz [92] (see also [116]). In Theorem 2.7 we prove that the corresponding game parameter chon−line (M ) in fact equals to χ(M ). This generalizes to the on-line setting our Generalization of Seymour’s Theorem 1.7 (and in particular Seymour’s Theorem 1.5). In the last Section 3 we examine indicated coloring, introduced for graphs by Grytczuk (cf. [47]). Theorem 2.8 asserts that χi (M ) = χ(M ). We hope that our results add a new aspect of structural type, by showing that most of pathological phenomena appearing for graph coloring games 25 26 2. GAME COLORINGS OF A MATROID are no longer present in the realm of matroids. The results of consecutive Sections come from our papers [65], [66] (with Wojciech Lubawski), and [67] respectively. 1. Game Coloring In this Section we study a game-theoretic variant of the chromatic number of a matroid defined as follows. Two players, Alice and Bob, alternately color elements of the ground set E of a matroid M using a fixed set of colors C. The only rule that both players have to obey is that at any moment of the play, all elements in the same color must form an independent set (it is a proper coloring). The game ends when the whole matroid has been colored, or if they arrive to a partial coloring that cannot be further extended (what happens when trying to color any uncolored element with any possible color results in a monochromatic circuit of M ). Alice wins in the first case, while Bob in the second. The game chromatic number of a matroid M , denoted by χg (M ), is the minimum size of the set of colors C for which Alice has a winning strategy. The above game is a matroidal analog of a well studied graph game coloring, which was introduced by Brams (cf. [40]) for planar graphs with a motivation to give an easier proof of the Four Color Theorem. The game was independently reinvented by Bodlaender [17], and since then the topic developed into several directions leading to interesting results, sophisticated methods and challenging open problems (see a recent survey [11]). The first step in studying game chromatic number of matroids was made by Bartnicki, Grytczuk, and Kierstead [12], who proved that for every graphic matroid M inequality χg (M ) ≤ 3χ(M ) holds. In Theorem 2.1 we improve and extend this result by showing that for every loopless matroid M we have χg (M ) ≤ 2χ(M ). This gives a nearly tight bound, since in Theorem 2.5 we provide a class of matroids satisfying χ(Mk ) = k and χg (Mk ) = 2k −1 for every k ≥ 3. Our bounds remain true also for the fractional parameters, as well as for list version of the game chromatic number. The results of this Section come from our paper [65]. We pass to the proof of an upper bound. To achieve our goal we shall need a more general version of the matroid coloring game. Let M1 , . . . , Md be matroids on the same ground set E. The set of colors is restricted to {1, . . . , d}. As before, the players alternately color elements of E, but now for all i the set of elements colored with i must be independent in the matroid Mi . We call this game a coloring game on M1 , . . . , Md . The initial game on M with d colors coincides with the coloring game on M1 = · · · = Md = M . Theorem 2.1 (Lasoń, [65]). Let M1 , . . . , Md be matroids on the ground set E. If E has a 2-covering by sets V1 , . . . , Vd , with Vi independent in Mi , then Alice has a winning strategy in the coloring game on M1 , . . . , Md . In particular, for every loopless matroid M we have χg (M ) ≤ 2χ(M ). Proof. Let us fix a 2-covering of E by sets V1 , . . . , Vd independent in corresponding matroids. Denote by C = U1 ∪ · · · ∪ Ud the set of colored elements (at some moment of the play), where Ui is the set of elements colored with i. Alice will be coloring element e which belongs to Vi and Vj 1. GAME COLORING 27 always with color i or j, and she will try to keep the following invariant after each of her move: (♣) for each i the set Ui ∪ (Vi \ C) is independent in Mi . Observe first that if the condition (♣) holds and there is an uncolored element e (e ∈ Vi for some i), then the player can make an ‘obvious’ move, namely color e with i. After this ‘obvious’ move the condition (♣) remains true. To prove that Alice can keep the condition (♣) assume that the condition holded (remember sets Ui and C of this moment) and later Bob has colored an element e with color j. If (Uj ∪ e) ∪ (Vj \ C) is independent in Mj , then (♣) still holds and we use the above observation. When (Uj ∪ e) ∪ (Vj \ C) is dependent in Mj , then by using the augmentation axiom (see Definition 0.1) we extend the independent set Uj ∪ e from the independent set Uj ∪ (Vj \ C) in Mj . The extension equals to (Uj ∪ e) ∪ (Vj \ C) \ f for some f ∈ Vj \ C. Since sets V1 , . . . , Vd form a 2-covering we know that f ∈ Vl for some l 6= j. Now Alice has an admissible move, and her strategy is to color f with color l. It is easy to observe that after her move the condition (♣) is preserved. To get the second assertion suppose χ(M ) = k, so matroid M has a partition into independent sets V1 , . . . , Vk . We may take M1 = · · · = M2k = M and repeat twice each set Vi to get a 2-covering of E. By the first part of the assertion we infer that Alice has a winning strategy and therefore χg (M ) ≤ 2k.  As usually in this kind of games the following question seems to be natural and non trivial (for graph coloring game it was asked by Zhu [115]). Question 2.2. Suppose Alice has a winning strategy on a matroid M with k colors. Does she also have a winning strategy with l > k colors? We can define also a fractional version of the game chromatic number. Let M be a matroid on a ground set E. For given integers a and b consider a slight modification of the matroid coloring game on M . As before, Alice and Bob alternately properly color elements of E using a fixed set of b colors. The only difference is that each element has to receive a colors. The infimum of fractions ab for which Alice has a winning strategy is called the fractional game chromatic number of M , which we denote by χf,g (M ). Clearly, χf,g (M ) ≤ χg (M ) for every matroid M . Using Theorem 2.1 applied to 2b matroids equal to the a-th blow up of a loopless matroid M (each element of M is blown up to a copies, see Definition 0.19) we get that χf,g (M ) ≤ 2χf (M ). Motivated by the behaviour of fractional chromatic number we can ask two natural questions about fractional game chromatic number. The infimum in Definition 1.2 of fractional chromatic number is always reached (see the proof of Theorem 1.3). Is it also true in the case of fractional game chromatic number? The difference χ(M ) − χf (M ) is less than 1 (Theorem 1.3). Can the difference χg (M ) − χf,g (M ) be arbitrary large? Now we generalize our results to the list version. The rules of the game between Alice and Bob do not change, except that now each element has 28 2. GAME COLORINGS OF A MATROID its own list of colors. So instead of a fixed set of colors C, each element e ∈ E can be colored by Alice or Bob only with a color from its list L(e). The minimum number k for which Alice has a winning strategy for every assignment of lists of size k is called the game list chromatic number of M , and denoted by chg (M ). Clearly, chg (M ) ≥ χg (M ) for every matroid M , however the same upper bound as in Theorem 2.1 holds also for list parameter. Corollary 2.3. For every loopless matroid M we have chg (M ) ≤ 2χ(M ). Proof. Let χ(M ) = k, and suppose L : E → P({1, . . . , d}) is a list assignment of size 2k. Denote Qi = {e : i ∈ L(e)} and Mi = M |Qi . By Theorem 1.7 for ℓ ≡ 2k and w ≡ 2, there is a 2-covering of E by sets V1 , . . . , Vd with Vi is independent in the matroid Mi . So by Theorem 2.1 Alice has a winning strategy, and as a consequence chg (M ) ≤ 2k.  In the same way as we defined fractional game chromatic number one can define the fractional game list chromatic number chf,g (M ). Clearly, chf,g (M ) ≤ chg (M ) and chf,g (M ) ≥ χf,g (M ) holds for every matroid M . As before, using Theorem 2.1 it is easy to get that chf,g (M ) ≤ 2χf (M ). In view of Seymour’s Theorem 1.5 we ask the following natural question. Question 2.4. Does for every matroid M the game list chromatic number equal to the game chromatic number chg (M ) = χg (M )? Now we pass to the proof of a lower bound. We present a family of matroids Mk for k ≥ 3, with χ(Mk ) = k and χg (Mk ) ≥ 2k − 1. This slightly improves the lower bound from the paper [12] of Bartnicki, Grytczuk and Kierstead. They give an example of graphical matroids Hk with χ(Hk ) = k and χg (Hk ) ≥ 2k − 2 for each k. We believe that their family also satisfies χg (Hk ) ≥ 2k − 1 for k ≥ 3, however a proof would be much longer and more technical than in our case. Fix k ≥ 3. Let E = C ⊔ D1 ⊔ · · · ⊔ D3k(2k−1) , where C, D1 , . . . , D3k(2k−1) are disjoint sets, C = {c1,1 , . . . , ck,2k−1 } has k(2k − 1) elements, and each Di = {d1,i , . . . , dk,i } has k elements. Let Mk be a transversal matroid (see Definition 0.12) on a ground set E with multiset of subsets Ak consisting of sets D1 , . . . , D3k(2k−1) and (2k − 1) copies of E. Theorem 2.5 (Lasoń, [65]). For every k ≥ 3 matroid Mk satisfies χ(Mk ) = k and χg (Mk ) ≥ 2k − 1. Proof. To prove the first part of the assertion we can take k independent sets Vi = {ci,1 , . . . , ci,2k−1 , di,1 , . . . , di,3k(2k−1) }. Observe that also χf (Mk ) = k. To prove the second part notice that the rank of C ∪ Di equals to 2k, so if there are t elements from Di colored with i, then there are at most 2k − t elements from C colored with i. This suggests that Bob should try to color each Di with one color. Suppose the set of colors in the game is {1, . . . , h} with h ≤ 2k − 2. We will show a winning strategy for Bob. Alice will always loose the game because she will not be able to color all elements of the set C. Assume first that Alice colors only elements from the set C, and her goal is to color all of them. It is the main case to understand. Denote by di and ci 1. GAME COLORING 29 the number of elements colored with i in Di and C respectively. Bob wants to keep the following invariant after each of his move: (♥) di ≥ ci for any color i. It is easy to see that he can always do it, and in fact there is an equality di = ci for every color i. Bob just mimics Alice’s moves. Whenever she colors some c ∈ C with i he responds by coloring an element of Di with i. He can do it, because when after Alice’s move ci = di + 1 for some i, then ci + di ≤ 2k. But, then also ci + (di + 1) ≤ 2k, so elements colored with i are independent, and di + 1 ≤ k, so there was an uncolored element in Di . Observe that when Bob plays with this strategy, then for each color i we have ci + di ≤ 2k and ci ≤ di , so as a consequence ci ≤ k. This means that Alice can color only hk elements of C, thus she looses. It remains to justify that coloring elements of D1 ∪ · · · ∪ D3k(2k−1) by Alice can not help her in coloring elements of C. To see this we have to modify the invariant that Bob wants to keep. We assume k ≥ 4, because for k = 3 more careful case analysis is needed. Denote by di , fi the number of elements in Di ∪ Di+h ∪ Di+2h colored with i and with some other color respectively, and ci as before. Now the invariant that Bob wants to keep after each of his move is the following: (♠) di ≥ ci + fi or di ≥ k + 2 for any color i. Analogously to the previous case one can show that Bob can keep this invariant. He just have to obey one more rule. Whenever Alice colors an element of Di ∪ Di+h ∪ Di+2h , then Bob colors another element of this set with i always trying to keep ǫi , which is the number of sets among Di , Di+h , Di+2h with at least one element colored with i, as low as possible. This completes a description of Bob’s strategy. Observe that fi ≥ ǫi − 1. Condition (♠) gives the same consequence as (♥). Let Di be the union of those of Di , Di+h , Di+2h , which have an element colored with i. Matroid Mk |C∪Di has rank 2k − 1 + ǫi . If di ≥ ci + fi we have 2k − 1 + ǫi ≥ ci + di ≥ 2ci + fi ≥ 2ci + ǫi − 1. Otherwise, if di ≥ k + 2, then 2k + 2 ≥ 2k − 1 + ǫi ≥ ci + di ≥ ci + k + 2. Therefore there can be at most k elements in C colored with i.  Suppose that Alice wins the game on Mk (for k ≥ 3), with 2k − 1 colors, and the above strategy for Bob. Then exactly k elements of the set C will be colored with i, for every color i. Additionally there are equalities in the inequalities showing a contradiction in the proof of Theorem 2.5. So Mk can be viewed as extremal examples for which 2k − 1 colors suffice for Alice to win. Our results lead naturally to the following question. Question 2.6. Is it true that χg (M ) ≤ 2χ(M ) − 1 holds for arbitrary loopless matroid M ? 30 2. GAME COLORINGS OF A MATROID 2. On-line List Coloring In this Section we consider a coloring of a matroid from lists (as in Section 2 of Chapter 1), but in a situation when only part of information about colors in the lists is known. This can be modeled by a two person game in which Alice will be coloring properly elements of the ground set of a matroid from lists revealed during the game by Bob. As usually Alice wins if at the end of the game, when all colors in the lists are revealed, the whole matroid is colored. Let M be a loopless matroid on a ground set E, N+ = {1, 2, 3, . . . } be the set of colors and let k be a positive integer. In the first round Bob chooses an arbitrary non-empty subset B1 ⊂ E and inserts color 1 to the lists of all elements of B1 . Then Alice decides which elements with 1 on its list color with color 1. So she chooses some independent set A1 ⊂ B1 and colors its elements with 1. In the i-th round Bob picks an arbitrarily non-empty subset Bi ⊂ E and inserts color i to the lists of all elements of Bi . Then Alice chooses an independent subset Ai ⊂ Bi and colors its elements with color i. The game ends when all lists have exactly k elements. If at the end of the game the whole matroid is colored, then Alice is the winner. Bob wins in the opposite case. Let chon−line (M ) denote the minimum number k for which Alice has a winning strategy in the above game. It is called the on-line list chromatic number of M . Clearly, we have ch(M ) ≤ chon−line (M ) for every loopless matroid M . The above game is a matroidal analog of the on-line graph coloring game introduced by Schauz [92] (see also [116]). For several classes of graphs online list chromatic number equals to list chromatic number. In almost all cases proofs of upper bounds for ch(G) remain valid for chon−line (G) [94]. For example a simple modification of the argument of Thomassen [104] shows that for every planar graph G we have chon−line (G) ≤ 5. The same applies to the result of Galvin [39], which works also in the on-line setting. Schauz proved [93] even an on-line version of combinatorial Nullstellensatz of Alon [5] (we also generalize combinatorial Nullstellensatz in a different direction in [71]). It implies that if a bound ch(G) ≤ k is a consequence of using combinatorial Nullstellensatz, then also chon−line (G) ≤ k holds (for example chon−line (G) ≤ 3 for planar bipartite graph G, or asymptotic version of Dinitz’s conjecture by Häggkvist and Janssen [49]). Recently, Gutowski [48] showed the on-line version of a theorem of Alon, Tuza and Voigt [4]. Namely, that the fractional on-line list chromatic number equals to the fractional chromatic number. The only graphs for which it is known that parameters ch(G) and chon−line (G) differ is a certain family of graphs with ch(G) = 2 and chon−line (G) = 3 [116]. Surprisingly the best upper bound on chon−line (G) in terms of ch(G) is exponential as observed by Zhu [115]. It follows from a result of Alon [6]. Therefore the problem of bringing upper and lower bounds closer is one of the most challenging problems in the area of on-line list chromatic number of graphs. The main result of this Section, which comes from a joint paper with Wojciech Lubawski [66], asserts that for matroids we have equality between 2. ON-LINE LIST COLORING 31 these two parameters. The proof relies on a multiple symmetric exchange property. Actually we prove a more general Theorem 2.7 that can be seen as the on-line version of our Generalization of Seymour’s Theorem 1.7. Recall that Theorem 1.7 provides equivalent conditions for a matroid M to be w-colorable from any lists of size ℓ. Now we may generalize this notion to the on-line setting in the same way as in the game of Schauz, except that now sizes of lists are exactly ℓ and Alice wins when she ended with a w-coloring of M . In this case we say that matroid M is on-line w-colorable from lists of size ℓ if Alice has a winning strategy. Conditions from Theorem 1.7 are necessary for the above property to hold. Our aim is to prove that they are sufficient. Theorem 2.7 (Lasoń, Lubawski, [66]). (On-line Version of Generalization of Seymour’s Theorem) Let ℓ and w be list size and weight functions on a matroid M respectively. Then the following conditions are equivalent: (1) matroid M is w-colorable from lists Lℓ (e) = {1, . . . , ℓ(e)}, (2) matroid M is w-colorable from any lists of size ℓ, (3) matroid M is on-line w-colorable from lists of size ℓ. In particular, for every loopless matroid M we have chon−line (M ) = χ(M ). Proof. Implications (3) ⇒ (2) ⇒ (1) from the first part of the assertion are P obvious. We will prove (1) ⇒ (3) by induction on the number w(E) = e∈E w(e). If w(E) = 0, then w is the zero vector and the assertion holds trivially. Suppose now that w(E) ≥ 1 and the assertion of the theorem holds for all w′ with w′ (E) < w(E). Let V ⊂ E be the set B1 of elements picked by Bob in the first round of the game. So, all elements of V have color 1 in their lists. For a given subset U ⊂ E, denote by cU : E → {0, 1} the characteristic function of U , that is, cU (e) = 1 if e ∈ U , and cU (e) = 0, otherwise. To prove inductive step we have to show that there exists an independent set I ⊂ V such that matroid M is (w − cI )-colorable from lists Lℓ−cV (e) = {1, 2, . . . , ℓ(e) − cV (e)}. Then simply Alice takes A1 = I, so she colors elements of I with color 1. By condition (1) there exists a w-coloring c : E → P(N+ ) of M from lists Lℓ (e) = {1, . . . , ℓ(e)}. Let Ii denote the set of elements colored with i, for i = 1, . . . , k. Let X1 = V ∩ I1 . By Proposition 1.10 (Multiple Symmetric Exchange Property) there exists Y2 ⊆ I2 such that both I1′ = (I1 \ X1 ) ∪ Y2 and I2′′ = (I2 \ Y2 ) ∪ X1 are independent. In general let Xi = V ∩ Ii′′ . So again by Proposition 1.10 there exists Yi+1 ⊆ Ii+1 , such that sets Ii′ := ′′ := (I (Ii′′ \ Xi ) ∪ Yi+1 and Ii+1 i+1 \ Yi+1 ) ∪ Xi are independent. Let I = Xk . It is not hard to see that c′ : E ∋ e → {i : e ∈ Ii′ } is a (w − cI )-coloring from lists Lℓ−cV (e) = {1, 2, . . . , ℓ(e) − cV (e)}. The assertion of the theorem follows by induction. The second part of the assertion follows as a corollary of the first one by taking w ≡ 1 and ℓ ≡ χ(M ).  Hence it turns out that matroids are so nice structures that if it is possible to color a matroid using k colors, then it is not only possible to do it from any lists of size k, but one can even do it on-line. 32 2. GAME COLORINGS OF A MATROID 3. Indicated Coloring In this Section we consider the following asymmetric variant of the coloring game, which was originally proposed by Grytczuk for graphs. In each round of the game Alice only indicates an uncolored element of the ground set of a loopless matroid M to be colored. Then Bob chooses a color for that element from a fixed set of colors. The rule Bob has to obey is that it is a proper coloring. The goal of Alice is to achieve a proper coloring of the whole matroid. The least number of colors guaranteeing a win for Alice is denoted by χi (M ) and called the indicated chromatic number of M . Clearly χi (M ) ≥ χ(M ). The indicated variant of the usual chromatic number for graphs, denoted by χi (G), was studied by Grzesik in [47]. It exhibits rather strange behaviour. For bipartite graphs we have trivially χi (G) = χ(G), while there are 3-chromatic graphs G with χi (G) ≥ 4. We show in Theorem 2.8 that for matroids there is always an equality χi (M ) = χ(M ). The results of this Section come from our paper [67]. As before we will need a more general approach. Let M1 , . . . , Mk be a collection of matroids on the same ground set E. We consider a modification of the above game in which the set of colors is {1, . . . , k}. The other rules are the same, except that now elements of color i must form an independent set in the matroid Mi , for each i = 1, . . . , k. Theorem 2.8 (Lasoń, [67]). Let M1 , . . . , Mk be matroids on the same ground set E. Suppose that there is a partition E = V1 ∪ · · · ∪ Vk such that Vi is independent in Mi . Then Alice has a winning strategy in the generalized indicated coloring game. In particular for every loopless matroid M we have χi (M ) = χ(M ). Proof. The proof goes by induction on the number of elements of E. For a set E consisting of only one element assertion clearly holds. Thus we assume that |E| > 1 and that the assertion is true for all sets of smaller size. Denote the rank functions of matroids M1 , . . . , Mk by r1 , . . . , rk respectively. From the Matroid Union Theorem 0.20 we know that r1 (A) + · · · + rk (A) ≥ |A| for each A ⊂ E. We distinguish two cases: Case 1. There is a proper subset ∅ = 6 A ( E with the equality r1 (A) + · · · + rk (A) = |A| . Then by subtracting equality for A from the inequality for A ∪ B we get that for every subset B ⊂ E \ A holds r1 (A ∪ B) − r1 (A) + · · · + rk (A ∪ B) − rk (A) ≥ |A ∪ B| − |A| = |B| . The left hand side of this inequality is the sum of ranks of the set B in matroids M1 /A, . . . , Mk /A. Thus by the Matroid Union Theorem 0.20 the collection of matroids M1 /A, . . . , Mk /A on E \ A satisfies assumptions of the theorem. The collection M1 |A , . . . , Mk |A of matroids on A clearly also does. By inductive assumption Alice has a winning strategy in the game on matroids M1 |A , . . . , Mk |A on the set A, so she plays with this strategy. Let 3. INDICATED COLORING 33 Ui denote the set of elements colored by Bob with i after the game is finished. Now the play moves to the set E \ A, so the original game is on matroids Mi /Ui . But since the collection of matroids Mi /A satisfies assumptions of the theorem, the collection of matroids Mi /Ui also does, and Alice has a winning strategy by inductive assumption. As a result she wins the whole game on E. Case 2. All proper subsets ∅ = 6 A ( E satisfy the strict inequality r1 (A) + · · · + rk (A) > |A| . Then in the first round of the game Alice indicates an arbitrary element e ∈ E. Obviously e ∈ Vl for some l, thus Bob has an admissible move, namely he may color e with l. Suppose Bob colors e with color j. Now the original game is played on the matroids Mi |E\e for i 6= j and Mj /e, on the set E \ e. For these matroids the second condition of the Matroid Union Theorem 0.20 holds since rj (A) can be smaller only by one. Hence Alice has a winning strategy by the inductive assumption.  Part 2 Algebraic Problems on Matroids CHAPTER 3 f -Vectors of Simplicial Complexes The results of this Chapter come from our paper [68] and concern f vectors of simplicial complexes. These vectors encode the number of faces of a given dimension in the complex, and are characterized by the Kruskal-Katona inequality. A similar characterization for matroids, or more generally, for pure complexes (those with all maximal faces, called facets, of the same dimension) remains elusive. A pure simplicial complex is called extremal if its f -vector satisfies equality in the highest dimensional Kruskal-Katona inequality. That is, a pure simplicial complex of dimension d with n facets is extremal if it has the least possible number of (d − 1)-dimensional faces among all complexes with n faces of dimension d. Our main result – Theorem 3.14 asserts that every extremal simplicial complex is vertex decomposable. This extends a theorem of Herzog and Hibi asserting that extremal complexes have the Cohen-Macaulay property. It is a simple fact that the last property is implied by vertex decomposability. Moreover, our argument is purely combinatorial with the main inspiration coming from the proof of Kruskal-Katona inequality. This answers a question of Herzog and Hibi, who asked for such a proof. 1. Kruskal-Katona Inequality Simplicial complexes are one of basic mathematical structures, even more basic than matroids. They play important role in a number of mathematical areas, ranging from combinatorics to algebra and topology (see [82, 101]). Here we present only most important information from our point of view. Simplicial complexes can be viewed as discrete objects or as their geometric realizations. Our approach is combinatorial, thus we introduce abstract simplicial complex. Definition 3.1. (Simplicial Complex) A simplicial complex ∆ on a finite set V , called the set of vertices, is a set of subsets of V , called faces (or simplices), satisfying property that a subset of a face is a face. Cardinality of a face σ minus one is called its dimension, and it is denoted by dim σ. The dimension of a simplicial complex is the biggest dimension of its face. Maximal faces are called facets, and a simplicial complex is called pure if all its facets are of the same dimension. Matroids are regarded as simplicial complexes via their set of independent sets. Clearly, a matroid is a pure simplicial complex (see Definition 0.1), since all bases are of equal cardinality. Let ∆ be a simplicial complex of dimension d. Let fk be the number of k-dimensional faces of ∆, for k = −1, . . . , d. 37 38 3. f -VECTORS OF SIMPLICIAL COMPLEXES Definition 3.2. (f -Vector) The f -vector of a simplicial complex ∆ is the sequence (f0 , . . . , fd ). Observe that a simplicial complex with one face of dimension k − 1 has at least k faces of dimension k − 2. One of most natural questions concerning simplicial complexes is to determine the minimum number of (k − 2)dimensional faces in a simplicial complex with n faces of dimension k − 1. A slightly more general problem is to characterize sequences which are f -vectors of some simplicial complex. Both problems were resolved independently by Kruskal [62] and Katona [60] in 1960’s. For a nonnegative integer k, they enlisted all k-element subsets of integers in the following order, called the squashed order : A < B if max(A \ B) < max(B \ A). Let Sk (n) be the set of first n sets in this list. For a given set U of k-element sets, denote by δU the set of all (k − 1)-element subsets of sets from U. Note that δSk (n) = Sk−1 (m) for some m, that is subsets of fixed cardinality of sets from a prefix in squashed order form also a prefix in squashed order. The Kruskal-Katona theorem reads as follows. Theorem 3.3 (Kruskal, [62]; Katona, [60]). (Kruskal-Katona Inequality) Let U be a family of n sets of size k. Then |δU| ≥ |δSk (n)| . This result was further generalized by Clements and Lindström in [21]. Daykin [28, 29] gave two simple proofs, and later Hilton [53] gave another one. There is even an algebraic proof, for which we refer the reader to [9]. The cardinality of the set δSk (n) may be easily determined. For a positive integer k, each positive integer n can be uniquely expressed as       ak ak−1 at n= + + ··· + , k k−1 t with 1 ≤ t ≤ at and at < · · · < ak . Denote by δk−1 (n) = |δSk (n)|, we have       ak ak−1 at δk−1 (n) = + + ··· + . k−1 k−2 t−1 As one can easily see Kruskal-Katona Theorem 3.3 gives the full characterization of f -vectors of simplicial complexes. Corollary 3.4. A sequence (f0 , . . . , fd ) is the f -vector of a simplicial complex if and only if fk−1 ≥ δk (fk ) for all k = 1, . . . , d. A lot of work has been done to characterize f -vectors of pure simplicial complexes, but only some partial results are known (see [18, 88]). It is believed that they are too complicated to have an easy description. Even shape of f -vectors of matroids, which are a special class of pure simplicial complexes, is a mystery. There is no guess for a classification, only some conditions are expected to be necessary. Welsh [109] conjectured that f -vector of a matroid is unimodal, that is fk grows with k, and after reaching maximum value decreases (f -vector of a general pure simplicial complex need not be unimodal [88]). Mason [78] strengthen it by conjecturing that f -vector of a matroid is log-concave, that is fk2 ≥ fk−1 fk+1 for 2. VERTEX DECOMPOSABLE SIMPLICIAL COMPLEXES 39 every k. Recently there was a breakthrough made by the seminal paper of Huh [55]. He proves that the coefficients of the characteristic polynomial of a matroid representable over a field of characteristic zero form a sign-alternating log-concave sequence. The proof uses advanced tools of algebraic geometry. Soon after it was generalized by Huh and Katz [57] to representable matroids. Using this result Lenz [74] proved Mason’s conjecture for representable matroids. See also [56, 27, 99, 23] for conjectures and results concerning h-vector of a matroid – another fundamental invariant of a simplicial complex encoding the number of faces in each dimension. There are also classes of pure simplicial complexes for which a classification of f -vectors exists. For example, in the case of pure clique-complexes (faces are cliques contained in the graph) of chordal graphs (see [52]). Some research in the spirit of Kruskal-Katona theorem has been done also for other structures than simplicial complexes. Just to mention the result of Stanley (see [100, 38]), who proved using toric geometry a classification for convex polytopes. 2. Vertex Decomposable Simplicial Complexes Definition 3.5. (Link of a Face) Let σ be a face of a simplicial complex ∆. The simplicial complex {τ ∈ ∆ : τ ∩ σ = ∅ and τ ∪ σ ∈ ∆} is called the link of σ in ∆, and denoted by lk∆ σ. For a vertex x of ∆ by ∆ \ x we will denote the simplicial complex {τ ∈ ∆ : x ∈ / τ }. The definition of vertex decomposability is inductive. This notion was invented by Provan and Billera [89]. Definition 3.6. (Vertex Decomposable Pure Simplicial Complex) A pure simplicial complex ∆ is vertex decomposable if one of the following conditions is satisfied: (1) ∆ consists of empty set, (2) ∆ consists of a single vertex, (3) for some vertex x both lk∆ {x} and ∆ \ x are pure and vertex decomposable. In particular, a zero dimensional complex is vertex decomposable. It is a special case of a much more general fact. Proposition 3.7. Every matroid is vertex decomposable. Proof. The proof goes by induction on the size of the ground set. If |E| = 1, then matroid M consists of a vertex or of the empty set. If |E| > 1, then let x be any element of E. Now both lkM {x} = M/x and M \ x are matroids on the ground set E\x, hence the assertion follows by induction.  We are ready to define the Stanley-Reisner ring of a simplicial complex. Definition 3.8. (Stanley-Reisner Ring) Let ∆ be a simplicial complex on the set of vertices {1, . . . , n}, and let K be a field. The Stanley-Reisner ring of ∆ (or the face ring) is K[∆] := K[x1 , . . . xn ]/I, where I is the ideal generated by all square-free monomials xi1 · · · xil for which {i1 , . . . , il } is not a face in ∆ (it is enough to consider only circuits of ∆). 40 3. f -VECTORS OF SIMPLICIAL COMPLEXES When we say that a pure simplicial complex is Cohen-Macaulay we always mean that its Stanley-Reisner ring has this property (see [10] for the definition of Cohen-Macaulay rings). Cohen-Macaulay simplicial complexes have also combinatorial description. Namely, Reisner [91] proved that a pure simplicial complex ∆ is Cohen-Macaulay over K if and only if for any face σ ∈ ∆ all except top homologies of lk∆ σ over K are zero. For CohenMacaulay complexes we have a full classification of h-vectors. Macaulay [76] proved that they consist exactly of O-sequences (see e.g. [18] for a definition of O-sequence). It is a folklore result that vertex decomposability is a stronger property than being Cohen-Macaulay (we refer the reader e.g. to [14], see also [72, Section 3.2] for an interesting method of counting homologies). Proposition 3.9. Every vertex decomposable simplicial complex is CohenMacaulay over any field. Due to Propositions 3.7, 3.9 and a theorem of Reisner for σ = ∅ we get the following interesting property of matroids (see also [1, 13]). Corollary 3.10. All homologies of a matroid except top ones are zero. 3. Extremal Simplicial Complexes Definition 3.11. (Extremal Simplicial Complex) A pure simplicial complex ∆ of dimension d and n facets is called extremal if it has the least possible number of (d − 1)-dimensional faces among all complexes with n faces of dimension d. We get the following as a consequence of Kruskal-Katona Theorem 3.3. Corollary 3.12. A pure simplicial complex ∆ of dimension d > 0 with f -vector (f0 , . . . , fd ) is extremal if and only if fd−1 = δd (fd ). In particular, a zero dimensional complex is extremal, since all such complexes have exactly one (−1)-dimensional face, namely the empty set. Herzog and Hibi proved using algebraic methods (in particular results from [9, 31]) the following theorem. Theorem 3.13 (Herzog, Hibi, [50]). Every extremal simplicial complex is Cohen-Macaulay over any field. They also asked for a combinatorial proof. We give it by proving by only combinatorial means that an extremal simplicial complex is vertex decomposable. Due to Proposition 3.9 vertex decomposable complex is CohenMacaulay over any field. Theorem 3.14 (Lasoń, [68]). Every extremal simplicial complex is vertex decomposable. Proof. For a better understanding of extremal simplicial complexes we will use Hilton’s idea from his proof [53] of Kruskal-Katona Theorem 3.3. First we define sets similar to Sk (n). Let Ski (n) be the first n k-element subsets of integers in the squashed order (A < B if max(A\B) < max(B\A)) which do not contain an element i. We will denote the set {i ∪ A : A ∈ U} by i ⊎ U. 3. EXTREMAL SIMPLICIAL COMPLEXES 41 S Let U be a family of n k-element sets. Denote by V = U the underlying set of vertices, and by v its cardinality. For i ∈ V , let Bi = {A ∈ U : i ∈ / A}, Ci = {A \ i : i ∈ A ∈ U}, and let bi , ci be the respective cardinalities. Note that always ci 6= 0. We want to find a vertex i such that |δBi | > |Ci |. Lemma 3.15. Either there exists a vertex i such that |δBi | > |Ci |, or U consists of all possible k-element subsets of V . Proof. We are going to count the sum of cardinalities of both sets when i runs over all elements of V . We have X X |δBi | ≥ kn(v − k)/(v − k) = kn = |Ci | , i∈V i∈V since on the left hand side each A ∈ U gives exactly k distinct sets in its boundary, and each A belongs to exactly v − k sets Bi . Some sets in boundaries of sets from Bi can be the same, but their number is at most (v − 1) − (k − 1) = v − k. On the right hand side each A ∈ U is counted k times. Hence we can find a desired i, or the above bounds are tight. In the latter case, if A ∈ δBi , then all v − k possibilities of completing it to a k-element set not containing i have to be in U. If v > k + 1 this means that U consists of all possible k-element subsets of V . It is because from any set in U we can delete any element and insert any other. If v = k + 1 or v = k the statement is easy to verify by hand.  Lemma 3.16. If ∆ is an extremal simplicial complex of positive dimension, then there exists a vertex x such that both lk∆ {x} and ∆\x are extremal. Proof. Let ∆ be an extremal simplicial complex of dimension d−1 > 0, and let U be the set of all d-element faces in ∆. If U consists of all possible d-element subsets of a given v-element set, then the assertion of the lemma is clearly true. We can take any vertex. Otherwise, due to Lemma 3.15, there exists i ∈ V such that |δBi | > |Ci |. We have that δU = δBi ∪ Ci ∪ (i ⊎ δCi ). Since δBi and i ⊎ δCi are disjoint, it follows that |δU| ≥ |δBi | + |i ⊎ δCi | > |Ci | + |i ⊎ δCi | . So, by Kruskal-Katona Theorem 3.3, (♦) i |δU| ≥ δSdi (bi ) + i ⊎ δSd−1 (ci ) , and (♥) i i |δU| > Sd−1 (ci ) + i ⊎ δSd−1 (ci ) . i i Notice that δSdi (bi ) = Sd−1 (e) for some e, thus δSdi (bi ) and Sd−1 (ci ) are i i comparable in the inclusion order. If δSd (bi ) ⊂ Sd−1 (ci ), then by (♥) we get i i i |δU| > Sd−1 (ci ) + i ⊎ δSd−1 (ci ) = δ(Sdi (bi ) ∪ (i ⊎ Sd−1 (ci ))) , But this contradicts the assumption that ∆ is extremal, since the complex i (ci )) has also bi + ci = |U| maximal faces. generated by sets Sdi (bi ) ∪ (i ⊎ Sd−1 i i Thus δSd (bi ) ⊃ Sd−1 (ci ), and by (♦) we obtain that i i (ci ))) . (ci ) = δ(Sdi (bi ) ∪ (i ⊎ Sd−1 |δU| ≥ δSdi (bi ) + i ⊎ δSd−1 42 3. f -VECTORS OF SIMPLICIAL COMPLEXES i With equality if and only if Ci ⊂ δBi , |δBi | = δSdi (bi ) , |δCi | = δSd−1 (ci ) . Complex ∆ is extremal, so there is an equality. We get that Ci ⊂ δBi and (Bi ), (Ci ) are extremal, where (A) denotes the simplicial complex generated by the set of faces A. Observe that lk∆ {i} = (Ci ) and ∆ \ i = (Bi ). The first equality is obvious, while the second is not as clear. If σ = {v1 , . . . , vk } is a face in ∆ \ i, then it is a subface of some facet F = {v1 , . . . , vd }. If i does not belong to F then F ∈ (Bi ) and so σ does. Otherwise, F \ {i} ∈ Ci ⊂ δBi , so F \ i ∪ j ∈ Bi for some j and as a consequence σ ∈ (Bi ). Now take x = i to get the assertion.  The final step of the proof goes by induction on the dimension d of ∆ and secondly on the number of facets. If d = 0, then ∆ consists of points and almost by the definition it is vertex decomposable. When d > 0, then by Lemma 3.16 there exists a vertex x, such that both complexes lk∆ {x} and ∆ \ x are extremal. The first is of lower dimension and the second either has the same dimension as ∆ but fewer facets, or it has smaller dimension. By inductive hypothesis, both lk∆ {x} and ∆ \ x are vertex decomposable, as a consequence ∆ also is.  We provide an example which shows that the opposite implication to the one from Theorem 3.14 does not hold. Example 3.17. A path on n ≥ 3 vertices is a vertex decomposable simplicial complex which is not extremal. Theorem 3.14 can be treated as a characterization of f -vectors implying vertex decomposability. It is best possible in the following sense. Let ∆ be a pure simplicial complex of dimension d > 0 with f -vector (f0 , . . . , fd ), and with fd−1 = δd (fd ) + c, c ∈ N. Due to Corollary 3.12 Theorem 3.14 asserts that if c = 0, then ∆ is vertex decomposable. The example below shows that for c = 1 it is already not true. Example 3.18. Let ∆ be a pure simplicial complex of dimension d with exactly two disjoint facets. If U is the set of facets of ∆, then |δU| = 2d + 2, while δd (2) = 2d + 1. Complex ∆ is not connected, thus due to a theorem of Reisner it is not Cohen-Macaulay over any field, and as a consequence also not vertex decomposable. Unfortunately, there is no inclusion between matroids and extremal simplicial complexes. The simplicial complex generated by S2 (4) is extremal, but it is not a matroid. The matroid U1,1 ⊔ U1,3 (see Definitions 0.18, 0.6) is not an extremal simplicial complex. CHAPTER 4 The Conjecture of White Describing minimal generating set of a toric ideal is a well-studied and difficult problem in commutative algebra. In 1980 White conjectured [112] that for every matroid M its toric ideal IM is generated by quadratic binomials corresponding to symmetric exchanges. Despite big interest in algebraic combinatorics community the conjecture is still wide open. It is confirmed only for some very special classes of matroids like graphic matroids [16], sparse paving matroids [19], lattice path matroids (a subclass of transversal matroids) [96], or matroids of rank at most three [59]. In Theorem 4.3 we prove White’s conjecture for strongly base orderable matroids, which is already a large class of matroids. In particular it contains transversal matroids, for which it was known only that the toric ideal IM is generated by quadratic binomials [22], and also laminar matroids. However, the most important of our results is Theorem 4.4 – White’s conjecture up to saturation. We prove that the saturation with respect to the so-called irrelevant ideal of the ideal generated by quadratic binomials corresponding to symmetric exchanges equals to the toric ideal of a matroid. Our result has even more natural formulation in the language of algebraic geometry. Namely, it means that both ideals define the same projective scheme, while the conjecture asserts that their affine schemes are equal. To our knowledge it is the only result concerning White’s conjecture, which is valid for all matroids. In the last Section 4 we discuss the original formulation of the conjecture due to White, which was combinatorial. It asserts that if two multisets of bases of a matroid have equal union (as a multiset), then one can pass between them by a sequence of single element symmetric exchanges. In fact White defined three properties of a matroid of growing difficulty, and conjectured that all matroids satisfy them. We reveal relations between these properties, present a beautiful alternative combinatorial reformulation by Blasiak [16], and also mention how to extend our results to discrete polymatroids. The results of this Chapter come from the joint work [69] with Mateusz Michałek. 1. Introduction White’s conjecture can be expressed in the language of combinatorics, algebra and geometry. We decided to present the conjecture and our results in the algebraic language, since then they have the most natural and compact formulation. However ideas of our proofs are purely combinatorial. 43 44 4. THE CONJECTURE OF WHITE Let M be a matroid on a ground set E with the set of bases B ⊂ P(E). For a fixed field K, let SM := K[yB : B ∈ B] be a polynomial ring. There is a natural K-homomorphism Y ϕM : SM ∋ yB −→ xe ∈ K[xe : e ∈ E]. e∈B Definition 4.1. (Toric Ideal of a Matroid) The toric ideal of a matroid M , denoted by IM , is the kernel of the map ϕM . Suppose that a pair of bases D1 , D2 is obtained from a pair of bases B1 , B2 by a symmetric exchange. That is, D1 = B1 ∪f \e and D2 = B2 ∪e\f for some e ∈ B1 and f ∈ B2 . Then clearly the quadratic binomial yB1 yB2 − yD1 yD2 belongs to the ideal IM . We say that such a binomial corresponds to symmetric exchange. Conjecture 4.2 (White, [112]). For every matroid M its toric ideal IM is generated by quadratic binomials corresponding to symmetric exchanges. Since every toric ideal is generated by binomials it is not hard to rephrase the above conjecture in the combinatorial language. It asserts that if two multisets of bases of a matroid have equal union (as a multiset), then one can pass between them by a sequence of single element symmetric exchanges. In fact this is the original formulation due to White (cf. [103]). We immediately see that the conjecture does not depend on the field K. Actually White stated a bunch of conjectures of growing difficulty. The weakest asserts that the toric ideal IM is generated by quadratic binomials, the second one is Conjecture 4.2, and the most difficult is an analog of Conjecture 4.2 for the noncommutative polynomial ring SM . However, Conjecture 4.2 attracted the most attention because of its consequences in commutative algebra. We will discuss the original conjectures of White and show more subtle relations between them in Section 4. The most significant result towards Conjectue 4.2 was obtained by Blasiak in 2008 [16], he confirmed it for graphical matroids. In 2010 Kashiwabara [59] checked the case of matroids of rank at most 3. In 2011 Schweig [96] proved the case of lattice path matroids, which is a subclass of transversal matroids. Recently, in 2013 Bonin [19] solved the case of sparse paving matroids (all circuits have r(M ) + 1 or r(M ) elements). For transversal matroids it was known only, due to Conca in 2007 [22], that the toric ideal IM is generated by quadratic binomials. 2. White’s Conjecture for Strongly Base Orderable Matroids Here we prove White’s conjecture for strongly base orderable matroids (see Definition 0.11), which is already a large class of matroids. An advantage of this class of matroids is that it is characterized by a certain property instead of a specific presentation. Thus it is easier to check if a matroid is strongly base orderable or not. In particular we get that White’s conjecture is true for transversal matroids, and also for laminar matroids. Our argument uses an idea from the proof [95] of a theorem of Davies and McDiarmid [26]. It asserts that if the ground set E of two strongly base 2. WHITE’S CONJECTURE FOR STRONGLY BASE ORDERABLE MATROIDS 45 orderable matroids can be partitioned into n bases in each of them, then there exists also a common partition. Theorem 4.3 (Lasoń, Michałek, [69]). If M is a strongly base orderable matroid, then the toric ideal IM is generated by quadratic binomials corresponding to symmetric exchanges. Proof. The ideal IM , as every toric ideal, is generated by binomials. Thus it is enough to prove that quadratic binomials corresponding to symmetric exchanges generate all binomials of IM . Fix n ≥ 2. We are going to show by decreasing induction on the overlap function n X d(yB1 · · · yBn , yD1 · · · yDn ) = max Bi ∩ Dπ(i) π∈Sn i=1 that a binomial yB1 · · · yBn − yD1 · · · yDn ∈ IM is generated by quadratic binomials corresponding to symmetric exchanges. Clearly, the biggest value of d is r(M )n. If d(yB1 · · · yBn , yD1 · · · yDn ) = r(M )n, then there exists a permutation π ∈ Sn such that for each i holds Bi = Dπ(i) , so yB1 · · · yBn −yD1 · · · yDn = 0. Suppose the assertion holds for all binomials with the overlap function greater than d < r(M )n. Let yB1 · · · yBn − yD1 · · · yDn be a binomial of IM with the overlap function equal to d. Without loss of generality we can assume that the identity permutation realizes the maximum in the definition of the overlap function. Then there is i for which there exists e ∈ Bi \ Di . Clearly, yB1 · · · yBn −yD1 · · · yDn ∈ IM if and only if B1 ∪· · ·∪Bn = D1 ∪· · ·∪ Dn as multisets. Thus there exists j 6= i such that e ∈ Dj \ Bj . Without lost of generality we can assume that i = 1, j = 2. Since M is a strongly base orderable matroid, there exist bijections πB : B1 → B2 and πD : D1 → D2 with the multiple symmetric exchange property. We can assume that πB is the identity on B1 ∩ B2 , and similarly that πD is the identity on D1 ∩ D2 . Let G be a graph on a vertex set B1 ∪ B2 ∪ D1 ∪ D2 with edges {b, πB (b)} for all b ∈ B1 \ B2 and {d, πD (d)} for all d ∈ D1 \ D2 . Graph G is bipartite since it is a sum of two matchings. Split the vertex set of G into two independent (in the graph sense) sets S and T . Define: B1′ = (S ∩ (B1 ∪ B2 )) ∪ (B1 ∩ B2 ), B2′ = (T ∩ (B1 ∪ B2 )) ∪ (B1 ∩ B2 ), D1′ = (S ∩ (D1 ∪ D2 )) ∪ (D1 ∩ D2 ), D2′ = (T ∩ (D1 ∪ D2 )) ∪ (D1 ∩ D2 ). By the multiple symmetric exchange property of πB sets B1′ , B2′ are bases, and they are obtained from the pair B1 , B2 by a sequence of symmetric exchanges. Therefore the binomial yB1 yB2 yB3 · · · yBn − yB1′ yB2′ yB3 · · · yBn is generated by quadratic binomials corresponding to symmetric exchanges. Analogously the binomial yD1 yD2 yD3 · · · yDn −yD1′ yD2′ yD3 · · · yDn is generated by quadratic binomials corresponding to symmetric exchanges. Moreover, since S and T are disjoint we have that d(yB1′ yB2′ yB3 · · · yBn , yD1′ yD2′ yD3 · · · yDn ) > d(yB1 yB2 · · · yBn , yD1 yD2 · · · yDn ). Thus by inductive assumption yB1′ yB2′ yB3 · · · yBn − yD1′ yD2′ yD3 · · · yDn is generated by quadratic binomials corresponding to symmetric exchanges. Composing these three facts we get the inductive assertion.  46 4. THE CONJECTURE OF WHITE 3. White’s Conjecture up to Saturation Let m be the ideal generated by all variables in SM (so-called irrelevant ideal). Recall that I : m∞ = {a ∈ SM : amn ⊂ I for some n ∈ N} is called the saturation of an ideal I with respect to the ideal m. Let JM be the ideal generated by quadratic binomials corresponding to symmetric exchanges. Clearly, JM ⊂ IM and White’s conjecture asserts that JM and IM are equal. We will prove a slightly weaker property, namely that ideals JM and IM are equal up to saturation, that is JM : m∞ = IM : m∞ (cf. [81, Section 4]). In fact the ideal IM , as a prime ideal, is saturated so IM : m ∞ = IM . Now we will express our result in the language of algebraic geometry. From its point of view it is natural to compare schemes (main objects of study in algebraic geometry) defined by ideals instead of ideals themself. A homogeneous ideal (JM and IM are homogeneous, because every basis has the same number of elements) defines two kinds of schemes – affine and projective. Affine schemes are equal if and only if the ideals are equal. Thus White’s conjecture asserts equality of affine schemes of JM and IM . Whereas ideals define the same projective scheme if and only if their saturations with respect to the irrelevant ideal are equal. That is, we prove that for every matroid M projective schemes Proj(SM /JM ) and Proj(SM /IM ) are equal. Additionally, since both JM and IM are contained in m, this implies that both ideals have equal radicals, which in other words means they have the same affine set of zeros. The toric variety Proj(SM /IM ) (we refer the reader to [38, 24] for background of toric geometry) has been already studied before. White proved that it is projectively normal [111]. Theorem 4.4 (Lasoń, Michałek, [69]). White’s conjecture is true up to saturation. That is, for every matroid M we have JM : m∞ = IM . Proof. Since JM ⊂ IM we get that JM : m∞ ⊂ IM : m∞ = IM . To prove the opposite inclusion IM ⊂ JM : m∞ it is enough to consider binomials since IM , as any toric ideal, is generated by binomials. To prove that a binomial yB1 . . . yBn − yD1 . . . yDn ∈ IM belongs to JM : m∞ it is enough to show that for each basis B ∈ B we have (r(M )−1)n yB (yB1 · · · yBn − yD1 · · · yDn ) ∈ JM . It is because then (yB1 · · · yBn − yD1 · · · yDn )m(r(M )−1)n|B| ∈ JM . Let B ∈ B be a basis. The polynomial ring SM has a natural grading given by degree function deg(yB ′ ) = 1, for each variable yB ′ . We define the B-degree by degB (yB ′ ) = |B ′ \ B|, and extend this notion also to bases degB (B ′ ) = |B ′ \ B|. Notice that the ideal IM is homogeneous with respect to both gradings. Additionally B-degree of yB is zero, thus multiplying by yB does not change B-degree of a polynomial. Observe that if degB (B ′ ) = 1, then B ′ differs from B only by a single element. We call such basis, and a corresponding variable, balanced. A monomial or a binomial is called balanced if it contains only balanced variables. We will prove by induction on n the following claim. As argued before this will finish the proof. 3. WHITE’S CONJECTURE UP TO SATURATION 47 Claim 4.5. If a binomial b ∈ IM has B-degree equal to n, then we have degB (b)−deg(b) yB b ∈ JM . If degB (b) − deg(b) deg(b)−degB (b) divides b, yB < 0, then the statement of the claim means that and that the quotient belongs to JM . If n = 0, then the claim is obvious, since 0 is the only binomial in IM with B-degree equal to 0. Suppose n > 0. It is easier to work with balanced variables, therefore we provide the following lemma. Lemma 4.6. For every basis B ′ ∈ B there exist bases B1 , . . . , BdegB (B ′ ) which are balanced and satisfy degB (B ′ )−1 yB yB ′ − yB1 · · · yBdeg ′ B (B ) ∈ JM . Proof. The proof goes by induction on degB (B ′ ). If degB (B ′ ) = 0, 1, then the assertion is clear. If degB (B ′ ) > 1, then for e ∈ B ′ \ B from the symmetric exchange property applied to B, B ′ there exists f ∈ B such that both B1 = B ∪ e \ f and B ′′ = B ′ ∪ f \ e are bases. Now degB (B1 ) = 1 and degB (B ′′ ) = degB (B ′ ) − 1. Applying inductive assumption for B ′′ we get balanced bases B2 , . . . , BdegB (B ′ ) with degB (B ′ )−2 yB yB ′′ − yB2 · · · yBdeg ′ B (B ) ∈ JM . Hence, since yB yB ′ − yB1 yB ′′ corresponds to symmetric exchange, we get degB (B ′ )−1 yB yB ′ − yB1 · · · yBdeg degB (B ′ )−2 +yB1 (yB degB (B ′ )−2 ′ B (B ) = yB yB ′′ − yB2 · · · yBdeg (yB yB ′ − yB1 yB ′′ )+ ′ B (B ) ) ∈ JM .  deg (B ′ )−deg(B ′ ) yB ′ of a monoLemma 4.6 allows us to change each factor yB B degB (m)−deg(m) m to a product of balanced variables. Notice that the mial yB B-degree is preserved. Hence Claim 4.5 reduces to the following claim. Claim 4.7. If b = yB1 · · · yBn − yD1 · · · yDn ∈ IM is a balanced binomial of B-degree equal to n, then b ∈ JM . With a balanced monomial m = yB1 · · · yBn we associate a bipartite multigraph G(m) with vertex classes B and E \ B (where E is the ground set of matroid M ). For each variable yBi of monomial m (Bi = B ∪ e \ f for f ∈ B, e ∈ E \ B) we put an edge {e, f } in G(m). Let b = yB1 · · · yBn −yD1 · · · yDn ∈ IM be a balanced binomial of B-degree equal to n. Observe that b belongs to IM if and only if each vertex from E has the same degree with respect to G(yB1 · · · yBn ) and G(yD1 · · · yDn ). Thus we can apply the following lemma, which is obvious from graph theory. Lemma 4.8. Let G and H be bipartite multigraphs with the same vertex classes. Suppose that each vertex has the same degree with respect to G and H. Then the symmetric difference of multisets of edges of G and of H can be partitioned into alternating cycles. That is simple cycles of even length, with two consecutive edges belonging to different graphs. 48 4. THE CONJECTURE OF WHITE We get a simple cycle with vertices f1 , e1 , f2 , e2 , . . . , fr , er consecutively. So for each i we have Bi′ = B ∪ ei \ fi ∈ B and Di′ = B ∪ ei−1 \ fi ∈ B (with a cyclic numeration modulo r). Then yB1′ · · · yBr′ divides yB1 · · · yBn and yD1′ · · · yDr′ divides yD1 · · · yDn . Suppose r < n. The balanced binomial b′ = yB1′ · · · yBr′ − yD1′ · · · yDr′ belongs to IM and has B-degree less than n. From inductive assumption we get that b′ ∈ JM . There are balanced monomials m1 , m2 , m3 such that b = yB1 · · · yBn − yD1 · · · yDn = m1 b′ − yD1′ · · · yDr′ (m2 − m3 ) and m2 −m3 ∈ IM . Balanced binomial b′′ = m2 −m3 ∈ IM has B-degree less than n, so by inductive assumption b′′ ∈ JM , and as a consequence b ∈ JM . Suppose now that r = n. We can assume that E = {f1 , e1 . . . , fn , en }, since otherwise we can contract B \ {f1 , . . . , fn } and restrict our matroid to the set {f1 , e1 . . . , fn , en }. Obviously the property of being generated extends from such a minor to a matroid. We say that a monomial m2 is achievable from m1 if m1 − m2 ∈ JM . In this situation we say also that variables of m2 are achievable from m1 . Observe that if there is a variable different from yB that is achievable from both monomials yB1 · · · yBn and yD1 · · · yDn , then the assertion follows by induction. Indeed, if a variable yB ′ is achievable from both, then there are monomials m1 , m2 such that b = (yB1 · · · yBn − yB ′ m1 ) + (yB ′ m2 − yD1 · · · yDn ) + yB ′ (m1 − m2 ). Binomial b′ = m1 − m2 ∈ IM has B-degree less than n, thus by inductive deg (b′ )−deg(b′ ) ′ b ∈ JM . Hence b′ ∈ JM because assumption yB B degB (b′ ) − deg(b′ ) = degB (b) − deg(b) − degB (yB ′ ) + deg(yB ′ ) ≤ 0. Suppose contrary – no variable different from yB is achievable from both monomials of b, and we will reach a contradiction. The forthcoming part of the proof illustrates a sentence of Sturmfels “Combinatorics is a nanotechnology of mathematics”. For k, i ∈ Zn (cyclic group ({1, . . . , n}, +(mod n)) we define: Ski := B ∪ {ek , ek+1 , . . . , ek+i−1 } \ {fk , fk+1 , . . . , fk+i−1 }, Tki := B ∪ {ek−1 , ek , . . . , ek+i−2 } \ {fk , fk+1 , . . . , fk+i−1 }, Uki := B ∪ {ek−i } \ {fk }. Sets Ski and Tki differ only on the set {e1 , . . . , en } by a shift by one. Notice that Skn = Tkn′ for arbitrary k, k ′ ∈ Zn and Sk1 = Uk0 = Bk′ , Tk1 = Uk1 = Dk′ . Hence m1 = yB1′ · · · yBn′ = yS11 · · · ySn1 and m2 = yD1′ · · · yDn′ = yT11 · · · yTn1 are monomials of b. Lemma 4.9. Suppose that for a fixed 0 < i < n and every k ∈ Zn the following conditions are satisfied: (1) sets Ski and Tki are bases, Q i−1 yS i j6=k,...,k+i−1 ySj1 is achievable from m1 , (2) monomial yB k Q i−1 yT i j6=k,...,k+i−1 yT 1 is achievable from m2 . (3) monomial yB k j Then for every k ∈ Zn neither of the sets Uk−i , Uki+1 is a basis. 3. WHITE’S CONJECTURE UP TO SATURATION 49 Proof. Suppose contrary that Uk−i is a basis. Then yS 1 yS i −yT i yU −i k k+1 k+1 k by the definition belongs to JM . Thus, by the assumption, variable yT i k+1 would be achievable from both m1 and m2 , which is a contradiction. The  argument for Uki+1 is analogous, it differs by a shift by one. Lemma 4.10. Suppose that for a fixed 0 < i < n and every k ∈ Zn the following conditions are satisfied: (1) set Ski is a basis, Q i−1 yS i j6=k,...,k+i−1 ySj1 is achievable from m1 , (2) monomial yB k (3) set Uk−j is not a basis for any 0 < j ≤ i. i y Then for every k ∈ Zn the set Ski+1 is a basis and yB Ski+1 a monomial achievable from m1 . Q j6=k,...,k+i ySj1 is i it Proof. From the symmetric exchange property for ek ∈ Sk1 \ Sk+1 i i i ∪ ek \ x and = Sk+1 \ Sk1 such that S̃k+1 follows that there exists x ∈ Sk+1 S̃k1 = Sk1 ∪ x \ ek are also bases. Thus x ∈ {fk , ek+1 , ek+2 , . . . , ek+i }. Notice that if x = ek+j for some j, then S̃k1 = Uk−j contradicting condition (3). i = Ski+1 and S̃k1 = B. In particular the binomial Thus x = fk . Hence S̃k+1 yS i yS 1 − yB yS i+1 belongs to JM (condition (1) guarantees that variable k k+1 k  yS i exists). Thus the assertion follows from condition (2). k+1 Analogously we get the following shifted version of Lemma 4.10. Lemma 4.11. Suppose that for a fixed 0 < i < n and every k ∈ Zn the following conditions are satisfied: (1) set Tki is a basis, Q i−1 yT i j6=k,...,k+i−1 yT 1 is achievable from m2 , (2) monomial yB k j (3) set Ukj+1 is not a base for any 0 < j ≤ i. i y Then for every k ∈ Zn the set Tki+1 is a basis and yB Tki+1 a monomial achievable from m2 . Q j6=k,...,k+i yTj1 is We are ready to reach a contradiction by an inductive argument. Firstly we verify that for i = 1 the assumptions of Lemma 4.9 are satisfied. Suppose now that for some 1 < i < n the assumptions of Lemma 4.9 are satisfied for every 1 ≤ j ≤ i. Then, by Lemma 4.9 the assumptions of both Lemma 4.10 and Lemma 4.11 are satisfied for every 1 ≤ j ≤ i. Thus by the assertions of Lemmas 4.10 and 4.11, the assumptions of Lemma 4.9 are satisfied for all 1 ≤ j ≤ i+ 1. We obtain that the assumptions and the assertions of Lemmas 4.9, 4.10 and 4.11 are satisfied for every 1 ≤ i < n. For i = n − 1 we obtain n−1 n−1 n yT1n is achievable from both m1 and m2 , y S1 = y B that the monomial yB this gives a contradiction.  Corollary 4.12. White’s conjecture holds for matroid M if and only if the ideal JM is saturated. In particular, in order to prove White’s conjecture it is enough to show that the ideal JM is prime or radical. 50 4. THE CONJECTURE OF WHITE 4. Remarks We begin with the original formulation of conjectures stated by White in [112]. At the end we present a beautiful alternative combinatorial reformulation by Blasiak [16], which is very natural, but not obviously equivalent. Two sequences of bases B = (B1 , . . . , Bn ) and D = (D1 , . . . , Dn ) are compatible if B1 ∪· · ·∪Bn = D1 ∪· · ·∪Dn as multisets (that is if yB1 · · · yBn − yD1 · · · yDn ∈ IM ). White defines three equivalence relations. Two sequences of bases B and D of equal length are in relation: ∼1 if D may be obtained from B by a composition of symmetric exchanges. That is ∼1 is the transitive closure of the relation which exchanges a pair of bases Bi , Bj in a sequence into a pair obtained by a symmetric exchange. ∼2 if D may be obtained from B by a composition of symmetric exchanges and permutations of bases. ∼3 if D may be obtained from B by a composition of multiple symmetric exchanges. Let T E(i) denote the class of matroids such that for every n ≥ 2 all compatible sequences of bases B, D of length n satisfy B ∼i D (the notion T E(i) is the same as original in [112]). An algebraic meaning of the property T E(3) is that the toric ideal IM is generated by quadratic binomials. The property T E(2) means that the toric ideal IM is generated by quadratic binomials corresponding to symmetric exchanges, while the property T E(1) is its analog for the noncommutative polynomial ring SM . We are ready to formulate the original conjecture [112, Conjecture 12] of White. Conjecture 4.13 (White, [112]). We have the following equalities: (1) T E(1) = the class of all matroids, (2) T E(2) = the class of all matroids, (3) T E(3) = the class of all matroids. Clearly, Conjecture 4.2 coincides with Conjecture 4.13 (2). It is straightforward [112, Proposition 5] that: (1) T E(1) ⊂ T E(2) ⊂ T E(3), (2) classes T E(1), T E(2) and T E(3) are closed under taking minors and dual matroid, (3) classes T E(1) and T E(3) are closed under direct sum. White claims also that the class T E(2) is closed under direct sum, however unfortunately there is a gap in his proof. We believe that it is an open question. Corollary 4.16 will show some consequences of T E(2) being closed under direct sum to the relation between classes T E(1) and T E(2). Lemma 4.14. The following conditions are equivalent for a matroid M : (1) M ∈ T E(1), (2) M ∈ T E(2) and for any two bases (B1 , B2 ) ∼1 (B2 , B1 ) holds. Proof. Implication (1) ⇒ (2) is clear just from the definition. To get the opposite implication it is enough to recall that any permutation is a composition of transpositions.  4. REMARKS 51 Proposition 4.15. The following conditions are equivalent: (1) M ∈ T E(1), (2) M ⊔ M ∈ T E(1), (3) M ⊔ M ∈ T E(2). Proof. Implications (1) ⇒ (2) ⇒ (3) were already discussed. To get (3) ⇒ (1) suppose that a matroid M satisfies M ⊔ M ∈ T E(2). First we prove that M ∈ T E(2). Let B = (B1 , . . . , Bn ) and D = (D1 , . . . , Dn ) be compatible sequences of bases of M . If B is a basis of M then B′ = ([B1 , B], . . . ) and D′ = ([D1 , B], . . . ) are compatible sequences of bases of M ⊔ M (where [B ′ , B ′′ ] denotes a basis of M ⊔ M consisting of basis B ′ of M on the first copy and B ′′ on the second). From the assumption we have B′ ∼2 D′ . Notice that any symmetric exchange in M ⊔ M is on the bases appearing on the first coordinate either trivial or a symmetric exchange. Thus, the same symmetric exchanges certify that B ∼2 D in M . Now due to Lemma 4.14 in order to prove M ∈ T E(1) it is enough to show that for any two bases B1 , B2 of M relation (B1 , B2 ) ∼1 (B2 , B1 ) holds. Sequences of bases ([B1 , B1 ], [B2 , B2 ]) and ([B2 , B1 ], [B1 , B2 ]) in M ⊔ M are compatible. Thus by the assumption one can be obtained from the other by a composition of symmetric exchanges and permutations. By the symmetry, without loss of generality we can assume that permutation are not needed. Now the projection of this symmetric exchanges to the first coordinate shows that (B1 , B2 ) ∼1 (B2 , B1 ) in M .  As a corollary of Proposition 4.15 we get that for reasonable classes of matroids the ‘standard’ version of White’s conjecture is equivalent to the ‘strong’ one. Corollary 4.16. If a class of matroids C is closed under direct sums, then C ⊂ T E(1) if and only if C ⊂ T E(2). In particular: (1) strongly base orderable matroids belong to T E(1), (2) Conjectures 4.13 (1) and (2) are equivalent, (3) T E(2) is closed under direct sum if and only if T E(1) = T E(2). Now we proceed to the alternative combinatorial description of White’s conjectures by Blasiak [16]. A matroid M whose ground set E is a union of two disjoint bases is called a 2-matroid. With a 2-matroid M we associate a graph B2 (M ) whose vertices are pairs of bases (B1 , B2 ) of M whose union is E, and with edges between pairs of bases corresponding to symmetric exchange. This graph was already defined in 1985 by Farber, Richter and Shank [37], who conjectured that it is always connected. They proved it for graphic 2-matroids (see also [8] for a game version). Blasiak defined an analog of graph B2 (M ) for matroids whose ground set is a disjoint union of more bases. Let k ≥ 3. A matroid M whose ground set E is a union of k disjoint bases is called a k-matroid. With a k-matroid M we associate a graph Bk (M ) whose vertices are sets of k bases {B1 , . . . , Bk } of M whose union is E. An edge joins two vertices if they have non-empty intersection, that is if one of bases is in both partitions. Blasiak [16] proves the following. 52 4. THE CONJECTURE OF WHITE Proposition 4.17. Conjecture 4.13 (3) (‘weak’ version of White’s conjecture) is equivalent to the fact that for every k ≥ 3 and for every k-matroid M the graph Bk (M ) is connected. Conjecture 4.13 (1) (‘strong’ version of White’s conjecture) is equivalent to the fact that for every k ≥ 2 and for every k-matroid M the graph Bk (M ) is connected. In the same way as we associate a toric ideal with a matroid one can associate a toric ideal IP with a discrete polymatroid P . Herzog and Hibi [51] extend White’s conjecture to discrete polymatroids, they also ask if the toric ideal IP of a discrete polymatroid possesses a quadratic Gröbner basis (see also [102]). Remark 4.18. Theorem 4.3 and Theorem 4.4 are also true for discrete polymatroids. There are several ways to prove that our results hold also for discrete polymatroids. One possibility is to use Lemma 5.4 from [51] to reduce a question if a binomial is generated by quadratic binomials corresponding to symmetric exchanges from a discrete polymatroid to a certain matroid. Another possibility is to associate to a discrete polymatroid P ⊂ Zn a matroid MP on the ground set {1, . . . , r(P )} × {1, . . . , n} in which a set I is independent if there is v ∈ P such that |I ∩ {1, . . . , r(P )} × {i}| ≤ vi holds for all i. It is straightforward that compatibility of sequences of bases and generation are the same in P and in MP . Moreover, one can easily prove that a symmetric exchange in MP corresponds to at most two symmetric exchanges in P . Part 3 Applications of Matroids CHAPTER 5 Obstacles for Splitting Necklaces The results of this Chapter come from our paper [70] and concern the famous necklace splitting problem. Suppose we are given a colored necklace, and we want to cut it so that the resulting pieces can be fairly split into two (or more) parts. What is the minimum number of cuts we have to make? In the continuous version of the problem (from which a discrete one follows) a necklace is an interval, a coloring is a partition of a necklace into Lebesgue measurable sets, and a splitting is fair if each part captures the same measure of every color. A theorem of Goldberg and West [42] asserts that if the number of parts is two, then every k-colored necklace has a fair splitting using at most k cuts. This fact has a simple proof [2] using the celebrated Borsuk-Ulam theorem (cf. [79]). Alon [3] extended this result to arbitrary number of parts q by showing that in this case k(q − 1) cuts suffice. In a joint paper with Alon, Grytczuk, and Michałek [7] we studied a kind of opposite question, which was motivated by the problem of Erdős on strongly nonrepetitive sequences [35] (cf. [25, 30, 45]). Namely, for which values of k and t there is a k-coloring of the real line such that no interval has a fair splitting into two parts with at most t cuts? A theorem of Goldberg and West [42] implies that k > t is a necessary condition. We prove that k > t + 2 is sufficient. In this Chapter we generalize this theorem to Euclidean spaces of arbitrary dimension d, and to arbitrary number of parts q. In higher dimensional problem a role of an interval plays an axis-aligned cube, and the cuts are made by axis-aligned hyperplanes. We prove that k(q − 1) > t + d + q − 1 is a sufficient condition (Theorem 5.7). Our bound is of the same order as a necessary condition k(q − 1) > t, which is a consequence of a theorem of Alon [3] (or by a more general theorem of de Longueville and Živaljević [75]). This substantially improves a result of Grytczuk and Lubawski [46], which is exponential in d. Moreover for d = 1, q = 2 we get exactly the result of our previous paper [7]. Additionally, we prove that if a stronger inequality k(q − 1) > dt + d + q − 1 is satisfied, then we can avoid necklaces in Rd with a fair q-splitting using arbitrary t hyperplane cuts (Theorem 5.13). In an another result from [7] of a slightly different flavor we proved that there is a 5-coloring of the real line distinguishing intervals, that is no two intervals contain the same measure of every color (this solves a problem from [44]). We generalize it by proving that there is a (2d + 3)-coloring of Rd distinguishing axis-aligned cubes (Theorem 5.16). The main framework of our proofs is the topological Baire category theorem. However, the crucial part in the reasoning, which impacts on the bounds we achieve, is several application of algebraic matroids. 55 56 5. OBSTACLES FOR SPLITTING NECKLACES 1. Introduction The following problem is well-known as the necklace splitting problem: A necklace colored with k colors has been stolen by q thieves. They intend to share the necklace fairly, but they do not know what are the values of different colors. Therefore they want to cut it so that the resulting pieces can be fairly split into q parts, which means that each part captures the same amount of every color. What is the minimum number of cuts they have to make? To be more precise a discrete k-colored necklace is an interval in N colored with k colors. A fair q-splitting of size t is a division of the necklace using t cuts into t + 1 intervals, which can be split into q parts with equal number of integers of every color. Goldberg and West [42] proved that every discrete k-colored necklace with number of integers of every color divisible by 2 has a fair 2-splitting of size at most k (see also [2] for a short proof using the Borsuk-Ulam theorem, and [79] for other applications of the Borsuk-Ulam theorem in combinatorics). It is easy to see that this bound is tight. Consider a necklace consisting of k pairs of consecutive integers, each pair colored differently. A famous generalization of Alon [3] to the case of q thieves, asserts that every discrete k-colored necklace with number of integers of every color divisible by q has a fair q-splitting of size at most k(q−1). A similar argument shows that the number of cuts is optimal. All proofs of these discrete theorems go through a continuous version of the problem, in which a necklace is an interval, a coloring is a partition of a necklace into Lebesgue measurable sets, and a splitting is fair if each part captures the same measure of every color. Alon [3] got even a more general version in which a coloring is replaced by a collection of arbitrary continuous probability measures (this generalizes a theorem of Hobby and Rice [54]). Theorem 5.1 (Alon, [3]). Let µ1 , . . . , µk be continuous probability measures on the interval [0, 1]. Then there exists a division of [0, 1] using at most k(q − 1) cuts into subintervals, which can be split into q parts having equal measure µi for each i. Here we are interested in multidimensional necklace splitting problem. Notice that every axis-aligned hyperplane in Rd is perpendicular to exactly one of axes and parallel to others. Those perpendicular to i-th axis we call aligned to i-th axis. Hence any collection of t axis-aligned hyperplane cuts gives a natural partition t = t1 + · · · + td , where exactly ti hyperplanes are aligned to i-th axis. By projecting, Alon’s Theorem 5.1 implies that for any k continuous probability measures on [0, 1]d there exists a fair q-splitting using at most k(q − 1) hyperplanes aligned to 1-st axis. De Longueville and Živaljević generalized it to a version in which we can fix arbitrary partition. Theorem 5.2 (de Longueville, Živaljević, [75]). Let µ1 , . . . , µk be continuous probability measures on d-dimensional cube [0, 1]d . For any selection of non-negative integers t1 , . . . , td satisfying k(q − 1) = t1 + · · · + td there exists a division of [0, 1]d using only axis-aligned hyperplane cuts and at most ti perpendicular to i-th axis into cuboids, which can be split into q parts having equal measure µi for each i. 1. INTRODUCTION 57 For the case of arbitrary continuous probability measures it is also easy to see that the total number of hyperplanes needed, that is k(q − 1), is best possible. Indeed, just divide the diagonal of the cube into k intervals and consider one dimensional measures on them. However, this argument has still one dimensional nature. In this Chapter we restrict our attention from continuous probability measures to measurable colorings, that is we are interested in splitting fully colored d-dimensional necklaces. A k-colored d-dimensional necklace is a nontrivial axis-aligned cube in Rd partitioned into k Lebesgue measurable sets, which we call colors. A fair q-splitting of size t of a necklace is a division using t axis-aligned hyperplane cuts into cuboids, which can be split into q parts, each capturing exactly 1/q of the total measure of every color. A very interesting question concerning splitting multidimensional necklaces is to find the minimum number td,k,q of cuts such that any k-colored d-dimensional necklace has a fair q-splitting using at most td,k,q axis-aligned hyperplane cuts. Clearly, Alon’s Theorem 5.1 (or more generally Theorem 5.2 of de Longueville and Živaljević) implies that k(q − 1) ≥ td,k,q . In Theorem 5.12 we actually prove that for q = 2 this bound is tight. Together with Theorem 5.2 this fully characterizes the set of tuples (t1 , . . . , td ) ∈ Nd , such that any k-colored d-dimensional necklace has a fair 2-splitting using at most ti hyperplanes aligned to i-th axis for each i. Namely, it is the set {(t1 , . . . , td ) ∈ Nd : k ≤ t1 + · · · + td }. However, the most interesting problem is a bit more difficult. We want to find a measurable coloring of the whole space Rd , which avoids necklaces having a fair q-splitting of a bounded size t. Of course, we want to minimize the number of colors k we use. This problem was already formulated and partially solved in our joint paper [7]. For d = 1, q = 2 a theorem of Goldberg and West [42] (or more generally Theorem 5.1) implies that k > t colors are needed, and we prove in [7] that k > t + 2 colors suffice. Theorem 5.3 (Alon, Grytczuk, Lasoń, Michałek, [7]). For every integers k, t ≥ 1 if k > t + 2, then there exists a measurable k-coloring of R such that no necklace has a fair 2-splitting using at most t cuts. Grytczuk and Lubawski [46] generalized our result to higher dimensional spaces, by proving that the same holds in Rd (for q = 2), provided d  d  d  t t t +4 − +3 + + 2 − 2d + t + 2d + 2. k> d d d They expected the lower bound on the number of colors is far from optimal, as even for d = 1 it gives much worse result than the one from [7]. Grytczuk and Lubawski asked [46, Problem 9] if k ≥ t + O(d) colors suffice for the case of two parts. In Theorem 5.7 we prove that the answer is yes, namely k(q − 1) > t + d + q − 1 is a sufficient condition for q parts (for d = 1 we get even a better bound k(q − 1) > t + 2). Moreover, in this case the set of ‘good’ colorings is dense. Note that our bound is of the same order as a necessary condition k(q−1) > t, 58 5. OBSTACLES FOR SPLITTING NECKLACES which is a consequence of Theorem 5.1 of Alon (or of a more general Theorem 5.2 of de Longueville and Živaljević). For d = 1, q = 2 Theorem 5.7 coincides with the result of our previous paper [7]. We conjecture that for q = 2 our bound is tight, as it captures the idea of degrees of freedom of objects we want to split. In Section 4 we generalize our results to the situation when arbitrary hyperplane cuts are allowed. This is closely related to the mass equipartitioning problem: Does for every k continuous mass distributions µ1 , . . . , µk in Rd there exist t hyperplanes dividing the space into pieces which can be split into q parts of equal measure µi for each i? A necessary condition for the positive answer is k(q − 1) ≤ td. It can be shown by considering one dimensional measure along the moment curve. The most basic and well-known positive result in this area is the Ham Sandwich Theorem for measures, asserting that for (d, k, q, t) = (d, d, 2, 1) the answer is yes. It is a consequence of the Borsuk-Ulam theorem (see [79]), and has also discrete versions concerning equipartitions of a set of points in the space by a hyperplane. Several other things are known (see [15, 32, 77, 80, 90]), however even for some small values (eg. (4, 1, 16, 4)) the question remains open. In the problem we are studying, k continuous mass distributions in Rd are restricted to a measurable k-coloring of a given necklace. In this variant in Theorem 5.15 we reprove the necessary inequality for q = 2. In the same spirit as in the case of axis-aligned hyperplane cuts we investigate existence of measurable colorings of Rd avoiding necklaces having a fair q-splitting of a bounded size t. In Theorem 5.13 we prove a sufficient condition on the number of colors k(q − 1) > dt + d + q − 1. In the last Section 5 we generalize another result from [7] of a slightly different flavor. In [7] we proved that there is a measurable 5-coloring of the real line distinguishing intervals, that is no two intervals contain the same measure of every color (this solves a problem from [44]). In Theorem 5.16 we generalize it by proving that there is a measurable (2d + 3)-coloring of Rd distinguishing axis-aligned cubes. We suspect that this number is optimal. Recently Vrećica and Živaljević [108] inspired by our question solved the case of arbitrary continuous probability measures. They showed d + 1 measures distinguishing cubes, and proved that for any d continuous probability measures on Rd there are two (in fact any finite number) nontrivial axis-aligned cubes which are not distinguished. The proofs of our results are based on the topological Baire category theorem applied to the space of all measurable colorings of Rd , equipped with a suitable metric. Section 2 describes this background. The remaining part of the argument is algebraic, it uses algebraic independence over suitably chosen fields. The rank of corresponding algebraic matroids describe well the idea of degrees of freedom of considered geometric objects. We will demonstrate in full details a proof of our main result – Theorem 5.7. Proofs of other results fit the same pattern, so we will omit parts which can be rewritten with just minor changes and describe only main differences. 2. THE TOPOLOGICAL SETTING 59 2. The Topological Setting Recall that a set in a metric space is nowhere dense if the interior of its closure is empty. Sets which can be represented as a countable union of nowhere dense sets are said to be of first category. Theorem 5.4 (Baire, cf. [87]). If X is a complete metric space and a set A ⊂ X is of first category, then X \ A is dense in X (in particular is non-empty). All of our results assert existence of certain ‘good’ colorings. The main framework of the argument will mimic the one from [7]. Our plan is to construct a suitable metric space of colorings and prove that the subset of ‘bad’ colorings is of first category, resulting that ‘good’ colorings exist. To prove it we will use entirely different technique than in our previous paper [7]. Our approach will be purely algebraic, we will compare transcendence degrees of some sets of numbers over suitably chosen fields. We go towards a proof of our main result – Theorem 5.7, which asserts existence of a measurable k-coloring of Rd with no necklace having a fair q-splitting of size at most t, provided k(q − 1) > t + d + q − 1. We will demonstrate this proof in full details. Proofs of other results fit the same pattern, so we will omit parts which can be rewritten with just minor changes and describe only main differences. This Section describes a common part of all proofs of our results. Let k be a fixed positive integer, and let {1, . . . , k} be the set of colors. Let M be a space of all measurable k-colorings of Rd , that is the set of measurable maps f : Rd → {1, . . . , k} (for each i the set f −1 (i) is Lebesgue measurable). For f, g ∈ M , and a positive integer n let Dn (f, g) = {x ∈ [−n, n]d : f (x) 6= g(x)}. Clearly Dn (f, g) is a measurable set, and we may define the normalized distance between f and g on [−n, n]d by dn (f, g) = L(Dn (f, g)) , (2n)d where L(D) denotes the Lebesgue measure of a set D. We define the distance between any two colorings f, g from M by ∞ X dn (f, g) . d(f, g) = 2n n=1 To get a metric space we need to identify colorings whose distance is zero (they differ only on a set of measure zero), but they are indistinguishable from our point of view. Formally M is metric space consisting of equivalence classes of M with distance function d. Lemma 5.5. Metric space M is complete. This is a straightforward generalization of the fact that sets of finite measure in any measure space form a complete metric space, with measure of symmetric difference as the distance function (see [87]). A coloring f ∈ M is called a cube coloring on [−n, n]d if there exists N such that f is constant on any half open cube from the division of [−n, n]d 60 5. OBSTACLES FOR SPLITTING NECKLACES into N d equal size cubes. Let In denote the set of all colorings from M that are cube colorings on [−n, n]d . Lemma 5.6. Let f ∈ M, then for every ǫ > 0 and an integer n there exists a cube coloring g ∈ In such that d(f, g) < ǫ. Proof. Let Ci = f −1 (i) ∩ [−n, n]d and let Ci∗ ⊂ [−n, n]d be a finite union of cuboids such that ǫ L((Ci \ Ci∗ ) ∪ (Ci∗ \ Ci )) < 2 2k for each i = 1, . . . , k. Define a coloring h so that for each i = 1, . . . , k, the ∗ ) is filled with color i, the remaining part of the set Ci∗ \ (C1∗ ∪ · · · ∪ Ci−1 d cube [−n, n] is filled with any fixed color, and h agrees with f everywhere outside [−n, n]d . Then d(f, h) < ǫ/2 and clearly each h−1 (i) ∩ [−n, n]d is a finite union of cuboids. Let A1 , . . . , At be the whole family of these cuboids. Now divide the cube [−n, n]d into N d equal cubes B1 , . . . , BN d for N > 4dt(2n)d /ǫ. Define a new coloring g such that g|Bi ≡ h|Bi whenever Bi ⊂ Aj for some j, and g|Bi is constant equal to any color otherwise. Clearly g is a cube coloring on [−n, n]d . Moreover g differs from h on at most 2dt(N )d−1 small cubes of total volume at most 2dt(2n)d /N < ǫ/2, thus we get d(f, g) < ǫ.  3. Axis-aligned Hyperplane Cuts Theorem 5.7 (Lasoń, [70]). For every integers d, k, t ≥ 1 and q ≥ 2 if k(q − 1) > t + d + q − 1, then there exists a measurable k-coloring of Rd such that no d-dimensional necklace has a fair q-splitting using at most t axis-aligned hyperplane cuts. The set of such colorings is dense. For d = 1 the same holds provided k(q − 1) > t + 2. Proof. A splitting of C = [α1 , α1 + α0 ] × · · · × [αd , αd + α0 ] using t axisaligned hyperplane cuts can be described by a partition t1 + · · · + td = t such that for each i there are exactly ti hyperplanes aligned to i-th axis, by i-th coordinates of this hyperplanes αi = β0i ≤ β1i ≤ · · · ≤ βtii ≤ βtii +1 = αi + α0 , and by a q-labeling of obtained pieces. By granularity of the splitting (or more generally of a division) we mean the length of the shortest subinterval i ] over all i, j. [βji , βj+1 (t ) For (t1 , . . . , td ) and n we denote by Bn i the set of k-colorings from M for which there exists at least one d-dimensional necklace contained in [−n, n]d which has a fair q-splitting using for each i exactly ti hyperplanes aligned to i-th axis and with granularity at least 1/n. Finally let [ Bn = Bn(ti ) t1 +···+td ≤t be a set of ‘bad’ colorings. Let Gt be a subset of M consisting of those k-colorings for which any necklace does not have a fair q-splitting using at most t axis-aligned hyperplane cuts. Clearly we have ∞ [ Bn . Gt = M \ n=1 3. AXIS-ALIGNED HYPERPLANE CUTS 61 Our aim is to show that sets Bn are nowhere dense, provided S suitable relation between d, k, q and t holds. This will imply that the union ∞ n=1 Bn is of first category, so the set of ‘good’ colorings Gt , which we are looking for, is dense. Lemma 5.8. Sets Bn are closed in M. (t ) Proof. Since Bn is a finite union of Bn i over all t1 + · · · + td ≤ t it (t ) is enough to show that sets Bn i are closed. Let (fm )m∈N be a sequence of (t ) colorings from Bn i converging to f in M. For each fm there exists a necklace d Cm in [−n, n] and a fair q-splitting of it with granularity at least 1/n. Since [−n, n]d is compact we can assume that cubes Cm converge to some cube C, and that hyperplane cuts also converge (notice that the granularity of this division of C is at least 1/n, in particular C is a nontrivial). The only thing left is to define a q-labeling of pieces of the division to get q equal parts. But there is a finite number of such labelings for each Cm , so one of them appears infinitely many times in the sequence. It is easy to see that this (t ) gives a fair q-splitting of the necklace C, so f ∈ Bn i .  Lemma 5.9. If k(q − 1) > t + d + q − 1, then every Bn has empty interior. For d = 1 the same holds provided k(q − 1) > t + 2. Proof. Suppose the assertion of the lemma is false: there is some f ∈ Bn and ǫ > 0 for which K(f, ǫ) = {g ∈ M : d(f, g) < ǫ} ⊂ Bn . By Lemma 5.6 there is a cube coloring g ∈ In such that d(f, g) < ǫ/2, so K(g, ǫ/2) ⊂ Bn . The idea is to modify slightly the coloring g so that the new coloring will be still close to g, but there will be no necklaces inside [−n, n]d possessing a q-splitting of size at most t and granularity at least 1/n. Without loss of generality we may assume that there are equally spaced axis-aligned cubes Ci1 ,...,id for i1 , . . . , id ∈ {1, . . . , N } in [−n, n]d such that N > 4n2 , and each cube is filled with a unique color in the cube coloring g. Notice that edges of cubes are rational, equal to 2n N . Let δ > 0 be a rational  ǫ 2n d d number satisfying δ < min{ 2N d , N }. Inside each cube Ci1 ,...,id choose a volume δd axis-aligned cube Di1 ,...,id with corners in rational points Qd . Let us call ‘white’ the first color from our set of colors {1, . . . , k}. For each cube Di1 ,...,id and a color j = 2, . . . , k we choose small enough real number xj;i1,...,id > 0 and an axis-aligned cube Dj;i1,...,id ⊂ Di1 ,...,id such that: (1) length of the side of Dj;i1 ,...,id equals to xj;i1,...,id (2) cubes Dj;i1 ,...,id for j = 2, . . . , k are disjoint (3) one of corners of Dj;i1 ,...,id is rational (lies in Qd ) (4) the set {xj;i1 ,...,id : j ∈ {2, . . . , k}, i1 , . . . , id ∈ {1, . . . , N }} is algebraically independent over Q. Now we define a coloring h, which is a modification of g. Let h equals to g outside cubes Di1 ,...,id for i1 , . . . , id ∈ {1, . . . , N }. Inside the cube Di1 ,...,id let h be equal to ‘white’ everywhere except cubes Dj;i1 ,...,id , where it is j. Clearly d(g, h) < ǫ/2, so h ∈ Bn . Hence there is a necklace A = [α1 , α1 + α0 ] × · · · × [αd , αd + α0 ] ⊂ [−n, n]d which has a fair q-splitting of size at most t and granularity at least 1/n. Denote by β1 , . . . , βt the coordinates of axis-aligned hyperplane cuts (for H = {xi = β} we get β). 62 5. OBSTACLES FOR SPLITTING NECKLACES Since granularity of the splitting is at least 1/n, and edges of cubes Ci1 ,...,id are smaller or equal to 1/(2n), each piece P of the division contains a cube Ci1 ,...,id ⊂ P for some i1 , . . . , id (so in particular Ci1 ,...,id is not divided by any hyperplane cut). Each of q parts contains at least one piece and as a consequence at least one cube. For each part we fix such a cube Cil ,...,il for 1 d l = 1, . . . , q. Now the key observation is that the amount of color j different from ‘white’ contained in each part l of the q-splitting is a polynomial Wj,l from Q[αi , βm , xj;i1 ,...,id : i = 0, . . . , d; m = 1, . . . , t; (i1 , . . . , id ) ∈ {1, . . . , N }d ]. Comparing the amount of color j in the l-th part and the first one we get Wj,l (αi , βm , xj;i1 ,...,id ) = Wj,1 (αi , βm , xj;i1 ,...,id ). Since the cube Cil ,...,il is not divided by any hyperplane from our splitting, a 1 d variable xj;il ,...,il appears only in the component xdj;il ,...,il in the polynomial 1 d 1 d ′ = W − xd . We get polynomial equations (♠): Wj,l . Denote Wj,l j,l j;il ,...,il 1 (Wj,1 − ′ )(αi , βm , xj;i1 ,...,id Wj,l d : (i1 , . . . ) 6= (ir1 , . . . ) r = 2, . . . , q) = xdj;il ,...,il , 1 d with coefficients in Q, for all j = 2, . . . , k and l = 2, . . . , q. Let MK be the algebraic matroid for a field extension K ⊂ L = Q(αi , βm , xj;i1 ,...,id ) and a suitable (big enough) finite subset E ⊂ L. Denote by T the set {αi , βm : i = 0, . . . , d; m = 1, . . . , t}, and suppose its rank in the matroid MQ equals to t + d + 1 − c, for some c. We will show that c ≥ 1, and c ≥ q − 1 for d = 1 (notice that in the case d = 1 we can assume, without loss of generality, that t ≥ q − 1). Observe that the volume of each part of the splitting is a polynomial over Q in variables α0 , . . . αd , β1 , . . . , βt . All q parts have to be of the same volume, so we get q−1 nontrivial polynomial equations. Each of them shows that T is a dependent set in MQ , hence c ≥ 1. For us the ideal situation would be if equations were ‘independent’ leading to c = q − 1, since it would give a better lower bound k(q − 1) > t + d + 1. Unfortunately, it is not always the case, see Example 5.10. However in dimension one these equations are linear, and it is easy to see that they are linearly independent. Hence c ≥ q − 1 for d = 1. Denote by K the field Q(xj;i1,...,id : (i1 , . . . , id ) 6= (ir1 , . . . , ird ) r = 2, . . . , q). Clearly, inclusion Q ⊂ K implies that rMK (T ) ≤ rMQ (T ) = t + d + 1 − c. Denote by L the set of left sides of equations (♠), and by R the set of right sides of equations (♠). We have that L ⊂ clMK (T ) (see Definition 0.4 of closure), so rMK (L) ≤ rMK (T ) = t + d + 1 − c. From condition (4) of the choice of xj;i1 ,...,id it follows that the set R is independent in MK , that is rMK (R) = (k − 1)(q − 1). But obviously L = R, hence (k − 1)(q − 1) ≤ t + d (and (k − 1)(q − 1) ≤ t − q + 3 for d = 1) and we get a contradiction.  Now from Lemmas 5.8 and 5.9 it follows that if k(q − 1) > t + d + q − 1, or k(q − 1) > t + 2 for d = 1, then sets Bn are nowhere dense, which proves the assertion of the theorem.  3. AXIS-ALIGNED HYPERPLANE CUTS 63 Example 5.10. Consider d = 2, q = 4, t1 = 1, t2 = 1. Then equality of volumes of three parts of the splitting implies that the fourth one also has the same volume. This phenomena is caused by the geometry of Rd and axisaligned hyperplanes. It shows that in general it is easier to q-split fairly a whole necklace, than its Lebesgue measurable subset. In dimension one these situations do not happen. By modifying slightly the proof of Theorem 5.7 one can easily obtain a version for cuboids with inequality k(q − 1) > t + 2d + q − 2. The idea standing behind the proof is to count ‘degrees of freedom’ of moving a necklace and cuts splitting it and to compare it with the number of equations forced by existence of a fair splitting. If the number of degrees of freedom is less than the number of equations, then there exists a coloring with the property that no necklace can be fairly split. Even though the intuitive meaning of degree of freedom is clear, it is hard to define it rigorously to be able to make use of it. It turns out that the rank in some algebraic matroid of the set {αi , βm : i = 0, . . . , d; m = 1, . . . , t} works. As we will see in Section 4, if we allow arbitrary hyperplane cuts, then each such cut adds d degrees of freedom, not only one as an axis-aligned hyperplane cut. Therefore for arbitrary hyperplane cuts we will get roughly d times worse bound. The same reason explains the difference between results for cubes and cuboids, the latter have larger degree of freedom. For d = 1 and q = 2 Theorem 5.7 gives exactly Theorem 5.3, the result of our joint paper with Alon, Grytczuk and Michałek [7]. However, the technique of showing that sets Bn have empty interior is quite different. To show a contradiction we used in [7] only one color, which was different from colors of cut points and the end points of an interval. If d > 1 or q > 2 this is not always possible. Additionally our previous argument was not only algebraic, but contained also an analytic part. We used some inequalities on measure of colors. We ask if our bound from Theorem 5.7 for q = 2 is tight. Question 5.11. Fix d, k, t ≥ 1 such that k ≤ t + d + 1. Does for every measurable k-coloring of Rd there exist a d-dimensional necklace which has a fair 2-splitting using at most t axis-aligned hyperplane cuts? We suspect that the answer is positive, however we know that if we replace a k-coloring by arbitrary k continuous measures then it is negative. In Section 5 we provide other examples when the case of arbitrary continuous measures differs from the case of measurable coloring. The last part of this Section is devoted to colorings of a given fixed necklace that avoid a fair splitting of a certain size. Theorem 5.12 (Lasoń, [70]). For every integers d, k, t ≥ 1 and q ≥ 2 if k(q − 1) > t + q − 2, then there is a measurable k-coloring of a d-dimensional necklace for which there is no fair q-splitting using at most t axis-aligned hyperplane cuts. The set of such colorings is dense. Proof. For d, k ≥ 1, q ≥ 2 and a d-dimensional necklace C = [α1 , α1 + (t ) α0 ] × · · · × [αd , αd + α0 ] we define the set Bn i of k-colorings from M for which there is a fair q-splitting of C with exactly ti hyperplanes aligned to 64 5. OBSTACLES FOR SPLITTING NECKLACES i-th axis, for each i, and granularity at least 1/n. Finally let [ Bn = Bn(ti ) t1 +···+td ≤t be ‘bad’ colorings. The aim, as previously, is to show that sets Bn are nowhere dense, provided suitable relation between d, k, q and t holds. This S will imply that the union ∞ B n=1 n is of first category, so the set of colorings which we are looking for is dense. Analogously to Lemma 5.8 sets Bn are closed in M. Similarly to Lemma 5.9 if k(q − 1) > t + q − 2, then Bn has empty interior. The only difference is that we want the set {xj;i1 ,...,id : j ∈ {2, . . . , k}, i1 , . . . , id ∈ {1, . . . , N }} to be algebraically independent over Q(αi : i = 0, . . . , d). We also add α0 , . . . , αd to the base field K, so the rank in the matroid of the left side of equations is equal to t−c ≤ t−1, while of the right side it is still (k−1)(q−1). This proves the assertion of the theorem.  For q = 2 the above theorem is tight. It asserts that there are measurable (t+1)-colorings of a given necklace for which there is no fair 2-splitting using at most t axis-aligned hyperplane cuts. The set of such colorings is even dense. Due to a result of Goldberg and West [42] it is the minimum number of colors. Together with Theorem 5.2 of de Longueville and Živaljević this fully characterizes the set of tuples (t1 , . . . , td ) ∈ Nd , such that any k-colored d-dimensional necklace has a fair 2-splitting using at most ti hyperplanes aligned to i-th axis for each i. Namely, it is the set {(t1 , . . . , td ) ∈ Nd : k ≤ t1 + · · · + td }. 4. Arbitrary Hyperplane Cuts Theorem 5.13 (Lasoń, [70]). For every integers d, k, t ≥ 1 and q ≥ 2 if k(q − 1) > dt + d + q − 1, then there is a measurable k-coloring of Rd such that no d-dimensional necklace has a fair q-splitting using at most t arbitrary hyperplane cuts. The set of such colorings is dense. Proof. An arbitrary hyperplane (not necessarily axis-aligned) can be described by d numbers β 1 , . . . , β d , which we will call parameters. There are several ways of defining parameters, fortunately usually transition functions between different definitions are rational over Q. For example for a hyperplane not passing through O = (0, . . . , 0) we can define parameters to be coefficients of its normalized equation β 1 x1 + · · · + β d xd = 1. In general pick d + 1 rational points in general position P1 , . . . , Pd+1 . Then no hyperplane contains all of them. We describe a hyperplane H using coefficients of its normalized equation with O being moved to a point Pi which does not belong to H. We will make use of the following classical fact. Lemma 5.14. Volume of a bounded set in Rd which is an intersection of a finite number of half spaces is a rational function over Q of parameters of hyperplanes supporting these half spaces. Proof. Such a set S is a convex polytope. From Cramer’s formula it follows that coordinates of vertices of S are rational functions over Q 4. ARBITRARY HYPERPLANE CUTS 65 of parameters. We make the barycentric subdivision of S into simplices. Vertices of these simplices are barycenters of faces of S, hence they are also rational functions over Q of parameters. Volume of each simplex equals to the determinant of the matrix of coordinates of its vertices divided by d!. Thus it is a rational function over Q of parameters. Then so is the volume of the original polytope.  Granularity of the splitting is the largest g such that any piece of the splitting contains an axis-aligned cube with side of length g. We denote by Bn the set of k-colorings from M for which there exists at least one ddimensional necklace contained in [−n, n]d which has a fair q-splitting with at most t hyperplanes and granularity at least 1/n. These are ‘bad’ colorings. As in the previous Section 3 our aim is to show that sets Bn are nowhere dense, provided suitable relation between d, k, q and t holds. This will imply that the set of colorings which we are looking for is dense. The proof that sets Bn are closed in M goes exactly the same as the proof of Lemma 5.8. To show that if (k − 1)(q − 1) > dt + d, then every Bn has empty interior we repeat the proof of Lemma 5.9 and make small modifications. Instead of a single parameter βi of a hyperplane we have d parameters βi1 , . . . , βid . Due to Lemma 5.14 the amount of color j different from ‘white’ contained in each part l of the q-splitting is a rational function Rj,l from p Q(αi , βm , xj;i1 ,...,id : i = 0, . . . , d; m = 1, . . . , t; p = 1, . . . , d; (i1 , . . . , id )), instead of a polynomial Wj,l . Now the rank in the matroid MK for the field K = Q(xj;i1 ,...,id : (i1 , . . . , id ) 6= (ir1 , . . . , ird ) r = 2, . . . , q) of the set of left sides of equations analogous to (♠) is at most dt + d + 1 − 1, while of the set of right sides equals to (k − 1)(q − 1). Hence the hypothesis that Bn has non-empty interior implies inequality (k − 1)(q − 1) ≤ dt + d. It follows that if (k − 1)(q − 1) > dt + d, then sets Bn are nowhere dense. This proves the assertion of the theorem.  Modifying slightly the above argument one may obtain a version for cuboids with inequality k(q − 1) > dt + 2d + q − 2. Other modification, similar to the proof of Theorem 5.12, leads the following theorem. Theorem 5.15 (Lasoń, [70]). For every integers d, k, t ≥ 1 and q ≥ 2 if k(q − 1) > dt + q − 2, then there is a measurable k-coloring of a ddimensional necklace for which there is no fair q-splitting using at most t arbitrary hyperplane cuts. The set of such colorings is dense. For q = 2 we get a necessary condition (k(q − 1) ≤ dt) for mass equipartitioning problem. It is an interesting question if this condition is also sufficient for existence of a fair q-splitting of size at most t of a k-colored ddimensional necklace. For (d, k, q, t) = (d, k, 2t , t) the question coincides with the conjecture of Ramos [90]. A small support is the positive answer for (d, k, q, t) = (d, d, 2l , 2l −1), which follows directly from 2l −1 applications of the Ham Sandwich Theorem for measures [79]. 66 5. OBSTACLES FOR SPLITTING NECKLACES 5. Colorings Distinguishing Cubes Here we are interested in a problem of a slightly different flavor. We say that a measurable coloring of Rd distinguishes axis-aligned cubes if no two nontrivial axis-aligned cubes contain the same measure of every color. What is the minimum number of colors needed for such a coloring? We get our results in a similar way to the proof of Theorem 5.7. For each n we define a set Bn of all k-colorings from M for which there exist two ddimensional cubes A, B contained in [−n, n]d which have the same measure d of every color, and such that S A \ B contains a translation of the cube 0, n1 . The ‘bad’ colorings are B = ∞ n=1 Bn . Sets Bn are closed, and we can easily get an analog of Lemma 5.9. This shows the following theorem. Theorem 5.16 (Lasoń, [70]). For every d ≥ 1 there exists a measurable (2d + 3)-coloring of Rd distinguishing axis-aligned cubes. The set of such colorings is dense. For d = 1 this gives the second result of our previous paper [7], namely the existence of a measurable 5-coloring of R with no two equally colored intervals. We suspect that this number of colors is optimal. Question 5.17. Fix d ≥ 1. Does for every measurable (2d + 2)-coloring of Rd there exist two nontrivial axis-aligned cubes which have the same measure of every color? A slight modification of the argument leads to a version of Theorem 5.16 for cuboids with 4d + 1 colors. Vrećica and Živaljević [108] solved the case of arbitrary continuous probability measures instead of measurable colorings of Rd . They prove that d+1 measures are sufficient to distinguish axis-aligned cubes, and that it is tight. Theorem 5.18 (Vrećica, Živaljević, [108]). For every d continuous probability measures µ1 , . . . , µd on Rd there are two nontrivial axis-aligned cubes C and C ′ such that µi (C) = µi (C ′ ) holds for every i. Moreover, there exist d + 1 continuous probability measures on Rd distinguishing axis-aligned cubes. In fact Vrećica and Živaljević show that even one can find any finite number of axis-aligned cubes in Rd which are not distinguished by given d measures. They also get results analogous to Theorem 5.18 for cuboids in Rd with 2d − 1 colors. For sure, the minimum number of colors in the two cases (arbitrary continuous measures, and measurable colorings of Rd ) is different. One can show the following. Example 5.19. There are two measures on R distinguishing intervals, while no measurable 2-coloring of R distinguishes intervals. Bibliography [1] R. Aharoni, E. Berger, The intersection of a matroid and a simplicial complex, Trans. Amer. Math. Soc. 358 (2006), no. 11, 4895-4917. [2] N. Alon, D. West, The Borsuk-Ulam theorem and bisection of necklaces, Proc. Amer. Math. Soc. 98 (1986), 623-628. [3] N. Alon, Splitting necklaces, Advances in Math. 63 (1987), 247-253. [4] N. Alon, Z. Tuza, M. Voigt, Choosability and fractional chromatic numbers, Discrete Math. 165/166 (1997), 31-38. [5] N. Alon, Combinatorial Nullstellensatz, Combin. Probab. Comput. 8 (1999), 7-29. [6] N. Alon, Degrees and choice numbers, Random Struct. Algor. 16 (2000), 364-368. [7] N. Alon, J. Grytczuk, M. Lasoń, M. Michałek, Splitting necklaces and measurable colorings of the real line, Proc. Amer. Math. Soc. 137 (2009), no. 5, 1593-1599. [8] S. Andres, W. Hochstättler, M. Merkel, On a base exchange game on bispanning graphs, Technical Report feu-dmo029.11, FernUniversität in: Hagen Fakultät für Mathematik und Informatik. [9] A. Aramova, J. Herzog, T. Hibi, Gotzmann theorems for exterior algebras and combinatorics, J. Algebra 191 (1997), 174-211. [10] M. Atiyah, I. Macdonald, Introduction to commutative algebra, Addison-Wesley Publishing Co., Mass.-London-Don Mills, 1969. [11] T. Bartnicki, J. Grytczuk, H. Kierstead, X. Zhu, The map coloring game, Amer. Math. Monthly 114 (2007), no. 9, 793-803. [12] T. Bartnicki, J. Grytczuk, H. Kierstead, The game of arboricity, Discrete Math. 308 (2008), 1388-1393. [13] A. Björner, B. Korte, L. Lovász, Homotopy properties of greedoids, Adv. in Appl. Math. 6 (1985), no. 4, 447-494. [14] A. Björner, Topological methods, 1819-1872, in: R. Graham, M. Grötschel, L. Lovász, Handbook of Combinatorics, Elsevier, Amsterdam, 1995. [15] P. Blagojević, G. Ziegler, The ideal-valued index for a dihedral group action, and mass partition by two hyperplanes, Topology Appl. 158 (2011), no. 12, 1326-1351. [16] J. Blasiak, The toric ideal of a graphic matroid is generated by quadrics, Combinatorica 28 (2008), 283-297. [17] H. Bodlaender, On the complexity of some coloring games, Internat. J. Found. Comput. Sci. 2 (1991), no. 2, 133-147. [18] M. Boij, J. Migliore, R. Miró-Roig, U. Nagel, F. Zanello, On the shape of a pure O-sequence, Mem. Amer. Math. Soc. 218 (2012), no. 2024. [19] J. Bonin, Basis-exchange properties of sparse paving matroids, Adv. in Appl. Math. 50 (2013), no. 1, 6-15. [20] R. Brualdi, Comments on bases in dependence structures, Bull. Austral. Math. Soc. 1 (1969), 161-167. [21] G. Clements, B. Lindstöm, A generalization of a combinatorial theorem of Macaulay, J. Combin. Theory 7 (1969), 230-238. [22] A. Conca, Linear spaces, transversal polymatroids and ASL domains, J. Algebraic Combin. 25 (2007), 25-41. [23] A. Constantinescu, T. Kahle, M. Varbaro, Generic and special constructions of pure O-sequences, arXiv:1212.3426. [24] D. Cox, J. Little, H. Schenck, Toric varieties, Graduate Studies in Mathematics 124, American Mathematical Society, Providence, 2011. 67 68 BIBLIOGRAPHY [25] J. Currie, Unsolved problems: Open problems in pattern avoidance, Amer. Math. Monthly 100 (1993), no. 8, 790-793. [26] J. Davies and C. McDiarmid, Disjoint common transversals and exchange structures, J. London Math. Soc. (2) 14 (1976), no. 1, 55-62. [27] J. Dawson, A collection of sets related to the Tutte polynomial of a matroid, Graph Theory Singapore 1983, Lecture Notes in Mathematics 1073 (1984), 193-204. [28] D. Daykin, A simple proof of the Kruskal-Katona theorem, J. Combin. Theory Ser. A 17 (1974), 252-253. [29] D. Daykin, An algorithm for cascades giving Katona-type inequalities, Nanta Math. 8 (1975), no. 2, 78-83. [30] F. Dekking, Strongly nonrepetitive sequences and progression-free sets, J. Combin. Theory Ser. A 27 (1979), no. 2, 181-185. [31] J. Eagon, V. Reiner, Resolutions of Stanley-Reisner rings and Alexander duality, J. Pure Appl. Algebra 130 (1989), no. 3, 265-275. [32] H. Edelsbrunner, Algorithms in Combinatorial Geometry, volume 10 of EATCS Monographs in Theoretical Computer Science, Springer-Verlag, Berlin, 1987. [33] J. Edmonds, Minimum partition of a matroid into independent subsets, J. Res. Nat. Bur. Standards Sect. B 69B (1965), 67-72. [34] J. Edmonds, Lehman’s switching game and a theorem of Tutte and Nash-Williams, J. Research National Bureau of Standards B 69B (1965), 73-77. [35] P. Erdős, Some unsolved problems, Magyar Tud. Akad. Mat. Kutató Int. Kö zl. 6 (1961), 221-254. [36] P. Erdős, A. Rubin, H. Taylor, Choosability in graphs, Congr. Numer. 26 (1980), 122-157. [37] M. Farber, B. Richter, H. Shank, Edge-disjoint spanning trees: A connectedness theorem, J. Graph Theory 8 (1985), no. 3, 319-324. [38] W. Fulton, Introduction to toric varieties, Annals of Mathematics Studies 131, Princeton University Press, Princeton, 1993. [39] F. Galvin, The list chromatic index of a bipartite multigraph, J. Combin. Theory Ser. B 63 (1995), no. 1, 153-158. [40] M. Gardner, Mathematical Games, Scientific American, 1981. [41] M. Goemans, Lecture notes on matroid optimization, MIT 2009. [42] C. Goldberg, D. West, Bisection of circle colorings, SIAM J. Algebraic Discrete Methods 6 (1985), no. 1, 93-106. [43] J. Greene, A new short proof of Kneser’s conjecture, Amer. Math. Monthly 109 (2002), no. 10, 918-920. [44] J. Grytczuk, W. Śliwa, Non-repetitive colorings of infinite sets, Discrete Math. 265 (2003), no. 1-3, 365-373. [45] J. Grytczuk, Thue type problems for graphs, points, and numbers, Discrete Math. 308 (2008), no. 19, 4419-4429. [46] J. Grytczuk, W. Lubawski, Splitting multidimensional necklaces and measurable colorings of Euclidean spaces, arXiv:1209.1809. [47] A. Grzesik, Indicated coloring of graphs, Discrete Math. 312 (2012), 3467-3472. [48] G. Gutowski, Mr. Paint and Mrs. Correct go fractional, Electron. J. Comb. 18 (2011), no. 1, RP140. [49] R. Häggkvist, J. Janssen, New bounds on the list-chromatic index of the complete graph and other simple graphs, Combin. Probab. Comput. 6 (1997), no. 3, 295-313. [50] J. Herzog, T. Hibi, Componentwise linear ideals, Nagoya Math. J. 153 (1999), 141153. [51] J. Herzog, T. Hibi, Discrete Polymatroids, J. Algebr. Combin. 16 (2002), 239-268. [52] J. Herzog, T. Hibi, S. Murai, N. Trung, X. Zheng, Kruskal-Katona type theorems for clique complexes arising from chordal and strongly chordal graphs, Combinatorica 28 (2008), no. 3, 315-323. [53] A. Hilton, A simple proof of the Kruskal-Katona theorem and of some associated binomial inequalities, Period. Math. Hungar. 10 (1979), no. 1, 25-30. [54] C. Hobby, J. Rice, A moment problem in L1 approximation, Proc. Amer. Math. Soc. 16 (1965), 665-670. BIBLIOGRAPHY 69 [55] J. Huh, Milnor numbers of projective hypersurfaces and the chromatic polynomial of graphs, J. Amer. Math. Soc. 25 (2012), no. 3, 907-927. [56] J. Huh, h-vectors of matroids and logarithmic concavity, arXiv:1201.2915. [57] J. Huh, E. Katz, Log-concavity of characteristic polynomials and the Bergman fan of matroids, Math. Ann. 354 (2012), no. 3, 1103-1116. [58] J. Kahn, Asymptotics of the list chromatic index for multigraphs, Random Struct. Algor. 17 (2000), no. 2, 117-156. [59] K. Kashiwabara, The toric ideal of a matroid of rank 3 is generated by quadrics, Electron. J. Combin. 17 (2010), no. 1, RP28. [60] G. Katona, A theorem of finite sets, Theory of Graphs (Proc. Colloq. Tihany, 1966), 187-207, Academic Press, New York, 1968. [61] V. Keränen, Abelian squares are avoidable on 4 letters, Lecture Notes in Comput. Sci. 623 (1992), 41-52. [62] J. Kruskal, The number of simplices in a complex, Mathematical Optimalization Techniques, 251-278, University of California Press, Berkeley, 1963. [63] J. Kung, Basis-Exchange Properties, in: N. White, Theory of matroids, Encyclopedia Math. Appl. 26, Cambridge University Press, Cambridge, 1986. [64] M. Lasoń, List coloring of matroids and base exchange properties, European J. Combin. 49 (2015), 265-268. [65] M. Lasoń, The coloring game on matroids, Discrete Math. 340 (2017), no. 4, 796-799. [66] M. Lasoń, W. Lubawski, On-line list coloring of matroids, Discrete Appl. Math. 217 (2017), no. 2, 353-355. [67] M. Lasoń, Indicated coloring of matroids, Discrete Appl. Math. 179 (2014), 241-243. [68] M. Lasoń, f-vectors implying vertex decomposability, Discrete Comput. Geom. 49 (2013), no. 2, 296-301. [69] M. Lasoń, M. Michałek, On the toric ideal of a matroid, Advances in Math. 259 (2014), 1-12. [70] M. Lasoń, Obstacles for splitting multidimensional necklaces, Proc. Amer. Math. Soc. 143 (2015), no. 11, 4655-4668. [71] M. Lasoń, A generalization of Combinatorial Nullstellensatz, Electron. J. Combin. 17 (2010), N32. [72] M. Lasoń, M. Michałek, On the full, strongly exceptional collections on toric varieties with Picard number three, Coll. Math. 62 (2011), no. 3, 275-296. [73] A. Lehman, A solution to the Shannon switching game, J. SIAM 12 (1964), 687-725. [74] M. Lenz, The f-vector of a realizable matroid complex is strictly log-concave, arXiv:1106.2944. [75] M. de Longueville, R. Živaljević, Splitting multidimensional necklaces, Advances in Math. 218 (2008), no. 3, 926-939. [76] F. MacAulay, Some properties of enumeration in the theory of modular systems, Proc. London Math. Soc. 26 (1927), no. 1, 531-555. [77] P. Mani-Levitska, S. Vrećica, R. Živaljević, Topology and combinatorics of partitions of masses by hyperplanes, Advances in Math. 207 (2006), no. 1, 266-296. [78] J. Mason, Matroids: unimodal conjectures and Motzkin’s theorem, 207-220, in: D. Welsh, D. Woodall, Combinatorics, Proc. Conf. Combinatorial Math., Math. Inst., Oxford, 1972. [79] J. Matoušek, Using the Borsuk-Ulam theorem: Lectures on Topological Methods in Combinatorics and Geometry, Springer-Verlag, Berlin, 2003. [80] B. Matschke, A note on mass partitions by hyperplanes, arXiv 1001.0193. [81] M. Michałek, Constructive degree bounds for group-based models, J. Combin. Theory Ser. A 120 (2013), no. 7, 1672-1694. [82] E. Miller, B. Sturmfels, Combinatorial commutative algebra, Graduate Texts in Mathematics 227, Springer-Verlag, New York, 2005. [83] M. Minsky, Steps toward artificial intelligence, Proc. IRE 49 (1961), no. 1, 8-30. [84] C. Nash-Williams, Decomposition of finite graphs into forests, J. London Math. Soc. 39 (1964), 12. [85] J. Oxley, Matroid Theory, Oxford Science Publications, Oxford University Press, Oxford, 1992. 70 BIBLIOGRAPHY [86] J. Oxley, What is a matroid?, Cubo Mat. Educ. 5 (2003), no. 3, 179-218. [87] J. Oxtoby, Measure and Category, Graduate Texts in Mathematics 2, Springer-Verlag, New York-Berlin, 1980. [88] A. Pastine, F. Zanello, Two unfortunate properties of pure f-vectors, arXiv:1211.3708. [89] J. Provan, L. Billera, Decompositions of simplicial complexes related to diameters of convex polyhedra, Math. Oper. Res. 5 (1980), no. 4, 576-594. [90] E. Ramos, Equipartition of mass distributions by hyperplanes, Discrete Comput. Geom. 15 (1996), no. 2, 147-167. [91] G. Reisner, Cohen-Macaulay quotients of polynomial rings, Advances in Math. 21 (1976), no. 1, 30-49. [92] U. Schauz, Mr. Paint and Mrs. Correct, Electron. J. Combin. 16 (2009), no. 1, RP77. [93] U. Schauz, A Paintability Version of the Combinatorial Nullstellensatz, and List Colorings of k-partite k-uniform Hypergraphs, Electron. J. Combin. 17 (2010), RP176. [94] U. Schauz, Flexible Color Lists in Alon and Tarsi’s Theorem, and Time Scheduling with Unreliable Participants, Electron. J. Combin. 17 (2010), no. 1, RP13. [95] A. Schrijver, Combinatorial Optimization, Polyhedra and Efficiency, Springer-Verlag, New York, 2003. [96] J. Schweig, Toric ideals of lattice path matroids and polymatroids, J. Pure Appl. Algebra 215 (2011), no. 11, 2660-2665. [97] P. Seymour, Decomposition of Regular Matroids, J. Combin. Theory Ser. B 28 (1980), no. 3, 305-359. [98] P. Seymour, A note on list arboricity, J. Combin. Theory Ser. B 72 (1998), 150-151. [99] R. Stanley, Cohen-Macaulay complexes, 51-62 in: M. Aiger, Higher Combinatorics, 31 NATO Adv. Study Inst. Ser. C: Math. Phys. Sci. 1977. [100] R. Stanley, The number of faces of a simplicial convex polytope, Advances in Math. 35 (1980), no. 3, 236-238. [101] R. Stanley, Combinatorics and Commutative Algebra, Second Ed., Progress in Mathematics 41, Birkhäuser Boston, Boston, 1996. [102] B. Sturmfels, Gröbner Bases and Convex Polytopes, Univ. Lecture Series 8, American Mathematical Society, Providence, 1995. [103] B. Sturmfels, Equations defining toric varieties, Proc. Sympos. Pure Math. 62 (1997), 437-449. [104] C. Thomassen, Every planar graph is 5-choosable, J. Combin. Theory Ser. B 62 (1994), no. 1, 180-181. [105] W. Tutte, A homotopy theorem for matroids I, II, Trans. Amer. Math. Soc. 88 (1958), 144-174. [106] V. Vizing, On an estimate of the chromatic class of a p-graph, Diskret. Analiz. 3 (1964), 25-30 [in Russian]. [107] V. Vizing, Coloring the vertices of a graph in prescribed colors, Diskret. Analiz. 29 (1976), 3-10 [in Russian]. [108] S. Vrećica, R. Živaljević, Measurable patterns, necklaces, and sets indiscernible by measure, arXiv:1305.7474. [109] D. Welsh, Combinatorial problems in matroid theory, 291-306, in: Combinatorial Mathematics and its Applications (Proc. Conf., Oxford, 1969), Academic Press, London, 1971. [110] D. West, Introduction to Graph Theory, Prentice Hall, New York, 2001. [111] N. White, The basis monomial ring of a matroid, Advances in Math. 24 (1977), no. 3, 292-297. [112] N. White, A unique exchange property for bases, Linear Algebra and its App. 31 (1980), 81-91. [113] H. Whitney, On the abstract properties of linear dependence, Amer. J. Math. 57 (1935), no. 3, 509-533. [114] D. Woodall, An exchange theorem for bases of matroids, J. Combin. Theory Ser. B 16 (1974), 227-228. [115] X. Zhu, The game coloring number of planar graphs, J. Combin. Theory Ser. B 75 (1999), no. 2, 245-258. [116] X. Zhu, On-line list colouring of graphs, Electron. J. Comb. 16 (2009), no. 1, RP127.
0math.AC
arXiv:1701.06124v1 [math.RA] 22 Jan 2017 THE RADICAL OF THE KERNEL OF A CERTAIN DIFFERENTIAL OPERATOR AND APPLICATIONS TO LOCALLY ALGEBRAIC DERIVATIONS WENHUA ZHAO Abstract. Let R be a commutative ring, A an R-algebra (not necessary commutative) and V an R-subspace or R-submodule of A. By the radical of V we mean the set of all elements a ∈ A such that am ∈ V for all m ≫ 0. We derive (and show) some necessary conditions satisfied by the elements in the radicals of the kernels of some (partial) differential operators such as all differential operators of commutative algebras; the differential operators P (D) of (noncommutative) A with certain conditions, where P (·) is a polynomial in n commutative free variables and D = (D1 , D2 , . . . , Dn ) are either n commutating locally finite R-derivations or n commutating R-derivations of A such that for each 1 ≤ i ≤ n, A can be decomposed as a direct sum of the generalized eigen-subspaces of Di ; etc. We then apply the results mentioned above to study Rderivations of A that are locally algebraic or locally integrable over R. In particular, we show that if R is an integral domain of characteristic zero and A is reduced and torsion-free as an R-module, then A has no nonzero locally algebraic R-derivations. We also show a formula for the determinant of a differential vandemonde matrix over commutative algebras. This formula not only provides some information for the radicals of the kernels of ordinary differential operators of commutative algebras, but also is interesting on its own right. 1. Background and Motivation Let R be a commutative ring and A an R-algebra (not necessary commutative). A derivation D of A is a map from A to A such that D(ab) = D(a)b + aD(b) for all a, b ∈ A. If D is also R-linear, we call it an R-derivation of A. Date: February 25, 2018. 2000 Mathematics Subject Classification. 47F05, 47E05, 16W25, 16D99. Key words and phrases. The radical, the kernel of a differential operator, locally algebraic or integral derivations, a differential vandemonde determinant, Mathieu subspaces (Mathieu-Zhao spaces). The author has been partially supported by the Simons Foundation grant 278638. 1 2 WENHUA ZHAO For each a ∈ A, we denote by ℓa the map from A to A that maps b ∈ A to ab. We call the associative algebra generated by ℓa (a ∈ A) and all derivations of A the Weyl algebra of A, and denote it by W(A). The subalgebra of W(A) generated by ℓa (a ∈ R) and all R-derivations of A will be denoted by WR (A). Elements of W(A) are called differential operators of A. For each Φ ∈ W(A), it is well-known and also easy to check that there exist some derivations D = (D1 , D2 , . . . , Dn ) of A and a polynomial P (ξ) ∈ A[ξ] in n noncommutative free variables ξ = (ξ1 , ξ2, . . . , ξn ) such that Φ = P (D), where P (D) throughout this paper is defined by first writing all the coefficients of P (ξ) on the left and then replacing ξi by Di for all 1 ≤ i ≤ n. Furthermore, if Φ ∈ WR (A), the same is true with Di (1 ≤ i ≤ n) being R-derivations of A and P (ξ) ∈ R[ξ]. We call the differential operator Φ = P (D) an ordinary differential operator of A, if P (ξ) is univariate, and a partial differential operator of A if P (ξ) is multivariate. Next, we recall the following two notions of associative algebras that were first introduced in [Z2, Z3]. Definition 1.1. An R-subspace V of an R-algebra A is said to be a Mathieu subspace (MS) of A if for all a, b, c ∈ A with am ∈ V for all m ≥ 1, we have bam c ∈ V for all m ≫ 0. Note that a MS is also called a Mathieu-Zhao space in the literature (e.g., see [DEZ, EN, EH], etc.), as suggested by A. van den Essen [E2]. The introduction of this notion is mainly motivated by the study in [M, Z1] of the well-known Jacobian conjecture (see [Ke, BCW, E1]). See also [DEZ]. But, a more interesting aspect of the notion is that it provides a natural but highly non-trivial generalization of the notion of ideals. Definition 1.2. [Z3, p. 247] Let V be an R-subspace (or a subset) of an R-algebra A. We define the radical r(V ) of V to be (1.1) r(V ) := {a ∈ A | am ∈ V for all m ≫ 0}. When A is commutative and V is an ideal of A, r(V ) coincides with the radical of V . So this new notion is also interesting on its own right. It is also crucial for the study of MSs. For example, it is easy to see that every R-subspace V of an R-algebra A with r(V ) = nil (A) is a MS of A, where nil (A) denotes the set of all nilpotent elements of A. We will frequently use this fact (implicitly) throughout this paper. Recent studies show that many MSs arise from the images of differential operators, especially, from the images of locally finite or locally nilpotent derivations, of certain associative algebras (e.g., see THE RADICAL OF THE KERNEL OF A DIFFERENTIAL OPERATOR 3 [Z1, Z2, EWZ, EZ] and [Z4]–[Z7], etc.). For some MSs arisen from the kernels of some ordinary differential operators of univariate polynomial algebras over a field, see [EN, EH]. In this paper we study the radicals of the kernels of some ordinary or partial differential operators of A and show that for certain differential operators Φ, the kernel Ker Φ is also a MS of A. We also apply some results proved in this paper to study R-derivations of A that are locally algebraic or locally integrable over R (see Definition 4.1). In particular, we show that if R is an integral domain of characteristic zero and A is reduced and torsion-free as an R-module, then A has no nonzero R-derivation that is locally algebraic over R (see Theorem 4.6). Furthermore, we also show a formula for the determinant of a differential vandemonde matrix over commutative algebras (see Proposition 5.1). This formula not only provides some information for the radicals of the kernels of ordinary differential operators of commutative algebras, but also is interesting on its own right. Arrangement and Content: In Section 2, we assume that A is commutative and derive some necessarily conditions for the elements in the radical of the kernel of an arbitrary differential operators of A (see Theorem 2.1 and Corollary 2.4). In particular, for every differential operator Φ ∈ W(A) such that Φ1A is not zero nor a zero-divisor of A, the kernel Ker Φ forms a MS of A. In Section 3, we drop the commutativity assumption on A but assume that (R, +) is torsion-free and A is reduced and torsion-free as an R-module. We first derive in Theorem 3.1 some necessary conditions satisfied by the elements in the radical of the kernel of a differential operator P (D) of A, where P (·) is a polynomial in n commutative free variables and D = (D1 , D2 , . . . , Dn ) are n commutating R-derivations of A such that for each 1 ≤ i ≤ n, A can be decomposed as a direct sum of the generalized eigen-subspaces of Di . We then show in Proposition 3.6 that if R also is an integral domain of characteristic zero, then the conclusions in Theorem 3.1 also hold for the differential operators of A which are multivariate polynomials in commuting locally finite R-derivations of A. Finally, we show in Proposition 3.7 that similar conclusions as those in Proposition 3.6 (with the same assumptions on R and A) also hold for all ordinary differential operators of A. In particular, for all the differential operators Φ in Theorem 3.1 and Propositions 3.6 and 3.7 with Φ1A 6= 0, Ker Φ forms a MS of A. In Section 4, we apply some results proved in Sections 2 and 3 to study some properties of R-derivation of A that are locally algebraic or 4 WENHUA ZHAO locally integral over R (see Definition 4.1). We first show in Theorem 4.3 that if A is commutative and (A, +) is torsion-free, then every locally integral D of A has its image in the nil-radical nil (A) of A. We also show in Theorem 4.6 that if R also is an integral domain of characteristic zero and A is reduced and torsion-free as an R-module, then A has no nonzero R-derivation that is locally algebraic over R. In Section 5, we assume that A is commutative and first show in Proposition 5.1 a formula for the determinant of a differential vandemonde matrix over A. We then apply this formula in Proposition 5.4 to derive more necessary conditions satisfied by the elements in the radicals of the kernels of all ordinary differential operators of A. we point out in Remark 5.3 that the formula derived in Proposition 5.1 can also be used to derive formulas for the determinants of several other families of matrices. 2. The Commutative Algebra Case In this section, unless stated otherwise, R denotes a unital commutative ring, A a commutative unital R-algebra and ξ = (ξ1 , ξ2, . . . , ξn ) n noncommutative free variables. We denote by A[ξ] the polynomial algebra in ξ over A, and ∂i (1 ≤ i ≤ n) the A-derivation ∂/∂ξi of A[ξ]. Once and for all, we fix in this section a nonzero P (ξ) ∈ A[ξ] and n R-derivations Di (1 ≤ i ≤ n) of A. Write D = (D1 , D2 , . . . , Dn ) and P P (ξ) = a0 + dk=1 Pk (ξ) for some a0 ∈ A, d ≥ 1 and homogeneous polynomials Pk (ξ) (1 ≤ k ≤ d) of degree k in ξ. For each u ∈ A, we set ∇D u := (D1 u, D2 u, . . . , Dn u), and call it the gradient of u with respect to D. When D is clear in the context, we will simply write ∇D u as ∇u. We define P (D) and P (∇u) by first writing P (ξ) as a polynomial in ξ with all the coefficients on the most left (of the monomials), and then replacing ξi by Di and Di u, respectively, for each 1 ≤ i ≤ n. The main result of this section is the following Theorem 2.1. With the setting as above, let u ∈ A such that um ∈ Ker P (D) for all 1 ≤ m ≤ d. Then (2.1) a0 ud = (−1)d d! Pd(∇u). Furthermore, if ud+1 also lies in Ker P (D), then (2.2) a0 ud+1 = 0. To show the theorem above, we first need the following two lemmas. The first lemma is well-known and can also be easily verified by using THE RADICAL OF THE KERNEL OF A DIFFERENTIAL OPERATOR 5 the mathematical induction, which is similar as the proof for the usual binomial formula. Lemma 2.2. Let u ∈ A and ℓu : A → A that maps a ∈ A to ua. Denote by adu : W(A) → W(A) that maps each Λ ∈ W(A) to [u, Λ] := ℓu Λ − Λℓu . Then for all Φ ∈ W(A) and k ≥ 1, we have   k k−i u Φ ◦ ℓiu . (adu ) (Φ) = (−1) i i=0 k (2.3) k X i Lemma 2.3. Let u ∈ A. Then the following statements hold: 1) there exists Q(ξ) ∈ A[ξ] with either Q(ξ) = 0 or deg Q(ξ) ≤ d − 2 such that (2.4) ad−u P (D) = n X (Di u)(∂i P )(D) + Q(D). i=1 (ad−u )d P (D) = d!P (∇u). 2) Proof: 1) First, if deg P (ξ) = 0, then the statement holds trivially, for A is commutative and hence ad−u P (ξ) = 0. So we assume deg P (ξ) ≥ 1. By the linearity and also the commutativity of A we may assume P (ξ) = ξi1 ξi2 · · · ξik with 1 ≤ ij ≤ n for all 1 ≤ j ≤ k. We use the induction on k ≥ 1. If k = 1, then ad−u Di1 = ℓDi1 u . Hence the statement holds by choosing Q(ξ) = 0. Assume that the statement holds for all 1 ≤ k ≤ m − 1 and consider the case k = m. Since ad−u is a derivation of W(A), we have ad−u P (D) = m X Di1 · · · (ad−u Dij ) · · · Dim = = j=1 ci · · · Dim + (ℓDij u )Di1 · · · D j Di1 · · · (ℓDij u ) · · · Dim j=1 j=1 m X m X m X [Di1 · · · Dij−1 , ℓDij u ]Dij+1 · · · Dim j=2 ci means that the term Di is omitted: Here D j j = m X j=1 ci · · · Dim + (Dij u)Di1 · · · D j m X (ad−Dij u (Di1 · · · Dij−1 ))Dij+1 · · · Dim . j=2 Applying the induction assumption to the terms ad−Dij u (Di1 · · · Dij−1 ) (2 ≤ j ≤ m) in the sum above we see that there exists Q(ξ) ∈ A[ξ] 6 WENHUA ZHAO with Q(ξ) = 0 or deg Q(ξ) ≤ m − 2 such that ad−u P (D) = = m X j=1 n X ci · · · Dim + Q(D) (Dij u)Di1 · · · D j (Di u)(∂i P )(D) + Q(D). i=1 Hence by the induction statement 1) follows. 2) First, by statement 1) it is easy to see that (ad−u )d P (D) = (ad−u )d Pd (D). Then by the linearity and also the commutativity of A we may assume Pd (ξ) = ξi1 ξi2 · · · ξid with 1 ≤ ij ≤ n for all 1 ≤ j ≤ d. Applying statement 1) (d times) we have X (Dk1 u)(Dk2 u) · · · (Dkd u)(∂k1 ∂k2 · · · ∂kd P ). (ad−u )d P (D) = 1≤k1 ,k2 ,...,kd ≤n Then by the equation above and the commutativity of A, it is easy to see that the equation in statement 2) follows. ✷ Now we can prove the main result of this section. Proof of Theorem 2.1: By Eq. (2.3) and Lemma 2.3, 2) we have   d X i d d (2.5) ud−i P (D) ◦ ℓiu . d!Pd (∇u) = (−1) (−1) i i=0 By applying both sides of the equation above to 1A ∈ A and then using the condition ui ∈ Ker P (D) (1 ≤ i ≤ d), we get d!Pd (∇u) = (−1)d ud P (D) · 1A . It is well-known and also easy to check that every derivation of a commutative ring annihilates the identity element of the ring. Hence d!Pd (∇u) = (−1)d ud a0 , i.e., Eq. (2.1) follows. Similarly, by applying Eq. (2.5) above to u ∈ A and using the condition ui ∈ Ker P (D) (1 ≤ i ≤ d + 1), we get d!Pd (∇u)u = 0. Then by Eq. (2.1) we get Eq. (2.2). ✷ One immediate consequence of Theorem 2.1 is the following Corollary 2.4. Let D, P (ξ), a0 be as in Theorem 2.1, and nil (A) the nil-radical of A, i.e., the set of all nilpotent elements of A. Then the following statements hold: 1) r(Ker Λ) ⊆ Ann (a0 ), where Ann (a0 ) is the set of the elements b ∈ A such that a0 b = 0; 2) if a0 is not zero nor a zero-divisor of A, then r(Ker P (D)) = nil (A) and Ker P (D) is a MS of A; THE RADICAL OF THE KERNEL OF A DIFFERENTIAL OPERATOR 7 3) if a0 = 0, then r(Ker P (D)) ⊆ {u ∈ A | Pd(∇u) = 0}. In particular, if n = 1, i.e., D is a single derivation of A, and the leading coefficient of P (ξ) is not a zero-divisor of A, then r(Ker P (D)) ⊆ {u ∈ A | Du ∈ nil (A)}. Example 2.5. Let R = C and A the C-algebra of all smooth complex d . Then for each nonzero valued functions f (x) over R. Let D = dx univariate polynomial P (ξ) ∈ C[ξ], Ker P (D) is the set of solutions f (x) ∈ A of the ordinary differential equation P (D)f = 0. Let λi (1 ≤ i ≤ k) be the set of all distinct roots of P (ξ) in C with multiplicity mi . Then it is well-known in the theory of ODE (e.g., see [L] or any other standard text book on ODE) that Ker P (D) is the C-subspace of A spanned by xj eλi x for all 1 ≤ i ≤ k and 1 ≤ j ≤ mi . From the fact above it is easy to verify directly that r(Ker P (D)) = {0}, if P (0) 6= 0; and r(Ker P (D)) = C, if P (0) = 0. Consequently, Theorem 2.1, Corollary 2.4 and also Proposition 5.4 in Section 5 all hold in this case. We end this section with the following two remarks. First, we will show in Propositions 3.7 and 5.4 that for the ordinary differential operators Φ of certain R-algebras A (not necessarily commutative), the radical r(Ker Φ) also satisfies some other necessarily conditions (other than those in Theorem 2.1). Second, Theorem 2.1 and Corollary 2.4 do not always hold for the differential operators of a noncommutative algebra, which can be seen from the following Example 2.6. Let X, Y be two noncommutative free variables and R[X, Y ] the polynomial algebra in X and Y over R. Let J be the two-sided ideal of R[X, Y ] generated by Y 2 and A := R[X, Y ]/J. Let D = ∂/∂X and P (ξ) = 1 − Xξ ∈ A[ξ]. Then P (D) = I − ℓX D, where I denotes the identity map of A, and ℓX the multiplication map by X from the left. Let f = XY ∈ A. Then it is easy to check that for all m ≥ 1, we have P (D)(f m) = 0 but P (D)(Xf m) = −Xf m 6= 0. Therefore, 0 6= f ∈ r(Ker P (D)) and Ker P (D) is not a MS of A. 3. Some Cases for Non-Commutative Algebras In this section, unless stated otherwise, R denotes a commutative ring such that the abelian group (R, +) is torsion-free, and A an Ralgebra (not necessarily commutative) that is torsion-free as an Rmodule. We denote by IA or simply I the identity map of A, and nil (A) the set of all nilpotent elements of A. We say A is reduced if nil (A) = {0}. 8 WENHUA ZHAO Furthermore, for each a ∈ A, we denote by Ann ℓ (a) the set of elements b ∈ A such that ab = 0. Let D be an R-derivation of A. We say that A is decomposable w.r.t. (with respect to) the R-derivation D if A can be written as a direct sum of the generalized eigen-subspaces of D. More precisely, P let H be thei set of all generalized eigenvalues of D in R and Aλ = ∞ i=1 Ker(D−λI) for each λ ∈ H. Then A = ⊕λ∈H Aλ . It is easy to verify inductively that for all m ≥ 1, a, b ∈ A and λ, µ ∈ R, we have m   X   m m (3.6) D − (λ + µ)I (ab) = (D − λI)i a (D − µI)m−i b . i i=0 Then by the identity above we have that AλP Aµ ⊆ Aλ+µ for all λ, µ ∈ H. i In other words, the decomposition Aλ = ∞ i=1 Ker(D − λI) above is actually an additive R-algebra grading of A. Some examples of R-derivations with respect to which A is decomposable are semi-simple R-derivations, for which Aλ (λ ∈ H) coincides with the eigen-space of D corresponding to the eigenvalue λ of D, and also locally finite derivations when the base ring R is an algebraically closed field. Once and for all, we let Di (1 ≤ i ≤ n) be n commuting R-derivations of A, i.e., Di Dj = Dj Di for all 1 ≤ i, j ≤ n, such that A is decomposable w.r.t. each Di . Then there exists a semi-subgroup Λ of the abelian group (Rn , +) such that A = ⊕λ∈Λ Aλ , (3.7) where for each λ = (k1 , k2 , . . . , kn ) ∈ Λ, (3.8) Aλ = n ∞ \ X i=1 In particular, (3.9) A0 = j=1  Ker(Di − ki I)j . n ∞ \ X i=1 j=1  Ker Dij . Note also that each Aλ (λ ∈ Λ) is invariant under Di (1 ≤ i ≤ n), and Aλ Aµ ⊆ Aλ+µ for all λ, µ ∈ Λ. Now, let ξ = (ξ1 , ξ2, . . . , ξn ) be n commutative free variables and 0 6= P (ξ) ∈ R[ξ]. We set D := (D1 , D2 , . . . , Dn ) and write and P (ξ) = Pd k=0 Pk (ξ) for some d ≥ 0 and homogeneous polynomials Pk (ξ) (1 ≤ k ≤ d) of degree k in ξ. We let P (D) be the differential operator of A obtained by replacing ξi by Di (1 ≤ i ≤ n). Since Di ’s commute with one anther, P (D) is well-defined. THE RADICAL OF THE KERNEL OF A DIFFERENTIAL OPERATOR 9 The first main result of this section is the following theorem which in some sense extends Theorem 2.1 to the differential operator P (D) of the R-algebra A which is not necessarily commutative. Theorem 3.1. With the setting as above, assume further that A is reduced. Then the following statements hold: 1) if P0 = P (0) = 0 and Pk (ξ) (1 ≤ k ≤ d) have no nonzero common zeros in Rn , then r(Ker P (D)) ⊆ A0 ; 2) if P0 = P (0) 6= 0, then r(Ker P (D)) = {0}, and Ker P (D) is a MS of A. In order to show the theorem above, we first need to show some lemmas. Lemma 3.2. Let R be an arbitrary commutative ring and A an Ralgebra that is torsion-free as an R-module. Let D and P (ξ) be fixed as above. Then the following statements hold: 1) Ker P (D) is homogeneous w.r.t. the grading of A in Eq. (3.7), i.e., M (3.10) Ker P (D) = (Aλ ∩ Ker P (D)). λ∈Λ 2) Let ZΛ (P ) be the set of λ ∈ Λ such that P (λ) = 0. Then M (3.11) Ker P (D) ⊆ Aλ . λ∈ZΛ (P ) Proof: 1) Since for each λ ∈ Λ, Aλ is preserved by Di (1 ≤ i ≤ n), and hence also is preserved by P (D), from which Eq. (3.10) follows. P 2) Let 0 6= u ∈ A and write u = ℓi=1 uλi for some distinct λi ∈ Λ (1 ≤ i ≤ n) and uλi ∈ Aλi . Then by Eq. (3.10) we have that u ∈ Ker P (D), if and only if uλi ∈ Ker P (D). So we may assume ℓ = 1 and u ∈ Aλ for some λ ∈ Λ. Write λ = (k1 , k2 , . . . , kn ). For each 1 ≤ j ≤ n, we define a nonnegative integer rj as follows. First, let rn be the greatest non-negative integer such that (Dn − kn I)rn u 6= 0, and inductively, for each 1 ≤ j ≤ n − 1, let rj be the  great Q n rs u 6= 0. non-negative integer such that (Dj −kj I)rj s=j+1 (Ds − ks I)  Q n rj u. Then 0 6= ũ ∈ Aλ , ũ ∈ Ker P (D), Set ũ := j=1 (Dj − kj I) and Dj ũ = kj ũ for all 1 ≤ j ≤ n. Hence 0 = P (D)ũ = P (λ)ũ. Since A is torsion-free as an R-module, we have P (λ) = 0, as desired. ✷ 10 WENHUA ZHAO Definition 3.3. Let A be a subset of Rn and λ ∈ A. We say λ is an extremal element of A if for all m ≥ 1, mλ can not be written as a linear combination of other elements of A with positive integer coefficients whose sum is less or equal to m. The following lemma should be well-known. But for the sake of completeness, we here include a direct proof. Lemma 3.4. Let R be a commutative ring such that the abelian group (R, +) is torsion free. Then every nonempty finite subset A of Rn has at least one extremal element. Proof: Write A = {λ1 , λ2 , . . . , λn } with λi 6= λj for all 1 ≤ i 6= j ≤ n. We use the induction on n. If n = 1, there is nothing to show. So we assume n ≥ 2. Consider first the case n = 2 with λ2 6= 0. If the lemma fails, then m1 λ1 = k1 λ2 and m2 λ2 = k2 λ1 for some mi , ki ≥ 1 with ki ≤ mi . Then m1 m2 λ2 = m1 (k2 λ1 ) = k2 (m1 λ1 ) = k1 k2 λ2 . Hence m1 m2 = k1 k2 , for λ2 6= 0 and (R, +) is torsion-free, from which we have m1 = k1 (and m2 = k2 ). By the fact that (R, +) is torsion-free again, we have λ1 = λ2 . Contradiction. Now assume the lemma holds for all 2 ≤ n ≤ k and consider the case n = k + 1. If λk+1 is an extremal point of A, then there is nothing to show. Assume otherwise. Then there exist m ≥ 1 and ci ∈ N (1 ≤ i ≤ k) such that (3.12) k X mλk+1 = ci λ i , i=1 (3.13) 1≤ k X ci ≤ m. i=1 By the induction assumption the set A′ := {λ1 , λ2 , . . . , λk } has an extremal element, say, λ1 . We claim that λ1 is also an extremal point of the set A. Otherwise, there exist q ≥ 1 and c′j ∈ N (2 ≤ j ≤ k + 1) such that (3.14) (3.15) qλ1 = c′k+1 λk+1 + 1 ≤ c′k+1 + X j=1 k X c′j λj , j=2 c′j ≤ q. THE RADICAL OF THE KERNEL OF A DIFFERENTIAL OPERATOR 11 Then by Eqs. (3.14) and (3.12) we have mqλ1 = mc′k+1 λk+1 + m (3.16) k X c′j λj j=1 = c′k+1 k X ci λ i + m i=1 k X c′j λj . j=1 For the sum of all the coefficients of the linear combination on the right hand side of the equation above, by Eqs. (3.13) and (3.15) we have (3.17) 1≤ c′k+1 k X ci + m i=1 = m(c′k+1 + X k X c′j ≤ c′k+1 m +m j=1 c′j ) k X c′j j=1 ≤ mq. j=1 Then by Eqs. (3.16) and (3.17), λ1 is not an extremal element of A′ , which contradicts to the choice of λ1 . Therefore λ1 is an extremal point of A, and the lemma follows. ✷ P Lemma 3.5. Let 0 6= u ∈ r(Ker P (D)) and write u = ℓi=1 uλi for some distinct λi ∈ Λ (1 ≤ i ≤ ℓ) and 0 6= uλi ∈ Aλi . Then for each extremal element λj of the set {λi | 1 ≤ i ≤ ℓ}, either uλj is nilpotent, or Pk (λj ) = 0 for all 0 ≤ k ≤ d. Proof: Assume that uλj is not nilpotent. Since λj is an extremal element of the set {λi | 1 ≤ i ≤ ℓ}, it is easy to see that for each m ≥ 1, the homogeneous component of um in Amλj is equal to um λj . m Since u ∈ Ker P (D) when m ≫ 0, by Lemma 3.2, 1) and 2) we have um λj ∈ Ker P (D) and P (mλj ) = 0 for all m ≫ 0. More explicitly, for all m ≫ 0, we have 0 = P (mλj ) = d X mk Pk (λj ). k=0 Since (R, +) is torsion-free, by the vandemonde determinant we have Pk (λj ) = 0 for all 0 ≤ k ≤ d. ✷ Proof of Theorem 3.1: Let 0 6= u ∈ r(Ker P (D)) and write u = Pℓ i=1 uλi for some distinct λi ∈ Λ (1 ≤ i ≤ ℓ) and 0 6= uλi ∈ Aλi . Let B be the set of all nonzero λi (1 ≤ i ≤ ℓ). If B 6= ∅, then by Lemma 3.4, B has at least one extremal element, say λj . Then by Definition 3.3, 12 WENHUA ZHAO λj is also an extremal element of the set B ∪ {0}. Since A is reduced, uλj is not nilpotent. Then by Lemma 3.5, Pk (λj ) = 0 for all 0 ≤ k ≤ d. If P0 = 0, and Pk (ξ) (1 ≤ k ≤ d) have no nonzero common zero in n R , then we have λj = 0, which is a contradiction. Therefore in this case B = ∅ and u ∈ A0 , whence the statement 1) follows. If P0 6= 0, then we also have B = ∅ and u ∈ A0 , for P0 (λj ) = P0 6= 0. Furthermore, since P (0) = P0 6= 0, by Lemma 3.2, 2) we have A0 ∩Ker P (D) = 0, whence u = 0. Contradiction. Therefore statement 2) follows. ✷ Next, we show that Theorem 3.1 with some extra conditions also holds for commuting locally finite R-derivations. Recall that an Rderivation δ of an R-algebra A is locally finite (over R) if for each u ∈ A, the R-submodule of A spanned by elements δ k u (k ≥ 0) over R is finitely generated as an R-module. Proposition 3.6. Assume that R is an integral domain of characteristic zero and A is a reduced R-algebra that is torsion-free as an R-module. Denote by KR the field of fractions of R and K̄R the algebraic closure of KR . Let P (ξ) ∈ R[ξ] and D = (D1 , D2 , . . . , Dn ) be n P commuting locally finite R-derivations of A. Write P (ξ) = di=0 Pk (ξ) with Pk (ξ) (0 ≤ k ≤ d) being homogeneous of degree k. Then the following statements hold: 1) if P (0) = 0 and Pk (ξ) (1 ≤ k ≤ d) have no nonzero common zeros in K̄Rn , then we have r(Ker P (D)) ⊆ A0 , where A0 = Tn P ∞ m i=1 ( m=1 Ker Di ); 2) if P (0) 6= 0, then r(Ker P (D)) = {0}, and Ker P (D) is a MS of A. Proof: Set Ā = K̄R ⊗R A. Since A is torsion-free as an R-module, the standard map A ≃ R ⊗R A → KR ⊗R A is injective, for by [AM, Prop. 3.3] KR ⊗R A is isomorphic to the localization S −1 A with S = R\{0}. Since every field is absolutely flat, the standard map KR ⊗R A → K̄R ⊗R A is also injective. Therefore, we may view A as an Rsubalgebra in the standard way and extend D K̄R -linearly to Ā, which we denote by D̄ = (D̄1 , D̄2 , . . . , D̄n ). Note that D̄i (1 ≤ i ≤ n) are n commuting K̄R -derivations of Ā, which are also locally finite over K̄R . Then Ā by [E1, Proposition 1.3.8]) is decomposable w.r.t. D̄i for each 1 ≤ i ≤ n. By applying Theorem 3.1 to P (D̄) and using the fact Ā0 ∩ A = A0 we see that the proposition follows. ✷ THE RADICAL OF THE KERNEL OF A DIFFERENTIAL OPERATOR 13 Next, we use the proposition above to show that Corollary 2.4 with some extra conditions can be extended to the ordinary differential operators of some noncommutative algebras. Proposition 3.7. Let R, A be as in Proposition 3.6 and D an arbitrary (single) R-derivation of A. Then for every univariate polynomial in 0 6= P (ξ) ∈ R[ξ], the following statements hold: P i 1) if P (0) = 0, then r(Ker P (D)) ⊆ r(A0 ), where A0 = ∞ i=1 Ker D ; 2) if P (0) 6= 0, then r(Ker P (D)) = {0}, and Ker P (D) is a MS of A. Proof: The case deg P (ξ) = 0 is trivial. So we assume deg P (ξ) ≥ 1. Let KR be the field of fractions of R with the algebraic closure K̄R , and set Ā = K̄R ⊗R A. As pointed out in the proof of Proposition 3.6 we may view A as an R-subalgebra of Ā in the standard way and extend D K̄R -linearly to Ā, which we denote by D̄. Let V = Ker P (D). Then V is an R-subspace of A preserved by D. Set V̄ = K̄R ⊗R V . Then D̄ |V̄ as a K̄R -linear map from V̄ to V̄ is algebraic over K̄R , for P (D |V ) = P (D) |V = 0 and hence P (D̄ |V̄ ) = 0. It is well-known (e.g., see [H, Proposition 4.2]) that V̄ can be decomposed as a direct sum of the generalized eigen-spaces of D̄ |V̄ . Let B̄ be the K̄R -subalgebra of Ā generated by elements of V̄ . Then B̄ is D̄-invariant. Furthermore, by Eq. (3.6) it is easy to see that B̄ is decomposable w.r.t. D̄ |B . Now let u ∈ r(Ker P (D)). Then there exists N ≥ 1 such that um ∈ Ker P (D), and hence is also in B̄, for all m ≥ N. Consequently, we also have um ∈ r(Ker P (D̄ |B̄ )) for all m ≥ N. Note that Pk (ξ) (1 ≤ k ≤ d) have no nonzero commons zero in K̄R , for P (ξ) is a univariate polynomial of degree greater or equal to 1. Therefore, if P (0) = 0, then by applying Proposition 1) to P (D̄ |B̄ ) (as a differential operator P3.6, ∞ of B̄), we have um ∈ Ker D̄ i for all m ≥ N. Since Ker D i = i=1 P i A ∩ Ker D̄ i for all i ≥ 1, we further have um ∈ A0 = ∞ i=1 D for all m ≥ N. Hence u ∈ r(A0 ) and statement 1) follows. If P (0) 6= 0, then by applying Proposition 3.6, 2) to P (D̄ |B̄ ) (as a differential operator of B̄), we have um = 0 for all m ≥ N. Since A is reduced, we have u = 0. Hence statement 2) also follows. ✷ We end this section with the following open problem which is worthy of further investigations. Open Problem 3.8. Let R be an arbitrary commutative ring and A an arbitrary unital noncommutative R-algebra. Let D = (D1 , D2 , . . . , Dn ) be n R-derivations of A, and Q(ξ) ∈ R[ξ] a polynomial in n noncommutative free variables ξ = (ξ1 , ξ2 , . . . , ξn ). Set a0 := Q(0) and denote 14 WENHUA ZHAO by Ann ℓ (a0 ) the set of all elements b ∈ A such that a0 b = 0. Decide  whether or not it is always true that r(Ker Q(D)) ⊆ r Ann ℓ (a0 ) ? 4. Some Applications to Locally Algebraic Derivations In this section we use some results proved in the last two sections to derive some properties of locally algebraic or locally integral derivations. Definition 4.1. Let R be a unital commutative ring, A an R-algebra and D an R-derivation of A. 1) We say D is algebraic over R if there exists a nonzero polynomial p(t) ∈ R[t] such that p(D) = 0. 2) We say D is locally algebraic over R if for each a ∈ A, there exists a D-invariant R-subalgebra A1 of A containing a, and a nonzero polynomial pa (t) ∈ R[t] such that pa (D) A1 = 0. If p(t) in statement 1) (resp., pa (t) in statement 2) for all a ∈ A) of the definition above can be chosen to be a monic polynomial, we say D is integral (resp., locally integral) over R. An example of a derivation that is locally algebraic but not algebraic is as follows. Example 4.2. Let xi (i ≥ 1) be a sequence of free commutative variables and C[xi | i ≥ 1] the polynomial algebra over C in xi (i ≥ 1). Let I be the ideal generated by xi+1 (i ≥ i P1) and A = C[xi | i ≥ 1]/I. Then it can be readily verified that D := ∞ i=1 xi ∂/∂xi is a well-defined C-derivation of A, which is locally algebraic but not (globally) algebraic over C. Theorem 4.3. Let R be a commutative ring and A a commutative R-algebra such that the abelian group (A, +) is torsion-free. Then for every R-derivation D of A that is locally integral over R, the image Im D := D(A) ⊆ nil (A), where nil (A) denotes the nil-radical of A, i.e., the set of nilpotent elements of A. Proof: Let a ∈ A, and A1 be a D-invariant R-subalgebra of A and pa (t) a monic polynomial in R[t] such that a ∈ A1 and pa (D) |A1 = 0. Then A1 ⊆ Ker pa (D). In particular, am ∈ Ker pa (D) for all m ≥ 1. Replacing pa (t) by tpa (t) we assume pa (0) = 0 and d := deg pa (t) ≥ 1. Then by Theorem 2.1 we have d!(Da)d = 0. Since (A, +) as an abelian group is torsion-free, we have (Da)d = 0, whence Da ∈ nil (A) and the theorem follows. ✷ Since every nilpotent R-derivation of A is locally integrable over R, by Theorem 4.3 we immediately have the following THE RADICAL OF THE KERNEL OF A DIFFERENTIAL OPERATOR 15 Corollary 4.4. Let R, A be as in Theorem 4.3 and D a nilpotent R-derivation of A. Then Im D ⊆ nil (A). Furthermore, from the proof of Theorem 4.3 it is also easy to see that we have the following Corollary 4.5. Let R and A be as in Theorem 4.3. Assume further that A is torsion-free as an R-module. Then for every R-derivation D of A that is locally algebraic over R, we have Im D ⊆ nil (A). Next, we consider the R-derivations of some reduced R-algebra A (not necessarily commutative) that are locally algebraic over R. Theorem 4.6. Let R be a unital integral domain of characteristic zero and A a unital reduced R-algebra (not necessarily commutative) that is torsion-free as an R-module. Then A has no nonzero R-derivations that are locally algebraic over R. In particular, A has no nonzero nilpotent R-derivations. Proof: Let D be an R-derivation of A that is locally algebraic over R. Let a ∈ A, and A1 be a D-invariant R-subalgebra of A and 0 6= pa (t) ∈ R[t] such that a ∈ A1 and pa (D) |A1 = 0. Then am ∈ A1 ⊆ Ker pa (D) for all m ≥ 1, whence a ∈ r(Ker pa (D)). Replacing pa (t) by tpa (t) we assume pa (0) = 0. Then by applying Proposition 3.7, 1) P to the differential operator pa (D), we have a ∈ i r(A0 ), where A0 = ∞ i=1 Ker D . Consequently, r(A0 ) = A. Then by [Z3, Lemma 2.4] we have A0 = A, i.e., D is locally nilpotent. Let KR be the field of fractions of R and B := KR ⊗R A. As pointed out in the proof of Proposition 3.6, we may view A as an R-subalgebra of B in the standard way and extend D KR -linearly to B, which we denote by D̄. Let a, pa (t) be fixed as above, and N ≥ 1 such that D N a = 0. Write pa (t) = tk h(t) for some k ≥ 1 and h(t) ∈ KR [t] with h(0) 6= 0. Then pa (D̄)a = 0 and D̄ N a = 0. Since gcd (pa (t), tN ) = tℓ in KR [t] with ℓ = min{k, N}, we have D̄ ℓ a = 0. Hence D k a = D̄ k a = 0 for all a ∈ A. Therefore D is nilpotent. Then by [Z4, Lemma 6.1] we have D = 0, whence the theorem follows. ✷ One remark on Theorem 4.6 is that without the characteristic zero condition, the theorem may be false, which can be seen from the following example. For more integral derivations of algebras over a field of characteristic p > 0, see [N]. Example 4.7. Let K be a field of characteristic p > 0, A = K[x] and D = d/dx. Then D p = 0. Hence D is a nonzero K-derivation of A that is algebraic over K. 16 WENHUA ZHAO One immediate consequence of Theorem 4.3, Corollary 4.5 and Theorem 4.6 is the following corollary which in some sense gives an affirm answer to the so-called LNED conjecture proposed in [Z4] for nilpotent, or locally integral, or locally algebraic derivations of certain algebras. Corollary 4.8. 1) Let R, A be as in Theorem 4.3 and D an Rderivation of A that is locally integral over R. Then D maps every R-subspace of A to a MS of A. 2) Let R, A be as in Corollary 4.5 or as in Theorem 4.3, and D an R-derivation of A that is locally algebraic over R. Then D maps every R-subspace of A to a MS of A. We end this section with the following Proposition 4.9. Let R be a commutative ring and A a reduced Ralgebra (not necessarily commutative) such that (A, +) is torsion-free. Let r ≥ 1, a ∈ A and D be an R-derivation of A such that D r am = 0 for all 1 ≤ m ≤ 2r−1 . Then a ∈ Ker D. Consequently, we have Ker D ⊆ r(Ker D k ) = r(Ker D). Note that when A is commutative, the lemma follows easily from Theorem 2.1. Here we give a proof independent on the commutativity of A. Proof of Proposition 4.9: The case r = 1 is obvious. So assume r ≥ 2. Then 2r − 2 ≥ r and for each 1 ≤ k ≤ 2r−2 , by the Leibniz rule we have 2r−2 X 2r − 2 2r−2 2k (D i ak )(D 2r−2−i ak ) 0=D a = i i=0 Since D i ak = 0 for all i ≥ r, there is only one term in the sum above that is not equal to 0, namely, the term with i = r − 1. Therefore 2r−2 (D r−1 ak )2 = 0. Since A is reduced and (A, +) is torsion-free, we r−1 have D r−1 uk = 0 for all 1 ≤ k ≤ 2(r−1)−1 . Repeating the procedure above we see that Da = 0, i.e., a ∈ Ker D. Now let u ∈ r(Ker D r ). Then there exists N ≥ 1 such that um ∈ Ker D r for all m ≥ N. Applying the result shown above to um (m ≥ N) we have um ∈ Ker D for all m ≥ N. Hence u ∈ r(Ker D), and r(Ker D r ) ⊆ r(Ker D). Conversely, since Ker D is an R-subalgebra of A and Ker D ⊆ Ker D r , we also have Ker D ⊆ r(Ker D r ) and r(Ker D) ⊆ r(Ker D r ). Hence the proposition follows. ✷ THE RADICAL OF THE KERNEL OF A DIFFERENTIAL OPERATOR 17 5. A Differential Vandemonde Determinant Throughout this section A stands for a commutative ring and D for a derivation of A. Proposition 5.1. Let A and D be fixed as above. Then for all f ∈ A and n ≥ 1, we have (5.1)  f f2  D(f ) D(f 2 )  2 D (f ) D 2 (f 2 ) det   .. ..  . . n−1 n−1 D (f ) D (f 2 ) Qn−1 k!. where αn = k=1  fn D(f n )   D 2 (f n )  = αn (Df ) 21 n(n−1) f n ,  ..  . · · · D n−1 (f n ) ··· ··· ··· The idea of the proof is to show that the matrix in Eq. (5.1) can be transformed by some elementary column operations to an upper triangular matrix whose the (i, i)th diagonal entry is equal to (i−1)!(Df )i−1 f for all 1 ≤ i ≤ n. For example, for the case n = 2, by subtracting from the second column the multiple of the first column by f we get     f 0 f f2 (5.2) =⇒ . D(f ) D(f 2 ) D(f ) f D(f ) To see this can be achieved for all n ≥ 2, it suffices to show the following lemma, from which Proposition 5.1 immediately follows. Lemma 5.2. Let D and f as in Proposition 5.1 and k ≥ 2. Then there exist αk,j ∈ A (1 ≤ j ≤ k − 1) such that for each 0 ≤ i ≤ k − 1, we have (5.3) i k D (f ) − k−1 X αk,j f k−j D i (f j ) = δi,k−1 (k − 1)!(Df )k−1f, j=1 where δi,k−1 is the Kronecker delta function. Proof: We use induction on k. If k = 2, then α2,1 = 1 solves the equations in Eq. (5.3), as already pointed out in Eq. (5.2) above. Assume that lemma holds for some k ≥ 2 and consider the case k +1. By writing f k+1 as f · f k and applying the Leibniz rule, we have for each 0 ≤ i ≤ k i−1   X i i k+1 i k (D i−ℓ f )D ℓ (f k ) D (f ) − f D (f ) = ℓ ℓ=0 18 WENHUA ZHAO Applying the induction assumption to D ℓ (f k ) and noticing that ℓ = k − 1, if and only if i = k = ℓ + 1, since ℓ ≤ i − 1 ≤ k − 1: ! k−1 i−1   X X i (D i−ℓ f ) αk,j f k−j D ℓ (f j ) = δi,k k(k − 1)!(Df )k f + ℓ j=1 ℓ=0   i−1 k−1 X i X (D i−ℓ f )D ℓ (f j ) = δi,k k!(Df )k f + αk,j f k−j ℓ j=1 ℓ=0 = δi,k k!(Df )k f + k−1 X j=1 k   αk,j f k−j D i (f j+1) − f D i (f j ) = δi,k k!(Df ) f + αk,k−1f D i (f k ) − αk,1 f k D(f ) + k−1 X (αk,j−1 − αk,j )f k+1−j D i (f j ). j=2 Set (5.4)   −αk,1 αk+1,j := αk,j−1 − αk,j  1 + α k,k−1 if j = 1; if 2 ≤ j ≤ k − 1; if j = k. Then αk+1,j (1 ≤ j ≤ k) solve the equations in Eq. (5.3) for the case k + 1 and 0 ≤ i ≤ k. Hence, by induction the lemma follows. ✷ Remark 5.3. One application of the formula in Eq. (5.1) is as follows. We first apply the formula to some special function f (x) and derivation D, and then evaluate x at a fixed point c. By doing so, we may get formulas for the determinants of several families of matrices, e.g., d letting f = xd , D = xm dx and c = ±1 for all d, m ∈ Z. In particular, if we choose d = −1, m = 0 and c = 1, then with a little more argument we get the following formula with 0! := 1 for all n ≥ 1: (5.5) n−1 Y 2  det (i + j − 2)! 1≤i,j≤n = k! . k=1 Another consequence of Proposition 5.1 is the following Proposition Let D be a derivation of A, ξ be a free variable and P5.4. d 0 6= P (ξ) = i=0 ci ξ i ∈ A[ξ]. Let f ∈ A such that f m ∈ Ker P (D) for all 1 ≤ m ≤ d + 1. Then for each 0 ≤ i ≤ d, we have (5.6) 1 αd+1 ci (Df ) 2 d(d+1) f d+1 = 0, THE RADICAL OF THE KERNEL OF A DIFFERENTIAL OPERATOR where αd+1 = 19 Qd k=1 k! Proof: Let B be the transpose of the matrix in Eq. (5.1) with n−1 = d. Since P (D)(f m ) = 0 for all 1 ≤ m ≤ d + 1, we have Bv = 0, where v denotes the column vector (c0 , c1 , . . . , cd )τ . Then from Eq. (5.1) the proposition follows. ✷ Corollary 5.5. Let D, f , P (ξ) be as in Proposition 5.4. Assume further that (A, +) is torsion-free and for some 0 ≤ i ≤ d, ci is not zero nor a zero-divisor of A. Then f Df is nilpotent. 1 Proof: By Proposition 5.4 we have αd+1 ci (Df ) 2 d(d+1) f d+1 = 0. Hence, 1 f d+1 (Df ) 2 d(d+1) = 0, for (A, +) is torsion-free and ci is not zero nor a zero-divisor of A. Then (f Df )m = 0 for all m ≥ max{d+1, 12 d(d+1)}, whence the corollary follows. ✷ References [AM] M. F. Atiyah and I. G. Macdonald, Introduction to Commutative Algebra. Addison-Wesley Publishing Co., 1969. [MR0242802]. [BCW] H. Bass, E. Connell and D. Wright, The Jacobian Conjecture, Reduction of Degree and Formal Expansion of the Inverse. Bull. Amer. Math. Soc. 7, (1982), 287–330. [DEZ] H. Derksen, A. van den Essen and W. Zhao, The Gaussian Moments Conjecture and the Jacobian Conjecture. To appear in Israel J. Math.. See also arXiv:1506.05192 [math.AC]. [E1] A. van den Essen, Polynomial Automorphisms and the Jacobian Conjecture. Prog. Math., Vol.190, Birkhäuser Verlag, Basel, 2000. [E2] A. van den Essen, Introduction to Mathieu Subspaces. “International ShortSchool/Conference on Affine Algebraic Geometry and the Jacobian Conjecture” at Chern Institute of Mathematics, Nankai University, Tianjin, China. July 14-25, 2014. [EH] A. van den Essen and L. C. van Hove, Mathieu-Zhao Spaces. To appear. [EN] A. van den Essen and S. Nieman, Mathieu-Zhao Spaces of Univariate Polynomial Rings with Non-zero Strong Radical. J. Pure Appl. Algebra, 220 (2016), no. 9, 3300–3306. [EZ] A. van den Essen and W. Zhao, Mathieu Subspaces of Univariate Polynomial Algebras. J. Pure Appl. Algebra. 217 (2013), no.7, 1316-1324. See also arXiv:1012.2017[math.AC]. [EWZ] A. van den Essen, D. Wright and W. Zhao, Images of Locally Finite Derivations of Polynomial Algebras in Two Variables. J. Pure Appl. Algebra 215 (2011), no.9, 2130-2134. [MR2786603]. See also arXiv:1004.0521[math.AC]. [H] J. E. Humphreys, (1972), Introduction to Lie Algebras and Representation Theory. Graduate Texts in Mathematics, Springer, 1972. [Ke] O. H. Keller, Ganze Gremona-Transformationen. Monats. Math. Physik 47 (1939), no. 1, 299-306. 20 [L] [M] [N] [Z1] [Z2] [Z3] [Z4] [Z5] [Z6] [Z7] WENHUA ZHAO S. Lefschetz, Differential Equations: Geometric Theory. Reprinting of the second edition. Dover Publications, Inc., New York, 1977. [MR0435481]. O. Mathieu, Some Conjectures about Invariant Theory and Their Applications. Algèbre non commutative, groupes quantiques et invariants (Reims, 1995), 263–279, Sémin. Congr., 2, Soc. Math. France, Paris, 1997. [MR1601155]. A. Nowicki, Integral Derivations. J. Alg. 110 (1987), no. 1, 262-276. W. Zhao, Images of Commuting Differential Operators of Order One with Constant Leading Coefficients. J. Alg. 324 (2010), no. 2, 231–247. [MR2651354]. See also arXiv:0902.0210 [math.CV]. W. Zhao, Generalizations of the Image Conjecture and the Mathieu Conjecture. J. Pure Appl. Alg. 214 (2010), 1200-1216. See also arXiv:0902.0212 [math.CV]. W. Zhao, Mathieu Subspaces of Associative Algebras. J. Alg. 350 (2012), no.2, 245-272. See also arXiv:1005.4260 [math.RA]. W. Zhao, Some Open Problems on Locally Finite or Locally Nilpotent Derivations and E-Derivations. Preprint. W. Zhao, Idempotents in Intersection of the Kernel and the Image of Locally Finite Derivations and E-derivations. Preprint. W. Zhao, The LFED and LNED Conjectures for Algebraic Algebras. Preprint. W. Zhao, Images of Ideals under Derivations and E-Derivations of Univariate Polynomial Algebras over a Field of Characteristic Zero. Preprint. Department of Mathematics, Illinois State University, Normal, IL 61761. Email: [email protected]
0math.AC
Brownian forgery of statistical dependences Vincent Wens∗ arXiv:1705.01372v2 [cond-mat.stat-mech] 11 Sep 2017 Laboratoire de Cartographie fonctionnelle du Cerveau, UNI – ULB Neurosciences Institute, Université libre de Bruxelles (ULB) & Magnetoencephalography Unit, Department of Functional Neuroimaging, Service of Nuclear Medicine, CUB – Hôpital Erasme, Brussels, Belgium. (Dated: January 15, 2018) The balance held by Brownian motion between temporal regularity and randomness is embodied in a remarkable way by Levy’s forgery of continuous functions. Here we describe how this property can be extended to forge arbitrary dependences between two statistical systems, and then establish a new Brownian independence test based on fluctuating random paths. We also argue that this result allows revisiting the theory of Brownian covariance from a physical perspective and opens the possibility of engineering nonlinear correlation measures from more general functional integrals. I. INTRODUCTION AND OVERVIEW The modern theory of Brownian motion provides an exceptionally successful example of how physical models can have far-reaching consequences beyond their initial field of development. Since its introduction as a model of particle diffusion, Brownian motion has indeed enabled the description of a variety of phenomena in cell biology, neuroscience, engineering, and finance [1]. Its mathematical formulation, based on the Wiener measure, also represents a fundamental prototype of continuous-time stochastic process and serves as powerful tool in probability and statistics [2, 3]. Following a similar vein, we develop in this note a new way of applying Brownian motion to the characterization of statistical independence. Our connection between Brownian motion and independence is motivated by recent developments in statistics, more specifically the unexpected coincidence of two different-looking dependence measures: distance covariance, which characterizes independence fully thanks to its built-in sensitivity to all possible relationships between two random variables [4], and Brownian covariance, a version of covariance that involves nonlinearities randomly generated by a Brownian process [5]. Their equivalence provides a realization of the aforementioned connection, albeit in a somewhat indirect way that conceals its naturalness. Our goal is to explicit how Brownian motion can reveal statistical independence by relying directly on the geometry of its sample paths. The brute force method to establish the dependence or independence of two real-valued random variables X and Y consists in examining all potential relations between them. More formally, it is sufficient to measure the covariances cov[f (X), g(Y)] associated with transformations f, g that are bounded and continuous (see, e.g., Theorem 10.1 in Ref. [6]). The question pursued here is whether using sample paths of Brownian motion in place of bounded continuous functions also allows to characterize independence, and we shall demonstrate that the answer is yes. In a nutshell, the statistical fluctuations of ∗ [email protected] Brownian paths B, B0 enable the stochastic covariance index cov[B(X), B0 (Y)] to probe arbitrary dependences between the random variables X and Y. Our strategy to realize this idea consists in establishing that, given any pair f, g of bounded continuous functions and any level of accuracy, the covariance cov[f (X), g(Y)] can be approximated generically by cov[B(X), B0 (Y)]. Crucially, the notion of genericity used here refers to the fact that the probability of picking paths B, B0 fulfilling this approximation is nonzero, which ensures that an appropriate selection of stochastic covariance can be achieved by finite sampling of Brownian motion. This core result of the paper will be referred to as the forgery of statistical dependences, in analogy with Levy’s classical forgery theorem [2]. Actually Levy’s remarkable theorem, which states that any continuous function can be approximated on a finite interval by generic Brownian paths, provides an obvious starting point of our analysis. Indeed, it stands to reason that if the paths B and B0 approach the functions f and g, respectively, then cov[B(X), B0 (Y)] should approach cov[f (X), g(Y)] as well. A technical difficulty, however, lies with the restriction to finite intervals since the random variables X and Y may be unbounded. Although it turns out that intervals can not be prolonged as such without ruining genericity, we shall describe first a suitable extension of Levy’s forgery that holds on infinite domains. Our forgery of statistical dependences will then follow. From a practical standpoint, using Brownian motion to establish independence turns out to be advantageous. Indeed, exploring all bounded continuous transformations exhaustively is realistically impossible. (This practical difficulty motivates the use of reproducing kernel Hilbert spaces, see, e.g., Ref. [7] for a review.) Generating all possible realizations of Brownian motion obviously poses the same problem, but this unwieldy task can be bypassed by averaging directly over sample paths. In this way, and quite amazingly, the measurement of an uncountable infinity of covariance indices can be replaced by a single functional integral. We shall discuss how this idea leads back to the concept of Brownian covariance and how the forgery of statistical dependences clarifies the way it does characterize independence, without reference 2 test trajectory f sample path in neighborhood (II.3) sample path out of neighborhood bottlenecks at t = ±T space to the equivalence with distance covariance. Brownian covariance represents a very promising tool for modern data analysis [8, 9] but appears to be still scarcely used in applications (with seminal exceptions for nonlinear time series [10] or brain connectomics [11]). Our approach based on random paths is both physically grounded and mathematically rigorous, so we believe that it may help further disseminate this method and establish it as a standard tool of statistics. 0 δ −T II. MAIN RESULTS Here we motivate and describe our main results, with sufficient precision to provide a self-contained presentation of the ideas introduced above while avoiding technical details. Their mathematical proofs are then developed in the dedicated Section III. We also use here assumptions that are slightly stronger than is necessary, and some generalizations are relegated to Appendix A. A. Extension of Levy’s forgery Imagine recording the movement of a free Brownian particle in a very large number of trials. In essence, Levy’s forgery ensures that one of these traces will follow closely a predefined test trajectory, at least for some time. To formulate this more precisely, let us focus for definiteness on standard Brownian motion B, whose initial value is set to B(0) = 0 and variance at time t is normalized to hB(t)2 i = |t|. We fix a real-valued continuous function f with f (0) = 0 (the test trajectory) and consider the uniform approximation event that a Brownian path B fits f tightly up to a constant distance δ > 0 on the time interval [−T, T ],  Uf,δ,T = |B(t) − f (t)| < δ , ∀ |t| ≤ T . (II.1) Levy’s forgery theorem states that this event is generic, i.e., it occurs with nonzero probability P(Uf,δ,T ) > 0 (see Chapter 1, Theorem 38 in Ref. [2]). This result requires both the randomness and the continuity of Brownian motion. Neither deterministic processes nor white noises satisfy this property. In all trials though, the particle will eventually drift away to infinity and thus deviate from any bounded test trajectory. Indeed, let us further assume that the function f is bounded and examine what happens when T T → ∞. If the limit event Uf,δ,∞ = T >0 Uf,δ,T occurs, the path B must be bounded too since the particle is forever trapped in a finite-size neighborhood of the test trajectory. However Brownian motion is almost surely (a.s.) unbounded at long times [2], so that P(Uf,δ,∞ ) = 0. Hence Levy’s forgery theorem does not work on infinite time domains. To accommodate this asymptotic behavior, we should thus allow the particle to diverge from the test trajectory, 0 time T FIG. 1. Extended forgery of continuous functions. This example depicts a test trajectory (smooth curve), its allowed neighborhood (shaded area) and two sample paths, one (solid random walk) illustrating the generic event (II.3) and the other (dotted random walk), the fact that arbitrary paths have low chances to enter the expanding neighborhoods through the bottlenecks. at least in a controlled way. Let us recall that the escape to infinity is a.s. slower for Brownian motion than for any movement at constant velocity (which is one way to state the law of large numbers [2]). This suggests adjoining to event (II.1) the loose approximation event  Ef,v,T = |B(t) − f (t)| < v|t| , ∀ |t| ≥ T (II.2) whereby the particle is confined to a neighborhood of the test trajectory that expands at finite speed v > 0. Asymptotic forgery theorem. Let f be bounded and continuous, and v, T > 0. Then P(Ef,v,T ) > 0. An elegant, albeit slightly abstract, proof rests on a short/long time duality between the classes of events (II.1) and (II.2), which maps Levy’s forgery and this asymptotic version onto each other (see Section III A). For a more concrete approach, let us focus on the large T limit that will be used to study statistical dependences. Since the path B(t) and the neighborhood size v|t| both diverge, the bounded term f (t) can be neglected in Eq. (II.2) and the event Ef,v,T thus merely requires not to outrun deterministic particles moving at speed v. The asymptotic forgery thus reduces to the law of large numbers, which ensures that P(Ef,v,T ) is close to one for all T  1. This probability decreases continuously as T is lowered [since the defining condition in Eq. (II.2) becomes stricter] but does not drop to zero until T = 0 is reached. This line of reasoning can be completed and also generalized to allow slower expansions (see the strong asymptotic forgery theorem, Appendix A). We now combine Levy’s forgery and the asymptotic version to obtain an extension valid at all timescales. Specifically, let us examine the joint approximation event T Jf,δ,T = Uf,δ,T Ef,v,T with v = δ/T . (II.3) In words, the particle is constrained to follow closely the test trajectory for some time but is allowed afterwards to deviate slowly from it (Fig. 1). P(Jf,δ,T ) > P(Uf,δ,T ) P(Ef,v,T ) . (II.4) The reason lies in the temporal continuity of Brownian motion. A particle staying in the uniform neighborhood while |t| ≤ T necessarily passes through the bottlenecks, and is thus more likely to remain within the expanding neighborhood than arbitrary particles, which have low chances to even meet the bottlenecks (Fig. 1). In other words, the proportion P(Jf,δ,T )/P(Uf,δ,T ) of sample paths B ∈ Ef,v,T among all those sample paths B ∈ Uf,δ,T should be larger than the unconstrained probability P(Ef,v,T ), hence the bound (II.4). B. Forgery of statistical dependences We now turn to the analysis of statistical relations using Brownian motion. Let us fix two random variables X, Y and a pair of bounded test trajectories f, g. Consider the covariance approximation event  CX,Y,f,g,ε = |cov[B(X), B0 (Y)] (II.5) − cov[f (X), g(Y)] < ε whereby the stochastic covariance cov[B(X), B0 (Y)], built by picking independently two sample paths B, B0 , coincides with the test covariance cov[f (X), g(Y)] up to a small error ε > 0 (Fig. 2). We argue that this event is generic too. The first step is to ensure that the set (II.5) is measurable so that its probability is meaningful. Physically, this technical issue is rooted once again in the escape of Brownian particles to infinity. The stochastic covariance can be expressed as a difference of two averages hB(X) B0 (Y)i and hB(X)ihB0 (Y)i (computed at fixed sample paths) involving the coordinates B(t), B0 (t0 ) at random moments t = X, t0 = Y. If long times and thus large coordinates are sampled too often, the two terms may diverge and lead to an ill-defined covariance, i.e., ∞ − ∞. To avoid this situation, we should therefore assume that asymptotic values of X and Y are unlikely enough. Actually we shall adopt hereafter the sufficient condition that X, Y are L2 , i.e., they have finite mean and variance (see Section III C for proof of measurability). 0 (a) space This result relies on the suitable integration of a “local” version of the theorem (see Section III B), but it can also be understood rather intuitively as follows. Imagine for a moment that the events (II.1) and (II.2) were independent. Their joint probability would merely be equal to the product of their marginal probabilities, which are positive by Levy’s forgery and the asymptotic forgery, and genericity would then follow. Actually they do interact because the associated neighborhoods are connected through the narrow bottlenecks at t = ±T (Fig. 1) but this should only increase their joint probability, i.e., B(X, Y) Y Extended forgery theorem. Let f be bounded and continuous with f (0) = 0, and δ, T > 0. Then P(Jf,δ,T ) > 0. probability density of stochastic covariance 3 ε X (b) time cov[f (X), g(Y)] cov[B(X), B0 (Y)] 0 covariance value FIG. 2. Forgery of statistical dependences. The distribution of the stochastic covariance (black histogram) and its standard deviation [B(X, Y)] are shown for simulated dependent random variables [n = 104 black dots, Insert (a)] as well as a test covariance (plain arrow) and a sample (dotted arrow) falling within the allowed error (shaded area). Insert (b) shows the associated functions. The case of independent variables is superimposed for comparative purposes. Forgery theorem of statistical dependences. Let X, Y be L2 random variables, f, g be bounded and continuous, and ε > 0. Then P(CX,Y,f,g,ε ) > 0. The idea is that one way of realizing the event (II.5) is to pick sample paths B ∈ Jf,δ,T , B0 ∈ Jg,δ,T that fit the test trajectories f, g tightly (δ  1) over a long time period (T  1), see Fig. 2(b). Indeed, as shown below, we have then cov[B(X), B0 (Y)] − cov[f (X), g(Y)] = O(δ, v) (II.6) with v = δ/T . This rough estimate explains why the event (II.5) must occur whenever δ and v are small enough [see Section III D, in particular Eqs. (III.22) and (III.23) for a more precise error bound and the nested forgery lemma, Eq. (III.24) for a full proof]. In turn, the extended forgery ensures the genericity of this selection of sample paths B, B0 and thus of the event CX,Y,f,g,ε as well [the necessary condition f (0) = g(0) = 0 can indeed be assumed without loss of generality, see Section III D, Eq. (III.26)]. To understand Eq. (II.6), imagine first that the random times were bounded with |X|, |Y| ≤ T . Then B(X), B0 (Y) differ from f (X), g(Y) by less than δ for all times X, Y [Eq. (II.1)] so the covariance error must be O(δ) at most. Now for unbounded random times the distance between sample paths and test trajectories may exceed δ and must actually diverge at long times, which could have led to an infinitely large covariance error if not for the fact that the occurence of |X| ≥ T or |Y| ≥ T is very unlikely. So the fit divergence v|t| [see Eq. (II.2)] is counterbalanced within averages by the fast decay of long times probability [e.g., P(|X| ≥ |t|) ≤ hX2 i/t2 ]. The contribution of |X| ≥ T or |Y| ≥ T to the covariance error is thus finite and scales as O(v), which leads to Eq. (II.6). This forgery theorem allows us to probe enough possible relationships to establish statistical dependence or 4 independence, as we explain now. Consider the two probability densities of stochastic covariance shown in Fig. 2, which were generated using simulations differing only by the presence or absence of coupling between X and Y. The distribution appears significantly wider for the dependent variables, so this suggests that width is the key indicator of a relation. Actually, for the independent variables the narrow peak observed reflects an underlying Dirac delta function (its nonzero width in Fig. 2 is due to finite sampling errors in the covariance estimates). Indeed the vanishing of all stochastic covariances is a necessary condition of independence. The impossibility of sampling nonzero values also turns out to be sufficient. Brownian independence test. Two L2 random variables X, Y are independent iff cov[B(X), B0 (Y)] = 0 a.s. To prove sufficiency, we show that the hypothesis cov[B(X), B0 (Y)] = 0 a.s. implies that all test covariances vanish, which is equivalent to the independence of X and Y (Theorem 10.1 of Ref. [6]). This can be understood concretely using the following thought experiment. Imagine that cov[f (X), g(Y)] 6= 0 for some pair of test trajectories f, g and let us fix, say, ε = |cov[f (X), g(Y)]|/4 (as in Fig. 2). We then generate sequentially samples of stochastic covariance until the approximation event CX,Y,f,g,ε [Eq. (II.5), shaded area in Fig. 2] occurs. The forgery of statistical dependences ensures that this sequence stops eventually and our choice of ε, that the last covariance sample is nonzero. However this contradicts our hypothesis, which imposes that all trials result in vanishing covariance. (See also Section III E for a settheoretic argument.) C. Bbn (X, Y)2 = cd ovn [B(X), B0 (Y)]2 (Xi , Yi )1≤i≤n fixed The forgery of statistical dependences provides an alternative approach to the theory of Brownian covariance, hereafter denoted B(X, Y). This dependence index emerges naturally in our context as the root mean square (r.m.s.) of the stochastic covariance (or equivalently its standard deviation since the mean hcov[B(X), B0 (Y)]i vanishes identically by symmetry B ←→ −B, see also Fig. 2). Thus (II.7) which is only a slight reformulation of the definition in Ref. [5]. For L2 random variables X, Y, the quadratic Gaussian functional integrals over the sample paths B, B0 can be computed analytically and the result reduces to distance covariance, so B(X, Y) inherits all its properties [5]. Alternatively, we argue here that the central results of the theory follow in a natural manner from Eq. (II.7). The first key property is that X, Y are independent iff B(X, Y) = 0. This mirrors directly the Brownian independence test since the r.m.s. (II.7) measures precisely the deviations of stochastic covariance from zero. This argument thus replaces the formal manipulations on the (II.8) determines an estimator Bbn (X, Y) of Brownian covariance. If X, Y are L2 random variables, this procedure allows to build the estimation theory of Brownian (and thus, distance) covariance from that of the elementary covariance. For instance, the rather intricate algebraic formula for the unbiased sample distance covariance [13, 14] is recovered by using, quite naturally from Eqs. (II.7) and (II.8), an unbiased estimator cd ovn [·, ·]2 of the covariance squared. (See Section III F for the explicit expressions of these estimators and a derivation of this statement.) The unbiasedness property hd covn [·, ·]2 i = cov[·, ·]2 is then automatically transferred to the corresponding estimator (II.8), i.e., Bbn (X, Y)2 = B(X, Y)2 , Brownian covariance revisited B(X, Y)2 = cov[B(X), B0 (Y)]2 , regularized singular integrals that underlie the theory of distance covariance [4]. Furthermore the forgery of statistical dependences clarifies how this works physically: Brownian motion fluctuates enough to make the functional integral (II.7) probe all the possible test covariances between X and Y. The second key aspect is the straightforward sample estimation of B(X, Y) using a parameter-free, algebraic formula. This is an important practical advantage over other dependency measures such as, e.g., mutual information [12]. Instead of relying on the sample formula of distance covariance [4, 5], Eq. (II.7) prompts us to estimate the stochastic covariance (or rather, its square) before averaging over the Brownian paths. So, given n joint samples Xi , Yi (i = 1, . . . , n) of X, Y and an expression for their sample covariance cd ovn , the functional integral (II.9) because the L2 convergence hypothesis is sufficient to ensure that averaging over the samples Xi , Yi commutes with the functional integration over the Brownian paths B, B0 . It is noteworthy that our construction of Brownian covariance and its estimator can be generalized by formally replacing the Brownian paths B, B0 by other stochastic processes or fields (in which case we may consider multivariate variables X, Y). This determines a simple rule to engineer a wide array of dependence measures via functional integrals, and opens the question of what processes allow to characterize independence. Our approach relied on the ability to probe generically all possible test covariances but, critically, the class of processes satisfying a forgery of statistical dependences might be relatively restricted. On the other hand, the original theory of Brownian covariance does extend to multidimensional Brownian fields or fractional Brownian motion [5], which are not continuous or Markovian (two properties central for the forgery theorems). So the forgery of statistical dependences provides a new and elegant tool to establish independence, but it may only represent a particular case of a more general theory of functional integral-based correlation measures. 5 III. MATHEMATICAL ANALYSIS We now proceed with a more detailed examination of the results presented in Section II. The most technical parts of the proofs are relegated to Appendix B. A. (II.2) except that their defining conditions are enforced only for 0 ≤ ±t ≤ T and ±t ≥ T , respectively. These events are generic too because they are constrained T less ± + − (e.g., Uf,δ,T contains Uf,δ,T = Uf,δ,T Uf,δ,T ) so that Asymptotic forgery and duality We sketched in Section II A a derivation of the asymptotic forgery theorem using the law of large numbers. A generalization can actually be developed fully (see Appendix A). However, this particular case enjoys a concise proof based on a symmetry argument. The short/long time duality. We start by establishing the duality relation P(Ef,v,T ) = P(Ufd ,v,1/T ) , where the dual fd of the function f is given by  |t|f (1/t) for t 6= 0 , fd (t) = 0 for t = 0 . (III.1) (III.2) Proof. By the time-inversion symmetry of Brownian motion [2], replacing B(t) by |t|B(1/t) in the right-hand side of Eq. (II.2) determines a new event with the same probability as Ef,v,T . Explicitly, this event is  |t|B(1/t) − f (t) < v|t| , ∀ |t| ≥ T  (III.3) = |B(t0 ) − fd (t0 )| < v , ∀ 0 < |t0 | ≤ 1/T , ± P(Uf,δ,T ) ≥ P(Uf,δ,T ) > 0 , (III.4) ± P(Ef,v,T ) (III.5) ≥ P(Ef,v,T ) > 0 , by monotonicity of the probability measure P(·), Levy’s forgery and the asymptotic forgery. We are going to focus below on the derivation of T + + P(Uf,δ,T Ef,v,T ) > 0 . (III.6) This is sufficient to prove the extended forgery because the backward-time (t ≤ 0) part of Brownian motion is merely an independent copy of its forward-time (t ≥ 0) part. Hence a similar result necessarily holds for the negative-time events and, in turn, Q T σ σ P(Jf,δ,T ) = σ=± P(Uf,δ,T Ef,v,T ) > 0 . (III.7) The integral formula. The first step in the demonstration of Eq. (III.6) relies on the following explicit expression for the joint probability as a functional integral. Let us introduce the family of auxiliary events  x Af,δ,T = |B(t) + x − f (T + t)| < δ + vt , ∀ t ≥ 0 (III.8) and denote their probability by x pf,δ,T (x) = P(Af,δ,T ) (III.9) where t0 = 1/t and Eq. (III.2) were used. The condition at t0 = 0 holds identically since B(0) = fd (0) = 0, so (III.3) coincides with Ufd ,v,1/T . This yields Eq. (III.1). for all x ∈ R. Then T + + P(Uf,δ,T Ef,v,T ) = 1U + Proof of the asymptotic forgery theorem. The theorem naturally follows from this duality and Levy’s forgery. The explicit formula (III.2) establishes that fd is continuous [for t 6= 0 this corresponds to the continuity of f , and for t = 0 to its boundedness supt∈R |f (t)| ≤ M since then |tf (1/t)| ≤ M |t| → 0 as t → 0]. Levy’s forgery theorem thus applies and shows that the right-hand side of Eq. (III.1) is nonzero. The first factor in this expectation value denotes the in+ dicator function of the event Uf,δ,T and enforces the constraint that all considered sample paths B must lie within the uniform neighborhood (II.1) for 0 ≤ t ≤ T . In the second factor the function (III.9) is evaluated at the position of the random path B(t) at t = T . As we explain below this function is Borel measurable, so pf,δ,T (B(T )) represents a proper random variable and the expectation value is well defined. B. Proof. The formula (III.10) is a direct consequence of the two following statements, each involving the conditional + probability P[Ef,v,T | B(T )]: Local extended forgery We now describe an analytical derivation of the extended forgery theorem that formalizes the intuitive argument given in Section II A. The ensuing bound for the probability of the joint event (II.3) does not quite reach that in Eq. (II.4) but is sufficient to ensure genericity. Restriction to positive-time events. As a preliminary, it will be useful to consider the positive- and negative-time ± ± events Uf,δ,T and Ef,v,T , which correspond to (II.1) and f,δ,T + P(Uf,δ,T T + Ef,v,T ) = 1U + pf,δ,T B(T ) f,δ,T  . (III.10) + P[Ef,v,T | B(T )] , (III.11) and + P[Ef,v,T | B(T )] = pf,δ,T B(T )  a.s. (III.12) They follow from fairly standard arguments about Brownian motion [2, 3] that we detail in Appendix B 1 and 6 Appendix B 2. The first implements the Markov prop+ erty that Ef,v,T depends on its past only via B(t) at the boundary time t = T . The second provides an explicit + representation of the random variable P[Ef,v,T | B(T )] (see also Chapter 1, Theorem 12 in Ref. [2]). The next step is to show that the integrand in the right-hand side of Eq. (III.10) cannot vanish identically, which provides a “local” version of the extended forgery. The full theorem will follow by integration. Extended forgery theorem (local version). There exists a subinterval J of the bottleneck f (T ) − δ < x < f (T ) + δ at t = T such that  T + + P Uf,δ,T {B(T ) ∈ J} ≥ P(Ug,`,T ) (III.13) for some continuous function g satisfying g(0) = 0 and some distance parameter ` > 0, and + pf,δ,T (x) > P(Ef,v,T ) (III.14) for all x ∈ J. By Eqs. (III.4) and (III.5), both lower bounds are positive. The two parts of the theorem are closely related to Levy’s forgery and the asymptotic forgery, and we treat them separately. Proof of (III.13). This property actually holds for arbitrary subintervals J, which we shall parameterize as x0 − ` < x < x0 + ` with |x0 − f (T )| ≤ δ − ` to ensure inclusion in the bottleneck. We also define the continuous function g by g(t) = f (t) + t[x0 − f (T )]/T . This setup + + + ensures that Ug,`,T ⊂ Uf,δ,T and Ug,`,T ⊂ {B(T ) ∈ J}, as the condition |B(t) − g(t)| < ` for all 0 ≤ t ≤ T implies |B(t) − f (t)| ≤ |B(t) − g(t)| + |g(t) − f (t)| < ` + |x0 − f (T )| ≤ δ (III.15) and |B(T ) − x0 | = |B(T ) − g(T )| < ` as g(T ) = x0 . The bound (III.13) follows from these two inclusions by the monotonicity of P(·). Two properties of the function pf,δ,T . For the second part it is useful to interject here the following simple statements about the function (III.9): (i) pf,δ,T vanishes identically outside the bottleneck, (ii) pf,δ,T is continuous within the bottleneck (and therefore, it is Borel measurable as well). The first claim rests on the observation that the set (III.8) is empty whenever |x−f (T )| ≥ δ (since its defining condition at t = 0 cannot be satisfied, as B(0) = 0), so we find pf,δ,T (x) = P(∅) = 0. The second claim may appear quite clear as well, since the set (III.8) should vary continuously with x, but this intuition is not quite right. For a more accurate statement, let us fix a point x in the bottleneck and consider an arbitrary sequence xn converging to it. Then the limit xn x event of Af,δ,T (assuming it exists) coincides with Af,δ,T modulo a zero-probability set, so by continuity of P(·) lim pf,δ,T (xn ) = pf,δ,T (x) . n→∞ (III.16) The full proof appears rather technical, so we only sketch the key ideas here and relegate the details to Appendix B 3. The limit event imposes that sample paths lie within the expanding neighborhood associated with (III.8) but are allowed to reach its boundary [essentially because the large n limit of the strict inequalities (III.8) at x = xn yields a nonstrict inequality]. However this hitting event a.s. never happens because the typical roughness of Brownian motion forbids meeting another trajectory without crossing it (see Appendix B 4 for this “boundary-crossing law”, which generalizes Lemma 1 on page 283 of Ref. [3] to time-dependent levels.). Proof of (III.14). The key observation here is that Z ∞ exp(−x2 /2T ) + √ P(Ef,v,T ) = x. pf,δ,T (x) , (III.17) 2πT −∞ which is merely the integral formula (III.10) with the con+ straint B ∈ Uf,δ,T removed [equivalently, this is Eq. (B.5) with S = R] and the expectation over B(T ) made ex+ plicit. If pf,δ,T (x) ≤ P(Ef,v,T ) everywhere, then for almost all x this inequality must saturate to an equality by Eq. (III.17) and thus pf,δ,T (x) > 0 by Eq. (III.5). This contradicts property (i) of pf,δ,T , so that Eq. (III.14) must hold for at least one point x in the bottleneck, and thus also on some subinterval J by the continuity property (ii). Proof of the extended forgery theorem. Finally we combine the local version of the theorem with the integral formula (III.10) to obtain T +  + + + ) P(Ef,v,T ) . (III.18) Ef,v,T > P(Ug,`,T P Uf,δ,T Indeed, further constraining the paths to B(T ) ∈ J allows to bound the right-hand side of (III.10) from below by  1U + T {B(T )∈J} pf,δ,T B(T ) f,δ,T  T + + > P Uf,δ,T {B(T ) ∈ J} P(Ef,v,T ) by (III.14), + + ≥ P(Ug,`,T ) P(Ef,v,T ) by (III.13). In passing through the first inequality we also used the identity that h1M i = P(M ) for any event M . The theorem (III.6) follows from the inequality (III.18) together with Eqs. (III.4) and (III.5). It is noteworthy that we stated and proved the extended forgery under the assumption that v = δ/T for convenience, but in the above arguments this restriction was actually artificial so the theorem holds for arbitrary parameters δ, v, T > 0. C. Measurability of the stochastic covariance index The premise of all the forgery theorems is that the associated approximation events are measurable. For the sets (II.1), (II.2) enforcing algebraic conditions on a time domain, this is a general property of separable stochastic 7 processes such as Brownian motion [3]. Indeed, the principles of probability theory only allow constraints on at most a countable infinity of time points but the notion of separability enables considering continuous time domains as well. In a nutshell, Brownian motion is separable because conditions such as a ≤ B(t) ≤ b imposed on a dense grid of rational time points t = k/n extend automatically to nonrational times by sample path continuity. The story is different for the set (II.5) because it involves a functional 0 0 0 cov[B(X), B (Y)] = hB(X)B (Y)i − hB(X)ihB (Y)i (III.19) of the sample paths B, B0 that depends in particular on the distribution of the random times X, Y. Here we provide convergence conditions that ensure the measurability of this functional and thus of the covariance approximation event (II.5). Sufficient conditions for measurability. The expectation values hB(X)i, hB0 (Y)i, and hB(X)B0 (Y)i are measurable functionals of the sample paths B, B0 whenever h|X|1/2 i, h|Y|1/2 i, and h|XY|1/2 i are finite, respectively. When these three conditions hold, the stochastic covariance (III.19) is thus measurable too. It is noteworthy that these conditions can be replaced by the stronger constraint that X, Y are L1 , i.e., h|X|i and h|Y|i are finite (since the Cauchy-Schwarz inequality implies, e.g., h|XY|1/2 i ≤ h|X|i1/2 h|Y|i1/2 ). In Section II B we used the even stronger assumption that X, Y are L2 to simplify and shorten our formulation of results. Proof. This is a direct consequence of the Fubini-Tonelli theorem [3]. Let us focus on the functional B → hB(X)i [the argument for hB0 (Y)i and hB(X)B0 (Y)i is completely similar]. In this setup, the theorem ensures its measurability if the iterated expectation value of B(X), computed first by averaging over sample paths B and then over random times X, is absolutely convergent. Since B(t) is a Gaussian random variable with zero mean and variance |t|, we find explicitly h|B(t)|it = X = k |X| 1/2 Z ∞ −∞ z. |z| e−z 2 /2 Nested forgery of statistical dependences We showed that the forgery of statistical dependences is induced by the extended forgery using the somewhat rough estimate (II.6) of the covariance error. We provide now an exact upper bound and a complete proof of the theorem. Error bound. For arbitrary sample paths B ∈ Jf,δ,T and B0 ∈ Jg,δ,T , we have cov[B(X), B0 (Y)] − cov[f (X), g(Y)] < εb (δ, v) (III.22) where v = δ/T and the error bound εb (δ, v) is given by εb (δ, v) = 2(Mf + Mg )δ + 2δ 2  + 2 Mg h|X|i + Mf h|Y|i v + 2h|X| + |Y|iδv (III.23)  + h|XY|i + h|X|ih|Y|i v 2 , with Mf = supt∈R |f (t)| and Mg = supt∈R |g(t)|. The derivation of this estimate relies on a relatively straightforward application of a series of inequalities and is relegated to Appendix B 5. Its significance rests on the fact that, when the polynomial coefficients in (III.23) are finite, the covariance error can be made arbitrarily small by taking δ, v → 0. Hence we obtain the following key intermediate result. The nested forgery lemma. Assume that Mf , Mg , h|X|i, h|Y|i, and h|XY|i are finite, and let ε > 0. Then there exists δ, T > 0 such that 0 Jf,δ,T × Jg,δ,T ⊂ CX,Y,f,g,ε , (III.24) where the prime on event Jg,δ,T indicates that it applies to the process B0 . Proof of the forgery theorem of statistical dependences. This is a direct corollary. The L2 hypothesis and the boundedness assumption on f, g ensure that the lemma (III.24) applies. Using the monotonicity of P(·) and the independence of B, B0 then yields 0 P(CX,Y,f,g,ε ) ≥ P(Jf,δ,T × Jg,δ,T ) , (III.20) where 1 k=√ 2π D. < ∞. (III.21) Thus the finiteness condition on h|X|1/2 i indeed fulfills the requirement of the Fubini-Tonelli theorem. Technically this application also needs the extra prerequisite that B(X) be a random variable, i.e., that the evaluation map (B, X) → B(X) be jointly measurable (and not just at fixed path B or fixed time X). This is a property of continuous stochastic processes such as Brownian motion [3]. = P(Jf,δ,T ) P(Jg,δ,T ) . (III.25) It is now tempting to invoke the extended forgery theorem, but here the conditions f (0) = g(0) = 0 were not imposed. This is however not an issue thanks to the translation invariance of covariance, i.e., cov[f (X), g(Y)] is equal to cov[f (X) − f (0), g(Y) − g(0)]. Thus P(CX,Y,f,g,ε ) = P(CX,Y,f −f (0),g−g(0),ε ) (III.26) ≥ P(Jf −f (0),δ,T ) P(Jg−g(0),δ,T ) > 0 , where we used Eq. (III.25) and the extended forgery theorem in the second line. Note that the L2 assumption used in Section II B was slightly stronger than needed, as made clear by the nested forgery lemma. 8 E. Brownian independence test and dichotomy To prove that the assertion cov[B(X), B0 (Y)] = 0 a.s. implies cov[f (X), g(Y)] = 0, we used in Section II B a discrete sampling method for which it is straightforward that the covariance approximation event (II.5) must occur simultaneously with the zero covariance event  ZX,Y = cov[B(X), B0 (Y)] = 0 . (III.27) Their compatibility, which was indeed central in our derivation, is actually a general property of probability theory and we use it here to provide an alternative, settheoretic argument. The dichotomy lemma. For arbitrary functions f, g and parameter ε, we have  T ZX,Y if |cov[f (X), g(Y)]| < ε CX,Y,f,g,ε ZX,Y = ∅ otherwise. (III.28) Proof. It is a direct consequence of Eqs. (II.5) and (III.27) that the intersection is characterized equivalently by the two conditions cov[B(X), B0 (Y)] = 0 and |cov[f (X), g(Y)]| < ε. Second proof of sufficiency for the Brownian independence test. By the forgery of statistical dependences the event CX,Y,f,g,ε is generic for f, g bounded and continuous and ε > 0, and by hypothesis the event ZX,Y occurs a.s. Since the intersection of a generic event and an almost sure event is never empty [if it were empty, the generic event would be a subset of a zero probability event (i.e., the complement of the almost sure event), which is forbidden by monotonicity of P(·)], the second possibility in Eq. (III.28) is ruled out so |cov[f (X), g(Y)]| < ε must hold true. Since ε > 0 was arbitrary, we conclude that cov[f (X), g(Y)] = 0. F. Unbiased estimation of Brownian covariance We now exemplify how an explicit estimator of Brownian covariance can be derived from the functional integral (II.8). Our construction enforces unbiasedness at finite sampling and allows to recover the unbiased sample formula of distance covariance [13, 14] that we review first. The unbiased estimator. Let us introduce the distance matrix aij = |Xi − Xj | between the samples Xi of X as well as its “U-centered” version P P k,l akl k (aik + akj ) , (III.29) Aij = aij − + n−2 (n − 1)(n − 2) where 1 ≤ i, j, k, l ≤ n. The analogous matrices for the corresponding samples of Y are denoted bij and Bij , respectively. With these notations and assuming n ≥ 4, X 1 Bbn (X, Y)2 = Aij Bij . (III.30) 4n(n − 3) distinct i,j This expression differs from the formula given in Refs. [13, 14] by a trivial factor 1/4 due to our use of the standard normalization for Brownian motion. The fact that Eq. (III.30) provides an unbiased estimation of Brownian covariance [see Eq. (II.9)] will follow directly from our derivation based on the functional integral (II.8), which starts naturally from an unbiased estimation of the covariance squared. Estimation of covariance squared. It is convenient to introduce the new random variables x = B(X), y = B0 (Y) and their samples xi = B(Xi ), yi = B0 (Yi ), all being defined at fixed sample paths B, B0 . An unbiased estimator for cov(x, y) itself is well-known from elementary statistics, but it can be checked by developing its square that the estimation of cov(x, y)2 is then hampered by systematic errors of order 1/n. Here, given n ≥ 4, we shall rather define  (n − 4)! X0 cd ovn (x, y)2 = xi xj yi yj − yi yk − yj yk + yk yl , n! (III.31) where the primed sum is taken over all distinct indices 1 ≤ i, j, k, l ≤ n. This expression differs from the aforementioned development by O(1/n) and is indeed free from finite sampling biases. This can be proven by averaging Eq. (III.31) over the n joint samples xi , yi . Indeed, using the identities hxi xj yi yj i = hxyi2 , hxi xj yi yk i = hxyihxihyi, and hxi xj yk yl i = hxi2 hyi2 and the fact that the primed sum contains n!/(n − 4)! terms, we find cd ovn (x, y)2 = hxyi2 − 2hxyihxihyi + hxi2 hyi2 = cov(x, y)2 . (III.32) Derivation of Eq. (III.30) from Eq. (II.8). We now average Eq. (III.31) with xi = B(Xi ) and yi = B0 (Yi ) over the independent random paths B, B0 , while the samples Xi , Yi are being kept constant. The computation factors into the functional integration of xi xj = B(Xi )B(Xj ) and yi yj = B0 (Yi )B0 (Yj ), so Bbn (X, Y)2 is obtained from the right-hand side of Eq. (III.31) by replacing xi xj and yi yj with the autocorrelation functions [2]  B(Xi )B(Xj ) = 21 |Xi | + |Xj | − aij ,  (III.33) B0 (Yi )B0 (Yj ) = 12 |Yi | + |Yj | − bij , respectively. This substitution rule can actually be simplified further to xi xj → −aij /2 because the terms involving |Xi | cancel out thanks to the algebraic identity X (yi yj − yi yk − yj yk + yk yl ) = 0 . (III.34) all distinct j,k,l6=i Similar cancellations also allow to use yi yj → −bij /2. We thus obtain the unbiased estimator (n − 4)! X0 Bbn (X, Y)2 = aij (bij − bik − bkj + bkl ) . 4n! (III.35) 9 The equivalence with Eq. (III.30) is not obvious at first sight. To make contact with it, let us fix i, j and consider the sum of the second factor over k, l, all indices being distinct. This contribution is the sum of the following four terms: P0 k,l bij = (n − 2)(n − 3)bij ,  P0 P − k,l bik = (n − 3) bij − k bik ,  P0 P − k,l bkj = (n − 3) bij − k bkj , P0 P P k,l bkl = k,l bkl − 2 k (bik + bkj ) + 2bij , the sums in the right-hand sides now running over unconstrained indices 1 ≤ k, l ≤ n. The total thus reduces to (n − 1)(n − 2)Bij , so Eq. (III.35) can be rewritten as Bbn (X, Y)2 = X 1 aij Bij . 4n(n − 3) where the growth function φ(t) is positive and continuous on t > 0, and satisfies the divergence condition √ t log log t lim = 0. (A.2) t→+∞ φ(t) The particular example φ(t) = vt leads back to Eq. (II.2). The asymptotic behavior (A.2), together with the law of the iterated logarithm, indeed ensures that the expansion described by φ(t) is a.s. faster than Brownian motion, i.e., lim |t|→∞ Finally we recover Eq. (III.30) because aij can be replaced by its U-centering Aij . Indeed the extra terms cancel in the sum thanks to the centering property P j6=i Bij = 0 (for all i fixed), which can be checked from the definition (III.29). Acknowledgments This research was carried over in the context of methodological developments for the MEG project at CUB – Hôpital Erasme (Université libre de Bruxelles, Brussels, Belgium), which is financially supported by the Fonds Erasme (Brussels, Belgium). a.s. , lim sup |t|→∞ |B(t)| |B(t)| ≤ lim sup p φ(|t|) 2|t| log log |t| |t|→∞ p 2|t| log log |t| . × lim sup φ(|t|) |t|→∞ Strong asymptotic forgery theorem. Let f be bounded and continuous, φ be a growth function as specified above, and T > 0. Then P(Ef,φ,T ) > 0. Proof. We revisit here the strategy of proof sketched in the main text. Let us start by demonstrating that lim P(Ef,φ,T ) = 1 . A STRONGER VERSION OF THE FORGERY THEOREMS The various forgery theorems obtained by extending Levy’s original forgery relied on approximations involving neighborhoods expanding at long times with a constant speed v. However it turns out that this assumption can be weakened. We now describe a stronger version of the asymptotic forgery theorem and then briefly discuss some implications. Slowly expanding neighborhoods. The key observation is that, in our discussion of the asymptotic forgery, the law of large numbers can be replaced almost verbatim by the more precise law of the iterated logarithm, which establishes that Brownian p motion diverges a.s. at long times not faster than 2|t| log log |t| [2]. Thus we naturally generalize the definition (II.2) using neighborhoods that expand strictly faster than almost all Brownian paths. Explicitly, we let  Ef,φ,T = |B(t) − f (t)| < φ(|t|) , ∀ |t| ≥ T , (A.1) (A.4) The law of the iterated logarithm states precisely that the first factor in the right-hand side equals a.s. to one, and the second factor tends to zero by definition (A.2). Therefore we obtain lim sup|t|→∞ |B(t)|/φ(|t|) = 0 a.s., which is equivalent to Eq. (A.3). (A.5) T →∞ Appendix A. (A.3) a fact that we shall use momentarily. To prove this assertion, let us start from the inequality (III.36) distinct i,j B(t) =0 φ(|t|) The event (A.1) increases monotonously when T is raised (since the defining condition becomes less strict) so S limT →∞ P(Ef,φ,T ) is equal to P( T >0 Ef,φ,T ) by continuity of the measure S P(·). Therefore we need to show that the limit event T >0 Ef,φ,T occurs a.s. This is the set-theoretic version of the statement that, for almost all sample paths B, we can find some T > 0 ensuring the inequality |B(t) − f (t)|/φ(|t|) < 1, ∀ |t| ≥ T . In turn this is a particular case of the slightly stronger assertion |B(t) − f (t)| =0 φ(|t|) |t|→∞ lim a.s. , (A.6) which holds thanks to the property (A.3) and the boundedness assumption on f [i.e., f (t)/φ(|t|) → 0]. Equation (A.5) follows. The limit (A.5) implies the existence of some value T? > 0 such that P(Ef,φ,T ) > 0 for all T ≥ T? . (A.7) It thus remains to prove the theorem for 0 < T < T? . 10 To that aim let us split ET f,φ,T at |t| = T? and identify it as a joint event Uf,φ,T,T? Ef,φ,T? where  Uf,φ,T,T? = |B(t) − f (t)| < φ(|t|) , (A.8) ∀ T ≤ |t| ≤ T? . Our strategy is to show first that these two events are generic, and then apply the techniques of Section III B to derive the genericity of their intersection. The case of Ef,φ,T? is taken care of by Eq. (A.7). For (A.8) observe that Uf,φ,T,T? ⊃ Uf,δ,T∗ whenever δ is chosen below the minimum value of the continuous function φ(|t|) over the compact time domain T ≤ |t| ≤ T∗ , since |B(t) − f (t)| < δ ≤ φ(|t|). If we pick a δ > 0 in this way (this is possible since φ is positive), we then obtain P(Uf,φ,T,T? ) ≥ P(Uf,δ,T∗ ) > 0 by monotonicity of P(·) and Levy’s forgery theorem. The last step is to ensure that the intersection is generic too. This is analogous to the extended forgery theorem (III.7) and can be proven along the exact same lines. Consequences for the other forgery theorems. Strong versions of the subsequent forgery theorems can be obtained by replacing the event (II.2) with its generalization (A.1). The end result is to weaken the convergence assumptions required for the random variables X, Y and therefore widen the applicability of the forgery of statistical dependences and the Brownian independence test. We gather the results here without repeating the derivations. The strong extended forgery theorem states that if f is bounded and continuous with f (0) = 0, φ is a growth function as above, and δ, T > 0, then T P(Uf,δ,T Ef,φ,T ) > 0 . (A.9) A covariance error inequality similar to Eq. T (III.22) then holds for arbitrary sample paths B ∈ U Ef,φ,T and f,δ,T T B0 ∈ Ug,δ,T Eg,φ,T with the bound εb (δ, φ) = 2(Mf + Mg )δ + 2δ 2   + 2 Mg hφ(|X|)i + Mf hφ(|Y|)i + 2hφ(|X|) + φ(|Y|)iδ + hφ(|X|)φ(|Y|)i + hφ(|X|)ihφ(|Y|)i (A.10) instead of (III.23). As a consequence, the forgery theorem of statistical dependences and the Brownian independence test both hold under the weaker convergence conditions that there exists a growth function φ for which hφ(|X|)i, hφ(|Y|)i, and hφ(|X|)φ(|Y|)i are finite. Appendix B. 1. COMPLETION OF PROOFS Markovian decomposition formula (III.11) This equation relies on the weak Markov property and is actually valid for general Markovian processes. Denot+ ing by Uf,δ,T the σ-algebra generated by the event Uf,δ,T , we can write + P(Uf,δ,T T + Ef,v,T ) = 1U + f,δ,T + P[Ef,v,T | Uf,δ,T ] (B.1) merely by definition of the conditional probability + P[Ef,v,T | Uf,δ,T ] [3]. Furthermore, Uf,δ,T is contained in the σ-algebra FT generated by the process B(t) with + 0 ≤ t ≤ T , since Uf,δ,T only involves conditions on this time interval. From this observation, + + P[Ef,v,T | Uf,δ,T ] = P[Ef,v,T | FT ] Uf,δ,T + = P[Ef,v,T | B(T )] Uf,δ,T , (B.2) where h . | Uf,δ,T i denotes the expectation value conditional to Uf,δ,T . In the first line we applied the tower property of conditioning that follows from Uf,δ,T ⊂ FT . In the second we used the weak Markov property that + events involving conditions for t ≥ T (such as Ef,v,T ) only depend on their past via the (σ-algebra generated by the) random variable B(T ). Combining Eqs. (B.1) and (B.2), we therefore find that the left-hand side of Eq. (III.11) is equal to + | B(T )] Uf,δ,T i . The conditional ex1U + hP[Ef,v,T f,δ,T pectation to Uf,δ,T can be dropped in this expression, again by mere definition of conditioning to Uf,δ,T , so we recover the right-hand side of (III.11). 2. + Representation (III.12) of P[Ef,v,T | B(T )] The defining property of the conditional probability + P[Ef,v,T | B(T )] is that P {B(T ) ∈ S} T + Ef,v,T  + = 1{B(T )∈S} P[Ef,v,T | B(T )] (B.3) for all Borel sets S ⊂ R, which only characterizes it modulo a zero-probability set. To prove the statement (III.12), it thus suffices to check that pf,δ,T (B(T )) satisfies the same property. This is a consequence of the fact that Brownian motion has independent increments themselves distributed as Brownian motions [3]. Indeed, the left-hand side of Eq. (B.3) is computed as the integral over all sample paths B satisfying both B(T ) ∈ S and |B(t) − f (t)| < vt for all t ≥ T . Defining the increment B0 (t0 ) = B(T + t0 ) − B(T ) with time step t0 = t − T , the latter constraint becomes |B0 (t0 ) + B(T ) − f (T + t0 )| < v(T + t0 ) = δ + vt0 (B.4) B(T ) for all t0 ≥ 0, or equivalently B0 ∈ Af,δ,T in view of the definition (III.8). Now as recalled above B0 is a Brownian motion independent of the position B(T ). An application of the Fubini-Tonelli theorem [3] thus enables us to inB(T ) tegrate first over all B0 ∈ Af,δ,T at fixed B(T ) and then 11 over B(T ) ∈ S. Therefore we obtain T +  P {B(T ) ∈ S} Ef,v,T It remains to derive these two inclusions. x = 1{B(T )∈S} h1{B0 ∈Af,δ,T } ix=B(T )  = 1{B(T )∈S} pf,δ,T B(T ) , (B.5) x x where we used the identity 1{B0 ∈Af,δ,T } = P(Af,δ,T ) and the definition (III.9) in passing through the second equality. This ends the demonstration of Eq. (III.12). 3. Continuity property (III.16) of pf,δ,T xn We shall prove that the limit of pf,δ,T (xn ) = P(Af,δ,T ) x exists and equals to pf,δ,T (x) = P(Af,δ,T ) whenever xn → x with x inside the bottleneck interval at t = T . We start from the difference formula xn xn x x P(Af,δ,T ) − P(Af,δ,T ) = P(Af,δ,T \Af,δ,T ) xn x − P(Af,δ,T \Af,δ,T ) (B.6) and argue that each of the two terms in the right-hand side converge to zero. Since sequences of probabilities are nonnegative, it is actually sufficient to show that xn xn x x lim sup P(Af,δ,T \Af,δ,T ) = P(Af,δ,T \Af,δ,T i.o.) (B.7) n→∞ and First inclusion (B.10). Let us consider an arbitrary samxn x ple path B satisfying B ∈ Af,δ,T \Af,δ,T i.o. Explicitly, this means that we can build a diverging sequence nk of xnk x indices such that B ∈ Af,δ,T and B ∈ / Af,δ,T , i.e., |B(tk ) + xnk − f (T + tk )| ≥ δ + vtk for some tk ≥ 0. At this point two cases emerge. Both will lead to the conclusion that B ∈ Z , which corresponds to the sought inclusion (B.10). (i) The set of times tk is bounded. In that situation, the Bolzano-Weierstrass theorem provides a subsequence tkm converging to some finite time τ . Assuming that B is continuous, let us evaluate Eq. (B.12) along k = km and take m → ∞. Using xnkm → x and the continuity of B and f , we find |B(τ ) + x − f (T + τ )| ≥ δ + vτ , which x contradicts the fact that B ∈ Af,δ,T . Therefore B cannot be continuous at τ , henceforth B ∈ Z . (ii) The set of times tk is unbounded. We can then divide both sides of Eq. (B.12) by tk and let k → ∞. Since tk → ∞ and both xn and f are bounded, it follows that lim supk→∞ |B(tk )|/tk ≥ v > 0. This shows that the sample path B cannot satisfy the law of large numbers, hence B ∈ Z . Second inclusion (B.11). Similarly, let us now consider xn x B ∈ Af,δ,T \Af,δ,T i.o., meaning that there exists a diverging sequence nk such that xn xn x x lim sup P(Af,δ,T \Af,δ,T ) = P(Af,δ,T \Af,δ,T i.o.) (B.8) |B(t) + xnk − f (T + t)| < δ + vt n→∞ vanish. Here i.o. stands for “infinitely often” and means that an infinite number of events in the sequence occur, and indeed allows evaluating superior limits [3]. To proceed with the proof, let us introduce the zeroprobability set Z of paths that are not continuous or do not satisfy the law of large numbers limt→∞ B(t)/t = 0, and the boundary-hitting event  x Bf,δ,T = |B(t) + x − f (T + t)| ≤ δ + vt (B.9) ∀ t ≥ 0, with equality at some t = τ for all t ≥ 0, and a set of times tk ≥ 0 such that that is closely related to (III.8) but imposes that sample paths actually reach the boundary of the associated neighborhood. This event turns out to also have probability zero, because a Brownian path hitting the boundary must a.s. also cross it and thus leave the neighborhood. For completeness we provide a derivation of this boundary-crossing law in Appendix B 4. With these definitions at hand, we claim that 4. xn x {Af,δ,T \Af,δ,T i.o.} ⊂ Z , S x xn x {Af,δ,T \Af,δ,T i.o.} ⊂ Z Bf,δ,T . (B.10) (B.11) They imply directly that the two limits (B.7) and (B.8) vanish by monotonicity and subadditivity of P(·), bex cause P(Z ) = P(Bf,δ,T ) = 0. Therefore the right-hand side of Eq. (B.6) does indeed tend to zero as n → ∞. (B.12) |B(tk ) + x − f (T + tk )| ≥ δ + vtk . (B.13) (B.14) Taking k → ∞ in Eq. (B.13) yields the nonstrict inequality |B(t) + x − f (T + t)| ≤ δ + vt. Repeating our analysis of Eq. (B.12) to Eq. (B.14) also shows that equality must x hold at some time τ , i.e., B ∈ Bf,δ,T , or else B ∈ Z , so we obtain the sought inclusion (B.11). x The boundary-crossing law for P(Bf,δ,T )=0 It is convenient to introduce the upper (+ sign) and lower (− sign) boundary curves γ± (t) = f (T + t) − x ± (δ + vt) . (B.15) They inherit the continuity of f , and since x lies within the bottleneck interval at t = T they also satisfy the condition γ− (0) < 0 < γ+ (0). The boundary-crossing law that we shall derive can be stated formally as follows:  P max[B(t) − γ+ (t)] = 0 = 0 , (B.16) t≥0  P min[B(t) − γ− (t)] = 0 = 0 . (B.17) t≥0 12 x This provides the sought estimate of P(Bf,δ,T ). Indeed, observing that the boundary-hitting event (B.9) implies either B(t) ≤ γ+ (t) with equality at some time, or else B(t) ≥ γ− (t) with equality at some time, we find  x Bf,δ,T ⊂ max[B(t) − γ+ (t)] = 0 t≥0 (B.18) S min[B(t) − γ− (t)] = 0 . t≥0 Applying the monotonicity and subadditivity of P(·) and x the boundary-crossing law then yield P(Bf,δ,T ) = 0. Proof of the boundary-crossing law. We shall focus here on the derivation of Eq. (B.16) [the case for (B.17) is completely similar]. To simplify the notations, let us introduce the running maximum Mτ = max [B(t) − γ+ (t)] 0≤t≤τ for 0 < τ ≤ ∞ Bound (III.22) for the covariance error Let us introduce the shorthand notations ∆cov = cov[B(X), B0 (Y)] − cov[f (X), g(Y)] , ∆B = B(X) − f (X) , ∆B0 = B0 (Y) − g(Y) . by monotonicity and subadditivity of P(·). To obtain Eq (B.16), we now evaluate the two terms in the righthand side and show that both vanish as τ → 0. For the first term it is sufficient to show that (B.21) τ →0 The limit event Mτ = 0 i.o. means explicitly that B(t) hits γ+ (t) for arbitrarily small times t > 0, which implies that limt→0 B(t) = γ+ (0) > 0 = B(0). Thus B cannot be continuous at t = 0, so the probability that Mτ = 0 occurs i.o. is zero. The second term vanishes for all 0 < τ < ∞. Noting that the increment process B0 (and thus, W) is independent of B(τ ) and applying the Fubini-Tonelli theorem, we can write explicitly Z ∞   exp(−w2 /2τ ) √ P W = −B(τ ) = w. P(W = −w) . 2πτ −∞ (B.22) But it is a general property of probability theory that P(W = −w) may be nonzero for at most a countable set of values w. Therefore P(W = −w) = 0 for almost all w and thus the integral (B.22) vanishes. (B.23) Using the bilinearity of covariance and the triangle inequality, we have ∆cov ≤ cov[∆B, g(Y)] + cov[f (X), ∆B0 ] + cov(∆B, ∆B0 ) . (B.24) We start by examining the first term. Developing the covariance, applying triangle inequalities, and using the boundedness of g, we find (B.19) (with the proviso that, say, Mτ = −∞ if a maximum is not attained for 0 ≤ t ≤ τ ). Thus Eq. (B.16) simply reads P(M∞ = 0) = 0. In what follows we fix τ < ∞ and we will eventually let τ → 0. The hitting (without crossing) event M∞ = 0 implies that either B(t) reaches γ+ (t) before t = τ , which corresponds to the event Mτ = 0, or else the increment B0 (t0 ) = B(τ + t0 ) − B(τ ) reaches γ+ (τ + t0 ) − B(τ ) for t0 = t − τ > 0, which corresponds to the hitting event W = −B(τ ), where W = maxt>0 [B0 (t) − γ+ (τ + t)]. It thus follows that   P(M∞ = 0) ≤ P(Mτ = 0) + P W = −B(τ ) (B.20) lim sup P(Mτ = 0) = P(Mτ = 0 i.o.) = 0 . 5. cov[∆B, g(Y)] = h∆B g(Y)i − h∆Bihg(Y)i ≤ h|∆B g(Y)|i + h|∆B|ih|g(Y)|i ≤ 2Mg h|∆B|i  < 2Mg δ + vh|X|i . (B.25) The last line was obtained by splitting the expectation value h|∆B|i into its contributions from the two regions · {|X| ≤ T } where |∆B| < δ [see Eq.(II.1)], and · {|X| > T } where |∆B| < v|X| [see Eq. (II.2)]. The second term is completely analogous,  cov[f (X), ∆B0 ] < 2Mf δ + vh|Y|i . (B.26) For the last term, we use four regions: · · · · {|X| ≤ T, |Y| ≤ T } {|X| ≤ T, |Y| > T } {|X| > T, |Y| ≤ T } {|X| > T, |Y| > T } where where where where |∆B ∆B0 | < δ 2 , |∆B ∆B0 | < δv|Y|, |∆B ∆B0 | < δv|X|, and |∆B ∆B0 | < v 2 |XY|. We then obtain in a similar way cov(∆B, ∆B0 ) ≤ h|∆B ∆B0 |i + h|∆B|ih|∆B0 |i < 2δ 2 + 2δvh|X|i + 2δvh|Y|i  + v 2 h|XY|i + h|X|ih|Y|i . (B.27) The covariance error inequality (III.22) with bound (III.23) is recovered by combining Eqs. (B.24) to (B.27). 13 [1] J. P. Sethna, Statistical Mechanics: Entropy, Order Parameters and Complexity (Oxford University Press, 2006). [2] D. Freeman, Brownian Motion and Diffusion (SpringerVerlag, New York, 1983). [3] I. I. Gikhman and A. V. Skorokhod, Introduction to the Theory of Random Processes (Dover, 1969). [4] G. J. Székely, M. L. Rizzo, and N. K. Bakirov, The Annals of Statistics, 35, 2769 (2007), arXiv:0803.4101 [math.ST]. [5] G. J. Székely and M. L. Rizzo, The Annals of Applied Statistics, 3, 1233 (2009), arXiv:1010.0297 [stat.AP]. [6] J. Jacod and P. E. Protter, Probability essentials, 2nd ed., Universitext (Springer-Verlag, Berlin, 2013). [7] A. Gretton and L. Györfi, J. Mach. Learn. Res., 11, 1391 (2010). [8] M. A. Newton, The Annals of Applied Statistics, 3, 1233 (2009), arXiv:1010.3575 [stat.AP]. [9] G. J. Székely and M. L. Rizzo, The Annals of Applied Statistics, 3, 1303 (2010), arXiv:1010.0844 [stat.AP]. [10] Z. Zhou, Journal of Time Series Analysis, 33, 438 (2012). [11] L. Geerligs, Cam-Can, and R. N. Henson, NeuroImage, 135, 16 (2016). [12] A. Kraskov, H. Stögbauer, and P. Grassberger, Phys. Rev. E, 69, 066138 (2004), arXiv:cond-mat 0305641 [cond-mat.stat-mech]. [13] G. J. Székely and M. L. Rizzo, Journal of Multivariate Analysis, 117, 193 (2013). [14] M. L. Rizzo and G. J. Székely, The Annals of Statistics, 42, 2382 (2014), arXiv:1310.2926 [stat.ME].
10math.ST
arXiv:1601.01633v2 [math.GR] 5 Apr 2016 COHOMOLOGY AND EXTENSIONS OF BRACES VICTORIA LEBED AND LEANDRO VENDRAMIN Abstract. Braces and linear cycle sets are algebraic structures playing a major role in the classification of involutive set-theoretic solutions to the YangBaxter equation. This paper introduces two versions of their (co)homology theories. These theories mix the Harrison (co)homology for the abelian group structure and the (co)homology theory for general cycle sets, developed earlier by the authors. Different classes of brace extensions are completely classified in terms of second cohomology groups. 1. Introduction A (left) brace is an abelian group (A, +) with an additional group operation ◦ such that for all a, b, c ∈ A, the following compatibility condition holds: (1.1) a ◦ (b + c) + a = a ◦ b + a ◦ c. The two group structures necessarily share the same neutral element, denoted by 0. Braces, in a slightly different but equivalent form, were introduced by Rump [24]; the definition above goes back to Cedó, Jespers, and Okniński [11]. To get a feeling of what braces look like, and to convince oneself that they are not as rare in practice as one might think, the reader is referred to Bachiller’s classification of braces of order p3 [1]. The growing interest into these structures is due to a number of reasons. First, braces generalize radical rings. Second, Catino-Rizzo and CatinoColazzo-Stefanelli [8, 6, 7] unveiled the role of an F -linear version of this notion into the classification problem for regular subgroups of affine groups over a field F . Third, braces are enriched cycle sets, and are therefore important in the study of set-theoretic solutions to the Yang-Baxter equation, as we now recall. A cycle set, as defined by Rump [23], is a set X with a binary operation · having bijective left translations X → X, a 7→ b · a, and satisfying the relation (1.2) (a · b) · (a · c) = (b · a) · (b · c). Rump showed that non-degenerate cycle sets (i.e., with invertible squaring map a 7→ a · a) are in bijection with non-degenerate involutive set-theoretic solutions to the Yang-Baxter equation. Such solutions form a combinatorially rich class of structures, connected with many other domains of algebra: semigroups of I-type, Bieberbach groups, Hopf algebras, Garside groups, etc. The cycle set approach turned out to be extremely fruitful for elucidating the structure of such solutions and obtaining classification results (see, for instance, [19, 20, 24, 25, 18, 10, 26, 12, 2010 Mathematics Subject Classification. 16T25, 20N02, 55N35, 20E22. Key words and phrases. Brace, cycle set, Yang-Baxter equation, extension, cohomology. The work of L. V. is partially supported by CONICET, PICT-2014-1376, MATH-AmSud, and ICTP. V. L. thanks the program ANR-11-LABX-0020-01 and Henri Lebesgue Center (University of Nantes) for support. The authors are grateful to the reviewer for useful remarks and interesting suggestions for a further development of the subject. 1 2 VICTORIA LEBED AND LEANDRO VENDRAMIN 11, 13, 16, 30, 28, 27] and references therein). In spite of the intensive ongoing research on cycle sets, their structure is still far from being completely understood. This can be illustrated by numerous conjectures and open questions in the area, many of which were formulated by Gateva-Ivanova and Cameron [15, 17] and by Cedó, Jespers, and del Rı́o [9]. Etingof, Schedler, and Soloviev [14] initiated the study of the structure group of a solution to the YBE—and in particular of a cycle set. These ideas were further explored in [22] and [29] for non-involutive solutions. Concretely, the structure group G(X,·) of a cycle set (X, ·) is the free group on the set X, modulo the relations (a · b)a = (b · a)b for all a, b ∈ X 1. In [14], the structure group of a non-degenerate cycle set (X, ·) was shown to be isomorphic, as a set, to the free abelian group Z(X) on X; see also [21] for an explicit graphical form of this isomorphism. The group G(X,·) thus carries a second, abelian, group structure—the one pushed back from Z(X) —and becomes a brace. Moreover, G(X,·) inherits a cycle set structure from X, and yields a key example of the following notion. A linear cycle set is a cycle set (A, ·) with an abelian group operation + satisfying, for all a, b, c ∈ A, the compatibility conditions (1.3) a · (b + c) = a · b + a · c, (1.4) (a + b) · c = (a · b) · (a · c). This structure also goes back to Rump [24], who showed it to be equivalent to the brace structure, via the relation a · b = a−1 ◦ (a + b). Understanding structure groups and certain classes of their quotients is often regarded as a reasonable first step towards understanding cycle sets. Even better: Bachiller, Cedó, and Jespers [3] recently reduced the classification problem for cycle sets to that for braces. This explains the growing interest towards braces and linear cycle sets. As pointed out by Bachiller, Cedó, Jespers, and Okniński [4], an extension theory for braces would be crucial for classification purposes, as well as for elaborating new examples. This served as motivation for our paper. A cohomology theory for general cycle sets was developed by the authors in [21]. The second cohomology groups were given particular attention: they were shown to encode central cycle set extensions. Here we propose homology and cohomology theories for linear cycle sets, and thus for braces. As usual, central linear cycle set extensions turn out to be classified by the second cohomology groups. For pedagogical reasons, we first study extensions that are trivial on the level of abelian groups, together with a corresponding (co)homology theory (Sections 2-3). Such extensions are still of interest, since it is often the cycle set operation that is the most significant part of the linear cycle set structure (as in the example of structure groups). On the other hand, they are technically much easier to handle than the general extensions (Sections 4-5). We therefore found it instructive to present this “reduced” case before the general one. When finishing this paper, we learned that an analogous extension theory was independently developed by Bachiller [2], using the language of braces. Some fragments of it in the F -linear setting also appeared in the work of Catino-ColazzoStefanelli [6]. 1Some authors prefer an alternative relation set a(a · b) = b(b · a), which defines an isomorphic group. COHOMOLOGY OF BRACES 3 An alternative approach to extensions was suggested earlier by Ben David and Ginosar [5]. Concretely, they studied the lifting problem for bijective 1-cocycles— which is yet another avatar of braces. Their work was translated into the language of braces by Bachiller [1]. Our choice of the linear cycle set language leads to more transparent constructions. Moreover, it made possible the development of a full cohomology theory extending the degree 2 constructions motivated by the extension analysis. Such a theory was missing in all the previous approaches. 2. Reduced linear cycle set cohomology From now on we work with linear cycle sets (=LCS). As explained in the introduction, all constructions and results can be directly translated into the language of braces. We will perform this translation for major results only. Take a LCS (A, ·, +) and an abelian group Γ. For n > 0, let RCn (A; Γ) denote ×n the abelian group Γ ⊗Z ZA×n ≃ Γ(A ) , modulo the linearity relation (2.1) γ(a1 , . . . , an−1 , an + a′n ) = γ(a1 , . . . , an−1 , an ) + γ(a1 , . . . , an−1 , a′n ) for the last copy of A. Denote by RCnD (A; Γ) the abelian subgroup of RCn (A; Γ) generated by the degenerate n-tuples, i.e. γ(a1 , . . . , an ) with ai = 0 for some 1 6 i 6 n. Consider also the quotient RCnN (A; Γ) = RCn (A; Γ)/RCnD (A; Γ). Further, define the maps ∂n : ΓA×n → ΓA×(n−1) , n > 1, as the linearizations of ∂n (a1 , . . . , an ) =(a1 · a2 , . . . , a1 · an ) Xn−2 (2.2) (−1)i (a1 , . . . , ai + ai+1 , . . . , an ) + i=1 + (−1)n−1 (a1 , . . . , an−2 , an ). Complete this family of maps by ∂1 = 0. Dually, for n > 0, let RC n (A; Γ) denote the set of maps f : A×n → Γ linear in the last coordinate: (2.3) f (a1 , . . . , an−1 , an + a′n ) = f (a1 , . . . , an−1 , an ) + f (a1 , . . . , an−1 , a′n ), n and let RCN (A; Γ) ⊂ RC n (A; Γ) comprise the maps vanishing on all degenerate n-tuples. Define the maps ∂ n : Fun(A×n , Γ) → Fun(A×(n+1) , Γ), n > 1, by (∂ n f )(a1 , . . . , an+1 ) =f (a1 · a2 , . . . , a1 · an+1 ) Xn−1 (2.4) (−1)i f (a1 , . . . , ai + ai+1 , . . . , an+1 ) + i=1 + (−1)n f (a1 , . . . , an−1 , an+1 ). These formulas resemble the group (co)homology construction for (A, +). We will now show that they indeed define a (co)homology theory. Proposition 2.1. Let (A, ·, +) be a linear cycle set and Γ be an abelian group. (1) The maps ∂• above ✓ square to zero: ∂n−1 ∂n = 0 for all n > 1; ✓ induce maps RCn (A; Γ) → RCn−1 (A; Γ); D ✓ and further restrict to maps RCnD (A; Γ) → RCn−1 (A; Γ). • (2) The maps ∂ above ✓ square to zero: ∂ n+1 ∂ n = 0 for all n > 1; ✓ restrict to maps RC n (A; Γ) → RC n+1 (A; Γ); n+1 n (A; Γ) → RCN (A; Γ). ✓ and further restrict to maps RCN 4 VICTORIA LEBED AND LEANDRO VENDRAMIN The induced or restricted maps from the proposition will be abusively denoted by the same symbols ∂• , ∂ • . In the proof we shall need the special properties of the zero element of a LCS. Lemma 2.2. In any LCS A, the relations a · 0 = 0, 0 · a = a hold for all a ∈ A. Proof. By the LCS axioms, one has a·0 = a·(0 + 0) = a·0 + a·0 and hence a·0 = 0. Similarly, 0 · a = (0 + 0) · a = (0 · 0) · (0 · a) = 0 · (0 · a), and the relation a = 0 · a follows by cancelling out 0 (recall that the left translation 0· is bijective).  Proof of Proposition 2.1. We treat only the homological statements here; they imply the cohomological ones by duality. Pn−1 The maps ∂n can be presented as signed sums ∂n = i=0 (−1)i ∂n;i , where (2.5) ∂n;0 (a1 , . . . , an ) = (a1 · a2 , . . . , a1 · an ), (2.6) ∂n;i (a1 , . . . , an ) = (a1 , . . . , ai + ai+1 , . . . , an ), 1 6 i 6 n − 2, (2.7) ∂n;n−1 (a1 , . . . , an ) = (a1 , . . . , an−2 , an ). The relation ∂n−1 ∂n = 0 then classically reduces to the “almost commutativity” ∂n−1;j ∂n;i = ∂n−1;i ∂n;j+1 for all i 6 j. In the case i > 0 this latter relation is either tautological, or follows from the associativity of +. For i = 0 < j, it follows from the left distributivity (1.3) for A. For i = 0 = j, it is a consequence of the second LCS relation (1.4) for A. Further, using the linearity (1.3) of the left translations an 7→ a1 · an , one sees that when applied to expressions of type (. . . , an−1 , an + a′n ) − (. . . , an−1 , an ) − (. . . , an−1 , a′n ), all the maps ∂n;i yield expressions of the same type. Hence their signed sums ∂• induce a differential on RC• . The possibility to further restrict to RC•D is guaranteed by Lemma 2.2.  Proposition 2.1 legitimizes the following definition: Definition 2.3. The reduced (normalized) cycles / boundaries / homology groups of a linear cycle set (A, ·, +) with coefficients in an abelian group Γ are the cycles / boundaries / homology groups of the chain complex (RC• (A; Γ), ∂• ) (respectively, (RC•N (A; Γ), ∂• )) above. Dually, the reduced (normalized) cocycles / coboundaries / cohomology groups of (A, ·, +) are those of the complex (RC • (A; Γ), ∂ • ) (re• spectively, (RCN (A; Γ), ∂ • )). We use the usual notations RQn (A; Γ), RQN n (A; Γ), n n RQ (A; Γ), RQN (A; Γ), where Q is one of the letters Z, B, or H. Remark 2.4. In the proof of Proposition 2.1 we actually showed that our (co)homology constructions can be refined into (co)simplicial ones. Example 2.5. Recall from the introduction that for a non-degenerate cycle set (X, ·), the free abelian group (Z(X) , +) can be seen as a linear cycle set, with the cycle set operation induced from ·. In this case RC1 (Z(X) ; Γ) is simply the abelian group Γ⊗Z Z(X) = Γ(X) , and for a1 , a2 ∈ Z(X) one calculates ∂1 (a1 , a2 ) = a1 ·a2 −a2 . Standard arguments from LCS theory then yield RH1 (Z(X) ; Γ) ∼ = Γ(Orb(X)) , where Orb(X) is the set of orbits of X, i.e., classes for the equivalence relation generated by a1 · a2 ∼ a2 for all a1 , a2 ∈ X. Similarly, one calculates the first reduced cohomology group: RH 1 (Z(X) ; Γ) ∼ = Fun(Orb(X), Γ). COHOMOLOGY OF BRACES 5 We finish with a comparison between the (co)homology of a LCS (A, ·, +) and the (co)homology of its underlying cycle set (A, ·), as defined in [21]. Recall that ×n the homology HnCS (A; Γ) of (A, ·) is computed by the complex (Γ(A ) , ∂nCS ), where Xn−1 (−1)i−1 ((ai · a1 , . . . , ai · ai−1 , ai · ai+1 , . . . , ai · an ) ∂nCS (a1 , . . . , an ) = i=1 −(a1 , . . . , ai−1 , ai+1 , . . . , an )). n Dually, the cohomology HCS (A; Γ) of (A, ·) is computed by the complex (Fun(A×n , Γ), ×n n n CS ∂CS ), with ∂CS f = f ◦ ∂n+1 . Define the map Sn : Γ(A ) → RCn (A; Γ) as the com- position of the anti-symmetrization map X γ(a1 , . . . , an ) 7→ (−1)σ γ(aσ(1) , . . . , aσ(n−1) , an ) σ∈Symn−1 (where (−1) is the sign of the permutation σ) and the obvious projection Γ(A RCn (A; Γ). σ ×n ) ։ Proposition 2.6. Let (A, ·, +) be a linear cycle set and Γ be an abelian group. The map S defined above yields a map of chain complexes Sn : (Γ(A ×n ) , ∂nCS ) → (RCn (A; Γ), ∂n ). Proof. One has to compare the evaluations of the maps ∂n Sn and Sn−1 ∂nCS on Pn−1 γ(a1 , . . . , an ). For this it is convenient to use the decomposition ∂n = i=0 (−1)i ∂n;i from (2.5)-(2.7). For 0 < i < n − 1, the map ∂n;i Sn is zero: in its evaluation, the terms ±γ(. . . , aj +ak , . . .) and ∓γ(. . . , ak +aj , . . .) (with the sum at the ith position) cancel. A careful sign inspection yields ∂n;0 Sn (γ(a1 , . . . , an )) = Xn−1 (−1)i−1 Sn−1 (γ(ai · a1 , . . . , ai · ai−1 , ai · ai+1 , . . . , ai · an )), i=1 ∂n;n−1 Sn (γ(a1 , . . . , an )) = Xn−1 (−1)n−1−i Sn−1 (γ(a1 , . . . , ai−1 , ai+1 , . . . , an )), i=1 hence the maps ∂n Sn and Sn−1 ∂nCS coincide.  As a consequence, one obtains the dual map n S n : (RC n (A; Γ), ∂ n ) → (Fun(A×n , Γ), ∂CS ) of cochain complexes, and the induced maps in (co)homology. 3. Cycle-type extensions vs. reduced 2-cocycles We now turn to a study of the reduced 2-cocycles of a linear cycle set (A, ·, +), i.e., maps f : A × A → Γ (where Γ is an abelian group) satisfying (3.1) f (a, b + c) = f (a, b) + f (a, c), (3.2) f (a + b, c) = f (a · b, a · c) + f (a, c) for all a, b, c ∈ A. The last relation, together with the commutativity of +, yields (3.3) f (a · b, a · c) + f (a, c) = f (b · a, b · c) + f (b, c), 2 (f ) = 0, so our f is necessarily a cocycle of the cycle set (A, ·). implying ∂CS Among the reduced 2-cocycles we distinguish the reduced 2-coboundaries ∂ 1 (θ)(a, b) = θ(a · b) − θ(b), where the map θ : A → Γ is linear. 6 VICTORIA LEBED AND LEANDRO VENDRAMIN Example 3.1. Let A and Γ be abelian groups. Consider the trivial linear cycle set structure a ·tr b = b over A. A map f : A × A → Γ is a reduced 2-cocycle of this LCS if and only if f is a bicharacter, in the sense of the bilinearity relations f (a + b, c) = f (a, c) + f (b, c), f (a, b + c) = f (a, b) + f (a, c). The reduced 2-coboundaries are all trivial in this case. Thus RH 2 (A; Γ) is the abelian group of bicharacters of A with values in Γ. Observe that for the cycle set n 2 (A, ·tr ), all the differentials ∂CS vanish. The second cohomology group HCS (A; Γ) of this cycle set thus comprises all the maps f : A × A → Γ, and is strictly larger than RH 2 (A; Γ). Example 3.2. Let A = {0, 1, 2, 3} = Z/4 be the cyclic group of 4 elements written additively. Then A is a brace with a ◦ b = a + b + 2ab, a−1 = (2a − 1)a. The corresponding linear cycle set structure on A is given by the operation a · b = a−1 ◦ (a + b) = (1 + 2a)b, which is b when one of a, b is even, and b + 2 otherwise. Take Γ = {0, 1} = Z/2. For a map f : Z/4 × Z/4 → Z/2, relation (3.1) means that f is of the form f (a, b) = bφ(a) (where the product is taken in Z/2, and b is reduced modulo 2), for some φ : Z/4 → Z/2. Relation (3.2) then translates as φ(a + b) = φ(b + 2ab) + φ(a). The substitution b = 0 yields φ(0) = 0. Analyzing other values of a and b, one sees that φ(1) and φ(3) can be chosen arbitrarily, and φ(2) has to equal φ(1)+φ(3). The reduced 2-coboundaries are again trivial: a linear map θ : Z/4 → Z/2 is necessarily of the form θ(a) = at for some constant t ∈ Z/2, yielding θ(a · b) = (a · b)t = (1 + 2a)bt = bt = θ(b) (since 2a = 0 in Z/2). Summarizing, one gets RH 2 (Z/4; Z/2) ≃ Z/2 × Z/2. Let us now turn to the underlying cycle set (Z/4, ·). Playing with condition (3.3), one verifies that its 2-cocycles are maps f : Z/4 × Z/4 → Z/2 verifying 3 linear relations f (0, 1) + f (0, 3) = 0, f (2, 1) + f (2, 3) = 0, f (1, 1) + f (1, 3) + f (3, 1) + f (3, 3) = 0. Its only non-trivial 2-coboundary is f (a, b) = ab mod 2. This implies 2 ZCS (Z/4; Z/2) ≃ (Z/2)4×4−3 = (Z/2)13 , 2 HCS (Z/4; Z/2) ≃ (Z/2)12 . We will now construct extensions of our LCS A by Γ out of 2-cocycles, show that any central cycle-type extension is isomorphic to one of this type, and that reduced 2-cocycles, modulo reduced 2-coboundaries, classify such extensions. Lemma 3.3. Let (A, ·, +) be a linear cycle set, Γ be an abelian group, and f : A × A → Γ be a map. Then the abelian group Γ ⊕ A with the operation (γ, a) · (γ ′ , a′ ) = (γ ′ + f (a, a′ ), a · a′ ), γ, γ ′ ∈ Γ, a, a′ ∈ A is a linear cycle set if and only if f is a reduced 2-cocycle, f ∈ RZ 2 (A; Γ). Notation 3.4. The LCS from the lemma is denoted by Γ ⊕f A. COHOMOLOGY OF BRACES 7 Proof. The left translation invertibility for Γ ⊕f A follows from to the same property for A. Properties (1.3) and (1.4) are equivalent for Γ ⊕f A to, respectively, properties (3.1) and (3.2) from the definition of a 2-cocycle for f . The cycle set property (1.2) follows from (1.4) and the commutativity of +.  Lemma 3.3 and the correspondence between linear cycle sets and braces yield the following result. Lemma 3.5. Let (A, ◦, +) be a brace, Γ be an abelian group, and f : A × A → Γ be a map. Then the abelian group Γ ⊕ A with the product (γ, a) ◦ (γ ′ , a′ ) = (γ + γ ′ + f (a, a′ ), a ◦ a′ ), γ, γ ′ ∈ Γ, a, a′ ∈ A, is a brace if and only if for the corresponding linear cycle set (A, ·, +), the map f (a, b) = f (a, a · b) is a reduced 2-cocycle. Before introducing the notion of LCS extensions, we need some preliminary definitions. Definition 3.6. A morphism between linear cycle sets A and B is a map ϕ : A → B preserving the structure, i.e. for all a, a′ ∈ A one has ϕ(a + a′ ) = ϕ(a) + ϕ(a′ ) and ϕ(a · a′ ) = ϕ(a) · ϕ(a′ ). The kernel of ϕ is defined by Ker ϕ = ϕ−1 (0). The notions of the image Im ϕ = ϕ(A), of a short exact sequence of linear cycle sets, and of linear cycle subsets, are defined in the obvious way. A linear cycle subset A′ of A is called central if for all a ∈ A, a′ ∈ A′ , one has a · a′ = a′ and a′ · a = a. For a LSC morphism ϕ : A → B, Ker ϕ and Im ϕ are clearly linear cycle subsets of A and B respectively. Lemma 2.2 can be rephrased by stating that {0} is a central linear cycle subset of A. Definition 3.7. A central cycle-type extension of a linear cycle set (A, ·, +) by an abelian group Γ is the datum of a short exact sequence of linear cycle sets ι π (3.4) 0 → Γ → E → A → 0, where Γ is endowed with the trivial cycle set structure γ · γ ′ = γ ′ , its image ι(Γ) is central in E (in the sense of Definition 3.6), and the short exact sequence of abelian groups underlying (3.4) splits. The adjective cycle-type refers here to the fact that our extensions are interesting on the level of the cycle set operation · only, and trivial on the level of the additive operation +, since we require the short exact sequences to linearly split. More general extensions—those taking into account the additive operation as well—are postponed until the next section. Cycle-type extensions are important, for example, for comparing the LCS structures on the structure group of a cycle set before and after a cycle set extension (cf. the introduction for more detail on structure groups, and [21] for the cycle set extension theory). The LCS Γ ⊕f A from Lemma 3.3 is an extension of A by Γ in the obvious way. We now show that this example is essentially exhaustive. ι π ι′ Definition 3.8. Two central cycle-type LCS extensions Γ ֌ E ։ A and Γ ֌ π′ E ′ ։ A are called equivalent if there exists a LCS isomorphism ϕ : E → E ′ making 8 VICTORIA LEBED AND LEANDRO VENDRAMIN the following diagram commute: 55 E ❘❘❘ ❘❘π❘ ❧❧❧ ❘❘❘ ❧ι❧❧ ❧ ❧ ❘)) )) 55❧ (3.5) Γ ))❘❘ ′ ∼ ϕ ❧55 55 A ❘❘❘ι π ′❧❧❧❧ ❘❘❘❘  ❧ ❧ )) ′ ❧❧❧ E The set of equivalence classes of central cycle-type extensions of A by Γ is denoted by CTExt(A, Γ). ι π Lemma 3.9. Let Γ ֌ E ։ A be a central cycle-type LCS extension, and s : A → E be a linear section of π. Then the map fe: A × A → E, (a, a′ ) 7→ s(a) · s(a′ ) − s(a · a′ ) takes values in ι(Γ) and defines a reduced cocycle f ∈ RZ 2 (A; Γ). Extensions E and Γ ⊕f A are equivalent. Furthermore, a cocycle f ′ obtained from another section s′ of π is cohomologous to f . Proof. The computation π(fe(a, a′ )) = πs(a) · πs(a′ ) − πs(a · a′ ) = a · a′ − a · a′ = 0 yields Im fe ⊆ Ker π = Im ι (by the definition of a short exact sequence). Hence the map f : A × A → Γ can be defined by the formula f = ι−1 fe. It remains to check relations (3.1)-(3.2) for this map. The linearity of s and of the left translations tb : a 7→ b · a gives fe(a, b + c) = s(a) · s(b + c) − s(a · (b + c)) = s(a) · (s(b) + s(c)) − s(a · b + a · c) = s(a) · s(b) + s(a) · s(c) − s(a · b) − s(a · c) = fe(a, b) + fe(a, c). hence f (a, b + c) = f (a, b) + f (a, c), by the linearity of ι. Similarly, one has fe(a + b, c) = s(a + b) · s(c) − s((a + b) · c) = (s(a) + s(b)) · s(c) − s((a · b) · (a · c)) = (s(a) · s(b)) · (s(a) · s(c)) + fe(a · b, a · c) − s(a · b) · s(a · c) = fe(a · b, a · c) + (fe(a, b) + s(a · b)) · (s(a) · s(c)) − s(a · b) · s(a · c) (1) = fe(a · b, a · c) + s(a · b) · (s(a) · s(c)) − s(a · b) · s(a · c) = fe(a · b, a · c) + s(a · b) · (s(a) · s(c) − s(a · c)) = fe(a · b, a · c) + s(a · b) · fe(a, c) (2) = fe(a · b, a · c) + fe(a, c). In (1) we got rid of fe(a, b) ∈ ι(Γ) since the centrality of ι(Γ) yields (fe(a, b) + x) · y = (fe(a, b) · x) · (fe(a, b) · y) = x · y for all x, y ∈ E. This centrality was also used in (2). Relation f (a + b, c) = f (a · b, a · c) + f (a, c) is now obtained from the corresponding relation for fe by applying ι−1 . We will next show that the linear map ϕ : Γ ⊕f A → E, γ ⊕ a 7→ ι(γ) + s(a) yields an equivalence of extensions. It is bijective, the inverse given by the map x 7→ ι−1 (x − sπ(x)) ⊕ π(x) (this map is well defined since x − sπ(x) ∈ Ker π = Im ι). COHOMOLOGY OF BRACES 9 Let us check that ϕ entwines the cycle set operations. One has ϕ((γ ⊕ a) · (γ ′ ⊕ a′ )) = ϕ((γ ′ + f (a, a′ )) ⊕ a · a′ ) = ι(γ ′ + f (a, a′ )) + s(a · a′ ) = ι(γ ′ ) + fe(a, a′ ) + (s(a) · s(a′ ) − fe(a, a′ )) = ι(γ ′ ) + s(a) · s(a′ ) = s(a) · ι(γ ′ ) + s(a) · s(a′ ) = s(a) · (ι(γ ′ ) + s(a′ )) = (ι(γ) + s(a)) · (ι(γ ′ ) + s(a′ )) = ϕ(γ ⊕ a) · ϕ(γ ′ ⊕ a′ ). We used the centrality of ι(γ ′ ) and ι(γ). The commutativity of the diagram (3.5) is obvious, and completes the proof. Suppose now that the reduced cocycles f and f ′ are obtained from the sections s and s′ respectively. Put θe = s − s′ : A → E. This is a linear map with its image contained in Ker π = Im ι. Hence it defines a linear map θ : A → Γ. To show that f and f ′ are cohomologous, we establish the property f ′ − f = ∂ 1 θ. One computes (fe − fe′ )(a, a′ ) = fe(a, a′ ) − fe′ (a, a′ ) = s(a) · s(a′ ) − s(a · a′ ) − s′ (a) · s′ (a′ ) + s′ (a · a′ ) e · a′ ) = s(a) · s(a′ ) − s(a) · s′ (a′ ) − θ(a e · a′ ) = s(a) · (s(a′ ) − s′ (a′ )) − θ(a e ′ ) − θ(a e · a′ ) (1) e ′ ) − θ(a e · a′ ). = s(a) · θ(a = θ(a e ′ ). The desired relation is obtained by applying In (1) we used the centrality of θ(a −1 ι .  We now compare extensions constructed out of different 2-cocycles. Lemma 3.10. Let (A, ·, +) be a linear cycle set, Γ be an abelian group, and f, f ′ ∈ RZ 2 (A; Γ) be two reduced 2-cocycles. The linear cycle set extensions Γ ⊕f A and Γ ⊕f ′ A are equivalent if and only if f and f ′ are cohomologous. Proof. Suppose that a linear map ϕ : Γ ⊕f A → Γ ⊕f ′ A provides an equivalence of extensions. The commutativity of the diagram (3.5) forces it to be of the form ϕ(γ ⊕ a) = (γ + θ(a)) ⊕ a for some linear map θ : A → Γ. Further, one computes ϕ((γ ⊕ a) · (γ ′ ⊕ a′ )) = ϕ((γ ′ + f (a, a′ )) ⊕ a · a′ ) = (γ ′ + f (a, a′ ) + θ(a · a′ )) ⊕ a · a′ , ϕ(γ ⊕ a) · ϕ(γ ′ ⊕ a′ ) = ((γ + θ(a)) ⊕ a) · ((γ ′ + θ(a′ )) ⊕ a′ ) = (γ ′ + θ(a′ ) + f ′ (a, a′ )) ⊕ a · a′ . Thus the map ϕ entwines the cycle set operations if and only if f ′ − f is the coboundary ∂ 1 θ. In the opposite direction, take cohomologous cocycles f and f ′ . This means that the relation f ′ − f = ∂ 1 θ holds for a linear map θ : A → Γ. Repeating the arguments above, one verifies that the map ϕ(γ⊕a) = (γ+θ(a))⊕a is an equivalence  of extensions Γ ⊕f A → Γ ⊕f ′ A. Put together, the preceding lemmas yield: Theorem 3.11. Let (A, ·, +) be a linear cycle set and Γ be an abelian group. The construction from Lemma 3.9 yields a bijective correspondence 1:1 CTExt(A, Γ) ←→ RH 2 (A; Γ). 10 VICTORIA LEBED AND LEANDRO VENDRAMIN We finish this section by observing that in degree 2, the normalization brings nothing new to the reduced LCS cohomology theory: Proposition 3.12. All reduced 2-cocycles of a linear cycle set (A, ·, +) are normalized. Moreover, one has an isomorphism in cohomology: 2 RH 2 (A; Γ) ∼ (A; Γ). = RHN Proof. Putting a = b = 0 in the defining relation (3.2) for a reduced 2-cocycle f , and using the properties of the element 0 from Lemma 2.2, one gets f (0, c) = 0 for all c ∈ A. Moreover, f (c, 0) = 0 by linearity. So f is normalized, hence the identification 2 RZ 2 (A; Γ) = RZN (A; Γ). In degree 1 the normalized and usual complexes coincide, yielding the desired cohomology group isomorphism in degree 2.  4. Full linear cycle set cohomology The previous section treated linear cycle set extensions of the form Γ⊕f A. They can be thought of as the direct product Γ ⊕ A of LCS with the cycle set operation · deformed by f . From now on we will handle a more general situation: the additive operation + on Γ ⊕ A will be deformed as well. Most proofs in this general case are analogous to but more technical than those from the previous sections. Take a linear cycle set (A, ·, +) and an abelian group Γ. For i > 0, j > 1, let ×(i+j) ) ShCi,j (A; Γ) be the abelian subgroup of Γ(A , generated by the partial shuffles X σ (4.1) (−1) γ(a1 , . . . , ai , ai+σ−1 (1) , . . . , ai+σ−1 (j) ) σ∈Shr,j−r taken for all 1 6 r 6 j − 1, ak ∈ A, γ ∈ Γ. Here Shr,j−r is the subset of all the permutations σ of j elements satisfying σ(1) 6 · · · 6 σ(r), σ(r + 1) 6 · · · 6 σ(j). ×(i+j) ) The term shuffle is used when i = 0. Put Ci,j (A; Γ) = Γ(A /ShCi,j (A; Γ). Recall the notations (4.2) ∂n;0 (a1 , . . . , an ) = (a1 · a2 , . . . , a1 · an ), (4.3) ∂n;i (a1 , . . . , an ) = (a1 , . . . , ai + ai+1 , . . . , an ), 1 6 i 6 n − 1, from the proof of Proposition 2.1, and consider the coordinate omitting maps ′ (4.4) ∂n;i (a1 , . . . , an ) = (a1 , . . . , ai−1 , ai+1 , . . . , an ), 1 6 i 6 n. Combine (the linearizations of) these maps into what we will show to be horizontal and vertical differentials of a bicomplex: Xi−1 ′ h (4.5) (−1)k ∂i+j;k + (−1)i ∂i+j;i , i > 1, j > 1; ∂i,j = ∂i+j;0 + k=1 Xj−1 ′ v ′ (4.6) −∂i,j (−1)k ∂i+j;i+k + (−1)j ∂i+j;i+j , i > 0, j > 2. = ∂i+j;i+1 + k=1 D Here the empty sums are zero by convention. As before, Ci,j (A; Γ) denotes the ×(i+j) ) abelian subgroup of Γ(A generated by the degenerate (i + j)-tuples, and ×(i+j) N ) D Ci,j (A; Γ) is the quotient Γ(A /(Ci,j (A; Γ) + ShCi,j (A; Γ)). ×(i+j) h v Dually, for f ∈ Fun(A , Γ), put ∂hi,j f = f ◦ ∂i+1,j and ∂vi,j f = f ◦ ∂i,j+1 , ×(i+j) ) where i > 0, j > 1, and f is extended to Z(A by linearity. Let C i,j (A; Γ) be ×(i+j) the abelian group of the maps A → Γ whose linearization vanishes on all the i,j partial shuffles (4.1) (with γ omitted), and let CN (A; Γ) ⊆ C i,j (A; Γ) comprise the maps which are moreover zero on all the degenerate (i + j)-tuples. We now assemble these data into a chain and a cochain bicomplex structures with normalization. COHOMOLOGY OF BRACES 11 Theorem 4.1. Let (A, ·, +) be a linear cycle set and Γ be an abelian group. ×(i+j) ) (1) The abelian groups Γ(A , i > 0, j > 1, together with the linear maps h v ∂i,j and ∂i,j above, form a chain bicomplex. In other words, the following relations are satisfied: h h (4.7) ∂i−1,j ∂i,j = 0, i > 2, j > 1; (4.8) v v ∂i,j−1 ∂i,j = 0, (4.9) h v ∂i,j−1 ∂i,j = i > 0, j > 3; v h ∂i−1,j ∂i,j , i > 1, j > 2. D Moreover, these maps restrict to the subgroups ShCi,j (A; Γ) and Ci,j (A; Γ), N and thus induce chain bicomplex structures on Ci,j (A; Γ) and on Ci,j (A; Γ). (2) The linear maps ∂hi,j and ∂vi,j yield a cochain bicomplex structure for the abelian groups Fun(A×(i+j) , Γ), i > 0, j > 1. This structure restricts to i,j C i,j (A; Γ) and further to CN (A; Γ). We abusively denote the induced or restricted maps from the theorem by the same symbols ∂•h , ∂•v , etc. The proof of the theorem relies on the following interpretation of our bicomplex. Its jth row is almost the complex from Proposition 2.1, with a slight modification: the last entry in an n-tuple, to which the ∂n;i with i > 0 did nothing and on which ∂n;0 acted by a left translation an 7→ a1 · an , is replaced with the j-tuple of last elements behaving in the same way. In the ith column, the first i entries of A×(i+•) v are never affected, and on the remaining entries the vertical differentials ∂i,• act as the differentials from Proposition 2.1 computed for the trivial cycle set operation a · b = b. Alternatively, the ith column can be seen as the Hochschild complex for (A, +) with coefficients in A×i , on which A acts trivially on both sides. Modding out ShCi,j (A; Γ) means passing from the Hochschild to the Harrison complex in each column. Proof. As usual, it suffices to treat only the homological statements. Due to the observation preceding the proof, the horizontal relation (4.7) and the vertical relation (4.8) follow from Proposition 2.1. For the mixed relation (4.9), note that the horizontal and vertical differentials involved affect, respectively, the first i and the last j entries of an (i + j)-tuple, with the exception of the ∂n;0 component of ∂ h . However, this component also commutes with ∂ v because of the linearity (with respect to +) of the left translation a1 · involved. Applying a left translation a· to each entry of a partial shuffle (4.1), one still gets a partial shuffle. Consequently, the horizontal differentials ∂ h restrict to ShCi,j (A; Γ). In order to show that the ∂ v restrict to ShCi,j (A; Γ) as well, it suffices to check that the expression X (−1)σ (aσ−1 (2) , . . . , aσ−1 (j) ) σ∈Shr,j−r + j−1 X (−1)k k=1 X (−1)σ (aσ−1 (1) , . . . , aσ−1 (k) + aσ−1 (k+1) , . . . , aσ−1 (j) ) σ∈Shr,j−r + (−1)j X σ∈Shr,j−r (−1)σ (aσ−1 (1) , . . . , aσ−1 (j−1) ) 12 VICTORIA LEBED AND LEANDRO VENDRAMIN is a linear combination of shuffles for all j > 1, 1 6 r 6 j − 1, ak ∈ A. Denote by S1 , S2 , and S3 the three sums in the expression above. We also use the classical notation X (−1)σ (aσ−1 (1) , . . . , aσ−1 (j) ) r,j−r (a1 , . . . , aj ) =   σ∈Shr,j−r  for shuffles, and the convention 0,j = j,0 = Id. Recall also notations (4.3)-(4.4). The sums Si then rewrite as ′ ′ S1 = r−1,j−r ∂j;1 + (−1)r r,j−r−1 ∂j;r+1 ,  r   ′ r−1,j−r ∂j;r S3 = (−1) Xr−1 (−1)k S2 = k=1  + (−1)j r−1,j−r  ∂j;k + r,j−r−1 Xj−1 ′ ∂j;j , k=r+1 (−1)k  r,j−r−1 ∂j;k , with empty sums declared to be zero. The decomposition for S1 follows from the analysis of the two possibilities for σ −1 (1) with σ ∈ Shr,j−r , namely, σ −1 (1) = 1 and σ −1 (1) = r +1. The decomposition for S3 corresponds to the dichotomy σ −1 (j) = r or σ −1 (j) = j. In S3 , the summands with σ −1 (k) = u 6 r < v = σ −1 (k + 1) and σ −1 (k) = v, σ −1 (k + 1) = u appear with opposite signs and can therefore be discarded. The remaining ones can be divided into two classes: those with σ −1 (k) < σ −1 (k + 1) 6 r and those with r < σ −1 (k) < σ −1 (k + 1), giving the decomposition above. Our Si are thus signed sums of shuffles, with the exception ′ ′ of the cases r ∈ {1, j − 1}. For r = 1, the non-shuffle terms ∂j;1 and −∂j;1 appear in S1 and S3 respectively; they annihilate each other in the total sum. The case r = j − 1 is treated similarly. D The possibility to restrict all the ∂ h and ∂ v to Ci,j (A; Γ) is taken care of, as usual, by Lemma 2.2. As a consequence, one obtains a chain bicomplex structure N on Ci,j (A; Γ).  We are now in a position to define the full (co)homology of a linear cycle set: Definition 4.2. The (normalized) cycles / boundaries / homology groups of a linear cycle set (A, ·, +) with coefficients in an abelian group Γ are the cycles / boundaries / homology groups of the total chain complex h v + (−1)i ∂i,j ) ( Cn (A; Γ) = ⊕i+j=n Ci,j (A; Γ), ∂n |Ci,j = ∂i,j N (respectively, CnN (A; Γ) = (⊕i+j=n Ci,j (A; Γ), ∂n )) of the bicomplex above. Dually, the (normalized) cocycles / coboundaries / cohomology groups of (A, ·, +) are ∗ those of the complex (C n (A; Γ) = ⊕i+j=n C i,j (A; Γ), ∂ n = ∂n+1 ) (respectively, i,j n (CN (A; Γ) = ⊕i+j=n CN (A; Γ), ∂ n )). We use the usual notations Qn (A; Γ) etc., where Q is one of the letters Z, B, or H. Remark 4.3. In fact our (co)chain bicomplex constructions can be refined into bisimplicial ones. Remark 4.4. Instead of considering the total complex of our bicomplex, one could v start by, say, computing the homology Hi,• of each column. The horizontal differv entials then induce a chain complex structure on each row H•,j . Observe that the first row is precisely the complex from Proposition 2.1. Its homology is then the reduced homology of our linear cycle set. 5. General linear cycle set extensions Our next step is to describe what a 2-cocycle looks like for the full version of linear cycle set cohomology theory. Such a 2-cocycle consists of two components COHOMOLOGY OF BRACES 13 f, g : A × A → Γ, seen as elements of C 1,1 (A; Γ) = Fun(A × A, Γ) and C 0,2 (A; Γ) = Sym(A × A, Γ) respectively. Here Sym denotes the abelian group of symmetric maps, in the sense of (5.1) g(a, b) = g(b, a). These maps should satisfy three identities, one for each component of C 3 (A; Γ) = C 2,1 (A; Γ) ⊕ C 1,2 (A; Γ) ⊕ C 0,3 (A; Γ). Explicitly, these identities read (5.2) f (a + b, c) = f (a · b, a · c) + f (a, c), (5.3) f (a, b + c) − f (a, b) − f (a, c) = g(a · b, a · c) − g(b, c), (5.4) g(a, b) + g(a + b, c) = g(b, c) + g(a, b + c). In particular, f is a 2-cocycle of the cycle set (A, ·), and g is a symmetric 2-cocycle of the group (A, +). The reduced cocycles are precisely those with g = 0. Further, the 2-coboundaries are couples of maps (5.5) f (a, b) = θ(a · b) − θ(b), (5.6) g(a, b) = θ(a + b) − θ(a) − θ(b) for some θ : A → Γ. We next give some elementary properties of 2-cocycles and 2-coboundaries. Lemma 5.1. Let (f, g) be a 2-cocycle of a linear cycle set (A, ·, +) with coefficients in an abelian group Γ. (1) For all x ∈ A, one has f (0, x) = f (x, 0) = 0, g(0, x) = g(x, 0) = g(0, 0). (2) The 2-cocycle (f, g) is normalized if and only if one has g(0, 0) = 0. Proof. Let us prove the first claim. The relation f (0, x) = 0 follows from (5.2) by choosing a = 0. Similarly, the relation f (x, 0) = 0 is (5.3) specialized at b = c = 0. Substitutions b = 0 and either a = 0 or c = 0 in (5.4) yield the last relation. Now the second claim directly follows from the previous point.  Lemma 5.2. Let (A, ·, +) be a linear cycle set, Γ be an abelian group, and f, g : A× A → Γ be two maps. Then the set Γ × A with the operations (γ, a) + (γ ′ , a′ ) = (γ + γ ′ + g(a, a′ ), a + a′ ), (γ, a) · (γ ′ , a′ ) = (γ ′ + f (a, a′ ), a · a′ ) for γ, γ ′ ∈ Γ, a, a′ ∈ A, is a linear cycle set if and only if (f, g) is a 2-cocycle, (f, g) ∈ Z 2 (A; Γ). Notation 5.3. The LCS from the lemma is denoted by Γ ⊕f,g A. Proof. The left translation invertibility for Γ ⊕f,g A follows from the same property for A. Properties (1.3) and (1.4) for Γ ⊕f,g A are equivalent to, respectively, properties (5.3) and (5.2) for (f, g). The associativity and the commutativity of + on Γ ⊕f,g A are encoded by property (5.4) for Γ ⊕f,g A and the symmetry of g respectively. Finally, if (f, g) is a 2-cocycle, then Lemma 5.1 implies that (−g(0, 0), 0) is the zero element for (Γ ⊕f,g A, +), and the opposite of (γ, a) is (−γ − g(0, 0) − g(a, −a), −a).  14 VICTORIA LEBED AND LEANDRO VENDRAMIN As we did in Lemma 3.5, we now translate Lemma 5.2 into the language of braces. Lemma 5.4. Let (A, ◦, +) be a brace, Γ be an abelian group, and f, g : A × A → Γ be two maps. Then the set Γ × A with the operations (γ, a) + (γ ′ , a′ ) = (γ + γ ′ + g(a, a′ ), a + a′ ), (γ, a) ◦ (γ ′ , a′ ) = (γ + γ ′ + f (a, a′ ), a ◦ a′ ) for γ, γ ∈ Γ, a, a ∈ A, is a brace if and only if for the corresponding linear cycle set (A, ·, +), the maps (5.7) f (a, b) = −f (a, a · b) + g(a, b) ′ ′ and g form a 2-cocycle (f , g) ∈ Z 2 (A; Γ). Proof. Recall the correspondence a · b = a−1 ◦ (a + b) between the corresponding brace and LCS operations. It can also be rewritten as a ◦ b = a + a ∗ b, where the map a 7→ a ∗ b is the inverse of the left translation a 7→ a · b. Now, given any (f , g) ∈ Z 2 (A; Γ), the formulas from Lemma 5.2 describe a LCS structure on Γ × A. Its operation ∗ reads (γ, a) ∗ (γ ′ , a′ ) = (γ ′ − f (a, a ∗ a′ ), a ∗ a′ ). The operations (γ, a) + (γ ′ , a′ ) = (γ + γ ′ + g(a, a′ ), a + a′ ), (γ, a) ◦ (γ ′ , a′ ) = (γ, a) + (γ, a) ∗ (γ ′ , a′ ) = (γ + γ ′ − f (a, a ∗ a′ ) + g(a, a ∗ a′ ), a ◦ a′ ) then yield a brace structure on Γ × A. These formulas have the desired form, with f (a, a′ ) = −f (a, a ∗ a′ ) + g(a, a ∗ a′ ), which, through the substitution b = a ∗ a′ , is equivalent to (5.7). Conversely, starting from a brace structure on Γ × A of the desired form, one sees that its associated LCS structure is as described in Lemma 5.2 with some (f , g) ∈ Z 2 (A; Γ). Repeating the argument above, one obtains the relation (5.7) connecting f , f , and g.  Definition 5.5. A central extension of a linear cycle set (A, ·, +) by an abelian group Γ is the datum of a short exact sequence of linear cycle sets ι π (5.8) 0 → Γ → E → A → 0, where Γ is endowed with the trivial cycle set structure, and its image ι(Γ) is central in E (in the sense of Definition 3.6). The notion of equivalence for central cycle-type LCS extensions (Definition 3.8) transports verbatim to these general extensions. The set of equivalence classes of central extensions of A by Γ is denoted by Ext(A, Γ). The LCS Γ ⊕f,g A from Lemma 5.2 is an extension of A by Γ in the obvious way. We now show that this example is essentially exhaustive. ι π Lemma 5.6. Let Γ ֌ E ։ A be a central LCS extension, and s : A → E be a set-theoretic section of π. (1) The maps fe, e g : A × A → E defined by fe: (a, a′ ) 7→ s(a) · s(a′ ) − s(a · a′ ), g : (a, a′ ) 7→ s(a) + s(a′ ) − s(a + a′ ) e COHOMOLOGY OF BRACES 15 both take values in ι(Γ) and determine a cocycle (f, g) ∈ Z 2 (A; Γ). (2) The cocycle above is normalized if and only if s is such, in the sense of s(0) = 0. (3) Extensions E and Γ ⊕f,g A are equivalent. (4) A cocycle (f ′ , g ′ ) obtained from another section s′ of π is cohomologous to (f, g). If both cocycles are normalized, then they are cohomologous in the normalized sense. Lemma 5.7. Let (A, ·, +) be a linear cycle set, Γ be an abelian group, and (f, g) and (f ′ , g ′ ) ∈ Z 2 (A; Γ) be 2-cocycles. The linear cycle set extensions Γ ⊕f,g A and Γ ⊕f ′ ,g′ A are equivalent if and only if (f, g) − (f ′ , g ′ ) is a normalized 2-coboundary. Recall that a normalized 2-coboundary is a couple of maps of the form ∂ 1 θ, where the map θ : A → Γ is normalized, in the sense of θ(0) = 0. The proof of these lemmas is technical but conceptually analogous to the proofs of Lemmas 3.9 and 3.10, and will therefore be omitted. Put together, the preceding lemmas prove Theorem 5.8. Let (A, ·, +) be a linear cycle set and Γ be an abelian group. The construction from Lemma 5.6 yields a bijective correspondence 1:1 2 Ext(A, Γ) ←→ HN (A; Γ). In other words, the central extensions of LCS (and thus of braces) are completely determined by their second normalized cohomology groups. References [1] D. Bachiller. Classification of braces of order p3 . J. Pure Appl. Algebra, 219(8):3568–3603, 2015. [2] D. Bachiller. Examples of simple left braces. arXiv:1511.08477, Nov. 2015. [3] D. Bachiller, F. Cedo, and E. Jespers. Solutions of the Yang-Baxter equation associated with a left brace. arXiv:1503.02814, Mar. 2015. [4] D. Bachiller, F. Cedó, E. Jespers, and J. Okniński. A family of irretractable square-free solutions of the Yang-Baxter equation. arXiv:1511.07769, Nov. 2015. [5] N. Ben David and Y. Ginosar. On groups of I-type and involutive Yang-Baxter groups. arXiv:1403.5740, Mar. 2014. [6] F. Catino, I. Colazzo, and P. Stefanelli. On regular subgroups of the affine group. Bull. Aust. Math. Soc., 91(1):76–85, 2015. [7] F. Catino, I. Colazzo, and P. Stefanelli. Regular subgroups of the affine group and asymmetric product of radical braces. Journal of Algebra, 455:164 – 182, 2016. [8] F. Catino and R. Rizzo. Regular subgroups of the affine group and radical circle algebras. Bull. Aust. Math. Soc., 79(1):103–107, 2009. [9] F. Cedó, E. Jespers, and Á. del Rı́o. Involutive Yang-Baxter groups. Trans. Amer. Math. Soc., 362(5):2541–2558, 2010. [10] F. Cedó, E. Jespers, and J. Okniński. Retractability of set theoretic solutions of the YangBaxter equation. Adv. Math., 224(6):2472–2484, 2010. [11] F. Cedó, E. Jespers, and J. Okniński. Braces and the Yang-Baxter equation. Comm. Math. Phys., 327(1):101–116, 2014. [12] F. Chouraqui. Garside groups and Yang-Baxter equation. Comm. Algebra, 38(12):4441–4460, 2010. [13] P. Dehornoy. Set-theoretic solutions of the Yang–Baxter equation, RC-calculus, and Garside germs. Adv. Math., 282:93–127, 2015. [14] P. Etingof, T. Schedler, and A. Soloviev. Set-theoretical solutions to the quantum YangBaxter equation. Duke Math. J., 100(2):169–209, 1999. [15] T. Gateva-Ivanova. A combinatorial approach to the set-theoretic solutions of the YangBaxter equation. J. Math. Phys., 45(10):3828–3858, 2004. 16 VICTORIA LEBED AND LEANDRO VENDRAMIN [16] T. Gateva-Ivanova. Set-theoretic solutions of the Yang-Baxter equation, Braces, and Symmetric groups. arXiv:1507.02602, July 2015. [17] T. Gateva-Ivanova and P. Cameron. Multipermutation solutions of the Yang-Baxter equation. Comm. Math. Phys., 309(3):583–621, 2012. [18] T. Gateva-Ivanova and S. Majid. Matched pairs approach to set theoretic solutions of the Yang-Baxter equation. J. Algebra, 319(4):1462–1529, 2008. [19] T. Gateva-Ivanova and M. Van den Bergh. Semigroups of I-type. J. Algebra, 206(1):97–112, 1998. [20] E. Jespers and J. Okniński. Monoids and groups of I-type. Algebr. Represent. Theory, 8(5):709–729, 2005. [21] V. Lebed and L. Vendramin. Homology of left non-degenerate set-theoretic solutions to the Yang-Baxter equation. arXiv:1509.07067, Sept. 2015. [22] J.-H. Lu, M. Yan, and Y.-C. Zhu. On the set-theoretical Yang-Baxter equation. Duke Math. J., 104(1):1–18, 2000. [23] W. Rump. A decomposition theorem for square-free unitary solutions of the quantum YangBaxter equation. Adv. Math., 193(1):40–55, 2005. [24] W. Rump. Braces, radical rings, and the quantum Yang-Baxter equation. J. Algebra, 307(1):153–170, 2007. [25] W. Rump. Semidirect products in algebraic logic and solutions of the quantum Yang-Baxter equation. J. Algebra Appl., 7(4):471–490, 2008. [26] W. Rump. The brace of a classical group. Note Mat., 34(1):115–144, 2014. [27] A. Smoktunowicz. A note on set-theoretic solutions of the Yang-Baxter equation. arXiv:1512.06642, Dec. 2015. [28] A. Smoktunowicz. On Engel groups, nilpotent groups, rings, braces and the Yang-Baxter equation. arXiv:1509.00420, Sept. 2015. [29] A. Soloviev. Non-unitary set-theoretical solutions to the quantum Yang-Baxter equation. Math. Res. Lett., 7(5-6):577–596, 2000. [30] L. Vendramin. Extensions of set-theoretic solutions of the Yang–Baxter equation and a conjecture of Gateva-Ivanova. J. Pure Appl. Algebra, 220(5):2064–2076, 2016. Laboratoire de Mathématiques Jean Leray, Université de Nantes, 2 rue de la Houssinière, BP 92208 F-44322 Nantes Cedex 3, France E-mail address: [email protected], [email protected] Depto. de Matemática, FCEN, Universidad de Buenos Aires, Pabellón I, Ciudad Universitaria (1428) Buenos Aires, Argentina E-mail address: [email protected]
4math.GR
Adaptive Submodularity: Theory and Applications in Active Learning and Stochastic Optimization Daniel Golovin ANTISPAM . GOLOVIN @ GMAIL . COM California Institute of Technology Pasadena, CA 91125, USA arXiv:1003.3967v5 [cs.LG] 6 Dec 2017 Andreas Krause ANTISPAM . KRAUSEA @ ETHZ . CH ETH Zurich 8092 Zurich, Switzerland Abstract Many problems in artificial intelligence require adaptively making a sequence of decisions with uncertain outcomes under partial observability. Solving such stochastic optimization problems is a fundamental but notoriously difficult challenge. In this paper, we introduce the concept of adaptive submodularity, generalizing submodular set functions to adaptive policies. We prove that if a problem satisfies this property, a simple adaptive greedy algorithm is guaranteed to be competitive with the optimal policy. In addition to providing performance guarantees for both stochastic maximization and coverage, adaptive submodularity can be exploited to drastically speed up the greedy algorithm by using lazy evaluations. We illustrate the usefulness of the concept by giving several examples of adaptive submodular objectives arising in diverse AI applications including management of sensing resources, viral marketing and active learning. Proving adaptive submodularity for these problems allows us to recover existing results in these applications as special cases, improve approximation guarantees and handle natural generalizations. Keywords: Adaptive Optimization, Stochastic Optimization, Submodularity, Partial Observability, Active Learning, Optimal Decision Trees 1. Introduction In many problems arising in artificial intelligence one needs to adaptively make a sequence of decisions, taking into account observations about the outcomes of past decisions. Often these outcomes are uncertain, and one may only know a probability distribution over them. Finding optimal policies for decision making in such partially observable stochastic optimization problems is notoriously intractable (see, e.g., Littman et al. (1998)). A fundamental challenge is to identify classes of planning problems for which simple solutions obtain (near-) optimal performance. In this paper, we introduce the concept of adaptive submodularity, and prove that if a partially observable stochastic optimization problem satisfies this property, a simple adaptive greedy algorithm is guaranteed to obtain near-optimal solutions. In fact, under reasonable complexity-theoretic assumptions, no polynomial time algorithm is able to obtain better solutions in general. Adaptive submodularity generalizes the classical notion of submodularity1 , which has been successfully used to develop approximation algorithms for a variety of non-adaptive optimization problems. Submodularity, informally, is an intuitive notion of diminishing returns, which states that adding an element to a small set helps more than adding that same element to a larger (super-) set. A celebrated result of the work of Nemhauser et al. (1978) guarantees that for such submodular functions, a simple greedy algorithm, which adds the element that maximally increases the objective value, selects a near-optimal set of k elements. Similarly, it is guaranteed to find a set of near-minimal cost that achieves a desired quota of utility (Wolsey, 1982), using near-minimum average time to do so (Streeter and Golovin, 0. This work appeared in the Journal of Artificial Intelligence Research (Golovin and Krause, 2011a), and an eariler extended abstract appeared in the International Conference on Learning Theory (Golovin and Krause, 2010). 1. For an extensive treatment of submodularity, see the books of Fujishige (2005) and Schrijver (2003). c 2012 Daniel Golovin and Andreas Krause. 2008). Besides guaranteeing theoretical performance bounds, submodularity allows us to speed up algorithms without loss of solution quality by using lazy evaluations (Minoux, 1978), often leading to performance improvements of several orders of magnitude (Leskovec et al., 2007). Submodularity has been shown to be very useful in a variety of problems in artificial intelligence (Krause and Guestrin, 2009b). The challenge in generalizing submodularity to adaptive planning — where the action taken in each step depends on information obtained in the previous steps — is that feasible solutions are now policies (decision trees or conditional plans) instead of subsets. We propose a natural generalization of the diminishing returns property for adaptive problems, which reduces to the classical characterization of submodular set functions for deterministic distributions. We show how these results of Nemhauser et al. (1978), Wolsey (1982), Streeter and Golovin (2008), and Minoux (1978) generalize to the adaptive setting. Hence, we demonstrate how adaptive submodular optimization problems enjoy theoretical and practical benefits similar to those of classical, nonadaptive submodular problems. We further demonstrate the usefulness and generality of the concept by showing how it captures known results in stochastic optimization and active learning as special cases, admits tighter performance bounds, leads to natural generalizations and allows us to solve new problems for which no performance guarantees were known. As a first example, consider the problem of deploying (or controlling) a collection of sensors to monitor some spatial phenomenon. Each sensor can cover a region depending on its sensing range. Suppose we would like to find the best subset of k locations to place the sensors. In this application, intuitively, adding a sensor helps more if we have placed few sensors so far and helps less if we have already placed many sensors. We can formalize this diminishing returns property using the notion of submodularity – the total area covered by the sensors is a submodular function defined over all sets of locations. Krause and Guestrin (2007) show that many more realistic utility functions in sensor placement (such as the improvement in prediction accuracy w.r.t. some probabilistic model) are submodular as well. Now consider the following stochastic variant: Instead of deploying a fixed set of sensors, we deploy one sensor at a time. With a certain probability, deployed sensors can fail, and our goal is to maximize the area covered by the functioning sensors. Thus, when deploying the next sensor, we need to take into account which of the sensors we deployed in the past failed. This problem has been studied by Asadpour et al. (2008) for the case where each sensor fails independently at random. In this paper, we show that the coverage objective is adaptive submodular, and use this concept to handle much more general settings (where, e.g., rather than all-or-nothing failures there are different types of sensor failures of varying severity). We also consider a related problem where the goal is to place the minimum number of sensors to achieve the maximum possible sensor coverage (i.e., the coverage obtained by deploying sensors everywhere), or more generally the goal may be to achieve any fixed percentage of the maximum possible sensor coverage. Under the first goal, the problem is equivalent to one studied by Goemans and Vondrák (2006), and generalizes a problem studied by Liu et al. (2008). As another example, consider a viral marketing problem, where we are given a social network, and we want to influence as many people as possible in the network to buy some product. We do that by giving the product for free to a subset of the people, and hope that they convince their friends to buy the product as well. Formally, we have a graph, and each edge e is labeled by a number 0 ≤ pe ≤ 1. We “influence” a subset of nodes in the graph, and for each influenced node, their neighbors get randomly influenced according to the probability annotated on the edge connecting the nodes. This process repeats until no further node gets influenced. Kempe et al. (2003) show that the set function which quantifies the expected number of nodes influenced is submodular. A natural stochastic variant of the problem is where we pick a node, get to see which nodes it influenced, then adaptively pick the next node based on these observations and so on. We show that a large class of such adaptive influence maximization problems satisfies adaptive submodularity. Our third application is in active learning, where we are given an unlabeled data set, and we would like to adaptively pick a small set of examples whose labels imply all other labels. The same problem arises in automated diagnosis, where we have hypotheses about the state of the system (e.g., what illness a patient has), and would like to perform tests to identify the correct hypothesis. In both domains we want to pick examples / tests to shrink the remaining version space (the set of consistent hypotheses) as quickly as possible. Here, we show that the reduction in version space probability mass is adaptive submodular, and use that observation to prove that the adaptive greedy algorithm is a near-optimal querying policy. Our results for active learning 2 Name A.S. Maximization A.S. Min Cost Coverage A.S. Min Sum Cover Data Dependent Bounds Accelerated Greedy Stochastic Submodular Maximization Stochastic Set Cover Adaptive Viral Marketing Active Learning Hardness in the absence of Adapt. Submodularity New Results Tight (1 − 1/e)-approx. for A.M.S. objectives Squared logarithmic approx. for A.M.S. objectives Tight 4-approx. for A.M.S. objectives Generalization to A.M.S. functions Generalization of lazy evaluations to the adaptive setting Generalization of the previous (1 − 1/e)-approx. to arbitrary per–item set distributions, and to item costs (ln(n) + 1)2 -approx. for arbitrary per-item set distributions, with item costs Adaptive analog of previous (1−1/e)-approx. for non-adaptive viral marketing, under more general reward functions; squared logarithmic approx. for the adaptive min cost cover version New analysis for generalized binary search and its approximate versions with and without item costs Ω(|E|1− )-approximation hardness for A.S. Maximization, Min Cost Coverage, and Min-Sum Cover, if f is not adaptive submodular. Location §5.1, page 10 §5.2, page 12 §5.3, page 15 §5.1, page 10 §4, page 9 §6, page 16 §7, page 18 §8, page 20 §9, page 24 §12, page 33 Table 1: Summary of our theoretical results. A.S. is shorthand for “adaptive stochastic”, and A.M.S. is shorthand for “adaptive monotone submodular.” and automated diagnosis are also related to recent results of Guillory and Bilmes (2010, 2011) who study generalizations of submodular set cover to an interactive setting. In contrast to our approach however, Guillory and Bilmes analyze worst-case costs, and use rather different technical definitions and proof techniques. We summarize our main contributions below, and provide a more technical summary in Table 1. At a high level, our main contributions are: • We consider a particular class of partially observable adaptive stochastic optimization problems, which we prove to be hard to approximate in general. • We introduce the concept of adaptive submodularity, and prove that if a problem instance satisfies this property, a simple adaptive greedy policy performs near-optimally, for both adaptive stochastic maximization and coverage, and also a natural min-sum objective. • We show how adaptive submodularity can be exploited by allowing the use of an accelerated adaptive greedy algorithm using lazy evaluations, and how we can obtain data-dependent bounds on the optimum. • We illustrate adaptive submodularity on several realistic problems, including Stochastic Maximum Coverage, Stochastic Submodular Coverage, Adaptive Viral Marketing, and Active Learning. For these applications, adaptive submodularity allows us to recover known results and prove natural generalizations. 1.1 Organization This article is organized as follows. In §2 (page 4) we set up notation and formally define the relevant adaptive optimization problems for general objective functions. For the reader’s convenience, we have also provided a reference table of important symbols on page 58. In §3 (page 6) we review the classical notion of submodularity and introduce the novel adaptive submodularity property. In §4 (page 9) we introduce the adaptive greedy policy, as well as an accelerated variant. In §5 (page 10) we discuss the theoretical guarantees that the adaptive greedy policy enjoys when applied to problems with adaptive submodular objectives. Sections 6 through 9 provide examples on how to apply the adaptive submodular framework to various applications, 3 namely Stochastic Submodular Maximization (§6, page 16), Stochastic Submodular Coverage (§7, page 18), Adaptive Viral Marketing (§8, page 20), and Active Learning (§9, page 24). In §10 (page 29) we report empirical results on two sensor selection problems. In §11 (page 31) we discuss the adaptivity gap of the problems we consider, and in §12 (page 33) we prove hardness results indicating that problems which are not adaptive submodular can be extremely inapproximable under reasonable complexity assumptions. We review related work in §13 (page 34) and provide concluding remarks in §14 (page 36). The Appendix (page 41) gives details of how to incorporate item costs and includes all of the proofs omitted from the main text. 2. Adaptive Stochastic Optimization We start by introducing notation and defining the general class of adaptive optimization problems that we address in this paper. For sake of clarity, we will illustrate our notation using the sensor placement application mentioned in §1. We give examples of other applications in §6, §7, §8, and §9. 2.1 Items and Realizations Let E be a finite set of items (e.g., sensor locations). Each item e ∈ E is in a particular (initially unknown) state from a set O of possible states (describing whether a sensor placed at location e would malfunction or not). We represent the item states using a function φ : E → O, called a realization (of the states of all items in the ground set). Thus, we say that φ(e) is the state of e under realization φ. We use Φ to denote a random realization. We take a Bayesian approach and assume that there is a known prior probability distribution p (φ) := P [Φ = φ] over realizations (e.g., modeling that sensors fail independently with failure probability), so that we can compute posterior distributions2 . We will consider problems where we sequentially pick an item e ∈ E, get to see its state Φ(e), pick the next item, get to see its state, and so on (e.g., place a sensor, see whether it fails, and so on). After each pick, our observations so far can be represented as a partial realization ψ, a function from some subset of E (i.e., the set of items that we already picked) to their states (e.g., ψ encodes where we placed sensors and which of them failed). For notational convenience, we sometimes represent ψ as a relation, so that ψ ⊆ E × O equals {(e, o) : ψ(e) = o}. We use the notation dom(ψ) = {e : ∃o.(e, o) ∈ ψ} to refer to the domain of ψ (i.e., the set of items observed in ψ). A partial realization ψ is consistent with a realization φ if they are equal everywhere in the domain of ψ. In this case we write φ ∼ ψ. If ψ and ψ 0 are both consistent with some φ, and dom(ψ) ⊆ dom(ψ 0 ), we say ψ is a subrealization of ψ 0 . Equivalently, ψ is a subrealization of ψ 0 if and only if, when viewed as relations, ψ ⊆ ψ 0 . Partial realizations are similar to the notion of “belief states” in Partially Observable Markov Decision Problems (POMDPs), as they encode the effect of all actions taken (items selected) and observations made, and determine our posterior belief about the state of the world (i.e., the state of all items e not yet selected, p (φ | ψ) := P [Φ = φ | Φ ∼ ψ]). 2.2 Policies We encode our adaptive strategy for picking items as a policy π, which is a function from a set of partial realizations to E, specifying which item to pick next under a particular set of observations (e.g., π chooses the next sensor location given where we have placed sensors so far, and whether they failed or not). We also allow randomized policies that are functions from a set of partial realizations to distributions on E, though our emphasis will primarily be on deterministic policies. If ψ is not in the domain of π, the policy terminates (stops picking items) upon observation of ψ. We use dom(π) to denote the domain of π. Technically, we require that dom(π) be closed under subrealizations. That is, if ψ 0 ∈ dom(π) and ψ is a subrealization of ψ 0 then ψ ∈ dom(π). We use the notation E(π, φ) to refer to the set of items selected by π under realization φ. Each deterministic policy π can be associated with a decision tree T π in a natural way (see Fig. 1 for an 2. In some situations, we may not have exact knowledge of the prior p (φ). Obtaining algorithms that are robust to incorrect priors remains an interesting source of open problems. We briefly discuss some robustness guarantees of our algorithm in §4 on page 9. 4 Figure 1: Illustration of a policy π, its corresponding decision tree representation, and the decision tree representation of π[2] , the level 2 truncation of π (as defined in §5.1). illustration). Here, we adopt a policy-centric view that admits concise notation, though we find the decision tree view to be valuable conceptually. Since partial realizations are similar to POMDP belief states, our definition of policies is similar to the notion of policies in POMDPs, which are usually defined as functions from belief states to actions. We will further discuss the relationship between the stochastic optimization problems considered in this paper and POMDPs in Section 13. 2.3 Adaptive Stochastic Maximization, Coverage, and Min-Sum Coverage We wish to maximize, subject to some constraints, a utility function f : 2E × OE → R≥0 that depends on which items we pick and which state each item is in (e.g., modeling the total area covered by the working sensors). Based on this notation, the expected utility of a policy π is favg (π) := E [f (E(π, Φ), Φ)] where the expectation is taken with respect to p (φ). The goal of the Adaptive Stochastic Maximization problem is to find a policy π ∗ such that π ∗ ∈ arg max favg (π) subject to |E(π, φ)| ≤ k for all φ, (1) π where k is a budget on how many items can be picked (e.g., we would like to adaptively choose k sensor locations such that the working sensors provide as much information as possible in expectation). Alternatively, we can specify a quota Q of utility that we would like to obtain, and try to find the cheapest policy achieving that quota (e.g., we would like to achieve a certain amount of information, as cheaply as possible in expectation). Formally, we define the average cost cavg (π) of a policy as the expected number of items it picks, so that cavg (π) := E [|E(π, Φ)|]. Our goal is then to find π ∗ ∈ arg min cavg (π) such that f (E(π, φ), φ) ≥ Q for all φ, (2) π i.e., the policy π ∗ that minimizes the expected number of items picked such that under all possible realizations, at least utility Q is achieved. We call Problem 2 the Adaptive Stochastic Minimum Cost Cover problem. We will also consider the problem where we want to minimize the worst-case cost cwc (π) := maxφ |E(π, φ)|. This worst-case cost cwc (π) is the cost incurred under adversarially chosen realizations, or equivalently the depth of the deepest leaf in T π , the decision tree associated with π. Yet another important variant is to minimize the average time required by a policy to obtain its utility. Formally, let u(π, t) be the expected utility obtained by π after t steps3 , let Q = E [fP (E, Φ)] be the maximum ∞ possible expected utility, and define the min-sum cost cΣ (π) of a policy as cΣ (π) := t=0 (Q − u(π, t)). We then define the Adaptive Stochastic Min-Sum Cover problem as the search for π ∗ ∈ arg min cΣ (π) . (3) π P Unfortunately, as we will show in §12, even for linear functions f , i.e., those where f (A, φ) = e∈A we,φ is simply the sum of weights (depending on the realization φ), Problems (1), (2), and (3) are hard to approximate 3. For a more formal definition of u(π, t), see §15.5 on page 56. 5 under reasonable complexity theoretic assumptions. Despite the hardness of the general problems, in the following sections we will identify conditions that are sufficient to allow us to approximately solve them. 2.4 Incorporating Item Costs Instead of quantifying the cost of a set E(π, φ) by the number of elements |E(π, φ)|, we Pcan also consider the case where each item e ∈ E has a cost c(e), and the cost of a set S ⊆ E is c(S) = e∈S c(e). We can then consider variants of Problems (1), (2), and (3) with the |E(π, φ)| replaced by c(E(π, φ)). For clarity of presentation, we will focus on the unit cost case, i.e., c(e) = 1 for all e, and explain how our results generalize to the non-uniform case in the Appendix. 3. Adaptive Submodularity We first review the classical notion of submodular set functions, and then introduce the novel notion of adaptive submodularity. 3.1 Background on Submodularity Let us first consider the very special case where p (φ) is deterministic or, equivalently, |O| = 1 (e.g., in our sensor placement applications, sensors never fail). In this case, the realization φ is known to the decision maker in advance, and thus there is no benefit in adaptive selection. Given the realization φ, Problem (1) is equivalent to finding a set A∗ ⊆ E such that A∗ ∈ arg max f (A, φ) such that |A| ≤ k. (4) A⊆E For most interesting classes of utility functions f , this is an NP-hard optimization problem. However, in many practical problems, such as those mentioned in §1, f (A) = f (A, φ) satisfies submodularity. A set function f : 2E → R is called submodular if, whenever A ⊆ B ⊆ E and e ∈ E \ B it holds that f (A ∪ {e}) − f (A) ≥ f (B ∪ {e}) − f (B), (5) i.e., adding e to the smaller set A increases f by at least as much as adding e to the superset B. Furthermore, f is called monotone, if, whenever A ⊆ B it holds that f (A) ≤ f (B) (e.g., adding a sensor can never reduce the amount of information obtained). A celebrated result by Nemhauser et al. (1978) states that for monotone submodular functions with f (∅) = 0, a simple greedy algorithm that starts with the empty set, A0 = ∅ and chooses Ai+1 = Ai ∪ {arg max f (Ai ∪ {e})} (6) e∈E\Ai guarantees that f (Ak ) ≥ (1 − 1/e) max|A|≤k f (A). Thus, the greedy set Ak obtains at least a (1 − 1/e) fraction of the optimal value achievable using k elements. Furthermore, Feige (1998) shows that this result is tight if P 6= NP; under this assumption no polynomial time algorithm can do strictly better than the greedy algorithm, i.e., achieve a (1 − 1/e + )-approximation for any constant  > 0, even for the special case of Maximum k-Cover where f (A) is the cardinality of the union of sets indexed by A. Similarly, Wolsey (1982) shows that the same greedy algorithm also near-optimally solves the deterministic case of Problem (2), called the Minimum Submodular Cover problem: A∗ ∈ arg min |A| such that f (A) ≥ Q. (7) A⊆E Pick the first set A` constructed by the greedy algorithm such that f (A` ) ≥ Q. Then, for integer-valued submodular functions, ` is at most |A∗ |(1 + log maxe f (e)), i.e., the greedy set is at most a logarithmic factor larger than the smallest set achieving quota Q. For the special case of Set Cover, where f (A) is the 6 cardinality of a union of sets indexed by A, this result matches a lower bound by Feige (1998): Unless NP ⊆ DTIME(nO(log log n) ), Set Cover is hard to approximate by a factor better than (1 − ε) ln Q, where Q is the number of elements to be covered. Now let us relax the assumption that p (φ) is deterministic. In this case, we may still want to find a nonadaptive solution (i.e., a constant policy πA that always picks set A independently of Φ) maximizing favg (πA ). If f is pointwise submodular, i.e., f (A, φ) is submodular in A for any fixed φ, the function f (A) = favg (πA ) is submodular, since nonnegative linear combinations of submodular functions remain submodular. Thus, the greedy algorithm allows us to find a near-optimal non-adaptive policy. That is, in our sensor placement example, if we are willing to commit to all locations before finding out whether the sensors fail or not, the greedy algorithm can provide a good solution to this non-adaptive problem. However, in practice, we may be more interested in obtaining a non-constant policy π, that adaptively chooses items based on previous observations (e.g., takes into account which sensors are working before placing the next sensor). In many settings, selecting items adaptively offers huge advantages, analogous to the advantage of binary search over sequential (linear) search4 . Thus, the question is whether there is a natural extension of submodularity to policies. In the following, we will develop such a notion – adaptive submodularity. 3.2 Adaptive Monotonicity and Submodularity The key challenge is to find appropriate generalizations of monotonicity and of the diminishing returns condition (5). We begin by again considering the very special case where p (φ) is deterministic, so that the policies are non-adaptive. In this case a policy π simply specifies a sequence of items (e1 , e2 , . . . , er ) which it selects in order. Monotonicity in this context can be characterized as the property that “the marginal benefit of selecting an item is always nonnegative,” meaning that for all such sequences (e1 , e2 , . . . , er ), items e and 1 ≤ i ≤ r it holds that f ({ej : j ≤ i} ∪ {e}) − f ({ej : j ≤ i}) ≥ 0. Similarly, submodularity can be viewed as the property that “selecting an item later never increases its marginal benefit,” meaning that for all sequences (e1 , e2 , . . . , er ), items e, and all i ≤ r, f ({ej : j ≤ i} ∪ {e}) − f ({ej : j ≤ i}) ≥ f ({ej : j ≤ r} ∪ {e}) − f ({ej : j ≤ r}). We take these views of monotonicity and submodularity when defining their adaptive analogues, by using an appropriate generalization of the marginal benefit. When moving to the general adaptive setting, the challenge is that the items’ states are now random and only revealed upon selection. A natural approach is thus to condition on observations (i.e., partial realizations of selected items), and take the expectation with respect to the items that we consider selecting. Hence, we define our adaptive monotonicity and submodularity properties in terms of the conditional expected marginal benefit of an item. Definition 1 (Conditional Expected Marginal Benefit) Given a partial realization ψ and an item e, the conditional expected marginal benefit of e conditioned on having observed ψ, denoted ∆(e | ψ), is   ∆(e | ψ) := E f (dom(ψ) ∪ {e} , Φ) − f (dom(ψ), Φ) Φ ∼ ψ (8) where the expectation is computed with respect to p (φ | ψ) = P [Φ = φ | Φ ∼ ψ]. Similarly, the conditional expected marginal benefit of a policy π is   ∆(π | ψ) := E f (dom(ψ) ∪ E(π, Φ), Φ) − f (dom(ψ), Φ) Φ ∼ ψ . (9) In our sensor placement example, ∆(e | ψ) quantifies the expected amount of additional area covered by placing a sensor at location e, in expectation over the posterior distribution pΦ(e) (o) := P [Φ(e) = o | Φ ∼ ψ] of whether the sensor will fail or not, and taking into account the area covered by the placed working 4. We provide a well–known example in active learning that illustrates this phenomenon crisply in §9; see Fig. 4 on page 24. We consider the general question of the magnitude of the potential benefits of adaptivity in §11 on page 31 . 7 sensors as encoded by ψ. Note that the benefit we have accrued upon observing ψ (and hence after having selected the items in dom(ψ)) is E [f (dom(ψ), Φ) | Φ ∼ ψ], which is the benefit term subtracted out in Eq. (8) and Eq. (9). Similarly, the expected total benefit obtained after observing ψ and then selecting e is E [f (dom(ψ) ∪ {e} , Φ) | Φ ∼ ψ]. The corresponding benefit for running π after observing ψ is slightly more complex. Under realization φ ∼ ψ, the final cumulative benefit will be f (dom(ψ) ∪ E(π, φ), φ). Taking the expectation with respect to p (φ | ψ) and subtracting out the benefit already obtained by dom(ψ) then yields the conditional expected marginal benefit of π. We are now ready to introduce our generalizations of monotonicity and submodularity to the adaptive setting: Definition 2 (Adaptive Monotonicity) A function f : 2E × OE → R≥0 is adaptive monotone with respect to distribution p (φ) if the conditional expected marginal benefit of any item is nonnegative, i.e., for all ψ with P [Φ ∼ ψ] > 0 and all e ∈ E we have ∆(e | ψ) ≥ 0. (10) Definition 3 (Adaptive Submodularity) A function f : 2E × OE → R≥0 is adaptive submodular with respect to distribution p (φ) if the conditional expected marginal benefit of any fixed item does not increase as more items are selected and their states are observed. Formally, f is adaptive submodular w.r.t. p (φ) if for all ψ and ψ 0 such that ψ is a subrealization of ψ 0 (i.e., ψ ⊆ ψ 0 ), and for all e ∈ E \ dom(ψ 0 ), we have ∆(e | ψ) ≥ ∆(e | ψ 0 ) . (11) From the decision tree perspective, the condition ∆(e | ψ) ≥ ∆(e | ψ 0 ) amounts to saying that for any decision tree T , if we are at a node v in T which selects an item e, and compare the expected marginal benefit of e selected at v with the expected marginal benefit e would have obtained if it were selected at an ancestor of v in T , then the latter must be no smaller than the former. Note that when comparing the two expected marginal benefits, there is a difference in both the set of items previously selected (i.e., dom(ψ) vs. dom(ψ 0 )) and in the distribution over realizations (i.e., p (φ | ψ) vs. p (φ | ψ 0 )). It is also worth emphasizing that adaptive submodularity is defined relative to the distribution p (φ) over realizations; it is possible that f is adaptive submodular with respect to one distribution, but not with respect to another. We will give concrete examples of adaptive monotone and adaptive submodular functions that arise in the applications introduced in §1 in §6, §7, §8, and §9. In the Appendix, we will explain how the notion of adaptive submodularity can be extended to handle non-uniform costs (since, e.g., the cost of placing a sensor at an easily accessible location may be smaller than at a location that is hard to get to). 3.3 Properties of Adaptive Submodular Functions It can be seen that adaptive monotonicity and adaptive submodularity enjoy similar closure properties as monotone submodular functions. In particular, if wP 1 , . . . , wm ≥ 0 and f1 , . . . , fm are adaptive monotone m submodular w.r.t. distribution p (φ), then f (A, φ) = i=1 wi fi (A, φ) is adaptive monotone submodular w.r.t. p (φ). Similarly, for a fixed constant c ≥ 0 and adaptive monotone submodular function f , the function g(E, φ) = min(f (E, φ), c) is adaptive monotone submodular. Thus, adaptive monotone submodularity is preserved by nonnegative linear combinations and by truncation. Adaptive monotone submodularity is also preserved by restriction, so that if f : 2E × OE → R≥0 is adaptive monotone submodular w.r.t. p (φ), then for any e ∈ E, the function g : 2E\{e} × OE → R≥0 defined by g(A, φ) := f (A, φ) for all A ⊆ E \ {e} and all φ is also adaptive submodular w.r.t. p (φ). Finally, if f : 2E × OE → R≥0 is adaptive monotone submodular w.r.t. p (φ) then for each partial realization ψ the conditional function g(A, φ) := f (A ∪ dom(ψ), φ) is adaptive monotone submodular w.r.t. p (φ | ψ) := P [Φ = φ | Φ ∼ ψ]. 3.4 What Problem Characteristics Suggest Adaptive Submodularity? Adaptive submodularity is a diminishing returns property for policies. Speaking informally, it can be applied in situations where there is an objective function to be optimized does not feature synergies in the benefits 8 of items conditioned on observations. In some cases, the primary objective might not have this property, but a suitably chosen proxy of it does, as is the case with active learning with persistent noise (Golovin et al., 2010; Bellala and Scott, 2010). We give example applications in §6 through §9. It is also worth mentioning where adaptive submodularity is not directly applicable. An extreme example of synergistic effects between items conditioned on observations is the class of “treasure hunting” instances used to prove Theorem 26 on page 33, where the (binary) state of certain groups of items encode the treasure’s location in a complex manner. Another problem feature which adaptive submodularity does not directly address is the possibility that items selection can alter the underlying realization φ, as is the case for the problem of optimizing policies for general POMDPs. 4. The Adaptive Greedy Policy The classical non-adaptive greedy algorithm (6) has a natural generalization to the adaptive setting. The greedy policy π greedy tries, at each iteration, to myopically increase the expected objective value, given its current observations. That is, suppose f : 2E × OE → R≥0 is the objective, and ψ is the partial realization indicating the states of items selected so far. Then the greedy policy will select the item e maximizing the expected increase in value, conditioned on the observed states of items it has already selected (i.e., conditioned on Φ ∼ ψ). That is, it will select e to maximize the conditional expected marginal benefit ∆(e | ψ) as defined in Eq. (8). Pseudocode of the adaptive greedy algorithm is given in Algorithm 1. The only difference to the classic, non-adaptive greedy algorithm studied by Nemhauser et al. (1978), is Line 1, where an observation Φ(e∗ ) of the selected item e∗ is obtained. Note that the algorithms in this section are presented for Adaptive Stochastic Maximization. For the coverage objectives, we simply keep selecting items as prescribed by π greedy until achieving the quota on objective value (for the min-cost objective) or until we have selected every item (for the min-sum objective). 4.1 Incorporating Item Costs The adaptive greedy algorithm can be naturally modified to handle non-uniform item costs by replacing its selection rule by ∆(e | ψ) e∗ ∈ arg max . c(e) e In the following, we will focus on the uniform cost case (c ≡ 1), and defer the analysis with costs to the Appendix. 4.2 Approximate Greedy Selection In some applications, finding an item maximizing ∆(e | ψ) may be computationally intractable, and the best we can do is find an α-approximation to the best greedy selection. This means we find an e0 such that ∆(e0 | ψ) ≥ 1 max ∆(e | ψ) . α e We call a policy which always selects such an item an α-approximate greedy policy. 4.3 Robustness & Approximate Greedy Selection As we will show, α-approximate greedy policies have performance guarantees on several problems. The fact that these performance guarantees of greedy policies are robust to approximate greedy selection suggests a particular robustness guarantee against incorrect priors p (φ). Specifically, if our incorrect prior p0 is such that when we evaluate ∆(e | ψ) we err by a multiplicative factor of at most α, then when we compute the greedy policy with respect to p0 we are actually implementing an α-approximate greedy policy (with respect to the true prior), and hence obtain the corresponding guarantees. For example, a sufficient condition for erring by at most a multiplicative factor of α is that there exists c ≤ 1 and d ≥ 1 with α = d/c such that c p (φ) ≤ p0 (φ) ≤ d p (φ) for all φ, where p is the true prior. 9 1 2 3 4 5 6 Input: Budget k; ground set E; distribution p (φ); function f . Output: Set A ⊆ E of size k begin A ← ∅; ψ ← ∅; for i = 1 to k do foreach e ∈ E \ A do compute ∆(e | ψ) = E [f (A ∪ {e} , Φ) − f (A, Φ) | Φ ∼ ψ] ; Select e∗ ∈ arg maxe ∆(e | ψ); Set A ← A ∪ {e∗ }; Observe Φ(e∗ ); Set ψ ← ψ ∪ {(e∗ , Φ(e∗ ))}; end Algorithm 1: The adaptive greedy algorithm, which implements the greedy policy. 4.4 Lazy Evaluations and the Accelerated Adaptive Greedy Algorithm The definition of adaptive submodularity allows us to implement an “accelerated” version of the adaptive greedy algorithm using lazy evaluations of marginal benefits as originally suggested for the non-adaptive case by Minoux (1978). The idea is as follows. Suppose we run π greedy under some fixed realization φ, and select items e1 , e2 , . . . , ek . Let ψ i := {(ej , φ(ej ) : j ≤ i)} be the partial realizations observed during the run of π greedy . The adaptive greedy algorithm computes ∆(e | ψ i ) for all e ∈ E and 0 ≤ i < k, unless e ∈ dom(ψ i ). Naively, the algorithm thus needs to compute Θ(|E|k) marginal benefits (which can be expensive to compute). The key insight is that i 7→ ∆(e | ψ i ) is nonincreasing for all e ∈ E, because of the adaptive submodularity of the objective. Hence, if when deciding which item to select as ei we know ∆(e0 | ψ j ) ≤ ∆(e | ψ i ) for some items e0 and e and j < i, then we may conclude ∆(e0 | ψ i ) ≤ ∆(e | ψ i ) and hence eliminate the need to compute ∆(e0 | ψ i ). The accelerated version of the adaptive greedy algorithm exploits this observation in a principled manner, by computing ∆(e | ψ) for items e in decreasing order of the upper bounds known on them, until it finds an item whose value is at least as great as the upper bounds of all other items. Pseudocode of this version of the adaptive greedy algorithm is given in Algorithm 2. In the non-adaptive setting, the use of lazy evaluations has been shown to significantly reduce running times in practice (Leskovec et al., 2007). We evaluated the naive and accelerated implementations of the adaptive greedy algorithm on two sensor selection problems, and obtained speedup factors that range from roughly 4 to 40 for those problems. See §10 on page 29 for details. 5. Guarantees for the Greedy Policy In this section we show that if the objective function is adaptive submodular with respect to the probabilistic model of the environment in which we operate, then the greedy policy inherits precisely the performance guarantees of the greedy algorithm for classic (non-adaptive) submodular maximization and submodular coverage problems, such as Maximum k-Cover and Minimum Set Cover, as well as min-sum submodular coverage problems, such as Min-Sum Set Cover. In fact, we will show that this holds true more generally: α–approximate greedy policies inherit precisely the performance guarantees of α–approximate greedy algorithms for these classic problems. These guarantees suggest that adaptive submodularity is an appropriate generalization of submodularity to policies. In this section we focus on the unit cost case (i.e., every item has the same cost). In the Appendix we provide the proofs omitted in this section, and show how our results extend to non-uniform item costs if we greedily maximize the expected benefit/cost ratio. 5.1 The Maximum Coverage Objective In this section we consider the maximum coverage objective, where the goal is to select k items adaptively to maximize their expected value. The task of maximizing expected value subject to more complex constraints, 10 Input: Budget k; ground set E; distribution p (φ); function f . Output: Set A ⊆ E of size k begin 1 A ← ∅; ψ ← ∅; Priority Queue Q ← EMPTY QUEUE; 2 foreach e ∈ E do Q. insert(e, +∞); 3 for i = 1 to k do 4 δmax ← −∞; emax ← NULL; 5 while δmax < Q. maxPriority( ) do 6 e ← Q. pop( ); 7 δ ← ∆(e | ψ) = E [f (A ∪ {e} , Φ) − f (A, Φ) | Φ ∼ ψ]; 8 Q. insert(e, δ); 9 if δmax < δ then 10 δmax ← δ; emax ← e; 11 A ← A ∪ {emax }; Q. remove(emax ); 12 Observe Φ(emax ); Set ψ ← ψ ∪ {(emax , Φ(emax ))}; end Algorithm 2: The accelerated version of the adaptive greedy algorithm. Here, Q. insert(e, δ) inserts e with priority δ, Q. pop( ) removes and returns the item with greatest priority, Q. maxPriority( ) returns the maximum priority of the elements in Q, and Q. remove(e) deletes e from Q. such as matroid constraints and intersections of matroid constraints, is considered in the work of Golovin and Krause (2011b). Before stating our result, we require the following definition. Definition 4 (Policy Truncation) For a policy π, define the level-k-truncation π[k] of π to be the policy obtained by running π until it terminates or until it selects k items, and then terminating. Formally, dom(π[k] ) = {ψ ∈ dom(π) : |ψ| < k}, and π[k] (ψ) = π(ψ) for all ψ ∈ dom(π[k] ). We have the following result, which generalizes the classic result of the work of Nemhauser et al. (1978) that the greedy algorithm achieves a (1 − 1/e)-approximation to the problem of maximizing monotone submodular functions under a cardinality constraint. By setting ` = k and α = 1 in Theorem 5, we see that the greedy policy which selects k items adaptively obtains at least (1 − 1/e) of the value of the optimal policy that selects k items adaptively, measured with respect to favg . For a proof see Theorem 38 in Appendix 15.3, which generalizes Theorem 5 to nonuniform item costs. Theorem 5 Fix any α ≥ 1. If f is adaptive monotone and adaptive submodular with respect to the distribution p (φ), and π is an α-approximate greedy policy, then for all policies π ∗ and positive integers ` and k,   ∗ favg (π[`] ) > 1 − e−`/αk favg (π[k] ).  In particular, with ` = k this implies any α-approximate greedy policy achieves a 1 − e−1/α approximation to the expected reward of the best policy, if both are terminated after running for an equal number of steps. If the greedy rule can be implemented only with small absolute error rather than small relative error, i.e., ∆(e0 | ψ) ≥ maxe ∆(e | ψ) − ε, an argument similar to that used to prove Theorem 5 shows that   ∗ favg (π[`] ) ≥ 1 − e−`/k favg (π[k] ) − `ε. This is important, since small absolute error can always be achieved (with high probability) whenever f can be evaluated efficiently, and sampling p (φ | ψ) is efficient. In this case, we can approximate N  1 X f (dom(ψ) ∪ {e} , φi ) − f (dom(ψ), φi ) , ∆(e | ψ) ≈ N i=1 where φi are sampled i.i.d. from p (φ | ψ). 11 5.1.1 DATA D EPENDENT B OUNDS For the maximum coverage objective, adaptive submodular functions have another attractive feature: they allow us to obtain data dependent bounds on the optimum, in a manner similar to the bounds for the non-adaptive case (Minoux, 1978). Consider the non-adaptive problem of maximizing a monotone submodular function f : 2A → R≥0 subject to the constraint |A| ≤ k. Let A∗ be an optimal solution, and fix any A ⊆ E. Then X f (A∗ ) ≤ f (A) + max (f (A ∪ {e}) − f (A)) (12) B:|B|≤k ∗ e∈B ∗ P because setting B = A we have f (A ) ≤ f (A ∪ B) ≤ f (A) P + e∈B (f (A ∪ {e}) − f (A)). Note that unlike the original objective, we can easily compute maxB:|B|≤k e∈B (f (A ∪ {e}) − f (A)) by computing δ(e) := f (A ∪ {e}) − f (A) for each e, and summing the k largest values. Hence we can quickly compute an upper bound on our distance from the optimal value, f (A∗ ) − f (A). In practice, such data-dependent bounds can be much tighter than the problem-independent performance guarantees of Nemhauser et al. (1978) for the greedy algorithm (Leskovec et al., 2007). Further note that these bounds hold for any set A, not just sets selected by the greedy algorithm. These data dependent bounds have the following analogue for adaptive monotone submodular functions. See Appendix 15.2 for a proof. Lemma 6 (The Adaptive Data Dependent Bound) Suppose we have made observations ψ after selecting dom(ψ). Let π ∗ be any policy such that |E(π ∗ , φ)| ≤ k for all φ. Then for adaptive monotone submodular f X ∆(π ∗ | ψ) ≤ max ∆(e | ψ) . (13) A⊆E,|A|≤k e∈A Thus, after running any policy π, we can efficiently compute a bound on the additional benefit that the optimal solution π ∗ could obtain beyond the reward of π. We do that by computing the conditional expected marginal benefits for all elements e, and summing the k largest of them. Note that these bounds can be computed on the fly when running the greedy algorithm, in a similar manner as discussed by Leskovec et al. (2007) for the non-adaptive setting. 5.2 The Min Cost Cover Objective Another natural objective is to minimize the number of items selected while ensuring that a sufficient level of value is obtained. This leads to the Adaptive Stochastic Minimum Cost Coverage problem described in §2, namely π ∗ ∈ arg minπ cavg (π) such that f (E(π, φ), φ) ≥ Q for all φ. Recall that cavg (π) is the expected cost of π, which in the unit cost case equals the expected number of items selected by π, i.e., cavg (π) := E [|E(π, Φ)|]. If the objective is adaptive monotone submodular, this is an adaptive version of the Minimum Submodular Cover problem (described on line (7) in §3.1). Recall that the greedy algorithm is known to give a (ln(Q) + 1)-approximation for Minimum Submodular Cover assuming the coverage function is integer-valued in addition to being monotone submodular (Wolsey, 1982). Adaptive Stochastic Minimum Cost Coverage is also related to the (Noisy) Interactive Submodular Set Cover problem studied by Guillory and Bilmes (2010, 2011), which considers the worst-case setting (i.e., there is no distribution over states; instead states are realized in an adversarial manner). Similar results for active learning have been proved by Kosaraju et al. (1999) and Dasgupta (2004), as we discuss in more detail in §9. We assume throughout this section that there exists a quality threshold Q such that f (E, φ) = Q for all φ, and for all S ⊆ E and all φ, f (S, φ) ≤ Q. Note that, as discussed in Section 3, if we replace f (S, φ) by a new function g(S, φ) = min(f (S, φ), Q0 ) for some constant Q0 , g will be adaptive submodular if f is. Thus, if f (E, φ) varies across realizations, we can instead use the greedy algorithm on the function truncated at some threshold Q0 ≤ minφ f (E, φ) achievable by all realizations. In contrast to Adaptive Stochastic Maximization, for the coverage problem additional subtleties arise. In particular, it is not enough that a policy π achieves value Q for the true realization; in order for π to terminate, it also requires a proof of this fact. Formally, we require that π covers f : 12 Definition 7 (Coverage) Let ψ = ψ(π, φ) be the partial realization encoding all states observed during the execution of π under realization φ. Given f : 2E × OE → R, we say a policy π covers φ with respect to f if f (dom(ψ), φ0 ) = f (E, φ0 ) for all φ0 ∼ ψ. We say that π covers f if it covers every realization with respect to f. Coverage is defined in such a way that upon terminating, π might not know which realization is the true one, but has guaranteed that it has achieved the maximum reward in every possible case (i.e., for every realization consistent with its observations). We obtain results for both the average and worst-case cost objectives. 5.2.1 M INIMIZING THE AVERAGE C OST Before presenting our approximation guarantee for the Adaptive Stochastic Minimum Cost Coverage, we introduce a special class of instances, called self–certifying instances. We make this distinction because the greedy policy has stronger performance guarantees for self–certifying instances, and such instances arise naturally in applications. For example, the Stochastic Submodular Cover and Stochastic Set Cover instances in §7, the Adaptive Viral Marketing instances in §8, and the Pool-Based Active Learning instances in §9 are all self–certifying. Definition 8 (Self–Certifying Instances) An instance of Adaptive Stochastic Minimum Cost Coverage is self–certifying if whenever a policy achieves the maximum possible value for the true realization it immediately has a proof of this fact. Formally, an instance (f, p (φ)) is self–certifying if for all φ, φ0 , and ψ such that φ ∼ ψ and φ0 ∼ ψ, we have f (dom(ψ), φ) = f (E, φ) if and only if f (dom(ψ), φ0 ) = f (E, φ0 ). One class of self–certifying instances which commonly arise are those in which f (A, φ) depends only on the state of items in A, and in which there is a uniform maximum amount of reward that can be obtained across realizations. Formally, we have the following observation. Proposition 9 Fix an instance (f, p (φ)). If there exists Q such that f (E, φ) = Q for all φ and there exists some g : 2E×O → R≥0 such that f (A, φ) = g ({(e, φ(e)) : e ∈ A}) for all A and φ, then (f, p (φ)) is self–certifying. Proof Fix φ, φ0 , and ψ such that φ ∼ ψ and φ0 ∼ ψ. Assuming the existence of g and treating ψ as a relation, we have f (dom(ψ), φ) = g(ψ) = f (dom(ψ), φ0 ). Hence f (dom(ψ), φ) = Q = f (E, φ) if and only if f (dom(ψ), φ0 ) = Q = f (E, φ0 ). For our results on minimum cost coverage, we also need a stronger monotonicity condition and a stronger submodularity condition: Definition 10 (Strong Adaptive Monotonicity) A function f : 2E × OE → R is strongly adaptive monotone with respect to p (φ) if, informally “selecting more items never hurts” with respect to the expected reward. Formally, for all ψ, all e ∈ / dom(ψ), and all possible outcomes o ∈ O such that P [Φ(e) = o | Φ ∼ ψ] > 0, we require E [f (dom(ψ), Φ) | Φ ∼ ψ] ≤ E [f (dom(ψ) ∪ {e} , Φ) | Φ ∼ ψ, Φ(e) = o] . (14) Strong adaptive monotonicity implies adaptive monotonicity, as the latter means that “selecting more items never hurts in expectation,” i.e., E [f (dom(ψ), Φ) | Φ ∼ ψ] ≤ E [f (dom(ψ) ∪ {e} , Φ) | Φ ∼ ψ] . To define strong adaptive submodularity, we first need the following extension of ∆(e | ψ): Definition 11 (Conditional Expected Marginal Benefit (Extended version)) Given partial realizations ψ ⊆ ψ 0 , let   ∆(e | ψ; ψ 0 ) := E f (dom(ψ) ∪ {e} , Φ) − f (dom(ψ), Φ) Φ ∼ ψ 0 (15) 13 Definition 12 (Strong Adaptive Submodularity) A function f : 2E × OE → R≥0 is strongly adaptive submodular with respect to distribution p (φ) if it is adaptive submodular and moreover the expected marginal benefit of any fixed item does not increase as more items are selected and their states are observed, conditioned on the (item, observation) pairs. Formally, f is adaptive submodular w.r.t. p (φ) if for all ψ and ψ 0 such that ψ is a subrealization of ψ 0 (i.e., ψ ⊆ ψ 0 ), and for all e ∈ E \ dom(ψ 0 ), we have ∆(e | ψ; ψ 0 ) ≥ ∆(e | ψ 0 ) . (16) In other words, conditioning on ψ 0 , adding items dom(ψ 0 ) \ dom(ψ) cannot increase the expected marginal benefit of e. A sufficient condition for strong adaptive submodularity with respect to p (φ) is that the function be adaptive submodular and pointwise submodular (i.e., f (A, φ) is submodular in A for any fixed φ), as we prove in Appendix 15.4. It is worth noting that pointwise submodularity is not sufficient to establish adaptive submodularity. A simple counterexample is f (S, φ) = | {e : e ∈ S, φ(e) = 1} |, with p (φ) = 1/2 if ∀e ∈ E, φ(e) = 1 and p (φ) = 1/2 if ∀e ∈ E, φ(e) = 0. In that case, ∆(e | ∅) = 1/2 yet ∆(e | {(e0 , 1)}) = 1 for any e0 6= e. We now state our main result for the average case cost cavg (π): Theorem 13 Suppose f : 2E × OE → R≥0 is strongly adaptive submodular and strongly adaptive monotone with respect to p (φ) and there exists Q such that f (E, φ) = Q for all φ. Let η be any value such that f (S, φ) > Q − η implies f (S, φ) = Q for all S and φ. Let δ = minφ p (φ) be the minimum probability of any ∗ realization. Let πavg be an optimal policy minimizing the expected number of items selected to guarantee every realization is covered. Let π be an α-approximate greedy policy with respect to the item costs. Then in general    2 Q ∗ cavg (π) ≤ α cavg (πavg ) ln +1 δη and for self–certifying instances    2 Q ∗ cavg (π) ≤ α cavg (πavg ) ln +1 . η Note that if range(f ) ⊂ Z, then η = 1 is a valid choice, so for general and self–certifying instances we have 2 2 ∗ ∗ cavg (π) ≤ α cavg (πavg ) (ln(Q/δ) + 1) and cavg (π) ≤ α cavg (πavg ) (ln(Q) + 1) , respectively. Historical Note: An earlier version of Theorem 13 claimed logarithmic approximation factors rather than the squared–logarithmic factors present here. Unfortunately, the proof was flawed as pointed out by Nan and Saligrama (2017). Determining whether the logarithmic bounds   an interesting open  hold  remains Q ∗ problem. In particular, it remains open whether cavg (π) ≤ α cavg (πavg ) ln δη + 1 for general instances     ∗ and cavg (π) ≤ α cavg (πavg ) ln Q + 1 for self–certifying instances under the conditions specified by η Theorem 13. It also remains open whether the strong adaptive submodularity condition is required. 5.2.2 M INIMIZING THE W ORST-C ASE C OST For the worst-case cost cwc (π) := maxφ |E(π, φ)|, strong adaptive monotonicity and strong submodularity are not required; adaptive monotonicity and adaptive submodularity suffice. We obtain the following result. Theorem 14 Suppose f : 2E × OE → R≥0 is adaptive monotone and adaptive submodular with respect to p (φ), and let η be any value such that f (S, φ) > f (E, φ) − η implies f (S, φ) = f (E, φ) for all S and φ. Let ∗ δ = minφ p (φ) be the minimum probability of any realization. Let πwc be the optimal policy minimizing the worst-case number of queries to guarantee every realization is covered. Let π be an α-approximate greedy policy. Finally, let Q := E [f (E, φ)] be the maximum possible expected reward. Then     Q ∗ cwc (π) ≤ α cwc (πwc ) ln +1 . δη 14 The proofs of Theorems 13 and 14 are given in Appendix 15.4. Thus, even though adaptive submodularity is defined w.r.t. a particular distribution, perhaps surprisingly, the adaptive greedy algorithm is competitive even in the case of adversarially chosen realizations, against a policy optimized to minimize the worst-case cost. Theorem 14 therefore suggests that if we do not have a strong prior, we can obtain the strongest guarantees if we choose a distribution that is “as uniform as possible” (i.e., maximizes δ) while still guaranteeing adaptive submodularity. 5.2.3 D ISCUSSION Note that the approximation factor for self–certifying instances in Theorem 14 reduces to the (ln(Q) + 1)approximation guarantee for the greedy algorithm for Set Cover instances with Q elements, in the case of a deterministic distribution p (φ). Moreover, with a deterministic distribution p (φ) there is no distinction between average-case and worst-case cost. Hence, an immediate corollary of the result of Feige (1998) mentioned in §3 is that for every constant  > 0 there is no polynomial time (1 − ) ln (Q/η) approximation algorithm for self–certifying instances of Adaptive Stochastic Min Cost Cover, under either the cavg (·) or the cwc (·) objective, unless NP ⊆ DTIME(nO(log log n) ). It remains open to determine whether or not Adaptive Stochastic Min Cost Cover with the worst-case cost objective admits a ln (Q/η) + 1 approximation for self–certifying instances via a polynomial time algorithm, and in particular whether the greedy policy has such an approximation guarantee. However, in Lemma 50 we show that Feige’s result also implies there is no (1 − ) ln (Q/δη) polynomial time approximation algorithm for general (non self-certifying) instances of Adaptive Stochastic Min Cost Cover under either objective, unless NP ⊆ DTIME(nO(log log n) ). In that sense, Theorem 14 is best-possible and Theorem 13 cannot be improved by more than a logarithmic factor and under reasonable complexity-theoretic assumptions. 5.3 The Min-Sum Cover Objective Yet another natural objective is the min-sum objective, in which an unrealized reward of x incurs a cost of x in each time step, and the goal is to minimize the total cost incurred. 5.3.1 BACKGROUND ON THE N ON - ADAPTIVE M IN -S UM C OVER P ROBLEM In the non-adaptive setting, perhaps the simplest form of a coverage problem with this objective is the Min-Sum Set Cover problem (Feige et al., 2004) in which the input is a set system (U, S), the output is a permutation of the sets hS1 , S2 , . . . , Sm i, and the goal is to minimize the sum of element coverage times, where the coverage time of u is the index of the first set that contains it (e.g., it is j if u ∈ Sj and u ∈ / Si for all i < j). In this problem and its generalizations the min-sum objective is useful in modeling processing costs in certain applications, for example in ordering diagnostic tests to identify a disease cheaply (Kaplan et al., 2005), in ordering multiple filters to be applied to database records while processing a query (Munagala et al., 2005), or in ordering multiple heuristics to run on boolean satisfiability instances as a means to solve them faster in practice (Streeter and Golovin, 2008). A particularly expressive generalization of min-sum set cover has been studied under the names Min-Sum Submodular Cover (Streeter and Golovin, 2008) and L1 -Submodular Set Cover (Golovin et al., 2008). The former paper extends the greedy algorithm to a natural online variant of the problem, while the latter studies a parameterized family of Lp -Submodular Set Cover problems in which the objective is analogous to minimizing the Lp norm of the coverage times for Min-Sum Set Cover instances. In the Min-Sum Submodular Cover problem, there is a monotone submodular function f : 2E → R≥0 defining the reward obtained from a collection of elements5 . There is an integral cost c(e) for each element, and the output is a sequence of all of the elements σ = he1 , e2 , . . . , en i. For each t ∈ R≥0 , we define the set of 5. To encode Min-Sum Set Cover instance (U, S), let E := S and f (A) := | ∪e∈A e|, where each e ∈ E is a subset of elements in U . 15 elements in the sequence σ within a budget of t:     X σ[t] := ei : c(ej ) ≤ t .   j≤i The cost we wish to minimize is then cΣ (σ) := ∞ X  f (E) − f (σ[t] ) . (17) t=0 Feige et al. (2004) proved that for Min-Sum Set cover, the greedy algorithm achieves a 4-approximation to the minimum cost, and also that this is optimal in the sense that no polynomial time algorithm can achieve a (4 − )-approximation, for any  > 0, unless P = NP. Interestingly, the greedy algorithm also achieves a 4-approximation for the more general Min-Sum Submodular Cover problem as well (Streeter and Golovin, 2008; Golovin et al., 2008). 5.3.2 T HE A DAPTIVE S TOCHASTIC M IN -S UM C OVER P ROBLEM In this article, we extend the result of Streeter and Golovin (2008) and Golovin et al. (2008) to an adaptive version of Min-Sum Submodular Cover. For clarity’s sake we will consider the unit-cost case here (i.e., c(e) = 1 for all e); we show how to extend adaptive submodularity to handle general costs in the Appendix. In the adaptive version of the problem, π[t] plays the role of σ[t] , and favg plays the role of f . The goal is to find a policy π minimizing cΣ (π) := ∞ X t=0 ∞ X  X  E [f (E, Φ)] − favg (π[t] ) = p (φ) f (E, φ) − f (E(π[t] , φ), φ) . φ (18) t=0 We call this problem the Adaptive Stochastic Min-Sum Cover problem. The key difference between this objective and the minimum cost cover objective is that here, the cost at each step is only the fractional extent that we have not covered the true realization, whereas in the minimum cost cover objective we are charged in full in each step until we have completely covered the true realization (according to Definition 7). We prove the following result for the Adaptive Stochastic Min-Sum Cover problem with arbitrary item costs in Appendix 15.5. Theorem 15 Fix any α ≥ 1. If f is adaptive monotone and adaptive submodular with respect to the distribution p (φ), π is an α-approximate greedy policy with respect to the item costs, and π ∗ is any policy, then cΣ (π) ≤ 4α cΣ (π ∗ ). 6. Application: Stochastic Submodular Maximization As our first application, consider the sensor placement problem introduced in §1. Suppose we would like to monitor a spatial phenomenon such as temperature in a building. We discretize the environment into a set E of locations. We would like to pick a subset A ⊆ E of k locations that is most “informative”, where we use a set function fˆ(A) to quantify the informativeness of placement A. Krause and Guestrin (2007) show that many natural objective functions (such as reduction in predictive uncertainty measured in terms of Shannon entropy with conditionally independent observations) are monotone submodular. Now consider the problem, where the informativeness of a sensor is unknown before deployment (e.g., when deploying cameras for surveillance, the location of objects and their associated occlusions may not be known in advance, or varying amounts of noise may reduce the sensing range). We can model this extension by assigning a state φ(e) ∈ O to each possible location, indicating the extent to which a sensor placed at location e is working. To quantify the value of a set of sensor deployments under a realization φ indicating to what 16 extent the various sensors are working, we first define (e, o) for each e ∈ E and o ∈ O, which represents the placement of a sensor at location e which is in state o. We then suppose there is a function fˆ : 2E×O → R≥0 which quantifies the informativeness of a set of sensor deployments in arbitrary states. (Note fˆ is a set function taking a set of (sensor deployment, state) pairs as input.) The utility f (A, φ) of placing sensors at the locations in A under realization φ is then f (A, φ) := fˆ({(e, φ(e)) : e ∈ A}). We aim to adaptively place k sensors to maximize our expected Q utility. We assume that sensor failures at each location are independent of each other, i.e., P [Φ = φ] = e∈E P [Φ(e) = φ(e)] , where P [φ(e) = o] is the probability that a sensor placed at location e will be in state o. Asadpour et al. (2008) studied a special case of our problem, in which sensors either fail completely (in which case they contribute no value at all) or work perfectly, under the name Stochastic Submodular Maximization. They proved that the adaptive greedy algorithm obtains a (1−1/e) approximation to the optimal adaptive policy, provided fˆ is monotone submodular. We extend their result to multiple types of failures by showing that f (A, φ) is adaptive submodular with respect to distribution p (φ) and then invoking Theorem 5. Fig. 2 illustrates an instance of Stochastic Submodular Maximization where f (A, φ) is the cardinality of union of sets index by A and parameterized by φ. Q Theorem 16 Fix a prior such that P [Φ = φ] = e∈E P [Φ(e) = φ(e)] and an integer k, and let the objective function fˆ : 2E×O → R≥0 be monotone submodular. Let π be any α-approximate greedy policy attempting to maximize f , and let π ∗ be any policy. Then for all positive integers `,   ∗ favg (π[`] ) ≥ 1 − e−`/αk favg (π[k] ). In particular, if π is the greedy policy (i.e., α = 1) and ` = k, then favg (π[k] ) ≥ 1 − 1 e  ∗ favg (π[k] ). Proof We prove Theorem 16 by first proving f is adaptive monotone and adaptive submodular in this model, and then applying Theorem 5. Adaptive monotonicity is readily proved after observing that f (·, φ) is monotone for each φ. Moving on to adaptive submodularity, fix any ψ, ψ 0 such that ψ ⊆ ψ 0 and any e∈ / dom(ψ 0 ). We aim to show ∆(e | ψ 0 ) ≤ ∆(e | ψ). Intuitively, this is clear, as ∆(e | ψ 0 ) is the expected marginal benefit of adding e to a larger base set than is the case with ∆(e | ψ), namely dom(ψ 0 ) as compared to dom(ψ), and the realizations are independent. To prove it rigorously, we define a coupled distribution µ over pairsQ of realizations φ ∼ ψ and φ0 ∼ ψ 0 such that φ(e0 ) = φ0 (e0 ) for all e0 ∈ / dom(ψ 0 ). Formally, 0 0 0 0 0 0 µ(φ, φ ) = e∈E\dom(ψ) P [Φ(e) = φ(e)] if φ ∼ ψ, φ ∼ ψ , and φ(e ) = φ (e ) for all e0 ∈ / dom(ψ 0 ); 0 0 0 0 0 0 otherwise µ(φ, φ ) = 0. (Note that µ(φ, φ ) > 0 implies φ(e )P= φ (e ) for all e ∈ dom(ψ) asPwell, since φ ∼ ψ, φ0 ∼ ψ 0 , and ψ ⊆ ψ 0 .) Also note that p (φ | ψ) = φ0 µ(φ, φ0 ) and p (φ0 | ψ 0 ) = φ µ(φ, φ0 ). Calculating ∆(e | ψ 0 ) and ∆(e | ψ) using µ, we see that for any (φ, φ0 ) in the support of µ, f (dom(ψ 0 ) ∪ {e} , φ0 ) − f (dom(ψ 0 ), φ0 ) = fˆ(ψ 0 ∪ {(e, φ0 (e))}) − fˆ(ψ 0 )) ≤ fˆ(ψ ∪ {(e, φ(e))}) − fˆ(ψ)) = f (dom(ψ) ∪ {e} , φ) − f (dom(ψ), φ) from the submodularity of fˆ. Hence P 0 0 0 0 0 ∆(e | ψ 0 ) = (φ,φ0 ) µ(φ, φ ) (f (dom(ψ ) ∪ {e} , φ ) − f (dom(ψ ), φ )) P 0 ≤ = (φ,φ0 ) µ(φ, φ ) (f (dom(ψ) ∪ {e} , φ) − f (dom(ψ), φ)) which completes the proof. 17 ∆(e | ψ) Figure 2: Illustration of part of a Stochastic Set Cover instance. Shown are the supports of two distributions over sets, indexed by items e (marked in blue) and e0 (yellow). 7. Application: Stochastic Submodular Coverage Suppose that instead of wishing to adaptively place k unreliable sensors to maximize the utility of the information obtained, as discussed in §6, we have a quota on utility and wish to adaptively place the minimum number of unreliable sensors to achieve this quota. This amounts to a minimum-cost coverage version of the Stochastic Submodular Maximization problem introduced in §6, which we call Stochastic Submodular Coverage. As in §6, in the Stochastic Submodular Coverage problem we suppose there is a function fˆ : 2E×O → R≥0 which quantifies the utility Q of a set of sensors in arbitrary states. Also, the states of each sensor are independent, so that P [Φ = φ] = e∈E P [Φ(e) = φ(e)]. n The goal is to obtain a quota o Q of utility at minimum cost. Thus, ˆ we define our objective as f (A, φ) := min Q, f ({(e, φ(e)) : e ∈ A}) , and want to find a policy π covering every realization and minimizing cavg (π) := E [|E(π, Φ)|]. We additionally assume that this quota can always be obtained using sufficiently many sensor placements; formally, this amounts to f (E, φ) = Q for all φ. We obtain the following result, whose proof we defer until the end of this section. Q Theorem 17 Fix a prior with independent sensor states s.t. P [Φ = φ] = e∈EP [Φ(e) = φ(e)], and let fˆ : 2E×O → R≥0 be a mon. submodular function. Fix Q ∈ R≥0 s.t. f (A, φ) := min Q, fˆ({(e, φ(e)) : e ∈ A}) satisfies f (E, φ) = Q for all φ. Let η be any value such that f (S, φ) > Q − η implies f (S, φ) = Q for all S and φ. Finally, let π be an α-approximate greedy policy for maximizing f , and let π ∗ be any policy. Then    2 Q cavg (π) ≤ α cavg (π ) ln +1 . η ∗ 7.1 A Special Case: The Stochastic Set Coverage Problem The Stochastic Submodular Coverage problem is a generalization of the Stochastic Set Coverage problem (Goemans and Vondrák, 2006). In Stochastic Set Coverage the underlying submodular objective fˆ is the number of elements covered in some input set system. In other words, there is a ground set U of n elements to be covered, and items E such that each item e is associated with a distribution over subsets of U . When an item is selected, a set is sampled from its distribution, as illustrated in Fig. 2. The problem is to adaptively select items until all elements of U are covered by sampled sets, while minimizing the expected number of items selected. Like us, Goemans and Vondrák also assume that the subsets are sampled independently for each item, and every element of U can be covered in every realization, so that f (E, φ) = |U | for all φ. Goemans and Vondrák primarily investigated the adaptivity gap (quantifying how much adaptive policies can outperform non-adaptive policies) of Stochastic Set Coverage, for variants in which items can be repeatedly 18 selected or not, and prove adaptivity gaps of Θ(log n) in the former case, and between Ω(n) and O(n2 ) in the latter. They also provide an n-approximation algorithm. More recently, Liu et al. (2008) considered a special case of Stochastic Set Coverage in which each item may be in one of two states. They were motivated by a streaming database problem, in which a collection of queries sharing common filters must all be evaluated on a stream element. They transform the problem to a Stochastic Set Coverage instance in which (filter, query) pairs are to be covered by filter evaluations; which pairs are covered by a filter depends on the (binary) outcome of evaluating it on the stream element. The resulting instances satisfy the assumption that every element of U can be covered in every realization. They study, among other algorithms, the adaptive greedy algorithm specialized toQ this setting, and show that if the subsetsPare sampled independently for n each item, so that P [Φ = φ] = e P [Φ(e) = φ(e)], then it is an Hn := x=1 x1 approximation. (Recall ln(n) ≤ Hn ≤ ln(n) + 1 for all n ≥ 1.) Moreover, Liu et al. report that it empirically outperforms a number of other algorithms in their experiments. The adaptive submodularity framework allows us to prove approximate results for richer item distributions over subsets of U than considered by Liu et al. (2008) as a corollary of Theorem 17. Specifically, we obtain a (ln(n) + 1)2 -approximation for the Stochastic Set Coverage problem with arbitrarily many outcomes for each stochastic set, where n := |U |. We model the Stochastic Set Coverage problem by letting φ(e) ⊆ U indicateQ the random set sampled from e’s distribution. Since the sampled sets are independent we have P [Φ = φ] = e P [Φ(e) = φ(e)]. For any A ⊆ E let f (A, φ) := | ∪e∈A φ(e)| be the number of elements of U covered by the sets sampled from items in A. As in the previous work mentioned above, we assume f (E, φ) = n for all φ. Therefore we may set Q = n. Since the range of f includes only integers, we may set η = 1. Applying Theorem 17 then yields the following result. Corollary 18 The adaptive greedy algorithm achieves a (ln(n) + 1)2 -approximation for Stochastic Set Coverage, where n := |U | is the size of the ground set. We now provide the proof of Theorem 17. Proof of Theorem 17: We will ultimately prove Theorem 17 by applying the bound from Theorem 13 for Stochastic Submodular Cover instances. The proof mostly consists of justifying this application. Without loss of generality we may assume fˆ is n o ˆ truncated at Q, otherwise we may use ĝ(S) = min Q, f (S) in lieu of fˆ. This removes the need to truncate f . Since we established the adaptive submodularity of f in the proof of Theorem 16, and by assumption f (E, φ) = Q for all φ, to apply Theorem 13 we need only show that f is strongly adaptive monotone and strongly adaptive submodular and that the instances under consideration are self–certifying. We begin by showing the strong adaptive monotonicity of f . Fix a partial realization ψ, an item e ∈ / dom(ψ) and a state o. Let ψ 0 = ψ ∪ {(e, o)}. Then treating ψ and ψ 0 as subsets of E × O, and using the monotonicity of fˆ, we obtain E [f (dom(ψ), Φ) | Φ ∼ ψ] = fˆ(ψ) ≤ fˆ(ψ 0 ) ≤ E [f (dom(ψ 0 ), Φ) | Φ ∼ ψ 0 ] , which is equivalent to the strong adaptive monotonicity condition. Next we show the strong adaptive adaptive submodularity of f by showing it is pointwise submodular (having already proven adaptive submodularity for it). This is clearly true, since for all φ, S 7→ fˆ({(e, φ(e)) : e ∈ S}) is monotone submodular by assumption. Finally we prove that these instances are self–certifying. Consider any ψ and φ, φ0 consistent with ψ. Then f (dom(ψ), φ) = fˆ(ψ) = f (dom(ψ), φ0 ). Since f (E, φ) = f (E, φ0 ) = Q by assumption, it follows that f (dom(ψ), φ) = f (E, φ) iff f (dom(ψ), φ0 ) = f (E, φ0 ), so the instance is self–certifying. We have shown that f and p (φ) satisfy the assumptions of Theorem 13 on this self–certifying instance. Hence we may apply it to obtain the claimed approximation guarantee.  19 Figure 3: Illustration of the Adaptive Viral Marketing problem. Left: the underlying social network. Middle: the people influenced and the observations obtained after one person is selected. 8. Application: Adaptive Viral Marketing For our next application, consider the following scenario. Suppose we would like to generate demand for a genuinely novel product. Potential customers do not realize how valuable the new product will be to them, and conventional advertisements are failing to convince them to try it. In this case, we may try to spur demand by offering a special promotional deal to a select few people, and hope that demand builds virally, propagating through the social network as people recommend the product to their friends and associates. Supposing we know something about the structure of the social networks people inhabit, and how ideas, innovation, and new product adoption diffuse through them, this begs the question: to which initial set of people should we offer the promotional deal, in order to spur maximum demand for our product? This, broadly, is the viral marketing problem. The same problem arises in the context of spreading technological, cultural, and intellectual innovations, broadly construed. In the interest of unified terminology we follow Kempe et al. (2003) and talk of spreading influence through the social network, where we say people are active if they have adopted the idea or innovation in question, and inactive otherwise, and that a influences b if a convinces b to adopt the idea or innovation in question. There are many ways to model the diffusion dynamics governing the spread of influence in a social network. We consider a basic and well-studied model, the independent cascade model, described in detail below. For this model Kempe et al. (2003) obtain a very interesting result; they show that the eventual spread of the influence f (i.e., the ultimate number of customers that demand the product) is a monotone submodular function of the seed set S of people initially selected. This, in conjunction with the results of Nemhauser et al. (1978) implies  that the greedy algorithm obtains at least 1 − 1e of the value of the best feasible seed set of size at most k, i.e., arg maxS:|S|≤k f (S), where we interpret k as the budget for the promotional campaign. Though Kempe et al. consider only the maximum coverage version of the viral marketing problem, their result in conjunction with that of Wolsey (1982) also implies that the greedy algorithm will obtain a quota Q of value at a cost of at most ln(Q) + 1 times the cost of the optimal set arg minS {c(S) : f (S) ≥ Q} if f takes on only integral values. 8.1 Adaptive Viral Marketing The viral marketing problem has a very natural adaptive analog. Instead of selecting a fixed set of people in advance, we may select a person to offer the promotion to, make some observations about the resulting spread of demand for our product, and repeat. See Fig. 3 for an illustration. In §8.2, we use the idea of 20 adaptive submodularity to obtain results analogous to those of Kempe et al. (2003) in the adaptive setting.  Specifically, we show that the greedy policy obtains at least 1 − 1e of the value of the best policy. Moreover, we extend this result by achieving that guarantee not only for the case where our reward is simply the number of influenced people, but also for any (nonnegative) monotone submodular function of the set of people influenced. In §8.3 we consider the minimum cost cover objective, and show that the greedy policy obtains a squared logarithmic approximation for it. To our knowledge, no approximation results for this adaptive variant of the viral marketing problem have been known. 8.1.1 I NDEPENDENT C ASCADE M ODEL In this model, the social network is a directed graph G = (V, A) where each vertex in V is a person, and each edge (u, v) ∈ A has an associated binary random variable Xuv indicating if u will influence v. That is, Xuv = 1 if u will influence v once it has been influenced, and Xuv = 0 otherwise. The random variables Xuv are independent, and have known means puv := E [Xuv ]. We will call an edge (u, v) with Xuv = 1 a live edge and an edge with Xuv = 0 a dead edge. When a node u is activated, the edges Xuv to each neighbor v of u are sampled, and v is activated if (u, v) is live. Influence can then spread from u’s neighbors to their neighbors, and so on, according to the same process. Once active, nodes remain active throughout the process, however Kempe et al. (2003) show that this assumption is without loss of generality, and can be removed. 8.1.2 T HE F EEDBACK M ODEL In the Adaptive Viral Marketing problem under the independent cascades model, the items correspond to people we can activate by offering them the promotional deal. How we define the states φ(u) depends on what information we obtain as a result of activating u. Given the nature of the diffusion process, activating u can have wide-ranging effects, so the state φ(u) has more to do with the state of the social network on the whole than with u in particular. Specifically, we model φ(u) as a function φu : A → {0, 1, ?}, where φu ((u, v)) = 0 means that activating u has revealed that (u, v) is dead, φu ((u, v)) = 1 means that activating u has revealed that (u, v) is live, and φu ((u, v)) = ? means that activating u has not revealed the status of (u, v) (i.e., the value of Xuv ). We require each realization to be consistent and complete. Consistency means that no edge should be declared both live and dead by any two states. That is, for all u, v ∈ V and a ∈ A, (φu (a), φv (a)) ∈ / {(0, 1), (1, 0)}. Completeness means that the status of each edge is revealed by some activation. That is, for all a ∈ A there exists u ∈ V such that φu (a) ∈ {0, 1}. A consistent and complete realization thus encodes Xuv for each edge (u, v). Let A(φ) denote the live edges as encoded by φ. There are several candidates for which edge sets we are allowed to observe when activating a node u. Here we consider what we call the Full-Adoption Feedback Model: After activating u we get to see the status (live or dead) of all edges exiting v, for all nodes v reachable from u via live edges (i.e., reachable from u in (V, A(φ)), where φ is the true realization. We illustrate the full-adoption feedback model in Fig. 3. 8.1.3 T HE O BJECTIVE F UNCTION ˆ In the simplest case,  the reward for influencing a set U ⊆ V of nodes is f (U ) := |U |. Kempe et al. (2003) obtain an 1 − 1e -approximation for the slightly more general case in which each node u has a weight wu P indicating its importance, and the reward is fˆ(U ) := u∈U wu . We generalize this result further, to include arbitrary nonnegative monotone submodular reward functions fˆ. This allows us, for example, to encode a value associated with the diversity of the set of nodes influenced, such as the notion that it is better to achieve 20% market penetration in five different (equally important) demographic segments than 100% market penetration in one and 0% in the others. 8.2 Guarantees for the Maximum Coverage Objective We are now ready to formally state our result for the maximum coverage objective. 21  Theorem 19 The greedy policy π greedy obtains at least 1 − 1e of the value of the best policy for the Adaptive Viral Marketing problem with arbitrary monotone submodular reward functions, in the independent cascade and full-adoption feedback models discussed above. That is, if σ(S, φ) is the set of all activated nodes when S is the seed set of activated nodes and φ is the realization, fˆ : 2V → R≥0 is an arbitrary monotone submodular function indicating the reward for influencing a set, and the objective function is f (S, φ) := fˆ(σ(S, φ)), then for all policies π and all k ∈ N we have   1 greedy favg (π[k] ) ≥ 1 − favg (π[k] ). e  ∗ More generally, if π is an α-approximate greedy policy then ∀` ∈ N, favg (π[`] ) ≥ 1 − e−`/αk favg (π[k] ). Proof Adaptive monotonicity follows immediately from the fact that f (·, φ) is monotonic for each φ. It thus suffices to prove that f is adaptive submodular with respect to the probability distribution on realizations p (φ), because then we can invoke Theorem 5 to complete the proof. We will say we have observed an edge (u, v) if we know its status, i.e., if it is live or dead. Fix any ψ, ψ 0 such that ψ ⊆ ψ 0 and any v ∈ / dom(ψ 0 ). We must show ∆(v | ψ 0 ) ≤ ∆(v | ψ). To prove this rigorously, we define a coupled distribution µ over pairs of realizations φ ∼ ψ and φ0 ∼ ψ 0 . Note that given the feedback model, the realization φ is a function of the random variables {Xuw : (u, w) ∈ A} indicating the status of each edge. For conciseness we use the notation X = {Xuw : (u, w) ∈ A}. We define µ implicitly in terms of a joint distribution µ̂ on X × X0 , where φ = φ(X) and φ0 = φ0 (X0 ) are the realizations induced by the two distinct sets of random edge statuses, respectively. Hence µ(φ(X), φ(X0 )) = µ̂(X, X0 ). Next, let us say a partial realization ψ observes an edge e if some w ∈ dom(ψ) has revealed its status as being live or dead. For edges (u, w) observed by ψ, the random variable Xuw is deterministically set to the status observed by ψ. Similarly, for edges (u, w) 0 observed by ψ 0 , the random variable Xuw is deterministically set to the status observed by ψ 0 . Note that since ψ ⊆ ψ 0 , the state of all edges which are observed by ψ are the same in φ and φ0 . All (X, X0 ) ∈ support(µ̂) have these properties. Additionally, we will construct µ̂ so that the status of all edges which are unobserved by 0 both ψ 0 and ψ are the same in X and X0 , meaning Xuw = Xuw for all such edges (u, w), or else µ̂(X, X0 ) = 0. The above constraints leave us with the following degrees of freedom: we may select Xuw for all (u, w) ∈ A which are unobserved by ψ. We select them independently, such that E [Xuw ] = puw as with the prior p (φ). Hence for all (X, X0 ) satisfying the above constraints, Y 1−Xuw uw µ̂(X, X0 ) = pX , uw (1 − puw ) (u,w) unobserved by ψ and otherwise µ̂(X, X0 ) = 0. Note that p (φ | ψ) = claim that for all (φ, φ0 ) ∈ support(µ) P φ0 µ(φ, φ0 ) and p (φ0 | ψ 0 ) = P φ µ(φ, φ0 ). We next f (dom(ψ 0 ) ∪ {v} , φ0 ) − f (dom(ψ 0 ), φ0 ) ≤ f (dom(ψ) ∪ {v} , φ) − f (dom(ψ), φ). (19) Recall f (S, φ) := fˆ(σ(S, φ)), where σ(S, φ) is the set of all activated nodes when S is the seed set of activated nodes and φ is the realization. Let B = σ(dom(ψ), φ) and C = σ(dom(ψ) ∪ {v} , φ) denote the active nodes before and after selecting v after dom(ψ) under realizations φ, and similarly define B 0 and C 0 with respect to ψ 0 and φ0 . Let D := C \B, D0 := C 0 \B 0 . Then Eq. (19) is equivalent to fˆ(B 0 ∪D0 )− fˆ(B 0 ) ≤ fˆ(B ∪D)− fˆ(B). By the submodularity of fˆ, it suffices to show that B ⊆ B 0 and D0 ⊆ D to prove the above inequality, which we will now do. We start by proving B ⊆ B 0 . Fix w ∈ B. Then there exists a path from some u ∈ dom(ψ) to w in (V, A(φ)). Moreover, every edge in this path is not only live but also observed to be live, by definition of the feedback model. Since (φ, φ0 ) ∈ support(µ), this implies that every edge in this path is also live under φ0 , as edges observed by ψ must have the same status under both φ and φ0 . It follows that there is a path from u to w in (V, A(φ0 )). Since u is clearly also in dom(ψ 0 ), we conclude w ∈ B 0 , hence B ⊆ B 0 . Next we show D0 ⊆ D. Fix some w ∈ D0 and suppose by way of contradiction that w ∈ / D. Hence there exists a path P from v to w in (V, A(φ0 )) but no such path exists in (V, A(φ)). The edges of P are all live 22 under φ0 , and at least one must be dead under φ. Let (u, u0 ) be such an edge in P . Because the status of this edge differs in φ and φ0 , and (φ, φ0 ) ∈ support(µ), it must be that (u, u0 ) is observed by ψ 0 but not observed by ψ. Because it is observed by ψ 0 , in our feedback model it must be that u is active after dom(ψ 0 ) is selected, i.e., u ∈ B 0 . However, this implies that all nodes reachable from u via edges in P are also active after dom(ψ 0 ) is selected, since all the edges in P are live. Hence all such nodes, including w, are in B 0 . Since D0 and B 0 are disjoint, this implies w ∈ / D0 , a contradiction. Having proved Eq. (19), we now proceed to use it to show ∆(v | ψ 0 ) ≤ ∆(v | ψ) as in §6. P 0 0 0 0 0 ∆(v | ψ 0 ) = (φ,φ0 ) µ(φ, φ ) (f (dom(ψ ) ∪ {v} , φ ) − f (dom(ψ ), φ )) P 0 ≤ = ∆(v | ψ) (φ,φ0 ) µ(φ, φ ) (f (dom(ψ) ∪ {v} , φ) − f (dom(ψ), φ)) which completes the proof. 8.2.1 C OMPARISON WITH S TOCHASTIC S UBMODULAR M AXIMIZATION It is worth contrasting the Adaptive Viral Marketing problem with the Stochastic Submodular Maximization problem of §6. In the latter problem, we can think of the items as being random independently distributed sets. In Adaptive Viral Marketing by contrast, the random sets (of nodes influenced when a fixed node is selected) depend on the random status of the edges, and hence may be correlated through them. Nevertheless, we can  obtain the same 1 − 1e approximation factor for both problems. A Comment on the Myopic Feedback Model. In the conference version of this article (Golovin and Krause, 2010), we considered an alternate feedback model called the myopic feedback model, in which after activating v we see the status of all edges exiting v in the social network, i.e., ∂+ (u) := {(u, v) : v ∈ V } ∩ A. We claimed that the objective f as defined previously is adaptive submodular in the independent cascade model with myopic feedback, and hence the greedy policy obtains a (1 − 1e ) approximation for it. We hereby retract this claim, and furthermore give a counterexample demonstrating that f is not adaptive submodular under myopic feedback. Consider a graph G = (V, E) with vertices V := {u, v, w}, and edges E := {(u, v), (v, w)}. The edge parameters are puv = 1 and pvw = 1 − . Let fˆ(U ) = |U | and construct f from fˆ accordingly. We let ψ = {(u, φu )}, where φu ((u, v)) = 1 and φu ((v, w)) = ?. Let ψ 0 = {(u, φu ), (v, φv )} where φv ((v, w)) = 0. Clearly, ψ ⊂ ψ 0 . Note ∆(w | ψ) = , since the marginal benefit of w over dom(ψ) is one if (v, w) is dead, and zero if it is live, and the former occurs with probability . In contrast, ∆(w | ψ 0 ) = 1, since ψ 0 contains the observation that (v, w) is dead. Hence ∆(w | ψ) < ∆(w | ψ 0 ), which violates adaptive submodularity. However, we conjecture that the greedy policy still obtains a constant factor approximation even in the myopic feedback model. 8.3 The Minimum Cost Cover Objective We may also wish to adaptively run our campaign until a certain level of market penetration has been achieved, e.g., a certain number of people have adopted the product. We can formalize this goal using the minimum cost cover objective. For this objective, we have an instance of Adaptive Stochastic Minimum Cost Cover, in which we are given a quota Q ≤ fˆ(V ) (quantifying the desired level of market penetration) and we must adaptively select nodes to activate until the set of all active nodes S satisfies fˆ(S) ≥ Q. We obtain the following result. Theorem 20 Fix a monotone submodular function fˆ : 2V → R≥0nindicating the reward for influencing a set, o and a quota Q ≤ fˆ(V ). Suppose the objective is f (S, φ) := min Q, fˆ(σ(S, φ)) , where σ(S, φ) is the set of all activated nodes when S is the seed set of activated nodes and φ is the realization. Let η be any value such that fˆ(S) > Q − η implies fˆ(S) ≥ Q for all S. Then any α-approximate greedy policy π on average 23 2    + 1 times the average cost of the best policy obtaining Q reward for the Adaptive costs at most α ln Q η Viral Marketing problem in the independent cascade model with full-adoption feedback as described above. 2    + 1 cavg (π ∗ ) for any π ∗ that covers every realization. That is, cavg (π) ≤ α ln Q η Proof We prove Theorem 20 by recourse to Theorem 13. We have already established that f is adaptive submodular, in the proof of Theorem 19. It remains to show that f is strongly adaptive monotone and strongly adaptive submodular, that these instances are self–certifying, and that Q and η equal the corresponding terms in the statement of Theorem 13. We start with strong adaptive monotonicity. Fix ψ, e ∈ / dom(ψ), and o ∈ O. We must show E [f (dom(ψ), Φ) | Φ ∼ ψ] ≤ E [f (dom(ψ) ∪ {e} , Φ) | Φ ∼ ψ, Φ(e) = o] . (20) Let V + (ψ) denote the active nodes after selecting dom(ψ) and observing ψ. By definition of the full adoption feedback model, V + (ψ) consists of precisely those nodes v for which there exists a path Puv from some u ∈ dom(ψ) to v via exclusively live edges. The edges whose status we observe consist of all edges exiting nodes in V + (ψ). It follows that every path from any u ∈ V + (ψ) to any v ∈ V \ V + (ψ) contains at least one edge which is observed by ψ to be dead. Hence, in every φ ∼ ψ, the set of nodes activated by selecting dom(ψ) is the same. Therefore E [f (dom(ψ), Φ) | Φ ∼ ψ] = fˆ(V + (ψ)). Similarly, if we define ψ 0 := ψ ∪ {(e, o)}, then E [f (dom(ψ) ∪ {e} , Φ) | Φ ∼ ψ, Φ(e) = o] = fˆ(V + (ψ 0 )). Note that once activated, nodes never become inactive. Hence, ψ ⊆ ψ 0 implies V + (ψ) ⊆ V + (ψ 0 ). Since fˆ is monotone by assumption, this means fˆ(V + (ψ)) ≤ fˆ(V + (ψ 0 )) which implies Eq. (20) and strong adaptive monotonicity. Next we establish strong adaptive submodularity. Given that we have already established adaptive submodularity, it is sufficient to also prove pointwise submodularity. For a fixed realization φ we have a set of live edges {(u, v) : Xuv = 1} which induce a set system in which u covers all nodes reachable from u via live edges. Let Su denote this set. It is straightforward to verify that a monotone submodular function fˆ on nodes induces a monotone submodular function on sets of those nodes. That is, g(A) := fˆ (∪u∈A Su ) is submodular whenever fˆ is. In particular, g is submodular if for every A and B we have g(A) + g(B) ≥ g(A ∩ B) + g(A ∪ B) however one can easily verify that this set of constraints is a subset of the corresponding submodularity constaints on fˆ. Next that these instances are self–certifying. Note that for every φ we have f (V, φ) = n we establish o min Q, fˆ(V ) = Q. From our earlier remarks, we know that f (dom(ψ), φ) = fˆ(V + (ψ)) for every φ ∼ ψ. Hence for all ψ and φ, φ0 consistent with ψ, we have f (dom(ψ), φ) = f (dom(ψ), φ0 ) and so f (dom(ψ), φ) = Q if and only if f (dom(ψ), φ0 ) = Q, which proves that the instance is self–certifying. Finally we show that Q and η equal the corresponding terms in the statement of Theorem 13. As noted ˆ ˆ earlier, f (V, φ) = Q for n all φ. n We defined o η as some o value such that f (S) > Q − η implies f (S) ≥ Q for all S. Since range(f ) = min Q, fˆ(S) : S ⊆ V , it follows that we cannot have f (S, φ) ∈ (Q − η, Q) for any S and φ, so that η satisfies the requirements of the corresponding term in Theorem 13. Hence we may apply Theorem 13 on this self–certifying instance with Q and η to obtain the claimed result. 9. Application: Automated Diagnosis and Active Learning An important problem in AI is automated diagnosis. For example, suppose we have different hypotheses about the state of a patient, and can run medical tests to rule out inconsistent hypotheses. The goal is to adaptively choose tests to infer the state of the patient as quickly as possible. A similar problem arises in active learning. Obtaining labeled data to train a classifier is typically expensive, as it often involves asking an expert. In active learning (c.f. Cohn et al. (1996); McCallum and Nigam (1998)), 24 Figure 4: Illustration of the Active Learning problem, in the simple special case of one-dimensional data and binary threshold hypotheses H = {hτ : τ ∈ R}, where hτ (x) = 1 if x ≥ τ and 0 otherwise. the key idea is that some labels are more informative than others: labeling a few unlabeled examples can imply the labels of many other unlabeled examples, and thus the cost of obtaining the labels from an expert can be avoided. As is standard, we assume that we are given a set of hypotheses H, and a set of unlabeled data points X where each x ∈ X is independently drawn from some distribution D. Let L be the set of possible labels. Classical learning theory yields probably approximately correct (PAC) bounds, bounding the number n of examples drawn i.i.d. from D needed to output a hypothesis h that will have expected error at most ε with probability at least 1 − δ, for some fixed ε, δ > 0. That is, if h∗ is the target hypothesis (with zero error), and error(h) := Px∼D [h(x) 6= h∗ (x)] is the error of h, we require P [error(h) ≤ ε] ≥ 1 − δ. The latter probability is taken with respect to D(X); the learned hypothesis h and thus error(h) depend on it. A key challenge in active learning is to avoid bias: actively selected examples are no longer i.i.d., and thus sample complexity bounds for passive learning no longer apply. If one is not careful, active learning may require more samples than passive learning to achieve the same generalization error. One natural approach to active learning that is guaranteed to perform at least as well as passive learning is pool-based active learning (McCallum and Nigam, 1998): The idea is to draw n unlabeled examples i.i.d. However, instead of obtaining all labels, labels are adaptively requested until the labels of all unlabeled examples are implied by the obtained labels. Now we have obtained n labeled examples drawn i.i.d., and classical PAC bounds still apply. The key question is how to request the labels for the pool to infer the remaining labels as quickly as possible. In the case of binary labels (or test outcomes) L = {−1, 1}, various authors have considered greedy policies which generalize binary search (Garey and Graham, 1974; Loveland, 1985; Arkin et al., 1993; Kosaraju et al., 1999; Dasgupta, 2004; Guillory and Bilmes, 2009; Nowak, 2009). The simplest of these, called generalized binary search (GBS) or the splitting algorithm, works as follows. Define the version space V to be the set of hypotheses consistent with the observed labels (here we assume P that there is no label noise). In the worst-case setting, GBS selects a query x ∈ X that minimizes h∈V h(x) . In the Bayesian setting we assume we are given a prior p over hypotheses; in this case GBS selects a query H P x ∈ X that minimizes p (h) · h(x) . Intuitively these policies myopically attempt to shrink a H h∈V measure of the version space (i.e., the cardinality or the probability mass) as quickly as possible. The former provides an O(log |H|)-approximation for the worst-case number of queries (Arkin et al., 1993), and the latter provides an O(log minh 1pH (h) )-approximation for the expected number of queries (Kosaraju et al., 1999; Dasgupta, 2004) and a natural generalization of GBS obtains the same guarantees with a larger set of labels (Guillory  and Bilmes, 2009). Kosaraju et al. also prove that running GBS on a modified prior p0H (h) ∝ max pH (h), 1/|H|2 log |H| is sufficient to obtain an O(log |H|)-approximation. Viewed from this perspective of the previous sections, shrinking the version space amounts to “covering” all false hypotheses with stochastic sets (i.e., queries), where query x covers all hypotheses that disagree with the target hypothesis h∗ at x. That is, x covers {h : h(x) 6= h∗ (x)}. As in §8, these sets may be correlated in complex ways determined by the set of possible hypotheses. As we will show, the reduction in version space mass is adaptive submodular, and this allows us to obtain a new analysis of GBS using adaptive submodularity, 25 which has a weaker approximation guarantee than is optimal, but is arguably more amenable to extensions and generalizations than previous analyses. Theorem 21 In the Bayesian setting in which there is a prior pH on a finite set of hypotheses H, the    2 generalized binary search algorithm makes OPT · ln minh 1pH (h) + 1 queries in expectation to identify a hypothesis drawn from pH , where OPT is the minimum expected number of queries made by any policy. If 0 2 minh pH (h) is sufficiently small, runningthe algorithm  on a modified prior pH (h) ∝ max pH (h), 1/|H| 2 improves the approximation factor to O (ln |H|) . We devote the better part of the remainder of this section to the proof of Theorem 21, which has several components. We first address the important special case of a uniform prior over hypotheses, i.e., pH (h) = 1/|H| for all h ∈ H, and then we reduce the case with a general prior to a uniform prior. We wish to appeal to Theorem 13, so we convert the problem into an Adaptive Stochastic Min Cost Cover problem. 9.1 The Reduction to Adaptive Stochastic Min Cost Cover Define a realization φh for each hypothesis h ∈ H. The ground set is E = X, and the outcomes are binary; we define O = {−1, 1} instead of using {0, 1} to be consistent with our earlier exposition. For all h ∈ H we set φh ≡ h, meaning φh (x) = h(x) for all x ∈ X. To define the objective function, we first need some notation. Given observed labels ψ ⊂ X × O, let V (ψ) denote the version space, i.e., the set of hypotheses for which h(x) = ψ(x) for all x ∈ dom(ψ). See Fig. 4 for an illustration P of an active learning problem in the case of indicator hypotheses. For a set of hypotheses V , let pH (V ) := h∈V pH (h) denote their total prior probability. Finally, let ψ(S, h) = {(x, h(x)) : x ∈ S} be the function with domain S that agrees with h on S. We define the objective function by f (S, φh ) := 1 − pH (V (ψ(S, h))) = pH ({h0 : ∃x ∈ S, h0 (x) 6= h(x)}) and use p (φh ) = pH (h) = 1/|H| for all h. Let π ∗ be an optimal policy for this Adaptive Stochastic Min Cost Cover instance. Note that there is an exact correspondence between policies for the original problem of finding the target hypothesis and our problem of covering the true realization; h∗ is identified as the target hypothesis if and only if the version space is reduced to {h∗ } which occurs if and only if φh∗ is covered. Hence cavg (π ∗ ) = OPT. Note that because we have assumed a uniform prior over hypotheses, we have f (X, φh ) = 1 − 1/|H| for all h. Furthermore, the instances are self–certifying. Lemma 22 The instances described above are self–certifying for arbitrary priors pH . Proof Intuitively, theses instances are self–certifying because to cover φh∗ a policy must identify φh∗ . More formally, these instances are self–certifying because for any φh and ψ such that φh ∼ ψ, we have that f (dom(ψ), φh ) = f (X, φh ) implies V (ψ) = {h}. This in turn means that φh is the only realization consistent with ψ, which trivially implies that any realization φ0 ∼ ψ also has f (dom(ψ), φ0 ) = f (X, φ0 ); hence the instance is self–certifying. 9.2 The Uniform Prior We next prove that the instances generated are adaptive submodular and strongly adaptive monotone under a uniform prior. Lemma 23 In the instances described above, f is strongly adaptive monotone and strongly adaptive submodular and with respect to a uniform prior pH . 26 Proof Demonstrating strong adaptive monotonicity under a uniform prior amounts to proving that adding labels cannot grow the version space, which is clear in our model. That is, each query x eliminates some subset of hypotheses, and as more queries are performed, the subset of hypotheses eliminated by x cannot grow. Moving on to adaptive submodularity, consider the expected marginal contribution of x under two partial realizations ψ, ψ 0 where ψ is a subrealization of ψ 0 (i.e., ψ ⊂ ψ 0 ), and x ∈ / dom(ψ 0 ). Let ψ[x/o] be the partial realization with domain dom(ψ) ∪ {x} that agrees with ψ on its domain, and maps x to o. For each o ∈ O, let ao := pH (V (ψ[x/o])), bo := pH (V (ψ 0 [x/o])). Since a hypothesis eliminated from the version space cannot later appear in the version space, we have ao ≥ bo for all o. Next, note the expected reduction in version space mass (and hence the expected marginal contribution) due to selecting x given partial realization ψ is P P X X  o0 6=o ao0  0 o6=o0 ao ao P ∆(x | ψ) = ao · P [Φ(x) 6= o | Φ ∼ ψ] = ao = P . (21) 0 0 o0 ao o0 ao o o∈O The corresponding quantity for ψ 0 has bo substituted for ao in Eq. (21), for each o ∈ O. To prove adaptive submodularity we must show ∆(x | ψ) ≥ ∆(x | ψ 0 ) and to do so it suffices show that ∂φ/∂zo ≥ 0 for each P  to  P P O 0 o and ~z ∈ ~c ∈ [0, 1] : o co > 0 , where φ(~z ) := / ( o0 zo0 ) has the same functional o6=o0 zo zo form as the expression for ∆(x | ψ) in Eq. (21). This is because ∂φ/∂zo ≥ 0 for each o implies that growing the version space in any manner cannot decrease the expected marginal benefit of query x, and hence shrinking it in any manner cannot increase the expected marginal benefit of x. It is indeed the case that ∂φ/∂zo ≥ 0 for each o. More specifically, it holds that P P 2 ∂φ b6=a zb + (b,c):b6=c,b6=a,c6=a zb zc ≥ 0, = P 2 ∂za ( b zb ) which can be derived through elementary calculus. To further show strong adaptive submodularity, we prove pointwise submodularity of the objective. For any fixed realization, h, the objective S 7→ f (S, φ) amounts to a weighted set cover problem on the incorrect hypotheses (plus a constant), and is thus submodular. Hence we can apply Theorem 13 to this self–certifying instance with maximum reward threshold Q = 2 1 − 1/|H|, and minimum gap η = 1/|H|, to obtain an upper bound of OPT (ln (|H| − 1) + 1) on the number of queries made by the generalized binary search algorithm (which corresponds exactly to the greedy policy for Adaptive Stochastic Min Cost Cover) under the assumption of a uniform prior over H. 9.3 Arbitrary Priors Now consider general priors over H. We construct the Adaptive Stochastic Min Cost Cover instance as before, only we change the objective function to f (S, φh ) := 1 − pH (V (ψ(S, h))) + pH (h). (22) First note that the instances remain self–certifying. The proof of Lemma 22 goes through completely unchanged by the modification of f . We proceed to show adaptive submodularity and strong adaptive monotonicity. Lemma 24 The objective function f as described in Eq. (22) is strongly adaptive monotone and strongly adaptive submodular with respect to arbitrary priors pH . Proof The modified objective is still adaptive submodular, because (S, φh ) 7→ pH (h) is clearly so, and because adaptive submodularity is defined via linear inequalities it is preserved under taking nonnegative linear combinations. Note that f (X, φh ) = 1 for all φh . It is still strongly adaptive submodular, because it is still pointwise submodular. 27 Showing f is strongly adaptive monotone requires slightly more work than before. Fix ψ, x ∈ / dom(ψ), and o ∈ O. We must show E [f (dom(ψ), Φ) | Φ ∼ ψ] ≤ E [f (dom(ψ) ∪ {x} , Φ) | Φ ∼ ψ, Φ(x) = o] . (23) Plugging in the definition of f , the inequality we wish to prove may be simplified to E [pH (Φ) | Φ ∼ ψ] − E [pH (Φ) | Φ ∼ ψ[x/o]] ≤ pH (V (ψ)) − pH (V (ψ[x/o])). (24) where Φ is the random realization of the hypothesis, and pH (φh ) = pH (h) for all h. Let Velim := V (ψ) − V (ψ[x/o]) be the set of hypotheses eliminated from the version space by the observation h(x) = o. Rewriting Eq. (24), we get X X pH (h)2 pH (h)2 − ≤ pH (Velim ). (25) pH (V (ψ)) pH (V (ψ[x/o])) h∈V (ψ) h∈V (ψ[x/o]) Let LHS25 denote the left hand side of Eq. (25). We prove Eq. (25) as follows. P 2 LHS25 ≤ [since pH (V (ψ[x/o])) ≤ pH (V (ψ))] h∈Velim pH (h) /pH (V (ψ)) P ≤ h∈Velim pH (h) · pH (V (ψ))/pH (V (ψ)) [since h ∈ V (ψ) ⇒ pH (h) ≤ pH (V (ψ))] = pH (Velim ) We conclude that f is adaptive submodular and strongly adaptive monotone. Hence we can apply Theorem 13 to this self–certifying instance with maximum reward threshold Q = 1, 2 and minimum gap η = 1/ minh pH (h). As a result we obtain an upper bound of OPT (ln (1/ minh pH (h)) + 1) on the number of queries made by generalized binary search for arbitrary priors, completing the proof of Theorem 21. 9.4 Improving the Approximation Factor for Highly Nonuniform Priors   2 To improve this to an O (log |H|) -approximation in the event that minh pH (h) is extremely small using the observation of Kosaraju et al. (1999), call a policy π  progressive if it eliminates at least one hypothesis from its version space in each query. Let p0H (h) = max pH (h), 1/|H|2 /Z be the modified prior, where P Z := h0 max pH (h0 ), 1/|H|2 is the P normalizing constant. Let c(π, h) be the cost (i.e., # of queries) of π under target h. Then cavg (π, p) := h c(π, h)p(h) is the expected cost of π under prior p. We will show that cavg (π, p0H ) is a good approximation to cavg (π, pH ). Call h rare if pH (h) < 1/|H|2 , and common  P |H| otherwise. First, note that h0 max pH (h0 ), 1/|H|2 ≤ 1 + 1/|H|, and so p0H (h) ≥ |H|+1 pH (h), for all |H| h. Hence for all π, we have cavg (π, p0H ) ≥ |H|+1 cavg (π, pH ). Next, we show cavg (π, p0H ) ≤ cavg (π, pH ) + 1. P 0 Consider the quantity cavg (π, pH ) − cavg (π, pH ) = h c(π, h) (p0H (h) − pH (h)). The positive contributions must come from rare hypotheses. However, the total probability mass of these under p0H is at most 1/|H|, and since π is progressive c(π, h) ≤ |H| for all h, hence the difference in costs is at most one. Let    2  2 α := ln minh 1p0 (h) + 1 ≤ ln |H|2 + |H| + 1 be the approximation factor for generalized binary H search when run on p0H . Let π be the policy of generalized binary search, and let π ∗ be an optimal policy under prior pH . Then |H| + 1 |H| + 1 |H| + 1 cavg (π, p0H ) ≤ α cavg (π ∗ , p0H ) ≤ α (cavg (π ∗ , pH ) + 1) . |H| |H| |H|   2 Thus for a general prior a simple modification of GBS yields an O (log |H|) -approximation. cavg (π, pH ) ≤ 28 9.5 Extensions to Arbitrary Costs, Multiple Classes, and Approximate Greedy Policies This result easily generalizes to handle the setting of multiple classes / test outcomes (i.e., |O| ≥ 2), and α-approximate greedy policies, where we lose a factor of α in the approximation factor. As we describe in the Appendix, we can generalize adaptive submodularity to incorporate costs on items, which allows us to extend this result to handle query costs as well. Recently, Gupta et al. (2010) showed how to simultaneously remove the dependence on both costs and probabilities from the approximation ratio. Specifically, within the context of studying an adaptive travelling salesman problem they investigated the Optimal Decision Tree problem, which is equivalent to the active learning problem we consider here. Using a clever, more complex algorithm than adaptive greedy, they achieve an O (log |H|)-approximation in the case of non-uniform costs and general priors. 9.6 Extensions to Active Learning with Noisy Observations Theorem 21 and the extensions mentioned so far are in the noise free case, i.e., the result of query x and observes h∗ (x), where h∗ is the target hypothesis. Many practical problems may have noisy observations. Nowak (2009) considered the case in which the outcomes are binary, i.e., O = {−1, 1}, the same query may be asked multiple times, and for each instance of each query the noise is independent. In this case he gives performance guarantees for generalized binary search. While this setting may be appropriate if the noise is due to measurement error, in some applications the noise is persistent, i.e., if query x is asked several times, the observation is always the same. Recently, Golovin et al. (2010) and Bellala and Scott (2010) have used the adaptive submodularity framework to obtain the first algorithms with provable (squared–logarithmic) approximation guarantees for active learning with persistent noise. 10. Experiments Greedy algorithms are often straightforward to develop and implement, which explains their popular use in practical applications, such as Bayesian experimental design and Active Learning, as discussed in §9 (also see the excellent introduction of Nowak (2009)) and Adaptive Stochastic Set Cover, e.g., for filter design in streaming databases as discussed in §7. Besides allowing us to prove approximation guarantees for such algorithms, adaptive submodularity provides the following immediate practical benefits: 1. The ability to use lazy evaluations to speed up its execution. 2. The ability to generate data-dependent bounds on the optimal value. In this section, we empirically evaluate their benefits within a sensor selection application, in a setting similar to the one described by Deshpande et al. (2004). In this application, we have deployed a network V of wireless sensors, e.g., to monitor temperature in a building or traffic in a road network. Since sensors are battery constrained, we must adaptively select k sensors, and then, given those sensor readings, predict, e.g., the temperature at all remaining locations. This prediction is possible since temperature measurements will typically be correlated across space. Here, we will consider the case where sensors can fail to report measurements due to hardware failures, environmental conditions or interference. 10.1 The Sensor Selection Problem with Unreliable Sensors More formally, we imagine every location v ∈ V is associated with a random variable Xv describing the temperature at that location, and there is a joint probability distribution p (xV ) := P [XV = xV ] that models the correlation between temperature values. Here, XV = [X1 , . . . , Xn ] is the random vector over all temperature values. We follow Deshpande et al. (2004) and assume that the joint distribution of the sensors is multivariate Gaussian. A sensor v can make a noisy observation Yv = Xv + εv , where εv is zero mean Gaussian noise with known variance σ 2 . If some measurements YA = yA are obtained at a subset of locations, then the conditional distribution p (xV | yA ) := P [XV = xV | YA = yA ] allows predictions at the unobserved locations, e.g., by predicting E[XV | YA = yA ] (which minimizes the mean squared error). Furthermore, this conditional distribution quantifies the uncertainty in the prediction: Intuitively, we would like to select sensors that 29 minimize the predictive uncertainty. One way to quantify the predictive uncertainty is to use the remaining Shannon entropy H (XV | YA = yA ) := E [− log2 (p (XV | yA ))] . We would like to adaptively select k sensors, to maximize the expected reduction in Shannon entropy (c.f. Sebastiani and Wynn (2000); Krause and Guestrin (2009a)). However, in practice, sensors are often unreliable, and might fail to report their measurements. We assume that after selecting a sensor, we find out whether it has failed or not before deciding which sensor to select next. We suppose that each sensor has an associated probability pfail (v) of failure, in which case no reading is reported, and that sensor failures are independent of each other and of the ambient temperature at v. Thus we have an instance of the Stochastic Maximization problem with E := V , O := {working, failed}, and  f (A, φ) := H (XV ) − H XV | y{v : φ(v)=working} . (26) For multivariate normal distributions, the entropy is given as H (XV | YA = yA ) = −1 1 ΣAV , ln(2πe)n ΣV A ΣAA + σ 2 I 2 where for sets A and B, ΣAB denotes the covariance (matrix) between random vectors XA and XB . Note that the predictive covariance does not depend on the actual observations yA , only on the set A of chosen locations. Thus, H (XV | YA = yA ) = H (XV | YA ) , where as usual, H (XV | YA ) = E [H (XV | YA = yA )]. As Krause and Guestrin (2005) show, the function g(A) := I (XV ; YA ) = H (XV ) − H (XV | YA ) (27) is monotone submodular, whenever the observations YV are conditionally independent given XV . This insight allows us to apply the result of §6 to show that the objective f defined in Eq. (26) is adaptive monotone submodular, using fˆ(S) := g({v : (v, working) ∈ S}) for any S ⊆ E × O. 10.1.1 DATA AND E XPERIMENTAL S ETUP Our first data set consists of temperature measurements from the network of 46 sensors deployed at Intel Research Berkeley, which were sampled at 30 second intervals for 5 consecutive days (starting Feb. 28th , 2004). We define our objective function with respect to the empirical covariance estimated from the data. We also use data from traffic sensors deployed along the highway I-880 South in California. We use traffic speed data for all working days from 6 AM to 11 AM for one month, from 357 sensors. The goal is to predict the speed on all 357 road segments. We again estimate the empirical covariance matrix. 10.1.2 T HE B ENEFITS OF L AZY E VALUATION For both data sets, we run the adaptive greedy algorithm, using both the naive implementation (Algorithm 1) and the accelerated version using lazy evaluations (Algorithm 2). We vary the probability of sensor failure, and evaluate the execution time and the number of evaluations of the function g (defined in Eq. (27)) each algorithm makes. Figures 5(a) and 5(b) plot execution time given a 50% sensor failure rate, on a computer with a 2.26 GHz dual core processor and 4 GB RAM. In these applications, function evaluations are the bottleneck in the computation, so the number of them serves as a machine-independent proxy for the running time. Figures 5(c) and 5(d) show the performance ratio in terms of this proxy. On the temperature data set, lazy evaluations speed up the computation by a factor of between roughly 3.5 and 7, depending on the failure probability. On the larger traffic data set, we obtain speedup factors between 30 and 38. We find that the benefit of the lazy evaluations increases with the problem size and with the failure probability. The dependence on problem size must ultimately be explained in terms of structural properties of the instances, which also benefit the nonadaptive accelerated greedy algorithm. The dependence on failure probability has a simpler 30 explanation. Note that in these applications, if the accelerated greedy algorithm selects v, which then fails, then it does not need to make any additional function evaluations to select the next sensor. Contrast this with the naive greedy algorithm, which makes a function evaluation for each sensor that has not been selected so far. 10.1.3 T HE B ENEFITS OF THE DATA D EPENDENT B OUND While adaptive submodularity allows us to prove worst-case performance guarantees for the adaptive greedy algorithm, in many practical applications it can be expected that these bounds are quite loose. For our sensor selection application, we use the data dependent bounds of Lemma 6 to compute an upper bound βavg on maxπ favg (π[k] ) as described below, and compare it with the performance guarantee of Theorem 5. For the accelerated greedy algorithm, we use the upper bounds on the marginal benefits stored in the priority queue instead of recomputing the marginal benefits, and thus expect somewhat looser bounds. We find that for our application, the bounds are tighter than the worst case bounds. We also find that the “lazy” data dependent bounds are almost as tight as the “eager” bounds using the eagerly recomputed marginal benefits ∆(e | ψ) for the latest and greatest ψ, though the former have slightly higher variance. Figures 5(e) and 5(f) show the performance of the greedy algorithm as well as the three bounds on the optimal value. Two subtleties arise  when using  the data-dependentPbounds to bound maxπ favg (π[k] ). The first is that ∗ Lemma 6 tells us that ∆ π[k] | ψ ≤ maxA⊆E,|A|≤k e∈A ∆(e | ψ), whereas we would like to bound the differenceh between the optimal reward and the algorithm’s i current expected reward, conditioned on seeing ∗ ψ, i.e., E f (E(π[k] , Φ), Φ) − f (dom(ψ), Φ) | Φ ∼ ψ . However, in our applications f is strongly adaptive monotone, and strong adaptive monotonicity implies that for any π ∗ we have h i   ∗ ∗ E f (E(π[k] , Φ), Φ) − f (dom(ψ), Φ) | Φ ∼ ψ ≤ ∆ π[k] |ψ . (28)   Hence, if we let OPT(ψ) := maxπ E f (E(π[k] , Φ), Φ) | Φ ∼ ψ , Lemma 6 implies that X OPT(ψ) ≤ E [f (dom(ψ), Φ) | Φ ∼ ψ] + max ∆(e | ψ) . A⊆E,|A|≤k (29) e∈A The second subtlety is that we obtain a sequence of bounds from Eq. (29). If we consider the (random) sequence of partial realizations observed by the adaptive greedy algorithm, ∅ = ψ 0 ⊂ ψ 1 ⊂ P· · · ⊂ ψ k , we obtain k+1 bounds β0 , . . . , βk , where βi := E [f (dom(ψ i ), Φ) | Φ ∼ ψ i ]+maxA⊆E,|A|≤k e∈A ∆(e | ψ i ). Taking the expectation over Φ, note that for any π, and any i, favg (π[k] ) ≤ E [OPT(ψ i )] ≤ E [βi ] . Therefore for any 0 ≤ i ≤ k , βi is a random variable whose expectation is an upper bound on the optimal expected reward of any policy. At this point we may be tempted to use the minimum of these, i.e., βmin := mini {βi } as our ultimate bound. However, a collection of random variables X0 , . . . , Xk with E [Xi ] ≥ τ for all i does not, in general, satisfy mini {Xi } ≥ τ . While it is possible in our case, with its independent sensor failures, to use concentration inequalities to bound mini {βi } − mini {E [βi ]} with high probability, and thus add an appropriate term to obtain a true upper bound from βmin , we take a different Pk 1 approach; we simply use the average bound βavg := k+1 i=0 βi . Of course, depending on the application, a particular bound βi (chosen independently of the sequence ψ 0 , ψ 1 , . . . , ψ k ) may be superior. For example, if g is modular, then β0 is best, whereas if g exhibits strong diminishing returns, then bounds βi with larger values of i may be significantly tighter. 11. Adaptivity Gap An important question in adaptive optimization is how much better adaptive policies can perform when compared to non-adaptive policies. This is quantified by the adaptivity gap, which is the worst-case ratio, over 31 Time for Standard and Accelerated Adaptive Greedy 0.45 0.4 Time for Standard and Accelerated Adaptive Greedy 40 35 Adaptive Greedy 0.35 Adaptive Greedy 30 0.3 25 0.25 20 0.2 15 0.15 Accelerated Adaptive Greedy 10 0.1 0.05 5 0 0 0 5 10 15 20 25 30 35 40 45 50 Accelerated Adaptive Greedy 0 50 100 150 200 250 300 350 (a) Temperature Data: Execution time (sec) for the naive vs accelerated implementations of adaptive greedy vs. the budget k on number of sensors selected, when pfail (v) = 0.5 for all v, plotted with standard errors. (b) Traffic Data: Execution time (sec) for the naive vs accelerated implementations of adaptive greedy vs. the budget k on number of sensors selected, when pfail (v) = 0.5 for all v, plotted with standard errors. 8 45 Redu ction in Function Evaluations vs. Pr[failure] 7 6 5 4 Reduction in Function Evaluations vs . Pr [failur e] 40 90% 80% 70% 60% 50% 40% 30% 20% 10% 75% 35 50% 25% 30 25 20 15 3 10 2 5 1 0 10 20 30 40 0 0 50 100 200 300 400 (c) Temperature Data: The ratio of function evaluations made by the naive vs accelerated implementations of adaptive greedy vs. the budget k on number of sensors selected, for various failure rates. Averaged over 100 runs. (d) Traffic Data: The ratio of function evaluations made by the naive vs accelerated implementations of adaptive greedy vs. the budget k on number of sensors selected, for various failure rates. Averaged over 10 runs. Reward for Adaptive Greedy, with Data−Dependent Bounds Reward for Adaptive Greedy, with Data−Dependent Bounds 100 150 90 Standard Bound Lazy Adaptive Bound Adaptive Bound Standard Bound Lazy Adaptive Bound Adaptive Bound 80 70 100 60 50 40 50 30 Adaptive Greedy Adaptive Greedy 20 10 0 0 5 10 15 20 25 30 35 40 45 0 50 (e) Temperature Data: Rewards & bounds on the optimal value when pfail (v) = 0.5 for all v vs. the budget k on number of sensors selected, plotted with standard errors. 0 50 100 150 200 250 300 350 400 (f) Traffic Data: Rewards & bounds on the optimal value when pfail (v) = 0.5 for all v vs. the budget k on number of sensors selected, plotted with standard errors. Figure 5: Experimental results. 32 problem instances, of the performance of the optimal adaptive policy to the optimal non-adaptive solution. Asadpour et al. (2008) show that in the Stochastic Submodular Maximization problem with independent failures (as considered in §6), the expected value of the optimal non-adaptive policy is at most a constant factor 1 − 1/e worse than the expected value of the optimal adaptive policy. While we currently do not have lower bounds for the adaptivity gap of the general Adaptive Stochastic Maximization problem (1), we can show that even in the case of adaptive submodular functions, the min-cost cover and min-sum cover versions have large adaptivity gaps, and thus there is a large benefit of using adaptive algorithms. In these cases, the adaptivity gap is defined as the worst-case ratio of the expected cost of the optimal non-adaptive policy divided by the expected cost of the optimal adaptive policy. For the Adaptive Stochastic Minimum Cost Coverage problem (2), Goemans and Vondrák (2006) show the special case of Stochastic Set Coverage without multiplicities has an adaptivity gap of Ω(|E|). Below we exhibit an adaptive stochastic optimization instance with adaptivity gap of Ω(|E|/ log |E|) for the Adaptive Stochastic Min-Sum Cover problem (3), which also happens to have the same adaptivity gap for Adaptive Stochastic Minimum Cost Coverage. Theorem 25 Even for adaptive submodular functions, the adaptivity gap of Adaptive Stochastic Min-Sum Cover is Ω(n/ log n), where n = |E|. Proof Suppose E = {1, . . . , n}. Consider the active learning problem where our hypotheses h : E → {−1, 1} are threshold functions, i.e., h(e) = 1 if e ≥ ` and h(e) = −1 if e < ` for some threshold `. There is a uniform distribution over thresholds ` ∈ {1, . . . , n + 1}. In order to identify the correct hypothesis with threshold `, our policy must observe at least one of ` − 1 or ` (and both of them if 1 < ` ≤ n). Let π̂ be an optimal non-adaptive policy for this problem. Note that π̂ can be represented as a permutation of E, because observing an element multiple times can only increase the cost while providing no benefit over observing it once, and each element must eventually be selected to guarantee coverage. For the min-sum cover objective, consider playing π̂ for n/4 time steps. Then P [` observed in n/4 steps] = n/4(n + 1). Likewise P [` − 1 observed in n/4 steps] = n/4(n + 1). Since at least one of these events must occur to identify the correct hypothesis, by a union bound n ≤ 1/2. P [π̂ identifies the correct hypothesis in n/4 steps] ≤ 2(n + 1) Thus a lower bound on the expected cost of π̂ is n/8, since for n/4 time steps a cost of at least 1/2 is incurred. Thus, for both the min-cost and min-sum cover objectives the cost of the optimal non-adaptive policy is Ω(n). As an example adaptive policy, we can implement a natural binary search strategy, which is guaranteed to identify the correct hypothesis after O(log n) steps, thus incurring cost O(log n), proving an adaptivity gap of Ω(n/ log n). 12. Hardness of Approximation In this paper, we have developed the notion of adaptive submodularity, which characterizes when certain adaptive stochastic optimization problems are well-behaved in the sense that a simple greedy policy obtains a constant factor or polylogarithmic factor approximation to the best policy. In contrast, we can also show that without adaptive submodularity, the adaptive stochastic optimization problems (1), (2), and (3) are extremely inapproximable, even with (pointwise) modular objective functions (i.e., those where for each φ, f : 2E × OE → R is modular/linear in the first argument): We cannot hope to achieve an O(|E|1−ε ) approximation ratio for these problems, unless the polynomial hierarchy collapses down to ΣP 2. Theorem 26 For all (possibly non-constant) β ≥ 1, no polynomial time algorithm for Adaptive Stochastic Maximization with a budget of βk items can approximate the reward of an optimal policy with a budget of only k items to within a multiplicative factor of O(|E|1−ε /β) for any ε > 0, unless PH = ΣP 2 . This holds even for pointwise modular f . 33 We provide the proof of Theorem 26 in Appendix 15.7. Note that by setting β = 1, we obtain O(|E|1−ε ) hardness for Adaptive Stochastic Maximization. It turns out that in the instance distribution we construct in the proof of Theorem 26 the optimal policy covers every realization (i.e., always finds the treasure) using a budget of k = O(|E|ε/2 ) items. Hence if PH 6= ΣP 2 then any randomized polynomial time algorithm wishing to cover this instance must have a budget β = Ω(|E|1−ε ) times larger than the optimal policy, in order to ensure the ratio of rewards, which is Ω(|E|1−ε /β), equals one. This yields the following corollary. Corollary 27 No polynomial time algorithm for Adaptive Stochastic Min Cost Coverage can approximate the cost of an optimal policy to within a multiplicative factor of O(|E|1−ε ) for any ε > 0, unless PH = ΣP 2 . This holds even for pointwise modular f . Furthermore, since in the instance distribution we construct the optimal policy π ∗ covers every realization using a budget of k, it has cΣ (π ∗ ) ≤ k. Moreover, since we have shown that under our complexity theoretic assumptions, any polynomial time randomized policy π with budget βk achieves at most o(β/|E|1−ε ) of the (unit) value obtained by the optimal policy with budget k, it follows that cΣ (π) = Ω(βk). Since we require β = Ω(|E|1−ε ) to cover any set of realizations constituting, e.g., half of the probability mass, we obtain the following corollary. Corollary 28 No polynomial time algorithm for Adaptive Stochastic Min-Sum Cover can approximate the cost of an optimal policy to within a multiplicative factor of O(|E|1−ε ) for any ε > 0, unless PH = ΣP 2 . This holds even for pointwise modular f . 13. Related Work There is a large literature on adaptive optimization under partial observability which relates to adaptive submodularity, which can be broadly organized into several different categories. Here, we only review relevant related work that is not already discussed elsewhere in the manuscript. 13.1 Adaptive Versions of Classic Non-adaptive Optimization Problems Many approaches consider stochastic generalizations of specific classic non-adaptive optimization problems, such as Set Cover (Goemans and Vondrák, 2006; Liu et al., 2008), Knapsack (Dean et al., 2008, 2005) and Traveling Salesman (Gupta et al., 2010). In contrast, in this paper our goal is to introduce a general problem structure – adaptive submodularity – that unifies a number of adaptive optimization problems. This is similar to how the classic notion of submodularity unifies various optimization problems such as Set Cover, Facility Location, nonadaptive Bayesian Experimental Design, etc. 13.2 Competitive Online Optimization Another active area of research in sequential optimization is the study of competitive online algorithms. A particularly relevant example is Online Set Cover Alon et al. (2009), where there is a known set system, an arbitrary sequence of elements is presented to the algorithm, and the algorithm must irrevocably select sets to purchase such that at all times the purchased sets cover all elements which have appeared so far. Alon et al. (2009) obtain a polylogarithmic approximation to this problem, via an online primal–dual framework which has been profitably applied to many other problems. Buchbinder and Naor (2009) provide a detailed treatment of this framework. Note that competitive analysis focuses on worst–case scenarios. In contrast, we assume probabilistic information about the world and optimize for the average case. 13.3 (Noisy) Interactive Submodular Set Cover Recent work by Guillory and Bilmes (2010, 2011) considers a class of adaptive optimization problems over a family of monotone submodular objectives {fh : h ∈ H}. In their problem, one must cover a monotone 34 submodular objective fh∗ which depends on the (initially unknown) target hypothesis h∗ ∈ H, by adaptively issuing queries and getting responses. Unlike traditional pool-based active learning, each query may generate a response from a set of valid responses depending on the target hypothesis. The reward is calculated by evaluating fh∗ on the set of (query, response) pairs observed, and the goal is to obtain some threshold Q of objective value at minimum total query cost, where queries may have nonuniform costs. In the noisy variant of the problem (Guillory and Bilmes, 2011), the set of (query, response) pairs observed need not be consistent with any hypothesis in H, and the goal is to obtain Q of value for all hypotheses that are “close” to being consistent with the observations. For both variants, Guillory and Bilmes consider the worst-case policy cost, and provide greedy algorithms optimizing clever hybrid objective functions. They prove an approximation guarantee of ln(Q|H|) + 1 for integer valued objective functions {fh }h∈H in the noise–free case, and similar logarithmic approximation guarantees for the noisy case. While similar in spirit to this work, there are several significant differences between the two. Guillory and Bilmes focus on worst-case policy cost, while we focus mainly on average-case policy cost. The structure of adaptive submodularity depends on the prior p (φ), whereas there is no such dependence in Interactive Submodular Set Cover. This dependence in turn allows us to obtain results, such as Theorem 13 for self– certifying instances, whose approximation guarantee does not depend on the number of realizations in the way that the guarantees for Interactive Submodular Set Cover depend on |H|. As Guillory and Bilmes prove, the latter dependence is fundamental under reasonable complexity-theoretic assumptions6 . An interesting open problem within the adaptive submodularity framework that is highlighted by the work on Interactive Submodular Set Cover is to identify useful instance-specific properties that are sufficient to improve upon the worst-case approximation guarantee of Theorem 14. 13.4 Greedy Frameworks for Adaptive Optimization The paper that is perhaps closest in spirit to this work is the one on Stochastic Depletion problems by Chan and Farias (2009), who also identify a general class of adaptive optimization problems than can be near-optimally solved using greedy algorithms (which in their setting give a factor 2 approximation). However, the similarity is mainly on a conceptual level: The problems and approaches, as well as example applications considered, are quite different. 13.5 Stochastic Optimization with Recourse A class of adaptive optimization problems studied extensively in operations research since Dantzig (1955) is the area of stochastic optimization with recourse. Here, an optimization problem, such as Set Cover, Steiner Tree or Facility Location, is presented in multiple stages. At each stage, more information is revealed, but costs of actions increase. A key difference to the problems studied in this paper is that in these problems, information gets revealed independently of the actions taken by the algorithm. There are general efficient, sampling based (approximate) reductions of multi-stage optimization to the deterministic setting; see, e.g.,Gupta et al. (2005). 13.6 Bayesian Global Optimization Adaptive Stochastic Optimization is also related to the problem of Bayesian Global Optimization (for a recent survey of the area, c.f. Brochu et al. (2009)). In Bayesian Global Optimization, the goal is to adaptively select inputs in order to maximize an unknown function that is expensive to evaluate (and can possibly only be evaluated using noisy observations). A common approach that has been successful in many applications (for a recent application in machine learning, c.f. Lizotte et al. (2007)), is to assume a prior distribution, such as a Gaussian process, over the unknown objective function. Several criteria for selecting inputs have been developed, such as the Expected Improvement (Jones et al., 1998) criterion. However, while recently performance guarantees where obtained in the no-regret setting (Grünewälder et al., 2010; Srinivas et al., 2010), we are not aware of any approximation guarantees for Bayesian Global Optimization. 6. They reduce to Set Cover and use the result of Feige (1998), which requires the assumption NP * DTIME(nO(log log n) ), but it suffices to assume only P 6= NP using the Set Cover approximation hardness result of Raz and Safra (1997) instead. 35 13.7 Probabilistic Planning The problem of decision making under partial observability has also been extensively studied in stochastic optimal control. In particular, Partially Observable Markov Decision Processes (Smallwood and Sondik, 1973), abbreviated as POMDPs, are a general framework that captures many adaptive optimization problems under partial observability. Unfortunately, solving POMDPs is PSPACE hard (Papadimitriou and Tsitsiklis, 1987), thus typically heuristic algorithms with no approximation guarantees are applied (Pineau et al., 2006; Ross et al., 2008). For some special instances of POMDPs related to Multi-armed Bandit problems, (near-)optimal policies can be found. These include the (optimal) Gittins-index policy for the classic Multi-armed Bandit problem (Gittins and Jones, 1979) and approximate policies for the Multi-armed Bandit problem with metric switching costs (Guha and Munagala, 2009) and special cases of the Restless Bandit problem (Guha et al., 2009). The problems considered in this paper can be formalized as POMDPs, albeit with exponentially large state space (where the world state represents the selected items and state/outcome of each item). Thus our results can be interpreted as widening the class of partially observable planning problems that can be efficiently approximately solved. 13.8 Previous Work by the Authors & Subsequent Developments This manuscript is an extended version of a paper that appeared in the Conference on Learning Theory Golovin and Krause (2010). More recently, Golovin and Krause (2011b) proved performance guarantees for the greedy policy for the problem of maximizing the expected value of a policy under constraints more complex than simply selecting at most k items. These include matroid constraints, where a policy can only select independent sets of items and the greedy policy obtains a 1/2–approximation for adaptive monotone submodular objectives, and more generally p-independence system constraints, where the greedy policy obtains a 1/(p + 1)–approximation. Golovin et al. (2010) and, shortly thereafter, Bellala and Scott (2010), used the adaptive submodularity framework to obtain the first algorithms with provable (squared logarithmic) approximation guarantees for the difficult and fundamental problem of active learning with persistent noise. Finally, Golovin et al. (2011) used adaptive submodularity in the context of a dynamic conservation planning, and obtain competitiveness guarantees for an ecological reserve design problem. 14. Conclusions Planning under partial observability is a central but notoriously difficult problem in artificial intelligence. In this paper, we identified a novel, general class of adaptive optimization problems under uncertainty that are amenable to efficient, greedy (approximate) solution. In particular, we introduced the concept of adaptive submodularity, generalizing submodular set functions to adaptive policies. Our generalization is based on a natural adaptive analog of the diminishing returns property well understood for set functions. In the special case of deterministic distributions, adaptive submodularity reduces to the classical notion of submodular set functions. We proved that several guarantees carried by the non-adaptive greedy algorithm for submodular set functions generalize to a natural adaptive greedy algorithm in the case of adaptive submodular functions, for constrained maximization and certain natural coverage problems with both minimum cost and minimum sum objectives. We also showed how the adaptive greedy algorithm can be accelerated using lazy evaluations, and how one can compute data-dependent bounds on the optimal solution. We illustrated the usefulness of the concept by giving several examples of adaptive submodular objectives arising in diverse AI applications including sensor placement, viral marketing, automated diagnosis and pool-based active learning. Proving adaptive submodularity for these problems allowed us to recover existing results in these applications as special cases and lead to natural generalizations. Our experiments on real data indicate that adaptive submodularity can provide practical benefits, such as significant speed ups and tighter data-dependent bounds. We believe that our results provide an interesting step in the direction of exploiting structure to solve complex stochastic optimization and planning problems under partial observability. 36 Acknowledgments This research was partially supported by ONR grant N00014-09-1-1044, NSF grant CNS-0932392, NSF grant IIS-0953413, DARPA MSEE grant FA8650-11-1-7156, a gift by Microsoft Corporation, an Okawa Foundation Research Grant, and by the Caltech Center for the Mathematics of Information. We wish to thank Vitaly Feldman and Jan Vondrák for providing an elegant proof of Lemma 47. References Noga Alon, Baruch Awerbuch, Yossi Azar, Niv Buchbinder, and Joseph Seffi Naor. The online set cover problem. SIAM Journal on Computing, 39:361–370, June 2009. ISSN 0097-5397. doi: 10.1137/060661946. URL http://portal.acm.org/citation.cfm?id=1654887.1654888. Esther M. Arkin, Henk Meijer, Joseph S. B. Mitchell, David Rappaport, and Steven S. Skiena. Decision trees for geometric models. In Proceedings of Symposium on Computational Geometry, pages 369–378, New York, NY, USA, 1993. ACM. ISBN 0-89791-582-8. doi: http://doi.acm.org/10.1145/160985.161167. Arash Asadpour, Hamid Nazerzadeh, and Amin Saberi. Stochastic submodular maximization. In WINE ’08: Proceedings of the 4th International Workshop on Internet and Network Economics, pages 477–489, Berlin, Heidelberg, 2008. Springer-Verlag. ISBN 978-3-540-92184-4. doi: http://dx.doi.org/10.1007/ 978-3-540-92185-1 53. G. Bellala and C. Scott. Modified group generalized binary search with near-optimal performance guarantees. Technical report, University of Michigan, 2010. Eric Brochu, Mike Cora, and Nando de Freitas. A tutorial on Bayesian optimization of expensive cost functions, with application to active user modeling and hierarchical reinforcement learning. Technical Report TR-2009-23, Department of Computer Science, University of British Columbia, November 2009. Niv Buchbinder and Joseph (Seffi) Naor. The design of competitive online algorithms via a primal–dual approach. Foundations and Trends in Theoretical Computer Science, 3:93–263, February 2009. ISSN 1551-305X. doi: 10.1561/0400000024. URL http://portal.acm.org/citation.cfm?id= 1576247.1576248. Carri W. Chan and Vivek F. Farias. Stochastic depletion problems: Effective myopic policies for a class of dynamic optimization problems. Mathematics of Operations Research, 34(2):333–350, 2009. ISSN 0364-765X. doi: http://dx.doi.org/10.1287/moor.1080.0364. David A. Cohn, Zoubin Gharamani, and Michael I. Jordan. Active learning with statistical models. Journal of Artificial Intelligence Research (JAIR), 4:129–145, 1996. George B. Dantzig. Linear programming under uncertainty. Management Science, 1:197–206, 1955. Sanjoy Dasgupta. Analysis of a greedy active learning strategy. In NIPS: Advances in Neural Information Processing Systems 17, pages 337–344. MIT Press, 2004. B. Dean, M.X. Goemans, and J. Vondrák. Adaptivity and approximation for stochastic packing problems. In Proceedings of the 16th ACM-SIAM Symposium on Discrete Algorithms,, pages 395–404, 2005. B. Dean, M.X. Goemans, and J. Vondrák. Approximating the stochastic knapsack problem: The benefit of adaptivity. Mathematics of Operations Research, 33:945–964, 2008. A. Deshpande, C. Guestrin, S. Madden, J. Hellerstein, and W. Hong. Model-driven data acquisition in sensor networks. In Proceedings of the International Conference on Very Large Data Bases (VLDB), pages 588–599, 2004. 37 Uriel Feige. A threshold of ln n for approximating set cover. Journal of the ACM, 45(4):634 – 652, July 1998. Uriel Feige and Carsten Lund. On the hardness of computing the permanent of random matrices. Computational Complexity, 6(2):101–132, 1997. ISSN 1016-3328. doi: http://dx.doi.org/10.1007/BF01262928. Uriel Feige, László Lovász, and Prasad Tetali. Approximating min sum set cover. Algorithmica, 40(4): 219–234, 2004. Vitaly Feldman and Jan Vondrák. Private communication. 2012. Satoru Fujishige. Submodular functions and optimization, volume 58. Annals of Discrete Mathematics, North Holland, Amsterdam, 2nd edition, 2005. ISBN 0-444-52086-4. M. R. Garey and Ronald L. Graham. Performance bounds on the splitting algorithm for binary testing. Acta Informatica, 3:347–355, 1974. J. C. Gittins and D. M. Jones. A dynamic allocation index for the discounted multiarmed bandit problem. Biometrika, 66(3):561–565, December 1979. Michel X. Goemans and Jan Vondrák. Stochastic covering and adaptivity. In Proceedings of 7th International Latin American Symposium on Theoretical Informatics, pages 532–543, 2006. Daniel Golovin and Andreas Krause. Adaptive submodularity: A new approach to active learning and stochastic optimization. In 23rd Annual Conference on Learning Theory, pages 333–345, 2010. Daniel Golovin and Andreas Krause. Adaptive submodularity: Theory and applications in active learning and stochastic optimization. Journal of Artificial Intelligence Research (JAIR), 42:427–486, 2011a. Daniel Golovin and Andreas Krause. Adaptive submodular optimization under matroid constraints. CoRR, abs/1101.4450, 2011b. Daniel Golovin, Anupam Gupta, Amit Kumar, and Kanat Tangwongsan. All-norms and all-Lp -norms approximation algorithms. In Ramesh Hariharan, Madhavan Mukund, and V Vinay, editors, IARCS Annual Conference on Foundations of Software Technology and Theoretical Computer Science (FSTTCS 2008), Dagstuhl, Germany, 2008. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, Germany. Daniel Golovin, Andreas Krause, and Debajyoti Ray. Near-optimal Bayesian active learning with noisy observations. In NIPS: Advances in Neural Information Processing Systems 23, pages 766–774. 2010. Daniel Golovin, Andreas Krause, Beth Gardner, Sarah J. Converse, and Steve Morey. Dynamic resource allocation in conservation planning. In AAAI ’11: Proceedings of the Twenty–Fifth AAAI Conference on Artificial Intelligence, pages 1331–1336. AAAI Press, 2011. Pranava R. Goundan and Andreas S. Schulz. Revisiting the greedy approach to submodular set function maximization. Technical report, Massachusetts Institute of Technology, 2007. Steffen Grünewälder, Jean-Yves Audibert, Manfred Opper, and John Shawe-Taylor. Regret bounds for Gaussian process bandit problems. In Proceedings of the 13th International Conference on Artificial Intelligence and Statistics, 2010. Sudipto Guha and Kamesh Munagala. Multi-armed bandits with metric switching costs. In Proceedings of the International Colloquium on Automata, Languages and Programming (ICALP), 2009. Sudipto Guha, Kamesh Munagala, and Peng Shi. Approximation algorithms for restless bandit problems. Technical Report 0711.3861v5, arXiv, 2009. Andrew Guillory and Jeff Bilmes. Average-case active learning with costs. In The 20th International Conference on Algorithmic Learning Theory, University of Porto, Portugal, October 2009. 38 Andrew Guillory and Jeff Bilmes. Interactive submodular set cover. In Proceedings of the International Conference on Machine Learning (ICML), number UWEETR-2010-0001, Haifa, Israel, 2010. Andrew Guillory and Jeff A. Bilmes. Simultaneous learning and covering with adversarial noise. In International Conference on Machine Learning (ICML), Bellevue, Washington, 2011. Anupam Gupta, Martin Pál, R. Ravi, and Amitabh Sinha. What about Wednesday? Approximation algorithms for multistage stochastic optimization. In Proceedings of the 8th International Workshop on Approximation Algorithms for Combinatorial Optimization Problems (APPROX), 2005. Anupam Gupta, Ravishankar Krishnaswamy, Viswanath Nagarajan, and R. Ravi. Approximation algorithms for optimal decision trees and adaptive TSP problems. In Proceedings of the International Colloquium on Automata, Languages and Programming (ICALP), volume 6198 of Lecture Notes in Computer Science, pages 690–701. Springer, 2010. ISBN 978-3-642-14164-5. Donald R. Jones, Matthias Schonlau, and William J. Welch. Efficient global optimization of expensive black-box functions. Journal of Global Optimization, 13:455–492, 1998. Haim Kaplan, Eyal Kushilevitz, and Yishay Mansour. Learning with attribute costs. In Proceedings of the 37th ACM Symposium on Theory of Computing, pages 356–365, 2005. David Kempe, Jon Kleinberg, and Éva Tardos. Maximizing the spread of influence through a social network. In KDD ’03: Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 137–146, New York, NY, USA, 2003. ACM. ISBN 1-58113-737-0. doi: http: //doi.acm.org/10.1145/956750.956769. S. Rao Kosaraju, Teresa M. Przytycka, and Ryan S. Borgstrom. On an optimal split tree problem. In Proceedings of the 6th International Workshop on Algorithms and Data Structures, pages 157–168, London, UK, 1999. Springer-Verlag. ISBN 3-540-66279-0. A. Krause and C. Guestrin. Near-optimal nonmyopic value of information in graphical models. In Proceedings of Uncertainty in Artificial Intelligence (UAI), 2005. Andreas Krause and Carlos Guestrin. Near-optimal observation selection using submodular functions. In Conference on Artificial Intelligence (AAAI) Nectar track, pages 1650–1654, 2007. Andreas Krause and Carlos Guestrin. Optimal value of information in graphical models. Journal of Artificial Intelligence Research (JAIR), 35:557–591, 2009a. Andreas Krause and Carlos Guestrin. Intelligent information gathering and submodular function optimization. Tutorial at the International Joint Conference in Artificial Intelligence, 2009b. Jure Leskovec, Andreas Krause, Carlos Guestrin, Christos Faloutsos, Jeanne VanBriesen, and Natalie Glance. Cost-effective outbreak detection in networks. In KDD ’07: Proceedings of the 13th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 420–429, New York, NY, USA, 2007. ACM. ISBN 978-1-59593-609-7. doi: http://doi.acm.org/10.1145/1281192.1281239. M. Littman, J. Goldsmith, and M. Mundhenk. The computational complexity of probabilistic planning. Journal of Artificial Intelligence Research, 9:1–36, 1998. Zhen Liu, Srinivasan Parthasarathy, Anand Ranganathan, and Hao Yang. Near-optimal algorithms for shared filter evaluation in data stream systems. In SIGMOD ’08: Proceedings of the 2008 ACM SIGMOD international conference on Management of data, pages 133–146, New York, NY, USA, 2008. ACM. ISBN 978-1-60558-102-6. doi: http://doi.acm.org/10.1145/1376616.1376633. 39 Daniel Lizotte, Tao Wang, Michael Bowling, and Dale Schuurmans. Automatic gait optimization with Gaussian process regression. In Proceedings of the Twentieth International Joint Conference on Artificial Intelligence (IJCAI), pages 944–949, 2007. Donald W. Loveland. Performance bounds for binary testing with arbitrary weights. Acta Informatica, 22(1): 101–114, 1985. ISSN 0001-5903. doi: http://dx.doi.org/10.1007/BF00290148. Andrew McCallum and Kamal Nigam. Employing EM and pool-based active learning for text classification. In Proceedings of the International Conference on Machine Learning (ICML), pages 350–358, 1998. Michel Minoux. Accelerated greedy algorithms for maximizing submodular set functions. In Proceedings of the 8th IFIP Conference on Optimization Techniques, pages 234–243. Springer, 1978. Kamesh Munagala, Shivnath Babu, Rajeev Motwani, Jennifer Widom, and Eiter Thomas. The pipelined set cover problem. In Proceedings of the Intl. Conf. on Database Theory, pages 83–98, 2005. F. Nan and V. Saligrama. Comments on the proof of adaptive stochastic set cover based on adaptive submodularity and its implications for the group identification problem in “group-based active query selection for rapid diagnosis in time-critical situations”. IEEE Transactions on Information Theory, 63(11):7612–7614, Nov 2017. ISSN 0018-9448. doi: 10.1109/TIT.2017.2749505. George L. Nemhauser, Laurence A. Wolsey, and Marshall L. Fisher. An analysis of approximations for maximizing submodular set functions - I. Mathematical Programming, 14(1):265–294, 1978. Rob Nowak. Noisy generalized binary search. In NIPS: Advances in Neural Information Processing Systems 22, pages 1366–1374. 2009. C. H. Papadimitriou and J. N. Tsitsiklis. The complexity of Markov decision processses. Mathematics of Operations Research, 12(3):441–450, 1987. Joelle Pineau, Geoff Gordon, and Sebastian Thrun. Anytime point-based approximations for large POMDPs. Journal of Artificial Intelligence Research (JAIR), 27:335–380, 2006. Ran Raz and Shmuel Safra. A sub-constant error–probability low–degree test, and a sub–constant errorprobability PCP characterization of NP. In STOC ’97: Proceedings of the twenty-ninth annual ACM Symposium on Theory of Computing, pages 475–484, New York, NY, USA, 1997. ACM. ISBN 0-89791888-6. doi: http://doi.acm.org/10.1145/258533.258641. Stéphane Ross, Joelle Pineau, Sébastien Paquet, and Brahim Chaib-draa. Online planning algorithms for POMDPs. Journal of Artificial Intelligence Research, 32:663–704, July 2008. ISSN 1076-9757. URL http://portal.acm.org/citation.cfm?id=1622673.1622690. Alexander Schrijver. Combinatorial optimization : polyhedra and efficiency. Volume B, Part IV, Chapters 39-49. Springer, 2003. P. Sebastiani and H. P. Wynn. Maximum entropy sampling and optimal Bayesian experimental design. Journal of the Royal Statistical Society, Series B, 62(1):145–157, 2000. R.D. Smallwood and E.J. Sondik. The optimal control of partially observable Markov decision processes over a finite horizon. Operations Research, 21:1071–1088, 1973. Niranjan Srinivas, Andreas Krause, Sham Kakade, and Matthias Seeger. Gaussian process optimization in the bandit setting: No regret and experimental design. In Proceedings of the International Conference on Machine Learning (ICML), 2010. Matthew Streeter and Daniel Golovin. An online algorithm for maximizing submodular functions. Technical Report CMU-CS-07-171, Carnegie Mellon University, 2007. 40 Matthew Streeter and Daniel Golovin. An online algorithm for maximizing submodular functions. In NIPS: Advances in Neural Information Processing Systems 21, pages 1577–1584, 2008. Laurence A. Wolsey. An analysis of the greedy algorithm for the submodular set covering problem. Combinatorica, 2(4):385–393, 1982. 15. Additional Proofs and Incorporating Item Costs In this appendix we provide all of the proofs omitted from the main text. For the results of §5, we do so by first explaining how our results generalize to the case where items have costs, and then proving generalizations which incorporate item costs. 15.1 Incorporating Costs: Preliminaries In this section we provide the preliminaries required to define and analyze the versions of our problems with non-uniform item costs. We suppose P each item e ∈ E has a cost c(e), and the cost of a set S ⊆ E is given by the modular function c(S) = e∈S c(e). We define the generalizations of problems (1), (2), and (3) in §15.3, §15.4, and §15.5, respectively. Our results are with respect to the greedy policy π greedy and α-approximate greedy policies. With costs, the greedy policy selects an item maximizing ∆(e | ψ) /c(e), where ψ is the current partial realization. Definition 29 (Approximate Greedy Policy with Costs) A policy π is an α-approximate greedy policy if for all ψ such that there exists e ∈ E with ∆(e | ψ) > 0,    ∆(e | ψ) 1 ∆(e0 | ψ) π(ψ) ∈ e : ≥ max , c(e) α e0 c(e0 ) and π terminates upon observing any ψ such that ∆(e | ψ) ≤ 0 for all e ∈ E. That is, an α-approximate greedy policy always obtains at least (1/α) of the maximum possible ratio of conditional expected marginal benefit to cost, and terminates when no more benefit can be obtained in expectation. A greedy policy is any 1-approximate greedy policy. It will be convenient to imagine the policy executing over time, such that when a policy π selects an item e, it starts to run e, and finishes running e after c(e) units of time. We next generalize the definition of policy truncation. Actually we require three such generalizations, which are all equivalent in the unit cost case. Definition 30 (Strict Policy Truncation) The strict level t truncation of a policy π, denoted by π[←t] , is obtained by runningn π for t time units, and unselecting items whose runs o have not finished by time t. Formally, P π[←t] has domain ψ ∈ dom(π) : c(π(ψ)) + e∈dom(ψ) c(e) ≤ t , and agrees with π everywhere in its domain. Definition 31 (Lax Policy Truncation) The lax level t truncation of a policy π, denoted by π[t→] , is obtained by t time units, and selecting the items running at time t. Formally, π[t→] has domain n running π forP o ψ ∈ dom(π) : e∈dom(ψ) c(e) < t , and agrees with π everywhere in its domain. Definition 32 (Policy Truncation with Costs) The level-t-truncation of a policy π, denoted by π[t] , is a randomized policy obtained by running π for t time units, and if some item e has been running for 0 ≤ τ < c(e) time at time t, selecting e independently with probability τ /c(e). Formally, π[t] is a randomized policy that agrees with π everywhere in its domain, has dom(π[←t] ) ⊆ dom(π[t] ) ⊆ dom(π[t→] ) with certainty, and includes   each ψ ∈ dom(π[t→] ) \ dom(π[←t] ) in its domain independently with probability P t − e∈dom(ψ) c(e) /c(π(ψ)). 41 In the proofs that follow, we will need a notion of the conditional expected cost of a policy, as well as an alternate characterization of adaptive monotonicity, based on a notion of policy concatenation. We prove the equivalence of our two adaptive monotonicity conditions in Lemma 36. Definition 33 (Conditional Policy Cost) The conditional policy cost of π conditioned on ψ, denoted c (π | ψ), is the expected cost of the items π selects under p (φ | ψ). That is, c (π | ψ) := E [c(E(π, Φ)) | Φ ∼ ψ]. Definition 34 (Policy Concatenation) Given two policies π1 and π2 define π1 @π2 as the policy obtained by running π1 to completion, and then running policy π2 as if from a fresh start, ignoring the information gathered7 during the running of π1 . Definition 35 (Adaptive Monotonicity (Alternate Version)) A function f : 2E × OE → R≥0 is adaptive monotone with respect to distribution p (φ) if for all policies π and π 0 , it holds that favg (π) ≤ favg (π 0 @π), where favg (π) := E [f (E(π, Φ), Φ)] is defined w.r.t. p (φ). Lemma 36 (Adaptive Monotonicity Equivalence) Fix a function f : 2E × OE → R≥0 . Then ∆(e | ψ) ≥ 0 for all ψ with P [Φ ∼ ψ] > 0 and all e ∈ E if and only if for all policies π and π 0 , favg (π) ≤ favg (π 0 @π). Proof Fix policies π and π 0 . We begin by proving favg (π 0 @π) = favg (π@π 0 ). Fix any φ and note that E(π 0 @π, φ) = E(π 0 , φ) ∪ E(π, φ) = E(π@π 0 , φ). Hence favg (π 0 @π) = E [f (E(π 0 @π, Φ), Φ)] = E [f (E(π@π 0 , Φ), Φ)] = favg (π@π 0 ). Therefore favg (π) ≤ favg (π 0 @π) holds if and only if favg (π) ≤ favg (π@π 0 ). We first prove the forward direction. Suppose ∆(e | ψ) ≥ 0 for all ψ and all e ∈ E. Note the expression favg (π@π 0 ) − favg (π) can be written as P a conical combination of (nonnegative) ∆(e | ψ) terms, i.e., for 0 some α ≥ 0, favg (π@π 0 ) − favg (π) = ψ,e α(ψ,e) ∆(e | ψ). Hence favg (π@π ) − favg (π) ≥ 0 and so 0 0 favg (π) ≤ favg (π@π ) = favg (π @π). We next prove the backward direction, in contrapositive form. Suppose ∆(e | ψ) < 0 for some ψ with P [Φ ∼ ψ] > 0 and e ∈ E. Let e1 , . . . , er be the items in dom(ψ) and define policies π and π 0 as follows. For i = 1, 2, . . . , r, both π and π 0 select ei and observe Φ(ei ). If either policy observes Φ(ei ) 6= ψ(ei ) it immediately terminates, otherwise it continues. If π succeeds in selecting all of dom(ψ) then it terminates. If π 0 succeeds in selecting all of dom(ψ) then it selects e and then terminates. We claim favg (π@π 0 ) − favg (π) < 0. Note that E(π@π 0 , φ) = E(π, φ) unless φ ∼ ψ, and if φ ∼ ψ then E(π@π 0 , φ) = E(π, φ) ∪ {e} and also E(π, φ) = dom(ψ). Hence favg (π@π 0 ) − favg (π) = E [f (E(π@π 0 , Φ), Φ) − f (E(π, Φ), Φ)] = E [f (E(π@π 0 , Φ), Φ) − f (E(π, Φ), Φ) | Φ ∼ ψ] · P [Φ ∼ ψ] = E [f (dom(ψ) ∪ {e} , Φ) − f (dom(ψ), Φ) | Φ ∼ ψ] · P [Φ ∼ ψ] = ∆(e | ψ) · P [Φ ∼ ψ] The last term is negative, as P [Φ ∼ ψ] > 0 and ∆(e | ψ) < 0 by assumption. Therefore favg (π) > favg (π@π 0 ) = favg (π 0 @π), which completes the proof. 7. Technically, if under any realization φ policy π2 selects an item that π1 previously selected, then π1 @π2 cannot be written as a function from a set of partial realizations to E, i.e., it is not a policy. This can be amended by allowing partial realizations to be multisets over elements of E × O, so that, e.g., if e is played twice then (e, ψ(e)) appears twice in ψ. However, in the interest of readability we will avoid this more cumbersome multiset formalism, and abuse notation slightly by calling π1 @π2 a policy. This issue arises whenever we run some policy and then run another from a fresh start. 42 15.2 Adaptive Data Dependent Bounds with Costs The adaptive data dependent bound has the following generalization with costs. Lemma 37 (The Adaptive Data Dependent Bound with Costs) Suppose we have made observations ψ after selecting dom(ψ). Let π ∗ be any policy. Then for adaptive monotone submodular f : 2E × OE → R≥0   ∆(e | ψ) ∗ ∗ ∆(π | ψ) ≤ Z ≤ c (π | ψ) max (30) e c(e) P P ∗ where Z = maxw e∈E we ∆(e | ψ) : e c(e)we ≤ c (π | ψ) and ∀e ∈ E, 0 ≤ we ≤ 1 . Proof Order the items in dom(ψ) arbitrarily, and consider the policy π that for each e ∈ dom(ψ) in order selects e, terminating if Φ(e) 6= ψ(e) and proceeding otherwise, and, should it succeed in selecting all of dom(ψ) without terminating (which occurs iff Φ ∼ ψ), then proceeds to run π ∗ as if from a fresh start, forgetting the observations in ψ. By construction the expected marginal benefit of running the π ∗ portion of π conditioned on Φ ∼ ψ equals ∆(π ∗ | ψ). For all e ∈ E, let w(e) = P [e ∈ E(π, Φ) | Φ ∼ ψ] be the probability that e is selected when running π, conditioned on Φ ∼ ψ. Whenever some e ∈ E \ dom(ψ) is selected by π, the current partial realization ψ 0 contains ψ as a subrealization; hence adaptive submodularity ∗ implies ∆(e | ψ 0 ) ≤ ∆(e | ψ). It follows that the total contribution of e to ∆(πP | ψ) is upper bounded by ∗ w(e) · ∆(e | ψ). Summing over e ∈ E \ dom(ψ), we get a bound of ∆(π | ψ) ≤ e∈E\dom(ψ) w(e)∆(e | ψ). Next,Pnote that each e ∈ E \ dom(ψ) contributes w(e)c(e) cost to c (π ∗ | ψ). Hence it must be the case that e∈E\dom(ψ) w(e)c(e) ≤ c (π ∗ | ψ). Obviously, w(e) ∈ [0, 1] for all e, since w(e) is a probability. P Hence ∆(π ∗ | ψ) ≤ e∈E\dom(ψ) w(e)∆(e | ψ) ≤ Z because setting we = w(e) is feasible for the the linear program for which Z is the optimal value. To show Z ≤ c (π ∗ | ψ) maxe (∆(e | ψ) /c(e)), consider any feasible solution w to the linear program defining Z. It attains objective value     X X X ∆(e | ψ) ∆(e | ψ) ∆(e | ψ) ∗ we ∆(e | ψ) ≤ we c(e) ≤ we c(e) max ≤ c (π | ψ) max e∈E e∈E c(e) c(e) c(e) e∈E since e∈E P e∈E e∈E we c(e) ≤ c (π ∗ | ψ) by the feasibility of w. A simple greedy algorithm can be used to compute Z; we provide pseudocode for it in Algorithm 3. The correctness of this algorithm is more readily discerned upon rewriting the linear program using variables xe = c(e)we to obtain ( ) X X ∗ xe ≤ c (π | ψ) and ∀e ∈ E, 0 ≤ xe ≤ c(e) . Z = max xe (∆(e | ψ) /c(e)) : x e∈E e Intuitively, it is clear that to optimize xP we should shift mass towards variables with the highest ∆(e | ψ) /c(e) ratio. Clearly, any optimal solution has e xe = c (π ∗ | ψ). Moreover, in any optimal solution, ∆(e | ψ) /c(e) > ∆(e0 | ψ) /c(e0 ) implies xe = c(e) or xe0 = 0, since otherwise it would be possible to shift mass from xe0 to xe and obtain an increase in objective value. If the ∆(e | ψ) /c(e) values are distinct for distinct items, there will be a unique solution satisfying these constraints, which Algorithm 3 will compute. Otherwise, we imagine perturbing each ∆(e | ψ) by independent random quantities e drawn uniformly from [0, ] to make them distinct. This changes thePoptimum value by at most |E|, which vanishes as we let  tend towards zero. Hence any solution satisfying e xe = c (π ∗ | ψ) and ∆(e | ψ) /c(e) > ∆(e0 | ψ) /c(e0 ) implies xe = c(e) or xe0 = 0 is optimal. Since Algorithm 3 outputs the value of such a solution, it is correct. 43 Input: Groundset E; Partial realization ψ; Costs c : E → N; Budget C = c (π ∗ | ψ); Conditional expected marginal P benefits ∆(e | ψ) for Pall e ∈ E. ∗ Output: Z = maxw e∈E we ∆(e | ψ) : e c(e)we ≤ c (π | ψ) and ∀e ∈ E, 0 ≤ we ≤ 1 begin ∆(e2 | ψ) ∆(en | ψ) 1 | ψ) Sort E by ∆(e | ψ) /c(e), so that ∆(e c(e1 ) ≥ c(e2 ) ≥ . . . ≥ c(en ) ; Set w ← 0; i ← 0; a ← 0; z ← 0; e ← NULL; while a < C do i ← i + 1; e ← ei ; we ← min {1, C − a}; a ← a + c(e)we ; z ← z + we ∆(e | ψ); Output z; end Algorithm 3: Algorithm to compute the data dependent bound Z of Lemma 37. 15.3 The Max-Cover Objective With item costs, the Adaptive Stochastic Maximization problem becomes one of finding some π ∗ ∈ arg max favg (π[k] ) (31) π where k is a budget on the cost of selected items, and we define favg (π) for a randomized policy π to be favg (π) := E [f (E(π, Φ), Φ)] as before, where the expectation is now over both Φ and the internal randomness of π which determines E(π, φ) for each φ. We prove the following generalization of Theorem 5. Theorem 38 Fix any α ≥ 1 and item costs c : E → N. If f is adaptive monotone and adaptive submodular with respect to the distribution p (φ), and π is an α-approximate greedy policy, then for all policies π ∗ and positive integers ` and k   ∗ favg (π[`] ) > 1 − e−`/αk favg (π[k] ). Proof The proof goes along the lines of the performance analysis of the greedy algorithm for maximizing a submodular function subject to a cardinality constraint of Nemhauser et al. (1978). An extension of that analysis to α-approximate greedy algorithms, which is analogous to ours but for the nonadaptive case, is shown by Goundan and Schulz (2007). For brevity, we will assume without loss of generality that π = π[`] ∗ and π ∗ = π[k] . Then for all i, 0 ≤ i < `  favg (π ∗ ) ≤ favg (π[i] @π ∗ ) ≤ favg (π[i] ) + αk favg (π[i+1] ) − favg (π[i] ) . (32) The first inequality is due to the adaptive monotonicity of f and Lemma 36, from which we may infer favg (π2 ) ≤ favg (π1 @π2 ) for any π1 and π2 . The second inequality may be obtained as a corollary of Lemma 37  as follows. Fix any partial realization ψ of the form (e, φ(e)) : e ∈ E(π[i] , φ) for some φ. Consider ∆(π ∗ | ψ), which equals the expected marginal benefit of the π ∗ portion of π[i] @π ∗ conditioned on Φ ∼ ψ. Lemma 37 allows us to bound it as E [∆(π ∗ | ψ)] ≤ E [c (π ∗ | ψ)] · max (∆(e | ψ) /c(e)) , e where the expectations are taken over the internal randomness of π ∗ , if there is any. Note that since 0 π ∗ has the form π[k] for some π 0 we know that for all φ, E [c(E(π ∗ , φ))] ≤ k, where the expectation is again taken over the internal randomness of π ∗ . Hence E [c (π ∗ | ψ)] ≤ k for all ψ. It follows that E [∆(π ∗ | ψ)] ≤ k · maxe (∆(e | ψ) /c(e)). By definition of an α-approximate greedy policy, π obtains at least (1/α) maxe (∆(e | ψ) /c(e)) ≥ E [∆(π ∗ | ψ)] /αk expected marginal benefit per unit cost in a step immediately following its observation of ψ. Next we take an appropriate convex combination of 44 the previousinequality with different values of ψ. Let Ψ be a random partial realization distributed as  pΨ (ψ) := P Ψ = ψ | ∃φ. ψ = (e, φ(e)) : e ∈ E(π[i] , φ) Then    1 ∆(e | Ψ) favg (π[i+1] ) − favg (π[i] ) ≥ E max α e c(e)   E [∆(π ∗ | Ψ)] ≥ E αk favg (π[i] @π ∗ ) − favg (π[i] ) = αk A simple rearrangement of terms then yields the second inequality in (32). Now define ∆i := favg (π ∗ ) − favg (π[i] ), so that (32) implies ∆i ≤ αk(∆i − ∆i+1 ), from which we infer   1 1 ` ∆i+1 ≤ 1 − αk ∆0 < e−`/αk ∆0 , where for this last inequality we have ∆i and hence ∆` ≤ 1 − αk  −x used the fact that 1 − x < e for all x > 0. Thus favg (π ∗ ) − favg (π[`] ) < e−`/αk favg (π ∗ ) − favg (π[0] ) ≤ e−`/αk favg (π ∗ ) so favg (π) > (1 − e−`/αk )favg (π ∗ ). 15.4 The Min-Cost-Cover Objective In this section, we provide arbitrary item cost generalizations of Theorem 13 and Theorem 14. With item costs the Adaptive Stochastic Minimum Cost Cover problem becomes one of finding, for some quota on utility Q, π ∗ ∈ arg min cavg (π) such that f (E(π, φ), φ) ≥ Q for all φ, (33) π where cavg (π) := E [c(E(π, Φ))]. Without loss of generality, we may take a truncated version of f , namely (A, φ) 7→ min {Q, f (A, φ)}, and rephrase Problem (33) as finding π ∗ ∈ arg min cavg (π) such that π covers φ for all φ. (34) π Hereby, recall that π covers φ if E [f (E(π, φ), φ)] = f (E, φ), where the expectation is over any internal randomness of π. We will consider only Problem (34) for the remainder. We also consider the worstcase variant of this problem, where we replace the expected cost cavg (π) objective with the worst-case cost cwc (π) := maxφ c(E(π, φ)). The definition of coverage (Definition 7 in §5.2 on page 13) requires no modification to handle item costs. Note, however, that coverage is all-or-nothing in the sense that covering a realization φ with probability less than one does not count as covering it. A corollary of this is that only items whose runs have finished help with coverage, whereas currently running items do not. For a simple example, consider the case where E = {e}, c(e) = 2, f (A, φ) = |A|, and policy π that selects e and then terminates. Then π[1] is a randomized policy which is π with probability 12 , and is the empty policy with probability 12 , so E [f (E(π, φ), φ)] = 12 < 1 = f (E, φ) for each φ. Hence, even though half the time π[1] covers all realizations, it is counted as not covering any. We begin with a claim relating pointwise submodularity to strong adaptive submodularity. Lemma 39 If f is adaptive submodular with respect to p (φ) and f is pointwise submodular meaning S 7→ f (S, φ) is submodular for all φ, then f is strongly adaptive submodular with respect to p (φ). Proof By assumption f is adaptive submodular with respect to p (φ), so it is sufficient to prove that Eq. (16) holds, i.e., ∆(e | ψ; ψ 0 ) ≥ ∆(e | ψ 0 ). Fix any ψ ⊆ ψ 0 and e ∈ E. Let δφ (e, S) := f (S ∪ {e} , φ) − f (S, φ). 45 From the definition of ∆(e | ψ; ψ 0 ) and ∆(e | ψ 0 ), we have X ∆(e | ψ; ψ 0 ) = P [φ | ψ 0 ] δφ (e, dom(ψ)) φ ≥ X = ∆(e | ψ 0 ) P [φ | ψ 0 ] δφ (e, dom(ψ 0 )) φ where we have used δφ (e, dom(ψ)) ≥ δφ (e, dom(ψ 0 )) by the pointwise submodularity of f . Now we provide an approximation guarantee for the average-case policy cost with arbitrary item costs. Theorem 40 Suppose f : 2E × OE → R≥0 is strongly adaptive submodular and strongly adaptive monotone with respect to p (φ) and there exists Q such that f (E, φ) = Q for all φ. Let η be any value such that f (S, φ) > Q − η implies f (S, φ) = Q for all S and φ. Let δ = minφ p (φ) be the minimum probability of any ∗ realization. Let πavg be an optimal policy minimizing the expected cost of items selected to guarantee every realization is covered. Let π be an α-approximate greedy policy with respect to the item costs. Then in general    2 Q ∗ cavg (π) ≤ α cavg (πavg ) ln +1 δη and for self–certifying instances cavg (π) ≤ ∗ α cavg (πavg )    2 Q ln +1 . η Note that if range(f ) ⊂ Z, then η = 1 is a valid choice, so for general and self–certifying instances we have 2 2 ∗ ∗ ) (ln(Q) + 1) , respectively. ) (ln(Q/δ) + 1) and cavg (π) ≤ α cavg (πavg cavg (π) ≤ α cavg (πavg Proof of Theorem 40 We first require some additional definitions. We extend f by defining f (A, ψ) via f (A, ψ) := E [f (A, Φ) | Φ ∼ ψ] (35) f (ψ) := f (dom(ψ), ψ) = E [f (dom(ψ), Φ) | Φ ∼ ψ] (36) Further, let Definition 41 (Execution Trace) Given a policy π and a realization φ, let the trace τ = τ (φ) of π under φ be the sequence of partial realizations specifying its set of observations through time, i.e., ψ 0 ⊂ ψ 1 ⊂ · · · ⊂ ψ ` , such that dom(ψ i ) \ dom(ψ i−1 ) consists precisely of the ith item selected by π. Definition 42 (Execution Arborescence) For a policy π, define its execution arborescence as the digraph (dom(π), A) with vertices dom(π) and edges A := {(ψ, ψ 0 ) : ∃φ, ψ 0 = ψ ∪ {(π(ψ), φ(π(ψ)))}} . In other words, its vertices are the partial realizations that may be encountered by π, and it has a directed edge from ψ to ψ 0 if it may immediately encounter ψ 0 after ψ in some execution trace. Definition 43 (Spanning Edges, Sources, Targets, and Successors) Given a policy π with execution arborescence with arcs A, and some positive x ∈ R, let Ax be the edges that span x expected benefit, i.e., Ax := {(ψ, ψ 0 ) : (ψ, ψ 0 ) ∈ A, f (ψ) < x ≤ f (ψ 0 )} 46 Further let Ux be the sources of edges in Ax and Vx be the targets. That is, Ux := {ψ : ∃ψ 0 , (ψ, ψ 0 ) ∈ Ax } and Vx := {ψ 0 : ∃ψ, (ψ, ψ 0 ) ∈ Ax }. Finally, the successors Sx (ψ) of some ψ ∈ Ux are the sources ψ 0 that are minimal supersets of ψ in Ux when interpreting partial realizations as set of (item, observation) pairs. Formally, Sx (ψ) := {ψ 0 : ψ 0 ∈ Ux , @ψ 00 , ψ ⊂ ψ 00 ⊂ ψ 0 } Definition 44 (Concatenative Pseudopolicies) Given two policies π and π ∗ and some nonnegative x ∈ R, we define the concatenative pseudopolicy π[x]@π ∗ as the stochastic process that, on each realization φ, runs policy π until it is just about to achieve x expected reward (formally, until reaching ψ (φ, x)), and then runs π ∗ as if from scratch. Hence on realization φ it will play dom(ψ (φ, x)) ∪ E(π ∗ , φ) where ψ (φ, x) is defined relative to π. Note that π[x]@π ∗ isn’t a proper policy, because it decides when to switch to executing π ∗ based on information about φ that it cannot infer from its current partial realization. Specifically, upon reaching ψ ∈ Ux it will select π ∗ (∅) if there exists ψ 0 such that (ψ, ψ 0 ) ∈ Ax and φ ∼ ψ 0 , and will select π(ψ) otherwise. T HE S TRATEGY Our overall strategy will be to bound the expected cost cavg (π) of π by bounding the price θ it pays per unit of expected reward gained as it runs (measured as E [f (dom(ψ), Φ) | Φ ∼ ψ] where ψ is the current partial realization) and then integrating over the run. To bound the price π pays, we define an alternative cost scheme ∗ for policies, bound the price π pays in terms of the alternative cost of an optimal policy πavg , and finally bound ∗ the alternative cost of πavg in terms of its actual cost. P ROGRESS BOUNDS First, observe that  ∗ ∆ πavg | ψ; ψ 0 ≥ Q − x for all (ψ, ψ 0 ) ∈ Ax ∗ x by definition of Ax , and f (E(πavg , φ), φ) =  ∗ maximum possible reward, if ∆ πavg | ψ; ψ 0 0 0 (37) ∗ πavg This holds because f (ψ) < Q for all φ since covers every realization. Since Q is the < Q − x then we can generate a ∗ violation of strong adaptive monotonicity by fixing some φ ∼ ψ , selecting E(πavg , φ0 ), and then selecting dom(ψ) to reduce the expected reward. The bound in Eq. (37) is useful when x is far from Q, however we will require a stronger bound for x very close to Q. We will use slightly different analyses for general instances and for self–certifying instances. We begin with general instances. For these, we will prove a stronger bound for obtaining the final δη reward, i.e., for x ∈ [Q − δη, Q].  ∗ ∆ πavg | ψ; ψ 0 ≥ δη for all (ψ, ψ 0 ) ∈ Ax (38) Fix ψ ∈ dom(π) and any φ0 ∼ ψ. We say ψ covers φ0 if π covers φ0 by the time it observes ψ. By definition of δ and η, if some φ0 ∼ ψ 0 is not covered by ψ then Q − f (dom(ψ), ψ 0 ) ≥ δη. Hence the last item that π selects, say upon observing ψ, must increase its conditional expected value from f (dom(ψ), ψ 0 ) ≤ Q − δη to Q for all φ ∼ ψ 0 . For self–certifying instances we show that  ∗ ∆ πavg | ψ; ψ 0 ≥ η for all (ψ, ψ 0 ) ∈ Ax (39) using a similar argument for x ∈ [Q − η, Q]. We first argue that the last item that π selects must increase its conditional expected value from at most Q − η to Q. For suppose π currently observes ψ, and has not achieved conditional value Q, i.e., f (ψ) < Q. Then some φ ∼ ψ is uncovered. Since the instance is self–certifying, every φ with φ ∼ ψ is uncovered, and has f (dom(ψ), φ) < f (E, φ) = Q. By definition of η, for each φ with φ ∼ ψ we then have f (dom(ψ), φ) ≤ Q − η, which implies f (dom(ψ), ψ 0 ) ≤ Q − η. 47 U PPER BOUNDING THE PRICE π PAYS WITH A LTERNATE C OSTS ∗ Next we upper-bound the price π pays in terms of an alternative pricing scheme, ĉ, for πavg . Note the cost of ∗ ∗ πavg under ĉ may be higher than cavg (πavg ). We will eventually bound the alternative cost in terms of the true cost. ∗ Fix x and any e that may be played by πavg under some realization φ. Consider any (ψ, ψ 0 ) ∈ Ax with 0 ∗ ∗ φ ∼ ψ . Fix some ψ e ∈ dom(πavg ) such that e = πavg (ψ e ). We charge e a price of c(e)/∆(e | ψ ∪ ψ e ) per-unit reward in this case. Hence, under realization φ, the alternative cost is   f (dom(ψ) ∪ dom(ψ e ) ∪ {e} , φ) − f (dom(ψ) ∪ dom(ψ e ), φ) c(e) ∆(e | ψ ∪ ψ e ) Taking the expectation over φ such that φ ∼ ψ 0 and φ ∼ ψ e (which is equivalent to φ ∼ ψ 0 ∪ ψ e ) yields an expected alternative cost of   ∆(e | ψ ∪ ψ e ; ψ 0 ∪ ψ e ) ĉe (ψ, ψ 0 ) := c(e) (40) ∆(e | ψ ∪ ψ e ) For brevity, we will define ∆a,e := ∆(e | ψ ∪ ψ e ; ψ 0 ∪ ψ e ) ∗ for arc a = (ψ, ψ 0 ) ∈ Ax and ψ e . Note we could have multiple ψ e such that πavg (ψ e ) = e, however for notational convenience we can imagine creating separate copies of e for each such partial realization and ∗ assume without loss of generality that for each item there is a unique partial realization that πavg plays it under; ∗ That is, for each e there is at most one ψ e such that πavg (ψ e ) = e. ∗ Next, for arc a = (ψ, ψ 0 ) define wa,e to be the probability that πavg plays e conditioned on ψ 0 :   ∗ wa,e := P e ∈ E(πavg , Φ) | Φ ∼ ψ 0 = P [ψ e | ψ 0 ]  ∗ ∗ Next, consider the cost ĉ πavg | ψ; ψ 0 which is the expected alternative cost paid by πavg starting from 0 0 dom(ψ) and conditioned on ψ for some fixed x and a = (ψ, ψ ) ∈ Ax . Note that since the alternative cost depends on x and a, this alternative cost does as well. Formally, X  ∗ ĉ πavg | ψ; ψ 0 := wa,e ĉe (ψ, ψ 0 ) e  P ∗ ∗ The expected benefit of running πavg starting from ψ and conditioning on ψ 0 is ∆ πavg | ψ; ψ 0 := e wa,e ∆a,e . Hence X  ∗ ĉ πavg | ψ; ψ 0 = wa,e ĉe (ψ, ψ 0 ) (41) e   ∆a,e = wa,e c(e) ∆(e | ψ ∪ ψ e ) e   X c(e0 ) ≥ wa,e ∆a,e min e0 ∆(e0 | ψ ∪ ψ e ) e   X c(e0 ) ≥ wa,e ∆a,e min e0 ∆(e0 | ψ) e    c(e0 ) ∗ 0 = ∆ πavg | ψ; ψ min e0 ∆(e0 | ψ) X Hence  min 0 e c(e0 ) ∆(e0 | ψ)  ≤ 48 ∗ ĉ πavg | ψ; ψ 0 (42) (43) (44) (45)  ∗ | ψ; ψ 0 ∆ πavg  (46) Next, note that Vx := {ψ 0 : ∃ψ, (ψ, ψ 0 ) ∈ Ax } partitions the set of realizations. We define the alternative ∗ cost of πavg as X  ∗ ∗ ĉavg (πavg ) := ĉ πavg | ψ; ψ 0 P [ψ 0 ] (47) (ψ,ψ 0 )∈Ax Note it too is implicitly dependent on x. B OUNDING THE A LTERNATE C OSTS ∗ ∗ Next we bound ĉavg (πavg ) in terms of cavg (πavg ) and problem parameters Q, η and δ. Ultimately we will show that for general instances     Q ∗ ∗ +1 (48) ĉavg (πavg ) ≤ cavg (πavg ) ln δη and for self–certifying instances ∗ ĉavg (πavg ) ≤ ∗ cavg (πavg )     Q ln +1 . η (49) It is sufficient to prove this bound element by element. That is, we define X ĉ(e) := ĉe (ψ, ψ 0 ) P [ψ 0 ] (ψ,ψ 0 )∈Ax         Q and prove ĉ(e)/c(e) ≤ ln δη + 1 for general instances, and ĉ(e)/c(e) ≤ ln Q + 1 for self– η certifying instances. ∗ To simplify the exposition we define some additional notation. Fix π, πavg , x, and e. Let q[ψ] := P [ψ | ψ e ] and let X q 0 [ψ] := q[ψ 0 ] ψ 0 :(ψ,ψ 0 )∈Ax Let ∆[ψ] := ∆(e | ψ ∪ ψ e ). Let δ[ψ] :=  X ψ 0 :(ψ,ψ 0 )∈Ax q[ψ 0 ] q 0 [ψ]  ∆(e | ψ ∪ ψ e ; ψ 0 ∪ ψ e ) be the conditional expected benefit of selecting e in the concatenative pseudopolicy π[x]@π ∗ immediately upon seeing ψ (see definition 44). By construction, {ψ 0 : ψ 0 ∈ Vx } partitions the set of all realizations (since for each φ there is a unique edge in the execution arborescence where the policy passes x expected reward in the trace τ (φ) as the policy runs). Furthermore, conditioned on ∃(ψ, ψ 0 ) ∈ Ax such that Φ ∼ ψ 0 , we pay c(e)δ[ψ]/∆[ψ]. Taking the expectation over ψ ∈ Ux , we obtain X q 0 [ψ]δ[ψ] ĉ(e) = (50) c(e) ∆[ψ] ψ∈Ux Next we need a subclaim about how ∆[ψ] relates to to ∆[ψ 0 ] for ψ ⊂ ψ 0 . Recall the definition of Sx (ψ) (from definition 43). ∗ Lemma 45 Fix π, πavg , x, and e and let pi , p0i , ∆[ψ], and δ[ψ] be as above. For any ψ ∈ Ux , if for some ∈R q 0 [ψ]δ[ψ] ≥ (q 0 [ψ] +  (q[ψ] − q 0 [ψ])) ∆[ψ] then E [∆[ψ 0 ] | ψ 0 ∈ Sx (ψ)] := X P [ψ 0 ∪ ψ e | ψ ∪ ψ e ] ∆[ψ 0 ] ≤ ∆[ψ] (1 − ) ψ 0 ∈Sx (ψ) 49 Proof Fix ψ. We claim that by strong adaptive submodularity, X q[ψ]∆[ψ] ≥ q 0 [ψ]δ[ψ] + q[ψ v ]∆[ψ v ] (51) ψ v ∈Sx (ψ) To see this, note that {ψ 0 : (ψ, ψ 0 ) ∈ Ax } ∪ {ψ v ∈ Sx (ψ)} partitions the set of realizations, and thus X q[ψ]∆[ψ] = q 0 [ψ]δ[ψ] + P [ψ v | ψ e ] ∆(e | ψ ∪ ψ e ; ψ v ∪ ψ e ) (52) ψ v ∈Sx (ψ) ≥ X 0 q [ψ]δ[ψ] + P [ψ v | ψ e ] ∆(e | ψ v ∪ ψ e ) (53) q[ψ v ]∆[ψ v ] (54) ψ v ∈Sx (ψ) ≡ X q 0 [ψ]δ[ψ] + ψ v ∈Sx (ψ) Given Eq. (51) and the assumption that q 0 [ψ]δ[ψ] ≥ (q 0 [ψ] + (q[ψ] − q 0 [ψ])) ∆[ψ], we have X q[ψ]∆[ψ] ≥ (q 0 [ψ] + (q[ψ] − q 0 [ψ])) ∆[ψ] + q[ψ v ]∆[ψ v ] ψ v ∈Sx (ψ) Hence X q[ψ v ]∆[ψ v ] ≤ (q[ψ] − q 0 [ψ])(1 − )∆[ψ] ψ v ∈Sx (ψ) Dividing by (q[ψ] − q 0 [ψ]) yields X P [ψ v | ψ ∪ ψ e ] ∆[ψ v ] = ψ v ∈Sx (ψ)  X ψ v ∈Sx (ψ) q[ψ v ] q[ψ] − q 0 [ψ]  ∆[ψ v ] ≤ (1 − )∆[ψ] which completes the proof. Recall we wish to upper-bound ĉ(e) c(e) using Eq. (50). Let ε(ψ) ∈ R be set such that q 0 [ψ]δ[ψ] = (q 0 [ψ] + ε(ψ) (q[ψ] − q 0 [ψ])) ∆[ψ] for all ψ. Then ĉ(e) c(e) = = X q 0 [ψ]δ[ψ] ∆[ψ] ψ∈Ux X (q 0 [ψ] + ε(ψ) (q[ψ] − q 0 [ψ])) (55) (56) ψ∈Ux = X q 0 [ψ] + ψ∈Ux = 1+ X ε(ψ) (q[ψ] − q 0 [ψ]) (57) ψ∈Ux X ε(ψ) (q[ψ] − q 0 [ψ]) (58) ε(ψ)+ (q[ψ] − q 0 [ψ]) (59) ψ∈Ux ≤ 1+ X ψ∈Ux (60) where ε(ψ)+ := max (0, ε(ψ)). 50 Lemma 46 For general instances X ε(ψ)+ (q[ψ] − q 0 [ψ]) ≤ ln  Q δη   Q η  ψ∈Ux and for self–certifying instances X 0 + ε(ψ) (q[ψ] − q [ψ]) ≤ ln ψ∈Ux Proof From Lemma 45, we have E [∆[ψ 0 ] | ψ 0 ∈ Sx (ψ) and ψ e ] ≤ ∆[ψ] (1 − ε(ψ)) (61) From adaptive submodularity, we also have ∆[ψ 0 ] ≤ ∆[ψ] for all ψ 0 ∈ Sx (ψ), and hence E [∆[ψ 0 ] | ψ 0 ∈ Sx (ψ) and ψ e ] ≤ ∆[ψ]. Hence  E [∆[ψ 0 ] | ψ 0 ∈ Sx (ψ) and ψ e ] ≤ ∆[ψ] 1 − ε(ψ)+ Note X ε(ψ)+ (q[ψ] − q 0 [ψ]) ≤ ψ∈Ux X ε(ψ)+ q[ψ] ψ∈Ux We will bound the latter sum by constructing an unbiased estimate of it, and taking the expectation of that estimate. Our unbiased estimator is based on a random trace τ (Φ) of π X R(Φ) := ε(ψ)+ (62) ψ∈Ux ,ψ∈τ (Φ) where Φ is sampled conditioned on Φ ∼ ψ e , i.e., each φ is sampled with probability P [φ | ψ e ]. For any ψ ∈ Ux , the probability of seeing ψ conditioned on e being selected (i.e. Φ ∼ ψ e ) is exactly q[ψ]. Hence X E [R(Φ) | Φ ∼ ψ e ] := q[ψ]ε(ψ)+ (63) ψ∈Ux ∩τ (Φ) Next consider the random sequence of partial realizations in Ux ∩τ (Φ), which we will denote as ψ 1 , ψ 2 , . . . , ψ k . We assume without loss of generality that no policy selects an item that will give it zero reward in expectation. Let Yi (Φ) be a random variable such that ∆[ψ i+1 ] = (1 − Yi )∆[ψ i ] By adaptive submodularity, Yi ≥ 0 for all i, and also E [Yi | Φ ∼ ψ e ] = ε(ψ i ) for all i from Eq. (61). Next, note that because the maximum possible reward is Q and the minimum increment is η, and the minimum probability of any realization is δ, we have for all ψ ∆[ψ] ≤ Q, and ∆[ψ] > 0 ⇒ ∆[ψ] ≥ δη. Hence ∆[ψ k ] = ∆[ψ 1 ] k−1 Y (1 − Yi ) i=1 implies k−1 Y (1 − Yi ) = ∆[ψ k ]/∆[ψ 1 ] ≥ δη/Q i=1 Using 1 − x ≤ ex , we conclude δη/Q ≤ k−1 Y k−1 X i=1 i=1 (1 − Yi ) ≤ exp − 51 ! Yi and hence Pk−1 i=1 Yi ≤ ln  Q δη  . Hence in general,  E [R(Φ) | Φ ∼ ψ e ] =  X E ε(ψ)+ | Φ ∼ ψ e  (64) ψ∈Ux ,ψ∈τ (Φ) = X E [Yi (Φ) | Φ ∼ ψ e ] (65) i≥1 =   X E Yi (Φ) | Φ ∼ ψ e  (66) i≥1  ≤ ln Q δη  (67) For self–certifying instances, the argument is similar, except that we have ∆[ψ] > 0 ⇒ ∆[ψ] ≥ η because the the minimum increment is η,and the reward is determined by the currentpartial realization. Hence  Pk−1 Q Q ∆[ψ k ] ≥ η, and so i=1 Yi ≤ ln η in this case, so E [R(Φ) | Φ ∼ ψ e ] ≤ ln η . P UTTING IT ALL TOGETHER We show how to obtain the result for general instances. The result for self–certifying instances is strictly analogous. From Eq. (37) and Eq. (38), we have that for all (ψ, ψ 0 ) ∈ Ax  ∗ ∆ πavg | ψ; ψ 0 ≥ max (Q − x, δη) . (68) From Eq. (46) and the fact that π is an α-approximate greedy policy by assumption we have that for all (ψ, ψ 0 ) ∈ Ax  ∗ αĉ πavg | ψ; ψ 0 c(π(ψ)) . ≤ (69) ∗ | ψ; ψ 0 ∆(π(ψ) | ψ) ∆ πavg Combining this with Eq. (68) yields  ∗ | ψ; ψ 0 αĉ πavg c(π(ψ)) ≤ ∆(π(ψ) | ψ) max (Q − x, δη) (70) for all (ψ, ψ 0 ) ∈ Ax . The expected price π pays per unit of expected reward at the instant it reaches x expected reward is θ(x) := X P [ψ 0 ] (ψ,ψ 0 )∈Ax ≤ X (ψ,ψ 0 )∈Ax = c(π(ψ)) ∆(π(ψ) | ψ)  ∗ αĉ πavg | ψ; ψ 0 P [ψ ] max (Q − x, δη) 0 ∗ αĉavg (πavg ) max (Q − x, δη) since {ψ 0 : ∃ψ, (ψ, ψ 0 ) ∈ Ax } partitions the set of realizations. Next, we can apply Lemma 47 to bound cavg (π) as follows. 52 (71) (72) (73) Z cavg (π) Q = θ(x)dx (74) x=0 Z Q ∗ αĉavg (πavg ) dx max (Q − x, δη) x=0 Z Q 1 ∗ = αĉavg (πavg ) dx x=0 max (Q − x, δη) Z Q    1 Q ∗ +1 dx ≤ αcavg (πavg ) ln δη x=0 max (Q − x, δη)    2 Q ∗ ≤ αcavg (πavg ) ln +1 δη ≤ (75) (76) (77) (78) This completes the proof for general instances (as mentioned, the result for self–certifying instances is strictly ∗ ∗ analogous, merely with different bounds used to relate ĉavg (πavg ) to cavg (πavg )), aside from proving Lemma 47. The elegant proof that we present here is due to Feldman and Vondrák (2012). Lemma 47 Suppose f : 2E × OE → R≥0 is adaptive submodular and strongly adaptive monotone with respect to p (φ) and there exists Q such that f (E, φ) = Q for all φ. Fix any policy π that covers every realization, and let θ(x) be defined as in the proof of Theorem 40. Then Z Q cavg (π) = θ(x)dx. x=0 Proof Recall θ(x) := X P [ψ 0 ] (ψ,ψ 0 )∈Ax c(π(ψ)) ∆(π(ψ) | ψ) We use an alternative charging scheme to compute the expected cost of π: When the policy observes ψ and selects e = π(ψ), we charge it based on the outcome o = φ(e). Specifically, we charge it proportionally to its gain in expected reward, as measured by ψ 7→ E [f (dom(ψ), Φ) | Φ ∼ ψ]. Formally, upon observing ψ 0 := ψ ∪ {(e, o)} after selecting e = π(ψ), the algorithm is charged   E [f (dom(ψ 0 ), Φ) | Φ ∼ ψ 0 ] − E [f (dom(ψ), Φ) | Φ ∼ ψ] Ce (ψ, ψ 0 ) := c(e) . (79) ∆(e | ψ) It is straightforward to show that E [Ce (ψ, ψ ∪ {(e, Φ(e))}) | Φ ∼ ψ] = c(e), using the basic fact that if Pevent A is the union of finitely many disjoint events B1 , . . . , Bm and X is any random variable, then i P [Bi | A] E [X | Bi ] = E [X | A]. Hence, in expectation this charging scheme charges the policy exactly cavg (π). Moreover, the rate at which it charges π upon observing ψ and selecting e = π(ψ) is exactly c(e)/∆(e | ψ) per unit expected gain. Let θ(ψ) := c(π(ψ))/∆(π(ψ) | ψ). Under this charging scheme the policy pays θ(ψ)dx to obtain an additional dx expected reward, given it already has x, the true realization is φ, and (ψ, ψ 0 ) ∈ Ax for some ψ 0 with φ ∼ ψ 0 . Let ψ (φ, x) be the partial realization in the execution trace of π under φ immediately before it obtains x reward in expectation. Formally ψ (φ, x) is the maximal element of dom(π) such that φ ∼ ψ and f (ψ) < x, where f (ψ) is defined as in Eq. (36). RQ The total alternative cost of π under φ is then x=0 θ(ψ (φ, x))dx, and we can compute the expected cost of π as "Z # Q cavg (π) = E θ(ψ (Φ, x))dx . x=0 53 where the expectation is taken with respect to Φ. We can then obtain the claimed equality by exchanging the integral and the expectation operators: Z Q cavg (π) = Z Q E [θ(ψ (Φ, x))] dx = x=0 θ(x)dx. x=0 We can justify this operator exchange in several ways, for example we can apply Tonelli’s theorem since the prices θ are always non-negative. Next we consider the worst-case cost. We generalize Theorem 14 by incorporating arbitrary item costs. Theorem 48 Suppose f : 2E × OE → R≥0 is adaptive monotone and adaptive submodular with respect to p (φ), and let η be any value such that f (S, φ) > f (E, φ) − η implies f (S, φ) = f (E, φ) for all S and φ. Let ∗ δ = minφ p (φ) be the minimum probability of any realization. Let πwc be the optimal policy minimizing the worst-case cost cwc (·) while guaranteeing that every realization is covered. Let π be an α-approximate greedy policy with respect to the item costs. Finally, let Q := E [f (E, Φ)] be the maximum possible expected reward. Then     Q ∗ cwc (π) ≤ α cwc (πwc ) ln +1 . δη ∗ ), let ` = αk ln (Q/δη), and apply Proof Let π be an α-approximate greedy policy. Let k = cwc (πwc Theorem 38 with these parameters to yield     δη ∗ −`/αk ∗ favg (πwc ). (80) favg (π[`] ) > 1 − e favg (πwc ) = 1 − Q ∗ ∗ Since πwc covers every realization by assumption, favg (πwc ) = E [f (E, Φ)] = Q, so rearranging terms of Eq. (80) yields Q − favg (π[`] ) < δη. Since favg (π[`] ) ≤ favg (π[`→] ) by the adaptive monotonicity of f , it follows that Q − favg (π[`→] ) < δη. By definition of δ and η, if some φ is not covered by π[`→] then Q − favg (π[`→] ) ≥ δη. Thus Q − favg (π[`→] ) < δη implies Q − favg (π[`→] ) = 0, meaning π[`→] covers every realization. We next claim that π[`→] has worst-case cost at most ` + αk. It is sufficient to show that the final item executed by π[`→] has cost at most αk for any realization. As we will prove, this follows from the facts that π ∗ is an α-approximate greedy policy and πwc covers every realization at cost at most k. The data dependent bound, Lemma 37 on page 43, guarantees that   ∗ ∗ ∆(πwc | ψ) ∆(πwc | ψ) ∆(e | ψ) ≥ ≥ . (81) max ∗ | ψ) e c(e) c (πwc k ∗ Suppose ψ ∈ dom(π). We would like to say that maxe ∆(e | ψ) ≤ ∆(πwc | ψ). Supposing this is true, any ∗ item e with cost c(e) > αk must have ∆(e | ψ) /c(e) < ∆(πwc | ψ) /αk, and hence cannot be selected by any α-approximate greedy policy upon observing ψ by Eq. (81), and thus the final item executed by π[`→] has cost ∗ at most αk for any realization. So we next show that maxe ∆(e | ψ) ≤ ∆(πwc | ψ). Towards this end, note that Lemma 49 implies max ∆(e | ψ) ≤ E [f (E, Φ) | Φ ∼ ψ] − E [f (dom(ψ), Φ) | Φ ∼ ψ] . e (82) ∗ and to prove maxe ∆(e | ψ) ≤ ∆(πwc | ψ) it suffices to show ∗ E [f (E, Φ) | Φ ∼ ψ] ≤ E [f (E(πwc , Φ) ∪ dom(ψ), Φ) | Φ ∼ ψ] . (83) Proving Eq. (83) is quite straightforward if f is strongly adaptive monotone. Given that f is only adaptive monotone, it requires some additional effort. So fix A ⊂ E and let πA be a non-adaptive policy that selects all 54 ∗ items in A in some arbitrary order. Let P := {ψ : dom(ψ) = A}. Apply Lemma 49 with π 0 = πA @πwc and any ψ ∈ P to obtain ∗ E [f (E(πwc , Φ) ∪ A, Φ) | Φ ∼ ψ] ≤ E [f (E, Φ) | Φ ∼ ψ] . P ∗ ψ]·E [fP (E(πwc , Φ) ∗ favg (πA @πwc ) (84) ∗ favg (πwc ) Note that ψ∈P P [Φ ∼ ∪ A, Φ) | Φ ∼ ψ] = ≥ = E [f (E, Φ)]. Since we know E [f (E, Φ)] = ψ∈P P [Φ ∼ ψ] · E [f (E, Φ) | Φ ∼ ψ], an averaging argument together with Eq. (84) then implies that for all ψ ∈ P ∗ E [f (E(πwc , Φ) ∪ A, Φ) | Φ ∼ ψ] = E [f (E, Φ) | Φ ∼ ψ] . (85) Since ψ was an arbitrary partial realization with dom(ψ) = A, and A ⊆ E was arbitrary, fix ψ ∈ dom(π) ∗ and let A = dom(ψ). With these settings, Eq. (85) implies Eq. (83), and thus maxe ∆(e | ψ) ≤ ∆(πwc | ψ), ∗ and thus an α-approximate greedy policy can never select an item with cost exceeding αk, where k = cwc (πwc ). Hence cwc (π[`→] ) − cwc (π[`] ) ≤ αk, and so cwc (π[`→] ) ≤ ` + αk. This completes the proof. Lemma 49 Fix adaptive monotone submodular objective f . For any policy π and any ψ ∈ dom(π) we have E [f (E(π, Φ), Φ) | Φ ∼ ψ] ≤ E [f (E, Φ) | Φ ∼ ψ] . Proof Augment π to a new policy π 0 as follows. Run π to completion, and let ψ 0 be the partial realization consisting of all of the states it has observed. If ψ ⊆ ψ 0 , then proceed to select all the remaining items in E in any order. Otherwise, if ψ * ψ 0 then terminate. Then E [f (E(π, Φ), Φ) | Φ ∼ ψ] ≤ E [f (E(π 0 , Φ), Φ) | Φ ∼ ψ] = E [f (E, Φ) | Φ ∼ ψ] (86) where the inequality is by repeated application of the adaptive monotonicity of f , and the equality is by construction. In §5.2 we described how the result of Feige (1998) implies that there is no polynomial time (1 − ) ln (Q/η) approximation algorithm for self–certifying instances of Adaptive Stochastic Min Cost Cover, unless NP ⊆ DTIME(nO(log log n) ). Here we show a related result for general instances. Lemma 50 For every constant  > 0, there is no (1 − ) ln (Q/δη) polynomial time approximation algorithm for general instances of Adaptive Stochastic Min Cost Cover, for either the average case objective cavg (·) or the worst-case objective cwc (·), unless NP ⊆ DTIME(nO(log log n) ). Proof We offer a reduction from the Set Cover problem. Fix a Set Cover instance with ground set U and sets {S1 , S2 , . . . , Sm } ⊆ 2U with unit-cost sets. Fix Q, η and δ such that 1/δ and Q/η are positive integers, Q and δη = |U |. Let E := {S1 , S2 , . . . , Sm }, and set the cost of each item to one. Partition U into 1/δ disjoint, equally sized subsets U1 , U2 , . . . , U1/δ . Construct a realization φi for each Ui . Let the set of states be O = {NULL}. Hence φi (e) = NULL for all i and e, so that no knowledge of the true realization is revealed by selecting items. We use a uniform distribution over realizations, i.e., p (φi ) = δ for all i. Finally, our objective is f (C, φi ) := | ∪S∈C (S ∩ Ui )|, i.e., the number of elements in Ui that we cover with sets in C. Since |O| = 1, every realization is consistent with every possible partial realization ψ. Hence for any ψ, we have E [f (dom(ψ), Φ) | Φ ∼ ψ] = δ fˆ(dom(ψ)), where fˆ(C) = | ∪S∈C S| is the objective function of the original set cover instance. Since fˆ is submodular, f is adaptive submodular. Likewise, since fˆ is monotone, and |O| = 1, f is strongly adaptive monotone. Now, to cover any realization, we must obtain the maximum possible value for all realizations, which means selecting a collection of sets C such that ∪S∈C S = U . Conversely, any C such that ∪S∈C S = U clearly covers f . Hence this instance of Adaptive Stochastic Min Cost Cover, with either the average case objective cavg (·) or the worst-case objective cwc (·), is equivalent to the original Set Cover instance. Therefore, the result from Feige (1998) implies that there is no polynomial time algorithm for obtaining a (1 − ) ln |U | = (1 − ) ln (Q/δη) approximation for Adaptive Stochastic Min Cost Cover unless NP ⊆ DTIME(nO(log log n) ). 55 15.5 The Min-Sum Objective In this section we prove Theorem 15, which appears on page 16, in the case where the items have arbitrary costs. Our proof resembles the analogous proof of Streeter and Golovin (2007) for the non-adaptive min-sum submodular cover problem, and, like that proof, ultimately derives from an extremely elegant performance analysis of the greedy algorithm for min-sum set cover due to Feige et al. (2004). The objective function cΣ (·) generalized to arbitrary cost items uses the strict truncation8 π[←t] in place of π[t] in the unit-cost definition: cΣ (π) := ∞ X ∞ X  X  E [f (E, Φ)] − favg (π[←t] ) = p (φ) f (E, φ) − f (E(π[←t] , φ), φ) . t=0 φ (87) t=0 We will prove that any α-approximate greedy policy π achieves a 4α-approximation for the min-sum objective, i.e., cΣ (π) ≤ 4α cΣ (π ∗ ) for all policies π ∗ . To do so, we require the following lemma. Lemma 51 Fix an α-approximate greedy  policy π for some adaptive monotone submodular function f and let si := α favg (π[i+1] ) − favg (π[i] ) . For any policy π ∗ and nonnegative integers i and k, we have ∗ favg (π[k] ) ≤ favg (π[←i] ) + k · si . ∗ ∗ Proof Fix π, π ∗ , i, and k. By adaptive monotonicity favg (π[k] ) ≤ favg (π[←i] @π[k] ). We next aim to prove ∗ favg (π[←i] @π[k] ) ≤ favg (π[←i] ) + k · si (88)  which is sufficient to complete theproof. Towards this end, fix a partial realization ψ of the form (e, φ(e)) : e ∈ E(π[←i] , φ)  ∗ ∗ ∗ for some φ. Consider ∆ π[k] | ψ , which equals the expected marginal benefit of the π[k] portion of π[←i] @π[k] conditioned on Φ ∼ ψ. Lemma 37 allows us to bound it as   h  i h  i ∆(e | ψ) ∗ ∗ , E ∆ π[k] | ψ ≤ E c π[k] | ψ · max e c(e) ∗ where the h expectations i are taken over the internal randomness of π , if there is any. Note that for all φ, we ∗ ∗ have E c(E(π[k] , φ)) ≤ k, where the expectation is again taken over the internal randomness of π[k] . Hence h  i h  i ∗ ∗ E c π[k] | ψ ≤ k for all ψ. It follows that E ∆ π[k] | ψ ≤ k · maxe (∆(e | ψ) /c(e)). By definition of an α-approximate greedy policy, π obtains at least h  i ∗ (1/α) max (∆(e | ψ) /c(e)) ≥ E ∆ π[k] | ψ /αk (89) e expected marginal benefit per unit cost in a step immediately following its observation of ψ. Next we take an appropriate convex combination  of the previous inequality with different values of ψ. Let Ψ be a random partial realization distributed as (e, Φ(e)) : e ∈ E(π[←i] , Φ) . Then taking the expectation of Eq. (89) over Ψ yields i   h  ∗ ∗ E ∆ π[k] |Ψ favg (π[←i] @π[k] ) − favg (π[←i] ) = favg (π[i+1] ) − favg (π[i] ) ≥ Eψ  . (90) αk αk  Multiplying Eq. (90) by αk, and substituting in si = α favg (π[i+1] ) − favg (π[i] ) , we conclude ksi ≥ ∗ favg (π[←i] @π[k] ) − favg (π[←i] ) which immediately yields Eq. (88) and concludes the proof. 56 Figure 6: An illustration of the inequality R∞ x=0 h(x)dx ≥ P i≥0 xi (yi − yi+1 ). Using Lemma 51, together with a geometric argument developed by Feige et al. (2004), we now prove Theorem 15. Proof of Theorem 15: Let Q := E [f (E, Φ)] be the maximum possible expected reward, where the expec tation is taken w.r.t. p (φ). Let π be an α-approximate greedy policy. Define Ri := Q − favg π[i] and  Pi ∗ define Pi := Q − favg π[←i] . Let xi := 2s , let yi := R2i , and let h(x) := Q − favg (π[x] ). We claim i   favg π[←i] ≤ favg π[i] and so Pi ≥ Ri . This clearly holds if π[←i] is the empty policy, and otherwise π can always select an item that contributes zero marginal benefit, namely an item it has already played previously. Hence an α-approximate greedy policy π can never expected marginal benefit, and  select  items with negative    ∗ so favg π[←i] ≤ favg π[i] . By Lemma 51, favg π[x ≤ f π + x s avg i i . Therefore [←i] i] Ri Pi ≥ = yi (91) 2 2     For similar reasons that favg π[←i] ≤ favg π[i] , we have favg π[i−1] ≤ favg π[i] , and so the sequence hy1 , y2 , . . .i is non-increasing. The adaptive monotonicity and adaptive submodularity of f imply that h(x) ∗ ∗ is non-increasing. Informally, this is because otherwise, if favg (π[x] ) > favg (π[x+1] ) for some x, then the optimal policy must be sacrificing immediate rewards at time x in exchange for greater returns later, and it can be shown that if such a strategy is optimal, R ∞ then adaptive P submodularity cannot hold. Eq. (91) and the monotonicity of h and i 7→ yi imply that x=0 h(x)dx ≥ i≥0 xi (yi − yi+1 ) (see Figure 6). The left hand bound for cΣ (π ∗ ), and because si = α (Ri − Ri+1 ) the right hand side simplifies to Pside is a lower 1 1 P = c (π), proving cΣ (π) ≤ 4α · cΣ (π ∗ ).  i≥0 i 4α 4α Σ h(xi ) ≥ Q − favg (π[←i] ) − xi · si = Pi − 15.7 Proof of Approximation Hardness in the Absence of Adaptive Submodularity We now provide the proof of Theorem 26 whose statement appears on page 33 in §12. Proof of Theorem 26: We construct a hard instance based on the following intuition. We make the algorithm go “treasure hunting”. There is a set of t locations {0, 1, , . . . , t − 1}, there is a treasure at one of these locations, and the algorithm gets unit reward if it finds it, and zero reward otherwise. There are m “maps,” each consisting of a cluster of s bits, and each purporting to indicate where the treasure is, and each map is stored in a (weak) secret-sharing way, so that querying few bits of a map reveals nothing about where it says the treasure is. Moreover, all but one of the maps are fake, and there is a puzzle indicating which map is the correct one indicating the treasure’s location. Formally, a fake map is one which is probabilistically independent of the location of the treasure, conditioned on the puzzle. 8. See Definition 30 on page 41. 57 15.6 A Symbol Table E, e ∈ E O, o ∈ O φ ψ Φ, Ψ ∼ p p(φ | ψ) π E(π, φ) ∆(e | ψ) ∆(π | ψ) ψ[e/o] k π[k] π[←k] π[k→] π@π 0 f favg c cavg cwc cΣ c (π | ψ) α Q η 1P Ground set of items, and an individual item. States an item may be in, or outcomes of selecting an item, and an individual state/outcome. A realization, i.e., a function from items to states. A partial realization, typically encoding the current set of observations; each ψ ⊂ E × O is a partial mapping from items to states. A random realization and a random partial realization, respectively. The consistency relation: φ ∼ ψ means ψ(e) = φ(e) for all e ∈ dom(ψ). The probability distribution on realizations. The conditional distribution on realizations: p(φ | ψ) := P [Φ = φ | Φ ∼ ψ]. A policy, which maps partial realizations to items. The set of all items selected by π when run under realization φ. The conditional expected marginal benefit of e conditioned on ψ: ∆(e | ψ) := E [f (dom(ψ) ∪ {e} , Φ) − f (dom(ψ), Φ) | Φ ∼ ψ]. The conditional expected marginal benefit of policy π conditioned on ψ: ∆(π | ψ) := E [f (dom(ψ) ∪ E(π, Φ), Φ) − f (dom(ψ), Φ) | Φ ∼ ψ]. Shorthand for ψ ∪ {(e, o)}. Budget on the cost of selected item sets. A truncated policy. See Definition 4 on page 11 (unit costs) and Definition 32 on page 41. A strictly truncated policy. See Definition 30 on page 41. A laxly truncated policy. See Definition 31 on page 41. Policies π and π 0 concatenated together. See Definition 34 on page 42. An objective function, of type f : 2E × OE → R≥0 unless stated otherwise. Average benefit: favg (π) := E [f (E(π, Φ), Φ)]. P Item costs c : E → N. Extended to sets via c(S) := e∈S c(e). Average cost of a policy: cavg (π) := E [c(E(π, Φ))]. Worst-case cost of a policy: cwc (π) := maxφ c(E(π, φ)).  P∞ Min-sum cost of a policy: cΣ (π) := t=0 E [f (E, Φ)] − favg (π[←t] ) . Conditional average policy cost: c (π | ψ) := E [c(E(π, Φ)) | Φ ∼ ψ]. Approximation factor for greedy optimization in an α-approximate greedy policy. Benefit quota. Often Q = E [f (E, Φ)]. Coverage gap: η = sup {η 0 : f (S, φ) > Q − η 0 implies f (S, φ) ≥ Q for all S, φ}. The indicator for proposition P , which equals one if P is true and zero if P is false. Table 2: Important symbols and notations used in this article 58 Our instance will have three types of items, E = ET ] EM ] EP , where |ET | = t encodes where the treasure is, |EM | = ms encodes the maps, and |EP | = n3 encodes the puzzle, where m, t, s and n are specified below. All outcomes are binary, so that O = {0, 1}, and we identify items with bit indices. Accordingly we say that φ(e) is the value of bit e. For all e ∈ EM ∪ EP , P [Φ(e) = 1] = .5 independently. The conditional distribution of Φ(ET ) given Φ(EM ∪ EP ) will be deterministic as specified below. Our objective function f is linear, and defined as follows: f (A, φ) = |{e ∈ A ∩ ET : φ(e) = 1}|. We now describe the puzzle, which is to compute i(P ) := (perm(P mod 2` for a suitable random P) modQp) n matrix P , and suitable prime p and integer `, where perm(P ) = σ∈Sn i=1 Piσ(i) is the permanent of P . We exploit Theorem 1.9 of Feige and Lund (1997) in which they show that if there exist constants η, δ > 0 such that a randomized polynomial time algorithm can compute (perm(P ) mod p) mod 2` correctly with n×n probability 2−` (1 + 1/nη ), where P is drawn uniformly at random from {0, 1, 2, . . . , p − 1} , p is any 1 P prime superpolynomial in n, and ` ≤ p 2 − δ , then PH = AM = Σ2 . To encode the puzzle, we fix a prime p ∈ [2n−2 , 2n−1 ] and use the n3 bits of φ(EP ) to sample P = P (φ) (nearly) P uniformly at random from n×n {0, 1, 2, . . . , p − 1} as follows. For a matrix P ∈ Zn×n , we let rep(P ) := ij Pij · p(i−1)n+(j−1) define a base p representation of P . Note rep(·) is one-to-one for n × n matrices with entries in Zp , so we can define 3 its inverse rep−1 (·). The encoding Pk(φ) interprets the bits φ(EP ) as an integer x in [2n ], and computes j 2 3 y = x mod (pn ). If x ≤ 2n /pn 2 2 pn , then P = rep−1 (y). Otherwise, P is the all zero matrix. This 2 3 2 latter event occurs with probability at most pn /2n ≤ 2−n , and in this case we simply suppose the algorithm 2 under consideration finds the treasure and so gets unit reward. This adds 2−n to its expected reward. So let us assume from now on that P is drawn uniformly atUrandom. m Next we consider the maps. Partition EM = i=1 Mi into m maps Mi , each consisting of s items. For each map Mi , partition its items into s/ log2 t groups of log2 t bits each, so that the bits of each group encode a log2 t bit binary string. Let vi ∈ {0, 1, . . . , t − 1} be the XOR of these s/ log2 t binary strings, interpreted as an integer (using any fixed encoding). We say Mi points to vi as the location of the treasure. A priori, each vi is uniformly distributed in {0, ..., t − 1}. For a particular realization of φ(EP ∪ EM ), define v(φ) := vi(P (φ)) . We set v(φ) to be the location of the treasure under realization φ, i.e., we label ET = {e0 , e1 , . . . , et−1 } and ensure φ(ej ) = 1 if j = vi(P (φ)) , and φ(e) = 0 for all other e ∈ ET . Note the random variable v = v(φ) is distributed uniformly at random in {0, 1, . . . , t−1}. Note that this still holds if we condition on the realizations of any set of s/ log2 t − 1 items in a map, because in this case there is still at least one group whose bits remain completely unobserved. Now consider the optimal policy with a budget of k = n3 + s + 1 items to pick. Clearly, its reward can be at most 1. However, given a budget of k, a computationally unconstrained policy can exhaustively sample EP , solve the puzzle (i.e., compute i(P )), read the correct map (i.e., exhaustively sample Mi(P ) ), decode the map (i.e., compute v = vi(P ) ), and get the treasure (i.e., pick ev ) thereby obtaining a reward of one. Now we give an upper bound on the expected reward R of any randomized polynomial time algorithm 3 A with a budget of βk items, assuming ΣP 2 6= PH. Fix a small constant γ > 0, and set s = n and 1/γ m = t = n . We suppose we give A the realizations φ(EM ) for free. We also replace its budget of βk items with a budget of βk specifically for map items in EM and an additional budget of βk specifically for the treasure locations in ET . Obviously, this can only help it. As noted, if it selects less than s/ log2 t bits from the map Mi(P ) indicated by P , the distribution over vi(P ) conditioned on those realizations is still uniform. Of course, knowledge of vi for i 6= i(P ) is useless for getting reward. Hence A can try at most βk log2 (t)/s = o(βk) maps in an attempt to find Mi(P ) . Note that if we have a randomized algorithm which n×n given a random P drawn from {0, 1, 2, . . . , p − 1} always outputs a set S of integers of size α such that P [i(P ) ∈ S] ≥ q, then we can use it to construct a randomized algorithm that, given P , outputs an integer x such that P [i(P ) = x] ≥ q/α, simply by running the first algorithm and then selecting a random item of S. If A does not find Mi(P ) , the distribution on the treasure’s location is uniform given its knowledge. Hence it’s budget of βk treasure locations can only earn it expected reward at most βk/t. Armed with these observations and Theorem 1.9 in the work of Feige and Lund (1997) and our complexity theoretic assumptions, we infer 59 2 E [R] ≤ o(βk) · 2−` (1 + 1/nη ) + βk/t + 2−n . Since s = n3 and m = t = n1/γ and γ = Θ(1) and η = 1 and ` = log2 m and k = n3 + s + 1 = 2n3 + 1, we have E [R] ≤ βk (1 + o(1)) = 2βn3−1/γ (1 + o(1)). t Next note that |E| = t + ms + n3 = n3+1/γ (1 + o(1)). Straightforward algebra shows that in order to ensure E [R] = o(β/|E|1−ε ), it suffices to choose γ ≤ ε/6. Thus, under our complexity theoretic assumptions, any polynomial time randomized algorithm A with budget βk achieves at most o(β/|E|1−ε ) of the value obtained by the optimal policy with budget k, so the approximation ratio is ω(|E|1−ε /β).  60
2cs.AI
Simultaneous Orthogonal Planarity? Patrizio Angelini1 , Steven Chaplick2 , Sabine Cornelsen3 , Giordano Da Lozzo4 , Giuseppe Di Battista4 , Peter Eades5 , Philipp Kindermann6 , Jan Kratochvı́l7 , Fabian Lipp2 , and Ignaz Rutter8 arXiv:1608.08427v1 [cs.CG] 30 Aug 2016 1 6 Universität Tübingen, Germany – [email protected] 2 Universität Würzburg, Germany – {steven.chaplick,fabian.lipp}@uni-wuerzburg.de 3 Konstanz University, Germany – [email protected] 4 Roma Tre University, Italy – {dalozzo,gdb}@dia.uniroma3.it 5 The University of Sydney, Australia – [email protected] FernUniversität in Hagen, Germany – [email protected] 7 Charles University, Czech Republic – [email protected] 8 Karlsruhe Institute of Technology, Germany – [email protected] Abstract. We introduce and study the OrthoSEFE-k problem: Given k planar graphs each with maximum degree 4 and the same vertex set, do they admit an OrthoSEFE, that is, is there an assignment of the vertices to grid points and of the edges to paths on the grid such that the same edges in distinct graphs are assigned the same path and such that the assignment induces a planar orthogonal drawing of each of the k graphs? We show that the problem is NP-complete for k ≥ 3 even if the shared graph is a Hamiltonian cycle and has sunflower intersection and for k ≥ 2 even if the shared graph consists of a cycle and of isolated vertices. Whereas the problem is polynomial-time solvable for k = 2 when the union graph has maximum degree five and the shared graph is biconnected. Further, when the shared graph is biconnected and has sunflower intersection, we show that every positive instance has an OrthoSEFE with at most three bends per edge. 1 Introduction The input of a simultaneous embedding problem consists of several graphs G1 = (V, E1 ), . . . , Gk = (V, Ek ) on the same vertex set. For a fixed drawing style S, the simultaneous embedding problem asks whether there exist drawings Γ1 , . . . , Γk of G1 , . . . , Gk , respectively, in drawing style S such that for any i and j the restrictions of Γi and Γj to Gi ∩ Gj = (V, Ei ∩ Ej ) coincide. The problem has been most widely studied in the setting of topological planar drawings, where vertices are represented as points and edges are represented as ? This research was initiated at the Bertinoro Workshop on Graph Drawing 2016. Research was partially supported by DFG grant Ka812/17-1, by MIUR project AMANDA, prot. 2012C4E3KT 001, by the grant no. 14-14179S of the Czech Science Foundation GACR, and by DFG grant WA 654/21-1. pairwise interior-disjoint Jordan arcs between their endpoints. This problem is called Simultaneous Embedding with Fixed Edges or SEFE-k for short, where k is the number of input graphs. It is known that SEFE-k is NP-complete for k ≥ 3, even in the restricted case of sunflower instances [26], where every pair of graphs shares the same set of edges, and even if such a set induces a star [3]. On the other hand, the complexity for k = 2 is still open. Recently, efficient algorithms for restricted instances have been presented, namely when (i) the shared graph G∩ = G1 ∩ G2 is biconnected [19,1] or a star-graph [1], (ii) G∩ is a collection of disjoint cycles [13], (iii) every connected component of G∩ is either subcubic or biconnected [26,11], (iv) G1 and G2 are biconnected and G∩ is connected [14], and (v) G∩ is connected and the input graphs have maximum degree 5 [14]; see the survey by Bläsius et al. [12] for an overview. For planar straight-line drawings, the simultaneous embedding problem is called Simultaneous Geometric Embedding and it is known to be NP-hard even for two graphs [17]. Besides simultaneous intersection representation for, e.g., interval graphs [20,14] and permutation and chordal graphs [21], it is only recently that the simultaneous embedding paradigm has been applied to other fundamental planarity-related drawing styles, namely simultaneous level planar drawings [2] and RAC drawings [4,8]. We continue this line of research by studying simultaneous embeddings in the planar orthogonal drawing style, where vertices are assigned to grid points and edges to paths on the grid connecting their endpoints [29]. In accordance with the existing naming scheme, we define OrthoSEFE-k to be the problem of testing whether k input graphs hG1 , . . . , Gk i admit a simultaneous planar orthogonal drawing. If such a drawing exists, we call it an OrthoSEFE of hG1 , . . . , Gk i. Note that it is a necessary condition that each Gi has maximum degree 4 in order to obtain planar orthogonal drawings. Hence, in the remainder of the paper we assume that all instances have this property. For instances with this property, at least when the shared graph is connected, the problem SEFE-2 can be solved efficiently [14]. However, there are instances of OrthoSEFE-2 that admit a SEFE but not an OrthoSEFE; see Fig. 1(a). Unless mentioned otherwise, all instances of OrthoSEFE-k and SEFE-k we consider are sunflower. Notice that instances with k = 2 are always sunflower. Let hG1 = (V, E1 ), G2 = (V, E2 )i be an instance of OrthoSEFE-2. We define the shared graph (resp. the union graph) to be the graph G∩ = (V, E1 ∩ E2 ) (resp. G∪ = (V, E1 ∪ E2 )) with the same vertex set as G1 and G2 , whose edge set is the intersection (resp. the union) of the ones of G1 and G2 . Also, we call the edges in E1 ∩ E2 the shared edges and we call the edges in E1 \ E2 and in E2 \ E1 the exclusive edges. The definitions of shared graph, shared edges, and exclusive edges naturally extend to sunflower instances for any value of k. One main issue is to decide how degree-2 vertices of the shared graph are represented. Note that, in planar topological drawings, degree-2 vertices do not require any decisions as there exists only a single cyclic order of their incident edges. In the case of orthogonal drawings there are, however, two choices for a degree-2 vertex: It can either be drawn straight, i.e., it is incident to two angles 2 C v3 v v4 v5 u (a) (b) Fig. 1. (a) A negative instance of OrthoSEFE-2. Shared edges are black, while exclusive edges are red and blue. The red edges require 270◦ angles on different sides of C. Thus, the blue edge (u, v) cannot be drawn. Note that the given drawing is a SEFE-2. (b) Examples of side assignments for the exclusive edges incident to degree-2 vertices of G∩ : orthogonality constraints are satisfied at v4 and v5 , while they are violated at v3 . of 180◦ , or bent, i.e., it is incident to one angle of 90◦ and to one angle of 270◦ . If v is a degree-2 vertex of the shared graph with neighbors u and w, and two exclusive edges e, e0 , say of G1 , are incident to v and are embedded on the same side of the path uvw, then v must be bent, which in turn implies that also every exclusive edge of G2 incident to v has to be embedded on the same side of uvw as e and e0 . In this way, the two input graphs of OrthoSEFE-2 interact via the degree-2 vertices. It is the difficulty of controlling this interaction that marks the main difference between SEFE-k and OrthoSEFE-k. To study this interaction in isolation, we focus on instances of OrthoSEFE-2 where the shared graph is a cycle for most of the paper. Note that such instances are trivial yes-instances of SEFE-k (provided the input graphs are all planar). Contributions and Outline. In Section 2 we provide our notation and we show that the existence of an OrthoSEFE of an instance of OrthoSEFE-k can be described as a combinatorial embedding problem. In Section 3, we show that OrthoSEFE-3 is NP-complete even if the shared graph is a cycle, and that OrthoSEFE-2 is NP-complete even if the shared graph consists of a cycle plus some isolated vertices. This contrasts the situation of SEFE-k where these cases are polynomially solvable [1,10,19,26]. In Section 4, we show that OrthoSEFE2 is efficiently solvable if the shared graph is a cycle and the union graph has maximum degree 5. Finally, in Section 5, we extend this result to the case where the shared graph is biconnected (and the union graph still has maximum degree 5). Moreover, we show that any positive instance of OrthoSEFE-k whose shared graph is biconnected admits an OrthoSEFE with at most three bends per edge. We close with some concluding remarks and open questions in Section 6. Full proofs can be found in the Appendix. 2 Preliminaries We will extensively make use of the Not-All-Equal 3-Sat (Nae3Sat) problem [25, p.187]. An instance of Nae3Sat consists of a 3-CNF formula φ with variables x1 , . . . , xn and clauses c1 , . . . , cm . The task is to find a Nae truth as3 signment, i.e., a truth assignment such that each clause contains both a true and a false literal. Nae3Sat is known to be NP-complete [27]. The variable–clause graph is the bipartite graph whose vertices are the variables and the clauses, and whose edges represent the membership of a variable in a clause. The problem Planar Nae3Sat is the restriction of Nae3Sat to instances whose variable– clause graph is planar. Planar Nae3Sat can be solved efficiently [23,28]. Embedding Constraints. Let hG1 , . . . , Gk i be an OrthoSEFE-k instance. A SEFE is a collection of embeddings Ei for the Gi such that their restrictions on G∩ are the same. Note that in the literature, a SEFE is often defined as a collection of drawings rather than a collection of embeddings. However, the two definitions are equivalent [22]. For a SEFE to be realizable as an OrthoSEFE it needs to satisfy two additional conditions. First, let v be a vertex of degree 2 in G∩ with neighbors u and w. If in any embedding Ei there exist two exclusive edges incident to v that are embedded on the same side of the path uvw, then any exclusive edge incident to v in any of the Ej 6= Ei must be embedded on the same side of the path uvw. Second, let v be a vertex of degree 3 in G∩ . All exclusive edges incident to v must appear between the same two edges of G∩ around v. We call these the orthogonality constraints. See Fig. 1(b). Theorem 1. An instance hG1 , . . . , Gk i of OrthoSEFE-k has an OrthoSEFE if and only if it admits a SEFE satisfying the orthogonality constraints. For the case in which the shared graph is a cycle C, we give a simpler version of the constraints in Theorem 1, which will prove useful in the remainder of the paper. By the Jordan curve Theorem, a planar drawing of cycle C divides the plane into a bounded and an unbounded region – the inside and the outside of C, which we call the sides of C. Now the problem is to assign the exclusive edges to either of the two sides of C so that the following two conditions are fulfilled. Planarity Constraints. Two exclusive edges of the same graph must be drawn on different sides of C if their endvertices alternate along C. Orthogonality Constraints. Let v ∈ V be a vertex that is adjacent to two exclusive edges ei and e0i of the same graph Gi , i ∈ {1, . . . , k}. If ei and e0i are on the same side of C, then all exclusive edges incident to v of all graphs G1 , . . . , Gk must be on the same side as ei and e0i . Note that this is a reformulation of the general orthogonality constraints. Further, the orthogonality constraints also imply that if ei and e0i are on different sides of C, then for each graph Gj that contains two exclusive edges ej and e0j incident to v, with j ∈ {1, . . . , k}, ej and e0j must be on different sides of C. The next theorem follows from Theorem 1 and from the following two observations. First, for a sunflower instance hG1 , . . . , Gk i whose shared graph is a cycle, any collection of embeddings is a SEFE [22]. Second, the planarity constraints are necessary and sufficient for the existence of an embedding of Gi [5]. Theorem 2. An instance of OrthoSEFE-k whose shared graph is a cycle C has an OrthoSEFE if and only if there exists an assignment of the exclusive edges to the two sides of C satisfying the planarity and orthogonality constraints. 4 Cj sj α j c4 = (x1, x3, x4) j δj t γj βj yaj ybj ycj j V54 sji uji vij wij zij rij C4 V44 Vi V34 V24 V14 V15 V25 V35 V45 V55 tji (a) (b) Fig. 2. (a) A clause gadget Cj (top) and a variable-clause gadget Vij (bottom); solid edges belong to the gadgets, dotted edges are optional, and dashed edges are transmission edges. (b) Illustration of instance hG1 , G2 , G3 i, focused on a clause c4 . Black edges belong to the shared graph G∩ . The red, blue, and green edges are the exclusive edges of G1 , G2 , and G3 , respectively. 3 Hardness Results We show that OrthoSEFE-k is NP-complete for k ≥ 3 for instances with sunflower intersection even if the shared graph is a cycle, and for k = 2 even if the shared graph consists of a cycle and isolated vertices. Theorem 3. OrthoSEFE-k with k ≥ 3 is NP-complete, even for instances with sunflower intersection in which (i) the shared graph is a cycle and (ii) k − 1 of the input graphs are outerplanar and have maximum degree 3. Proof sketch. The membership in NP directly follows from Theorem 2. To prove the NP-hardness, we show a polynomial-time reduction from the NP-complete problem Positive Exactly-Three Nae3Sat [24], which is the variant of Nae3Sat in which each clause consists of exactly three unnegated literals. Let x1 , x2 , . . . , xn be the variables and let c1 , c2 , . . . , cm be the clauses of a 3-CNF formula φ of Positive Exactly-Three Nae3Sat. We show how to construct an equivalent instance hG1 , G2 , G3 i of OrthoSEFE-3 such that G1 and G2 are outerplanar graphs of maximum degree 3. We refer to the exclusive edges in G1 , G2 , and G3 as red, blue, and green, respectively; refer to Fig. 2. For each clause cj , j = 1, . . . , m, we create a clause gadget C j as in Fig. 2(a) (top). For each variable xi , i = 1, . . . , n, and each clause cj , j = 1, . . . , m, we create a variable-clause gadget Vij as in Fig. 2(a) (bottom). Observe that the (dotted) green edge {wij , rij } in a variable-clause gadget is only part of Vij if xi does not occur in cj . Otherwise, there is a green edge {wij , yxj } connecting wij to one of the three vertices yaj , ybj , or ycj (dashed stubs) in the clause gadget. Observe that these three variable-clause edges per clause can be realized in such a way that there exist no planarity constraints between pairs of them. In Fig. 2(b), the variable-clause gadgets V14 , V34 , V44 are incident to variable-clause edges, while V24 and V54 contain edges {w24 , r24 } and {w54 , r54 }, respectively. 5 The gadgets are ordered as indicated in Fig. 2(b). The variable-clause gadgets Vij , with i = 1, . . . , n, always precede the clause gadget V j , for any j = 1, . . . , m. Further, if j is odd, then the gadgets V1j , . . . , Vnj appear in this order, otherwise they appear in reversed order Vnj , . . . , V1j . Finally, Vij and Vij+1 , for i = 1, . . . , n and j = 1, . . . , m − 1, are connected by an edge {wij , wij+1 }, which is blue if j is odd and red if j is even. We call these edges transmission edges. Assume hG1 , G2 , G3 i admits an OrthoSEFE. Planarity constraints and orthogonality constraints guarantee three properties: (i) If the edge {uji , vij } is , vij+1 }, i = 1, . . . , n, j = 1, . . . , m − 1. This is due to inside C, then so is {uj+1 i the fact that, by the planarity constraints, the two green edges incident to wij lie on the same side of C and hence, by the orthogonality constraints, the two transmission edges incident to wij also lie in this side. We call {u1i , vi1 } the truth edge of variable xi . (ii) Not all the three green edges a = {αj , β j }, b = {β j , γ j }, and c = {γ j , δ j } lie on the same side of C. Namely, the two red edges of the clause gadget C j must lie on opposite sides of C because of the interplay between the planarity and the orthogonality constraints in the subgraph of C j induced by the vertices between β j and γ j . Hence, if edges a, b, and c lie in the same side of C, then the orthogonality constraints at either β j or γ j are not satisfied. (iii) For each clause cj = (xa , xb , xc ), edge a = {αj , β j } lies in the same side of C as the truth edge of xa . This is due to the planarity constraints between each of these two edges and the variable-clause edge {waj , yaj }. Analogously, edge b (edge c) lies on the same side as the truth edge of xb (of xc ). Hence, setting xi = true (xi = false) if the truth edge of xi is inside C (outside C) yields a Nae3Sat truth assignment that satisfies φ. The proof for the other direction is based on the fact that assigning the truth edges to either of the two sides of C according to the Nae3Sat assignment of φ also implies a unique side assignment for the remaining exclusive edges that satisfies all the orthogonality and the planarity constraints. It is easy to see that G1 and G2 are outerplanar graphs with maximum degree 3, and that the reduction can be extended to any k > 3.  In the following we describe how to modify the construction in Theorem 3 to show hardness of OrthoSEFE-2. We keep only the edges of G1 and G3 . Variable-clause gadgets and clause gadgets remain the same, as they are composed only of edges belonging to these two graphs. We replace each transmission edge in G2 by a transmission path composed of alternating green and red edges, starting and ending with a red edge. This transformation allows these paths to traverse the transmission edges of G1 and the variable-clause edges of G3 without introducing crossings between edges of the same color. It is easy to see that the properties described in the proof of Theorem 3 on the assignments of the exclusive edges to the two sides of C also hold in the constructed instance, where transmission paths take the role of the transmission edges. Theorem 4. OrthoSEFE-2 is NP-complete, even for instances hG1 , G2 i in which the shared graph consists of a cycle and a set of isolated vertices. 6 e e δ Pδ β Pβ P γ γ α f f (a) e e G xδ Pα f (b) Gφ c0e xβ x γ xα (c) xe xδ c00e xβ x γ c0f xα 00 cf xf (d) Fig. 3. (a) Instance hG1 , G2 i satisfying the properties of Lemma 1, where the edges in E2 belonging to the components α, β, γ, and δ of H have different line styles. (b) e (d) Variable–clause graph Gφ . Polygons for the components of H. (c) Graph G. 4 Shared Graph is a Cycle In this section we give a polynomial-time algorithm for instances of OrthoSEFE2 whose shared graph is a cycle and whose union graph has maximum degree 5 (Theorem 5). In order to obtain this result, we present an efficient algorithm for more restricted instances (Lemma 1) and give a series of transformations (Lemma 2–3) to reduce any instance with the above properties to one that can be solved by the algorithm in Lemma 1. Lemma 1. OrthoSEFE-2 is in P for instances hG1 , G2 i such that the shared graph C is a cycle and G1 is an outerplanar graph with maximum degree 3. Proof. The algorithm is based on a reduction to Planar Nae3Sat, which is in P [23,28]. First note that since G1 is outerplanar there exist no two edges in E1 alternating along C. Hence, there are no planarity constraints for G1 . We now define an auxiliary graph H with vertex set E2 \ E1 and edges corresponding to pairs of edges alternating along C; see Fig 3(a). W.l.o.g. we may assume that H is bipartite, since G2 would not meet the planarity constraints otherwise [5]. Let B be the set of connected components of H, and for each component B ∈ B, fix a partition B1 , B2 of B into independent sets (possibly B2 = ∅ in case of a singleton B). Note that in any inside/outside assignment of the exclusive edges of G2 that meets the planarity constraints, for every B ∈ B, all edges of B1 lie in one side of C and all edges of B2 lie in the other side. Draw the cycle C as a circle in the plane. For a component B ∈ B, let PB be the polygon inscribed into C whose corners are the endvertices in V of the edges in E2 corresponding to the vertices of B; refer to Fig. 3(b). If B only contains one vertex (i.e., one edge of G2 ), we consider the digon PB as the straight-line segment connecting the vertices of this edge. If B has at least two vertices, we let PB be open along its sides, i.e. it will contain the corners and all inner points (in Fig. 3(b) we depict this by making the sides of PB slightly concave). One can easily show that for any two components B, D ∈ B, their polygons PB , PD e may share only some of their corners, but no inner points. Hence the graph G 7 obtained by placing a vertex xB inside the polygon PB , for B ∈ B, making xB adjacent to each corner of PB and adding the edges E1 , is planar; see Fig. 3(c). We construct a formula φ with variables xB , B ∈ B, such that φ is Naesatisfiable if and only if hG1 , G2 i admits an inside/outside assignment meeting all planarity and orthogonality constraints. The encoding of the truth assignment will be such that xB is true when the edges of B1 are inside C and the edges of B2 are outside, and xB is false if the reverse holds. Every assignment satisfying the planarity constraints for G2 defines a truth-assignment in the above sense. Let e = (v, w) be an exclusive edge of E1 and let e1v , e2v (e1w , e2w ) be the exclusive edges of E2 incident to v (to w, respectively); we assume that all such four edges of E2 exist, the other cases being simpler. Let B(u, i) be the component containing the edge eiu , for u ∈ {v, w} and i ∈ {1, 2}. Define the literal `iu to be xB(u,i) if eiu ∈ B1 (u, i) and ¬xB(u,i) if eiu ∈ B2 (u, i). With our interpretation of the truth assignment, an edge eiu is inside C if and only if `iu is true. Now, for the assignment to meet the orthogonality constraints, if `1v = `2v , say both are true, then e must be assigned inside C as well, which would cause a problem if and only if `1w = `2w = false. Hence the orthogonality constraints are described by Nae-satisfiability of the clauses ce = (`1v , `2v , ¬`1w , ¬`2w ), for each e ∈ E1 . To reduce to Nae3Sat, we introduce a new variable xe for each edge e ∈ E1 \ E2 and replace the clause ce by two clauses c0e = (`1v , `2v , xe ) and c00e = (¬xe , ¬`1w , ¬`2w ). A planar drawing of the variable–clause graph Gφ of the e (see Figs. 3(c) resulting formula φ is obtained from the planar drawing Γe of G and 3(d)) by (i) placing each variable xB , with B ∈ B, on the point where vertex xB lies in Γe, (ii) placing each variable xe , with e ∈ E1 , on any point of edge e in Γe, (iii) placing clauses c0e and c00e , for each edge e = (v, w) ∈ E1 , on the points where vertices v and w lie in Γe, respectively, and (iv) drawing the edges of Gφ as the corresponding edges in Γe. This implies that Gφ is planar and hence we can test the Nae-satisfiability of φ in polynomial time [23,28]. t u The next two lemmas show that we can use Lemma 1 to test in polynomial time any instance of OrthoSEFE-2 such that G∩ is a cycle and each vertex v ∈ V has degree at most 3 in either G1 or G2 . Lemma 2. Let hG1 , G2 i be an instance of OrthoSEFE-2 whose shared graph is a cycle and such that G1 has maximum degree 3. It is possible to construct in polynomial time an equivalent instance hG∗1 , G∗2 i of OrthoSEFE-2 whose shared graph is a cycle and such that G∗1 is outerplanar and has maximum degree 3. Proof sketch. We construct an equivalent instance hG01 , G02 i of OrthoSEFE-2 such that G0∩ is a cycle, G01 has maximum degree 3, and the number of pairs of edges in G01 that alternate along G0∩ is smaller than the number of pairs of edges in G1 that alternate along G∩ . Repeatedly applying this transformation yields an equivalent instance hG∗1 , G∗2 i satisfying the requirements of the lemma. Consider two edges e = (u, v) and f = (w, z) of G1 such that u, w, v, z appear in this order along cycle G∩ and such that the path Pu,z in G∩ between u and z that contains v and w has minimal length. If G1 is not outerplanar, then the edges e and f always exist. Fig. 4 illustrates the construction of hG01 , G02 i. 8 e u H1 eav eaw w H2 ebw v e eaw ebv z H3 H1 u f w0 eav z0 H2 w v u0 f0 ebw ebv e0 v0 H3 z f Fig. 4. Instances (left) hG1 , G2 i and (right) hG01 , G02 i for the proof of Lemma 2. Edges of G∩ (G0∩ ) are black. Exclusive edges of G1 (G01 ) are red and those of G2 (G02 ) are blue. h f e ue H1 v H2 uf e H3 uh ue f 00 e0 00 h f0 H1 ve vf h00 0 v0 H2 vh uf H3 uh Fig. 5. Illustration of the transformation for the proof of Lemma 3 to reduce the number of vertices incident to two exclusive edges in G1 . Edges e0 , f 0 of G2 and h0 of G1 (right) take the role of edges e, f of G1 and h of G2 (left), respectively. Thus, the orthogonality constraints at v 0 are equivalent to those at v. By the choice of e and f , and by the fact that G1 has maximum degree 3, there is no exclusive edge in G1 with one endpoint in the set H2 of vertices between w and v, and the other one not in H2 . Further, observe that in an OrthoSEFE of hG01 , G02 i edges f and f 0 (edges e and e0 ) must be on the same side. Further, e and f must be in different sides of G0∩ . It can be concluded that hG01 , G02 i has an OrthoSEFE if and only if hG1 , G2 i has an OrthoSEFE.  The proof of the next lemma is based on the replacement illustrated in Fig. 5. Afterwards, we combine these results to present the main result of the section. Lemma 3. Let hG1 , G2 i be an instance of OrthoSEFE-2 whose shared graph is a cycle and whose union graph has maximum degree 5. It is possible to construct in polynomial time an equivalent instance hG∗1 , G∗2 i of OrthoSEFE-2 whose shared graph is a cycle and such that graph G∗1 has maximum degree 3. Theorem 5. OrthoSEFE-2 can be solved in polynomial time for instances whose shared graph is a cycle and whose union graph has maximum degree 5. 5 Shared Graph is Biconnected We now study OrthoSEFE-k for instances whose shared graph is biconnected. In Theorem 6, we give a polynomial-time Turing reduction from instances of OrthoSEFE-2 whose shared graph is biconnected to instances whose shared graph is a cycle. In Theorem 7, we give an algorithm that, given a positive instance of OrthoSEFE-k such that the shared graph is biconnected together 9 with a SEFE satisfying the orthogonality constraints, constructs an OrthoSEFE with at most three bends per edge. We start with the Turing reduction, i.e., we develop an algorithm that takes as input an instance hG1 , G2 i of OrthoSEFE-2 whose shared graph G∩ = G1 ∩G2 is biconnected and produces a set of O(n) instances hG11 , G12 i,. . . ,hGh1 , Gh2 i of OrthoSEFE-2 whose shared graphs are cycles. The output is such that hG1 , G2 i is a positive instance if and only if all instances hGi1 , Gi2 i, i = 1, . . . , h, are positive. The reduction is based on the SEFE testing algorithm for instances whose shared graph is biconnected by Bläsius et al. [10,11], which can be seen as a generalized and unrooted version of the one by Angelini et al. [1]. We first describe a preprocessing step. Afterwards, we give an outline of the approach of Bläsius et al. [11] and present the Turing reduction in two steps. We assume familiarity with SPQR-trees [7,6]; for formal definitions see Appendix A. Lemma 4. Let hG1 , G2 i be an instance of OrthoSEFE-2 whose shared graph is biconnected. It is possible to construct in polynomial time an equivalent instance hG∗1 , G∗2 i whose shared graph is biconnected and such that each endpoint of an exclusive edge has degree 2 in the shared graph. We continue with a brief outline of the algorithm by Bläsius et al. [11]. First, the algorithm computes the SPQR-tree T of the shared graph. To avoid special cases, T is augmented by adding S-nodes with only two virtual edges such that each P-node and each R-node is adjacent only to S-nodes and Q-nodes. Then, necessary conditions on the embeddings of P-nodes and R-nodes are fixed up to a flip following some necessary conditions. Afterwards, by traversing all S-nodes, a global 2SAT formula is produced whose satisfying assignments correspond to choices of the flips that result in a SEFE. We refine this approach and show that we can choose the flips independently for each S-node, which allows us to reduce each of them to a separate instance, whose shared graph is a cycle. We now describe the algorithm of Bläsius et al. [11] in more detail. Consider a node µ of T . A part of skel(µ) is either a vertex of skel(µ) or a virtual edge of skel(µ), which represents a subgraph of G. An exclusive edge e has an attachment in a part x of skel(µ) if x is a vertex that is an endpoint of e or if x is a virtual edge whose corresponding subgraph contains an endpoint of e. An exclusive edge e of G1 or of G2 is important for µ if its endpoints are in different parts of skel(µ). It is not hard to see that, to obtain a SEFE, the embedding of the skeleton skel(µ) of each node µ has to be chosen such that for each exclusive edge e the parts containing the attachments of e share a face. It can be shown that any embedding choice for P-nodes and R-nodes that satisfies these conditions can, after possibly flipping it, be used to obtain a SEFE [1, Theorem 1]. The proof does not modify the order of exclusive edges around degree-2 vertices of G∩ , and therefore applies to OrthoSEFE-2 as well. Now let µ be an S-node. Let ε be a virtual edge of skel(µ), Gε be the subgraph represented by ε, and ν be the corresponding neighbor of µ in the SPQR-tree of G. An attachment of ν with respect to µ is an interior vertex of Gε that is incident to an important edge e for µ. If ν has such an attachment, then it is 10 u u u Pε skel(ν) Cε v (a) v (b) v (c) Fig. 6. (a) Skeleton of an S-node µ in which the R-node ν corresponding to the virtual edge ε = (u, v) is expanded to show its skeleton. (b) Replacing ε with cycle Cε . (c) Replacing Cε with path Pε ; vertices a1 , a2 , x1 , . . . , x4 , b1 , b2 are green boxes. a P- or R-node. It is a necessary condition on the embedding of Gε that each attachment x with respect to µ must be incident to a face incident to the virtual edge twin(ε) of skel(ν) representing µ, and that their clockwise circular order together with the poles of ε is fixed up to reversal [11, Lemma 8]. For the purpose of avoiding crossings in skel(µ), we can thus replace each virtual edge ε that does not represent a Q-node by a cycle Cε containing the attachments of ε with respect to µ and the poles of ε in the order Oε . We keep only the important edges of µ. Altogether this results in an instance hGµ1 , Gµ2 i of SEFE modeling the requirements for skel(µ); see Figs. 6(a) and 6(b). Lemma 5. Let hG1 , G2 i be an instance of OrthoSEFE-2 whose shared graph is biconnected. Then hG1 , G2 i admits an OrthoSEFE if and only if all instances hGµ1 , Gµ2 i admit an OrthoSEFE. Next, we transform a given instance hGµ1 , Gµ2 i of OrthoSEFE-2 as above into an equivalent instance hGµ1 , Gµ2 i whose shared graph is a cycle. Let Cεi be the cycles corresponding to the neighbor νi , i = 1, . . . , k of µ in hGµ1 , Gµ2 i. To obtain the instance hGµ1 , Gµ2 i, we replace each cycle Cεi with poles u and v by a path Pεi from u to v that first contains two special vertices a1 , a2 followed by the clockwise path from u to v (excluding the endpoints), then four special vertices x1 , . . . , x4 , then the counterclockwise path from u to v (excluding the endpoints), and finally two special vertices b1 , b2 followed by v. In addition to the existing exclusive edges (note that we do not remove any vertices), we add to G1 the exclusive edges (a2 , x3 ), (x1 , x3 ), (x2 , x4 ), (x2 , b1 ), and to G2 the exclusive edges (a1 , x3 ) and (x2 , b2 ) to G2 ; see Fig. 6(c). The above reduction together with the next lemma implies the main result. Lemma 6. hGµ1 , Gµ2 i admits an OrthoSEFE if and only if hGµ1 , Gµ2 i does. Theorem 6. OrthoSEFE-2 when the shared graph is biconnected is polynomialtime Turing reducible to OrthoSEFE-2 when the shared graph is a cycle. Also, the reduction does not increase the maximum degree of the union graph. Corollary 1. OrthoSEFE-2 can be solved in polynomial time for instances whose shared graph is biconnected and whose union graph has maximum degree 5. 11 to vn vi or (a) around v1 vn vi v1 to v1 (b) around v2 , . . . , vn−1 (c) around vn Fig. 7. Constructing a drawing with at most three bends per edge Observe that, from the previous results it is not hard to also obtain a SEFE satisfying the orthogonality constraints, if it exists. We show how to construct an orthogonal geometric realizations of such a SEFE. Theorem 7. Let hG1 , . . . , Gk i be a positive instance of OrthoSEFE-k whose shared graph is biconnected. Then, there exists an OrthoSEFE hΓ1 , Γ2 , . . . , Γk i of hG1 , . . . , Gk i in which every edge has at most three bends. Proof sketch. We assume that a SEFE satisfying the orthogonality constraints is given. We adopt the method of Biedl and Kant [9]. We draw the vertices with increasing y-coordinates with respect to an s-t-ordering [15] v1 , . . . , vn on the shared graph. We choose the face to the left of (v1 , vn ) as the outer face of the union graph. The edges will bend at most on y-coordinates near their incident vertices and are drawn vertically otherwise. Fig. 11 indicates, how the ports are assigned. We make sure that an edge may only leave a vertex to the bottom if it is incident to vn or to a neighbor with a lower index. Thus, there are exactly three bends on {v1 , vn }. Any other edge {vi , vj }, 1 ≤ i < j ≤ n has at most one bend around vi and at most two bends around vj .  6 Conclusions and Future Work In this work we introduced and studied the problem OrthoSEFE-k of realizing a SEFE in the orthogonal drawing style. While the problem is already NP-hard even for instances that can be efficiently tested for a SEFE, we presented a polynomial-time testing algorithm for instances consisting of two graphs whose shared graph is biconnected and whose union graph has maximum degree 5. We have also shown that any positive instance whose shared graph is biconnected can be realized with at most three bends per edge. We conclude the paper by presenting a lemma that, together with Theorem 6, shows that it suffices to only focus on a restricted family of instances to solve the problem for all instances whose shared graph is biconnected. Lemma 7. Let hG1 , G2 i be an instance of OrthoSEFE-2 whose shared graph G∩ is a cycle. An equivalent instance hG∗1 , G∗2 i of OrthoSEFE-2 such that (i) the shared graph G∗∩ is a cycle, (ii) graph G∗1 is outerplanar, and (iii) no two degree-4 vertices in G∗1 are adjacent, can be constructed in polynomial time. 12 References 1. P. Angelini, G. D. Battista, F. Frati, M. Patrignani, and I. Rutter. Testing the simultaneous embeddability of two graphs whose intersection is a biconnected or a connected graph. J. Discrete Algorithms, 14:150–172, 2012. 2. P. Angelini, G. Da Lozzo, G. Di Battista, F. Frati, M. Patrignani, and I. Rutter. Beyond level planarity. In M. Nöllenburg and Y. Hu, editors, GD ’16, LNCS. Springer, 2017. To appear. 3. P. Angelini, G. D. Lozzo, and D. Neuwirth. Advancements on SEFE and partitioned book embedding problems. Theor. Comput. Sci., 575:71–89, 2015. 4. E. N. Argyriou, M. A. Bekos, M. Kaufmann, and A. Symvonis. Geometric RAC simultaneous drawings of graphs. J. Graph Algorithms Appl., 17(1):11–34, 2013. 5. L. Auslander and S. V. Parter. On embedding graphs in the sphere. J. Math. Mech., 10(3):517–523, 1961. 6. G. D. Battista and R. Tamassia. On-line maintenance of triconnected components with SPQR-trees. Algorithmica, 15(4):302–318, 1996. 7. G. D. Battista and R. Tamassia. On-line planarity testing. SIAM J. Comput., 25(5):956–997, 1996. 8. M. A. Bekos, T. C. van Dijk, P. Kindermann, and A. Wolff. Simultaneous drawing of planar graphs with right-angle crossings and few bends. J. Graph Algorithms Appl., 20(1):133–158, 2016. 9. T. Biedl and G. Kant. A better heuristic for orthogonal graph drawings. Comput. Geom., 9(3):159–180, 1998. 10. T. Bläsius, A. Karrer, and I. Rutter. Simultaneous embedding: Edge orderings, relative positions, cutvertices. In S. K. Wismath and A. Wolff, editors, GD’13, volume 8242 of LNCS, pages 220–231. Springer, 2013. 11. T. Bläsius, A. Karrer, and I. Rutter. Simultaneous embedding: Edge orderings, relative positions, cutvertices. ArXiv e-prints, abs/1506.05715, 2015. 12. T. Bläsius, S. G. Kobourov, and I. Rutter. Simultaneous embedding of planar graphs. In R. Tamassia, editor, Handbook of Graph Drawing and Visualization. CRC Press, 2013. 13. T. Bläsius and I. Rutter. Disconnectivity and relative positions in simultaneous embeddings. Comput. Geom., 48(6):459–478, 2015. 14. T. Bläsius and I. Rutter. Simultaneous PQ-ordering with applications to constrained embedding problems. ACM Trans. Alg., 12(2):16, 2016. 15. U. Brandes. Eager st-ordering. In R. H. Möhring and R. Raman, editors, ESA’02, volume 2461 of LNCS, pages 247–256. Springer, 2002. 16. G. Di Battista and R. Tamassia. On-line graph algorithms with SPQR-trees. In M. S. Paterson, editor, ICALP’90, volume 443 of LNCS, pages 598–611. Springer, 1990. 17. A. Estrella-Balderrama, E. Gassner, M. Jünger, M. Percan, M. Schaefer, and M. Schulz. Simultaneous geometric graph embeddings. In S. Hong, T. Nishizeki, and W. Quan, editors, GD’07, volume 4875 of LNCS, pages 280–290. Springer, 2007. 18. C. Gutwenger and P. Mutzel. A linear time implementation of SPQR-trees. In J. Marks, editor, GD’00, volume 1984 of LNCS, pages 77–90. Springer, 2000. 19. B. Haeupler, K. R. Jampani, and A. Lubiw. Testing simultaneous planarity when the common graph is 2-connected. J. Graph Algorithms Appl., 17(3):147–171, 2013. 20. K. R. Jampani and A. Lubiw. Simultaneous interval graphs. In O. Cheong, K. Chwa, and K. Park, editors, ISAAC’10, volume 6506 of LNCS, pages 206–217, 2010. 13 21. K. R. Jampani and A. Lubiw. The simultaneous representation problem for chordal, comparability and permutation graphs. J. Graph Algorithms Appl., 16(2):283–315, 2012. 22. M. Jünger and M. Schulz. Intersection graphs in simultaneous embedding with fixed edges. J. Graph Algorithms Appl., 13(2):205–218, 2009. 23. B. M. E. Moret. Planar NAE3SAT is in P. ACM SIGACT News, 19(2):51–54, 1988. 24. B. M. E. Moret. Theory of Computation. Addison-Wesley-Longman, 1998. 25. C. H. Papadimitriou. Computational complexity. Academic Internet Publ., 2007. 26. M. Schaefer. Toward a theory of planarity: Hanani–Tutte and planarity variants. J. Graph Algorithms Appl., 17(4):367–440, 2013. 27. T. J. Schaefer. The complexity of satisfiability problems. In R. J. Lipton, W. A. Burkhard, W. J. Savitch, E. P. Friedman, and A. V. Aho, editors, Proceedings of the 10th Annual ACM Symposium on Theory of Computing (STOC’78), pages 216–226. ACM, 1978. 28. W. Shih, S. Wu, and Y. Kuo. Unifying maximum cut and minimum cut of a planar graph. IEEE Trans. Computers, 39(5):694–697, 1990. 29. R. Tamassia. On embedding a graph in the grid with the minimum number of bends. SIAM J. Comput., 16(3):421–444, 1987. 14 Appendix A Definitions for the appendix In Section 2 we already discussed how to assign the exclusive edges to either of the Sk two sides of C. We formalise this assignment by means of a function A : i=1 Ei \ E(G∩ ) → {l, r}, where A(e) = l (resp. A(e) = r) if edge e lies to the left (resp. to the right) of C, according to an arbitrary orientation of C. Connectivity and SPQR-trees. A graph G = (V, E) is connected if there is a path between any two vertices. A cutvertex is a vertex whose removal disconnects the graph. A separating pair {u, v} is a pair of vertices whose removal disconnects the graph. A connected graph is biconnected if it does not have a cutvertex and a biconnected graph is 3-connected if it does not have a separating pair. We consider st-graphs with two special pole vertices s and t. The family of st-graphs can be constructed in a fashion very similar to series-parallel graphs. Namely, an edge st is an st-graph with poles s and t. Now let Gi be an stgraph with poles si , ti for i = 1, . . . , k and let H be a planar graph with two designated adjacent vertices s and t and k + 1 edges st, e1 , . . . , ek . We call H the skeleton of the composition and its edges are called virtual edges; the edge st is the parent edge and s and t are the poles of the skeleton H. To compose the Gi ’s into an st-graph with poles s and t, we remove the edge st from H and replace each ei by Gi for i = 1, . . . , k by removing ei and identifying the poles of Gi with the endpoints of ei . In fact, we only allow three types of compositions: in a series composition the skeleton H is a cycle of length k + 1, in a parallel composition H consists of two vertices connected by k + 1 parallel edges, and in a rigid composition H is 3-connected. For every biconnected planar graph G with an edge st, the graph G − st is an st-graph with poles s and t [16]. Much in the same way as series-parallel graphs, the st-graph G − st gives rise to a (de-)composition tree T describing how it can be obtained from single edges. The nodes of T corresponding to edges, series, parallel, and rigid compositions of the graph are Q-, S-, P-, and R-nodes, respectively. To obtain a composition tree for G, we add an additional root Qnode representing the edge st. We associate with each node µ the skeleton of the composition and denote it by skel(µ). For a Q-node µ, the skeleton consists of the two endpoints of the edge represented by µ and one real and one virtual edge between them representing the rest of the graph. For a node µ of T , the pertinent graph pert(µ) is the subgraph represented by the subtree with root µ. For a virtual edge ε of a skeleton skel(µ), the expansion graph of ε is the pertinent graph pert(µ0 ) of the neighbor µ0 corresponding to ε when considering T rooted at µ. The SPQR-tree of G with respect to the edge st, originally introduced by Di Battista and Tamassia [16], is the (unique) smallest decomposition tree T for G. Using a different edge s0 t0 of G and a composition of G − s0 t0 corresponds to rerooting T at the node representing s0 t0 . It thus makes sense to say that 15 T is the SPQR-tree of G. The SPQR-tree of G has size linear in G and can be computed in linear time [18]. Planar embeddings of G correspond bijectively to planar embeddings of all skeletons of T ; the choices are the orderings of the parallel edges in P-nodes and the embeddings of the R-node skeletons, which are unique up to a flip. When considering rooted SPQR-trees, we assume that the embedding of G is such that the root edge is incident to the outer face, which is equivalent to the parent edge being incident to the outer face in each skeleton. We remark that in a planar embedding of G, the poles of any node µ of T are incident to the outer face of pert(µ). Hence, in the following we only consider embeddings of the pertinent graphs with their poles lying on the same face. B Omitted or Sketched Proofs from Section 2 Theorem 1. An instance hG1 , . . . , Gk i of OrthoSEFE-k has an OrthoSEFE if and only if it admits a SEFE satisfying the orthogonality constraints. Proof. For the if part, let E be the embedding of G∩ determined by the SEFE hE1 , . . . Ek i of hG1 , . . . , Gk i. Observe that the orthogonality constraints at each vertex define (i) whether a degree 2 vertex of G∩ has to be drawn straight or bent, and (ii) which face incident to a degree 3 vertex of G∩ has to be assigned the 180◦ angle. It is not hard to see that a planar orthogonal drawing Γ of G∩ in which the embedding of G∩ is E satisfying such requirements can be constructed. We draw the exclusive edges in each Ei as orthogonal polylines in Γ inside the face of E determined by the SEFE. The fact that the exclusive edges of each Ei can be drawn in Γ without introducing any crossings descends from the fact that Ei is a planar embedding of Gi . For the only if part, let v be a vertex in G∩ such that the orthogonality constraints are not satisfied at v. If v has exactly two neighbors u and w in G∩ , then we need to assign a port to two exclusive edges of the same graph (one for each of these edges) on one side of the path uvw and a port to at least one exclusive edge on the other side of the path uvw. If v has degree 3 in G∩ , then we need to assign a port to an exclusive edge between a pair of edges of G∩ and a port to an exclusive edge between a different pair of edges of G∩ . Hence, in both cases we need at least five ports, which is not possible on the grid. t u C Omitted or Sketched Proofs from Section 3 Theorem 3. OrthoSEFE-3 is NP-complete, even for instances hG1 = (V, E1 ), G2 = (V, E2 ), G3 = (V, E3 )i with sunflower intersection in which (i) the shared graph G∩ = (V, E1 ∩ E2 ∩ E3 ) is a cycle and (ii) G1 and G2 are outerplanar graphs with maximum degree 3. Proof. The membership in NP directly follows from Theorem 2, since an assignment A, which is a certificate for our problem, can be easily verified in polynomial time to satisfy all the planarity and the orthogonality constraints. 16 To prove that the problem is NP-hard, we show a reduction from the NPcomplete problem Positive Exactly-Three Nae3Sat [24], which is the variant of Nae3Sat in which each clause consists of exactly three unnegated literals. See Fig. 2. Let x1 , x2 , . . . , xn be the variables and let c1 , c2 , . . . , cm be the clauses of a 3-CNF formula φ of Positive Exactly-Three Nae3Sat. We show how to construct an equivalent instance hG1 = (V, E1 ), G2 = (V, E2 ), G3 = (V, E3 )i of OrthoSEFE-3; refer to Fig. 2(b). Assume, without loss of generality, that the literals in each clause cj = (xja , xjb , xjc ) are such that a > b > c, if j is odd, and a < b < c, otherwise. A variable-clause gadget Vij for a variable xi belonging to a clause cj is a subgraph of G∪ defined as follows. Gadget Vji contains a path (sji , uji , wij , vij , zij , rij , tji ) belonging to G∩ , and edges {uji , vij } and {wij , zij } belonging to E3 ; see Fig. 2(a). The clause gadget C j for a clause cj is a subgraph of G∪ defined as follows. Gadget C j contains a path (sj , αj , yaj , β j , ybj , dj1 , . . . , dj6 , γ j , ycj , δ j , tj ) belonging to G∩ , and edges {αj , β j }, {β j , γ j }, {γ j , δ j }, {dj1 , dj3 }, {dj2 , dj4 }, {dj3 , dj5 }, {dj4 , dj6 } belonging to E3 , and edges {β j , dj3 } and {dj4 , γ j } belonging to E1 ; see Fig. 2(a). Initialize G∪ to the union of Vij , for i = 1, . . . , n and j = 1, . . . , m, and of C j , for j = 1, . . . , m. Then, for j = 1, . . . , m and for i = 1, . . . , n − 1 identify vertex tji with vertex sji+1 , if j is odd, or identify vertex tji+1 with vertex sji , otherwise. Further, for j = 1, . . . , m (where m + 1 = 1), identify vertex tjn with vertex sj j j and vertex tj with vertex sj+1 n , if j is odd, or identify vertex t1 with vertex s j+1 and vertex tj with vertex s1 , otherwise. To complete the construction of hG1 , G2 , G3 i we add to G∪ exclusive edges as follows. For i = 1, . . . , n and for j = 1, . . . , m − 1, we add an edge {wij , wij+1 } to E2 , if j is odd, or to E1 , otherwise. We call these edges transmission edges. Further, for i = 1, . . . , n and for j = 1, . . . , m, we add an edge {wij , yij }, if xi ∈ cj , or an edge {wij , rij }, otherwise. Clearly, the construction of instance hG1 , G2 , G3 i can be completed in polynomial time. Graph G∩ is a cycle, as we already observed. Also, the transmission edges in E1 (in E2 ) do not alternate along G∩ , since the variable-clause gadgets appear along G∩ in the order V1j , . . . , Vnj , if j is odd, or in the order Vnj , . . . , V1j , otherwise. Also, no transmission edge in E1 alternates with edges (β j , d3 ) and (d4 , γ j ), for any j, and such edges do not alternate with each other by construction. Hence, G1 and G2 are outerplanar. The fact that G1 and G2 have maximum degree 3 also directly follows from the construction. Given a positive instance φ of Positive Exactly-Three Nae3Sat, we show that hG1 , G2 , G3 i is a positive instance of OrthoSEFE-3. Given a satisfying truth assignment T : X → {true, false} where X denotes the set of variables in φ, we construct an assignment A of the exclusive edges of E1 ∪ E2 ∪ E3 to the two sides of G∩ satisfying all the planarity and the orthogonality constraints. For i = 1, . . . , n and for j = 1, . . . , m, we set A(e) = l, for each exclusive edge e ∈ E1 ∪ E2 ∪ E3 incident to wij , if T (xi ) = true, or A(e) = r, otherwise. For i = 17 1, . . . , n and for j = 1, . . . , m, we set A(uji , vij ) = r, if T (xi ) = true, or A(uji , vij ) = l, otherwise. For each clause cj = (xja , xjb , xjc ), we set A(αj , β j ) = l, if T (xa ) = false, or A(αj , β j ) = r, otherwise; we set A(β j , γ j ) = l, if T (xb ) = false, or A(β j , γ j ) = r, otherwise; and we set A(γ j , δ j ) = l, if T (xc ) = false, or A(γ j , δ j ) = r, otherwise. Finally, for each clause cj = (xja , xjb , xjc ), consider the literal x◦ with ◦ ∈ {a, c} such that T (x◦ ) = T (xb ), if any, otherwise let x◦ = xa . Suppose that x◦ = xa ; set A(dj1 , dj3 ) = A(dj3 , dj5 ) = A(β j , dj3 ) = r and set A(dj2 , dj4 ) = A(dj4 , dj6 ) = A(dj4 , γ j ) = l, if T (x◦ ) = false, or set A(dj1 , dj3 ) = A(dj3 , dj5 ) = A(β j , dj3 ) = l and set A(dj2 , dj4 ) = A(dj4 , dj6 ) = A(dj4 , γ j ) = r, otherwise. Suppose that x◦ = xc ; set A(dj1 , dj3 ) = A(dj3 , dj5 ) = A(β j , dj3 ) = l and set A(dj2 , dj4 ) = A(dj4 , dj6 ) = A(dj4 , γ j ) = r, if T (x◦ ) = false, or set A(dj1 , dj3 ) = A(dj3 , dj5 ) = A(β j , dj3 ) = l and set A(dj2 , dj4 ) = A(dj4 , dj6 ) = A(dj4 , γ j ) = r, otherwise. We show that A satisfies the planarity constraints. First observe that the planarity constraints for the edges in E1 and E2 are trivially satisfied by A since G1 and G2 are outerplanar. As for the edges in E3 , we have that the only pairs of edges that alternate along G∩ are h(uji , vij ), (wij , zij )i, for i = 1, . . . , n and for j = 1, . . . , m, pairs h(waj , yaj ), (αj , β j )i, h(wbj , ybj ), (β j , γ j )i, and h(wcj , ycj ), (γ j , δ j )i, for j = 1, . . . , m, and the edges incident to the dummy vertices dj1 , . . . , dj6 , for j = 1, . . . , m. However, it is easy to verify that A assigns alternating edges to different sides of G∩ . We show that A satisfies the orthogonality constraints at every vertex. For all the vertices except for wij , β j , dj3 , dj4 , and γ j , for i = 1, . . . , n and for j = 1, . . . , m, this is true since they have only one incident exclusive edge. For vertices wij , dj3 , and dj4 , with i = 1, . . . , n and for j = 1, . . . , m, this is true since all the edges incident to wij are assigned to the same side of G∩ by A, by construction. For vertex βj and γj , we distinguish two cases based on whether there exists a ◦ ∈ {a, c} with T (xb ) = T (x◦ ): (i) If this is case, let ◦ = c without loss of generality; the case ◦ = a can be shown analogously. Then, A(β j , γ j ) = A(γ j , δ j ) = A(dj4 , γ j ), by construction, and hence the orthogonality constraints are satisfied at γ j . To prove that they are also satisfied at β j , it suffices to show that the two edges of E3 incident to β j are assigned to different sides of G∩ , given that β j has degree 3 in G1 and degree 2 in G2 . Namely, due to the fact that T is a Nae3Sat truth assignment we have that T (xa ) 6= T (xb ), and hence A(αj , β j ) 6= A(β j , γ j ). (ii) In the second case, T (xa ) 6= T (xb ) 6= T (xc ), hence we have that A(αj , β j ) 6= A(β j , γ j ) 6= A(γ j , δ j ). Since vertices β j and γ j have degree 4 in G3 , degree 3 in G1 , and degree 2 in G2 , this implies that the orthogonality constraints are satisfied at β j and at γ j . Suppose that hG1 = (V, E1 ), G2 = (V, E2 ), G3 = (V, E3 )i is a positive instance of OrthoSEFE-3 and let A be the corresponding assignment of the exclusive edges to the sides of G∩ . We show how to construct a Nae3Sat truth assignment T that satisfies φ. For i = 1, . . . , n, we set T (xi ) = true if and only if A(wi1 , zi1 ) = l. We start by proving that, for each i = 1, . . . , n, all the edges incident to wij , with 1 ≤ j ≤ m, are assigned to the same side of G∩ . Observe that, for each i = 1, . . . , n and for each for each j = 1, . . . , m, the 18 two edges in G3 incident to wij both alternate with edge (uji , vij ) along G∩ and hence are assigned to the same side of G∩ by the planarity constraints. Hence, by the orthogonality constraints at wij , all the exclusive edges in E1 ∪ E2 incident to wij lie on the same side of G∩ as (wij , zij ). Further, since any two vertices wij and wij+1 , are connected by a transmission edge in either E1 or in E2 , the statement follows. This property allows us to focus on each clause separately. Let cj = (xja , xjb , xjc ) be a clause in φ, with 1 ≤ j ≤ m. We show that T (xja ) = T (xjb ) = T (xjc ) does not hold. First, we show that A(β j , dj3 ) 6= A(dj4 , γ j ). Namely, by the planarity constraints, A(dj1 , dj3 ) = A(dj3 , dj5 ) 6= A(dj2 , dj4 ) = A(dj4 , dj6 ); then, by the orthogonality constraints at dj3 and at dj4 , we have that A(β j , dj3 ) = A(dj1 , dj3 ) = A(dj3 , dj5 ) and that A(dj4 , γ j ) = A(dj2 , dj4 ) = A(dj4 , dj6 ), and the statement follows. Second, A(αj , β j ) = A(β j , γ j ) = A(γ j , δ j ) does not hold, since A(β j , dj3 ) 6= A(dj4 , γ j ) and by the orthogonality constraints at β j and at γ j . This implies that A(waj , yaj ) = A(wbj , ybj ) = A(wcj , ycj ) does not hold, and hence A(waj , zaj ) = A(wbj , zbj ) = A(wcj , zcj ) does not hold, since all the edges incident to wij , with 1 ≤ j ≤ m, are assigned to the same side of G∩ . This concludes the proof that T (xja ) = T (xjb ) = T (xjc ) does not hold. It is easy to see that the reduction can be performed in polynomial time and that it can be extended to any k > 3 by subdividing two edges of G∩ for each additional graph Gi and by introducing an exclusive edge between these vertices only belonging to Gi . t u D Omitted or Sketched Proofs from Section 4 Lemma 2. Let hG1 , G2 i be an instance of OrthoSEFE-2 such that G∩ = (V, E1 ∩ E2 ) is a cycle and G1 has maximum degree 3. It is possible to construct in polynomial time an equivalent instance hG∗1 , G∗2 i of OrthoSEFE-2 such that G∗∩ = (V ∗ , E1∗ ∩ E2∗ ) is a cycle and G∗1 is an outerplanar graph with maximum degree 3. Proof. We describe how to construct an equivalent instance hG01 , G02 i of OrthoSEFE2 such that G0∩ is a cycle, G01 has maximum degree 3 and the number of pairs of edges in G01 that alternate along G0∩ is smaller than the number of pairs of edges in G1 that alternate along G∩ . Note that repeatedly performing this transformation eventually yields an equivalent instance hG∗1 , G∗2 i satisfying the requirements of the lemma. Consider two edges e = (u, v) and f = (w, z) of G1 such that u, w, v, z appear in this order along cycle G∩ and such that the path Pu,z in G∩ between u and z that contains v and w has minimal length. If G1 is not outerplanar, edges e and f always exist. 0 Initialize G0∩ = G∩ . Replace path Pu,z in G0∩ by a path Pu,z , as follows; refer to Fig. 8. Let H1 , H2 , and H3 be the sets of vertices between u and w, between 0 w and v, and between v and z in G∩ . Path Pu,z contains u, then the vertices 0 in H1 , then a dummy vertex w , then the vertices in H2 , then a dummy vertex 19 e u H1 eav eaw w H2 ebw v f H3 e eaw ebv z u H1 w0 eav H2 z0 w v u0 f0 ebw (a) ebv e0 v0 H3 z f (b) Fig. 8. Instances (a) hG1 , G2 i and (b) hG01 , G02 i for the proof of Lemma 2. Edges of the shared graph G∩ are black. Exclusive edges of G1 are red and those of G2 are blue. z 0 , then three dummy vertices x1 , x2 , x3 , then v, then four dummy vertices x4 , x5 , x6 , x7 , then w, then three dummy vertices x8 , x9 , x10 , then three dummy vertices u0 , x11 , and v 0 , then the vertices in H3 , and finally z. Note that G0∩ contains all the vertices of G∩ , plus a set of dummy vertices. We now describe the exclusive edges in E10 and E20 . Initialize E10 = E1 and E20 = E2 . Add edges e0 = (u0 , v 0 ) and f 0 = (w0 , z 0 ) to E10 . Also, add edges (z 0 , x2 ), (z 0 , x3 ), (x1 , v), (x3 , x4 ), (v, x6 ), (x5 , w), (x7 , x8 ), (w, x10 ), (x8 , u0 ), and (x9 , u0 ) to E20 . Finally, replace in E20 each edge (x, w) incident to w by an edge (x, w0 ) and each edge (x, v) incident to v with an edge (x, v 0 ). Before proving the statement, we observe an important property that will be used in the following, namely that there exists no exclusive edge in E1 , and hence in E10 , with an endpoint in H2 and the other one not in H2 . In fact, there exists no edge connecting a vertex of H2 to any of u, v, w, z, since these vertices are already incident to edges e and f , respectively, and since G1 has maximum degree 3. Also, there exists no exclusive edge g connecting a vertex of H2 to a vertex of H1 (of H3 ), since in this case g would alternate with f (with e), hence contradicting the minimality of path Pu,z . Finally, the existence of an exclusive edge connecting a vertex of H2 to any other vertex in V would immediately make the instance negative, since G1 would not be planar. We now prove that hG01 , G02 i satisfies the required properties. First, graph G0∩ is a cycle by construction. Second, G01 has maximum degree 3, since (i) every vertex in V ∩ V 0 is incident to the same edges in E10 as in E1 , (ii) dummy vertices xi , with i = 1, . . . , 11, have degree 2, and (iii) dummy vertices w0 , z 0 , u0 , and v 0 have degree 3. Third, the number of pairs of alternating edges in hG01 , G02 i is smaller than in hG1 , G2 i. In fact, (i) edge e0 does not alternate with any edge of E10 , since x11 is not incident to any exclusive edge in E10 , (ii) edge f 0 does not alternate with any edge in E10 , since there exists no exclusive edge in E10 with an endpoint in H2 and the other one not in H2 , and (iii) all pairs of edges in E1 ∩ E10 that alternate along G0∩ also alternate along G∩ , except for edges e and f , which alternate along G∩ but not along G0∩ . We now prove that hG01 , G02 i is equivalent to hG1 , G2 i. Suppose that hG1 , G2 i admits an OrthoSEFE hΓ1 , Γ2 i. By Theorem 2, hΓ1 , Γ2 i determines an assignment A of the exclusive edges of E1 and of E2 to the two sides of G∩ satisfying all the planarity and the orthogonality constraints. We 20 show how to construct an assignment A0 of the exclusive edges of E10 and of E20 to the two sides of G0∩ satisfying all the constraints. For each exclusive edge g ∈ E1 ∩ E10 , set A0 (g) = A(g). Also, set A0 (e0 ) = A(e) and A0 (f 0 ) = A(f ). For each exclusive edge g ∈ E2 ∩ E20 , set A0 (g) = A(g). Also, for each edge (x, w0 ) (resp. (x, v 0 )) incident to w0 (resp. to v 0 ), set A0 (x, w0 ) = A(x, w) (resp. A0 (x, v 0 ) = A(x, v)). Further, set A0 (x1 , v) = A0 (v, x6 ) = A0 (x7 , x8 ) = A0 (x8 , u0 ) = A0 (x9 , u0 ) = A(e) and set A0 (z 0 , x2 ) = A0 (z 0 , x3 ) = A0 (x3 , x4 ) = A0 (x5 , w) = A0 (w, x10 ) = A(f ). The planarity constraints for the edges of G01 are satisfied since any pair of edges that alternate along G0∩ also alternate in G∩ and since their assignment in A0 and in A are the same, by construction. We prove that the planarity constraints for the edges of G02 are satisfied by A . For the edges that are not incident to any dummy vertex, this is true for the same reason as for the edges of G01 . For each edge (x, w0 ) incident to w0 , this is true since A0 (x, w0 ) = A(x, w), and since (x, w0 ) alternates with an edge g ∈ E20 along G0∩ if and only if edge (x, w) alternates with an edge g ∗ along G∩ , where g ∗ = g if g is not incident to v 0 , while g ∗ = (y, v) if g = (y, v 0 ). Analogous arguments hold for each edge (x, v 0 ) incident to v 0 . Finally, the fact that the planarity constraints for each edge incident to two dummy vertices are satisfied by A0 can be easily verified; recall that A(e) 6= A(f ). 0 We now prove that the orthogonality constraints are satisfied by A0 at every vertex of V 0 . For the vertices in V 0 ∩V \{w, v}, this is true since they are satisfied by A and since for every exclusive edge g ∈ E10 ∪ E20 incident to these vertices, we have that g ∈ E1 ∪ E2 , by construction, and A0 (g) = A(g). For vertex w, this is true since A0 (x5 , w) = A0 (w, x10 ) = A(f ) = A0 (f ). For vertex v, this is true since A0 (x1 , v) = A0 (v, x6 ) = A(e) = A0 (e). For vertex u0 , this is true since A0 (x8 , u0 ) = A0 (x9 , u0 ) = A0 (e0 ) = A(e). For vertex z 0 , this is true since A0 (z 0 , x2 ) = A0 (z 0 , x3 ) = A0 (f 0 ) = A(f ). For vertex w0 , assume there exist two exclusive edges eaw , ebw ∈ E20 that are incident to w0 , the case in which there exists only one or none of them being trivial. Since A0 (eaw ) = A(eaw ), A0 (ebw ) = A(ebw ), and A0 (f 0 ) = A(f ), and since the orthogonality constraints at w are satisfied by A, the orthogonality constraints at w0 are satisfied by A0 . Analogously, the orthogonality constraints at v 0 between edges eav , ebv ∈ E20 , if any, and edge e0 ∈ E10 are satisfied by A0 since the same constraints at v between edges eav , ebv ∈ E2 and e ∈ E1 are satisfied by A. Since vertices xi , with i = 1, . . . , 11, have degree 2 in G01 , this concludes the proof that A0 satisfies the orthogonality constraints. Suppose that hG01 , G02 i admits OrthoSEFE hΓ10 , Γ20 i, and let A0 be the corresponding assignment of the exclusive edges of E10 and of E20 to the two sides of G0∩ . We show how to construct an assignment A of the exclusive edges of E1 and of E2 to the two sides of G∩ satisfying all the planarity and the orthogonality constraints. For each exclusive edge g ∈ E1 , set A(g) = A0 (g). For each exclusive edge g ∈ E2 ∩ E20 , set A(g) = A0 (g). Also, for each edge (x, w) (resp. (x, v)) incident to w (resp. to v), set A(x, w) = A0 (x, w0 ) (resp. A(x, v) = A0 (x, v 0 )). 21 We prove that the planarity constraints for the edges of G1 are satisfied by A. For each pair (e1 , e2 ) of exclusive edges in E1 such that {e1 , e2 } = 6 {e, f }, this is true since e1 and e2 alternate along G∩ if and only if they alternate along G0∩ , by construction. For pair (e, f ), this is true for the following reason. By planarity constraints, we have A0 (x1 , v) = A0 (v, x6 ) 6= A0 (x3 , x4 ); hence, by orthogonality constraints at vertex v, we have A0 (e) = A0 (x1 , v) = A0 (v, x6 ). Analogously, we have A0 (f ) = A0 (x5 , w) = A0 (w, x10 ) 6= A0 (x7 , x8 ). Since, by planarity constraints, A0 (v, x6 ) 6= A0 (x5 , w), we have A0 (e) 6= A0 (f ) and hence A(e) 6= A(f ). We prove that the planarity constraints for the edges of G2 are satisfied by A. For each pair (e1 , e2 ) of exclusive edges in E2 such that neither e1 nor e2 is incident to either of w and v, this is true since e1 and e2 alternate along G∩ if and only if they alternate along G0∩ , by construction. For each edge (x, w) incident to w, this is true since A(x, w) = A0 (x, w0 ), and since (x, w) alternates with an edge g ∈ E2 along G∩ if and only if edge (x, w) alternates with an edge g ∗ along G0∩ , where g ∗ = g if g is not incident to v, while g ∗ = (y, v 0 ) if g = (y, v). Analogous arguments hold for each edge (x, v) incident to v. We now prove that the orthogonality constraints are satisfied by A at every vertex of V . For vertices in V \{w, v}, this is true since they are satisfied by A0 and since for every exclusive edge g ∈ E1 ∪ E2 incident to these vertices, g ∈ E10 ∪ E20 , by construction, and A(g) = A0 (g). In order to prove that the constraints are satisfied also at w and v, we first argue that A(f ) = A0 (f 0 ) and A(e) = A0 (e0 ): Namely, by planarity constraints, A0 (z 0 , x2 ) 6= A0 (x1 , v) 6= A0 (z 0 , x3 ), and hence A0 (z 0 , x2 ) = A0 (z 0 , x3 ). Similarly, A0 (x5 , w) 6= A0 (x7 , x8 ) 6= A0 (w, x10 ), and hence A0 (x5 , w) = A0 (w, x10 ). Then, by using the longer chain of alternating edges we get A0 (z 0 , x3 ) 6= A0 (x1 , v) 6= A0 (x3 , x4 ) 6= A0 (v, x6 ) 6= A0 (x5 , w) and thus A0 (z 0 , x3 ) = A0 (x5 , w). Finally, by orthogonality constraints at z 0 and w, we get A0 (f 0 ) = A0 (z 0 , x3 ) and A0 (f ) = A0 (x5 , w). Since A(f ) = A0 (f ) we conclude A(f ) = A0 (f 0 ). The equality A(e) = A0 (e0 ) follows symmetrically. We now prove that the orthogonality constraints are satisfied at w and v. For vertex w, assume there exist two exclusive edges eaw , ebw ∈ E2 incident to w, the case in which there exists only one or none of them being trivial. Since A(eaw ) = A0 (eaw ), A(ebw ) = A0 (ebw ), and A(f ) = A0 (f 0 ), and since the orthogonality constraints at w0 between eaw , ebw and f 0 are satisfied by A0 , the orthogonality constraints at w between eaw , ebw and f are satisfied by A. Analogously, the orthogonality constraints at v between edges eav , ebv ∈ E2 , if any, and edge e ∈ E1 are satisfied by A since the same constraints at v 0 between eav , ebv ∈ E20 and e0 ∈ E10 are satisfied by A0 . This concludes the proof of the lemma. t u Lemma 3. Let hG1 , G2 i be an instance of OrthoSEFE-2 such that G∩ = (V, E1 ∩ E2 ) is a cycle and each vertex v ∈ V has degree at most 3 in either G1 or G2 . It is possible to construct in polynomial time an equivalent instance hG∗1 , G∗2 i of OrthoSEFE-2 such that G∗∩ = (V ∗ , E1∗ ∩ E2∗ ) is a cycle and graph G∗1 has maximum degree 3. Proof. We describe how to construct an equivalent instance hG01 , G02 i of OrthoSEFE2 such that G0∩ is a cycle, each vertex v ∈ V 0 has degree at most 3 in either G01 22 h f e ue H1 v H2 uf e H3 uh ue f 00 e0 00 ve (a) h f0 H1 vf h00 0 v0 H2 vh uf H3 uh (b) Fig. 9. Instances (a) hG1 , G2 i and (b) hG01 , G02 i for the proof of Lemma 3. Edges of the shared graph G∩ are black. Exclusive edges of G1 are red and those of G2 are blue. or G02 , and the number of degree-4 vertices in G01 is smaller than the number of degree-4 vertices in G1 . Note that repeatedly performing this transformation eventually yields an equivalent instance hG∗1 , G∗2 i satisfying the requirements of the lemma. Consider a vertex v ∈ V such that there exists two edges e = (v, ue ), f = (v, uf ) ∈ E1 incident to v. Assume without loss of generality that ue , v, and uf appear in this order along G∩ . Suppose that there exists an edge h = (v, uh ) ∈ E2 incident to v, the other case being simpler. We describe the construction for the case in which vertices ue , v, uf , uh appear in this order along G∩ ; the other cases are analogous. Initialize G0∩ = G∩ ; refer to Fig. 9. Replace v in G0∩ by a path Pv = x1 , x2 , ve , x3 , y1 , y2 , vf , y3 , v 0 , z1 , z2 , vh , z3 composed of dummy vertices. We now describe the exclusive edges in E10 and E20 . Set Ei0 , with i = 1, 2, contains all the exclusive edges in Ei that are not incident to v. Also, E10 contains edges e00 = (ve , ue ), f 00 = (vf , uf ), and h0 = (vh , v 0 ). Finally, E20 contains edges (x1 , ve ), (x2 , x3 ), (y1 , vf ), (y2 , y3 ), (z1 , vh ), (z2 , z3 ), and edges e0 = (ve , v 0 ), f 0 = (vf , v 0 ), and h00 = (vh , uh ). We prove that hG01 , G02 i satisfies the required properties. First, graph G0∩ is a cycle by construction. Second, the degree of the vertices in V \ V 0 is the same in G01 (resp. G02 ) as in G1 (resp. as in G2 ), while all the dummy vertices have degree at most 3 in G01 . Hence, every vertex in V 0 has degree at most 3 in either G01 or G02 ; also, the number of degree-4 vertices in G01 is smaller than the number of degree-4 vertices in G1 , since v ∈ / G01 . 0 0 We now prove that hG1 , G2 i is equivalent to hG1 , G2 i. Suppose that hG1 , G2 i admits an OrthoSEFE hΓ1 , Γ2 i, and let A be the corresponding assignment of the exclusive edges of E1 and of E2 to the two sides of G∩ , which exists by Theorem 2. We show how to construct an assignment A0 of the exclusive edges of E10 and of E20 to the two sides of G0∩ satisfying all the constraints. For each exclusive edge g ∈ E1 ∪ E2 incident to v, set A0 (g) = A(g). Also, set 0 00 A (e ) = A(e), A0 (f 00 ) = A(f ), and A0 (h0 ) = A(h). Finally, set A0 (x1 , ve ) = A0 (e0 ) = A(e) and A0 (x2 , x3 ) 6= A(e); set A0 (y1 , vf ) = A0 (f 0 ) = A(f ) and A0 (y2 , y3 ) 6= A(f ); and set A0 (z1 , vh ) = A0 (h00 ) = A(h) and A0 (z2 , z3 ) 6= A(h). 23 We prove that the planarity constraints for the edges of G01 are satisfied by A . Note that, by construction, edge h0 does not alternate with any edge of G01 along G0∩ . Also, edges e00 and f 00 do not alternate with each other along G0∩ . Further, if edge e00 (edge f 00 ) alternates with an edge g ∈ G01 along G0∩ , then edge e (edge f ) alternates with g along G∩ . Finally, any two edges not incident to any dummy vertex that alternate along G0∩ also alternate along G∩ . In all the described cases, the planarity constraints are satisfied by A0 since they are satisfied by A. 0 We prove that the planarity constraints for the edges of G02 are satisfied by A . Note that, by construction, edges e0 , f 0 , (x1 , ve ), (x2 , x3 ), (y1 , vf ), (y2 , y3 ), (z1 , vh ), and (z2 , z3 ) do not alternate with any edge of G02 that is not incident to a dummy vertex along G0∩ ; it easy to verify that A0 satisfies the planarity constraints among these edges. Also, edge h00 alternates with (z2 , z3 ), but A0 (h00 ) 6= A0 (z2 , z3 ) by construction. Further, if edge h00 alternates with an edge g 6= (z2 , z3 ) ∈ G02 along G0∩ , then edge h alternates with g along G∩ . Finally, any two edges not incident to any dummy vertex that alternate along G0∩ also alternate along G∩ . In all these cases, the planarity constraints are satisfied by A0 since they are satisfied by A. 0 We now prove that the orthogonality constraints are satisfied by A0 at every vertex in V 0 . For non-dummy vertices, this is true since they are satisfied by A and since all the edges incident to them have the same assignment in A as in A0 . For vertices xi , yi , and zi , with i = 1, 2, 3, this is true since they have degree 2 in G01 . For vertex ve , this is true since A0 (e00 ) = A0 (x1 , ve ) = A0 (e0 ) = A(e); similar arguments apply for vertices vf and vh . Finally, for vertex v 0 , this is true since (i) A0 (e0 ) = A(e), A0 (f 0 ) = A(f ), and A0 (h0 ) = A(h), (ii) e, f , and h are incident to v in G2 , and (iii) A satisfies the orthogonality constraints. This concludes the proof that A0 satisfies the orthogonality constraints. Suppose that hG01 , G02 i admits an OrthoSEFE hΓ10 , Γ20 i, and let A0 be the corresponding assignment of the exclusive edges of E10 and of E20 to the two sides of G0∩ . We show how to construct an assignment A of the exclusive edges of E1 and of E2 to the two sides of G∩ satisfying all the planarity and the orthogonality constraints. For each exclusive edge g ∈ E1 ∪ E2 not incident to v, set A(g) = A0 (g). Also, set A(e) = A0 (e00 ), A(f ) = A0 (f 00 ), and A(h) = A0 (h00 ). We prove that the planarity constraints for the edges of G1 and of G2 are satisfied by A. Consider any pair of edges hg1 , g2 i of the same graph Gi , with i = 1, 2, that alternate along G∩ . If none of g1 and g2 is incident to v, then they also alternate along G0∩ . Hence, the planarity constraints are satisfied by A since they are satisfied by A0 . Otherwise, assume g1 is incident to v; note that g2 is not incident to v, since g1 and g2 alternate along G∩ . If g1 = e (if g1 = f ; if g1 = h), then edge e00 (edge f 00 ; edge h00 ) alternates with g2 along G0∩ . Further, A(e) = A0 (e00 ), A(f ) = A0 (f 00 ), A(h) = A0 (h00 ), and A(g2 ) = A0 (g2 ). Hence, the planarity constraints for these edges are satisfied by A since they are satisfied by A0 . 24 We finally prove that the orthogonality constraints are satisfied by A. For the vertices in V \ {v, ue , uf , uh }, this is true since they are satisfied by A0 and since for every exclusive edge g ∈ E1 ∪E2 incident to these vertices, we have that g ∈ E10 ∪ E20 , by construction, and A(g) = A0 (g). For vertex ue , this is true since for each edge g incident to ue different from e, it holds that A(g) = A0 (g), since A(e) = A0 (e00 ), and since the orthogonality constraints at ue are satisfied by A0 . Analogous arguments hold for vertices uf and uh . To prove that this is true for v, we first argue that A(e) = A0 (e0 ), that A(f ) = A0 (f 0 ), and that A(h) = A0 (h0 ): Namely, by planarity constraints, we get A0 (e0 ) = A0 (x1 , ve ) since they both alternate with (x2 , x3 ); hence, by orthogonality constraints at ve , we get A0 (e00 ) = A0 (e0 ). Since A(e) = A0 (e00 ), by construction, we conclude A(e) = A0 (e0 ). The equalities A(f ) = A0 (f 0 ) and A(h) = A0 (h0 ) follow symmetrically. Hence, the orthogonality constraints at v are satisfied by A since they are satisfied at v 0 by A0 . This concludes the proof. t u Theorem 5. OrthoSEFE-2 can be solved in polynomial time for instances whose shared graph is a cycle and whose union graph has maximum degree 5. Proof. First apply Lemma 3 to obtain an equivalent instance hG01 , G02 i such that G0∩ is a cycle and graph G01 has maximum degree 3. Then, apply Lemma 2 to obtain an equivalent instance hG001 , G002 i such that G00∩ is a cycle and G001 is an outerplanar graph with maximum degree 3. Finally, apply Lemma 1 to test in polynomial time whether hG001 , G002 i, and hence hG1 , G2 i, is a positive instance. t u E Omitted or Sketched Proofs from Section 5 Lemma 4. Let hG1 , G2 i be an instance of OrthoSEFE-2 whose shared graph is biconnected. It is possible to construct in polynomial time an equivalent instance hG∗1 , G∗2 i whose shared graph is biconnected and such that each endpoint of an exclusive edge has degree 2 in the shared graph. Proof. We start with a simplification step that removes certain edges. An exclusive edge e = uv of G1 or of G2 is an intra-pole edge if its endpoints are adjacent in some skeleton of the SPQR-tree of the shared graph G∩ . If neither u nor v is incident to other exclusive edges, then uv is isolated. Let E10 and E20 be the isolated intra-pole edges of G1 and G2 , respectively. We claim that the instance hG1 − E10 , G2 − E20 i admits an OrthoSEFE if and only if hG1 , G2 i does. The if part is clear since we can simply remove the isolated intra-pole edges from an OrthoSEFE of hG1 , G2 i to obtain an OrthoSEFE. Conversely, Angelini et al. [1] show that the intra-pole edges can be reinserted into any SEFE, and thus also into an OrthoSEFE of hG1 − E10 , G2 − E20 i without crossings, i.e., the planarity constraints are satisfied for hG1 , G2 i. Since the edges in E10 ∪ E20 are isolated also the orthogonal constraints are trivially satisfied, and we obtain an OrthoSEFE of hG1 , G2 i. This finishes the proof of the claim. 25 v0 v v0 w3 x x u v w2 w1 u Fig. 10. Moving exclusive edges from a vertex with degree 3 in the shared graph to a new vertex with degree 2 in the shared graph. In the following, we assume that hG1 , G2 i has been preprocessed in this way, and it hence does not contain isolated intra-pole edges. Consider an exclusive edge e = uv in G1 or G2 , say in G1 , such that u has degree 3 in the shared graph. Assume that ux is an edge of G incident to u such that, in every OrthoSEFE of hG1 , G2 i the edge uv is embedded in a face of G incident to ux (we describe how to determine such an edge later). We perform the following transformation. We subdivide ux by three vertices w1 , w2 , w3 and add the edge w1 w3 . We further replace uv by w2 v and also, if it exists, the (unique) exclusive edge e0 = uv 0 (from G2 ) by w2 v 0 ; see Fig. 10. Call the resulting in0 0 stance hG1 , G2 i. It is not difficult to see that hG1 , G2 i admits an OrthoSEFE if 0 0 0 0 and only if hG1 , G2 i does. If hG1 , G2 i admits an OrthoSEFE, we can contract the vertices w1 , w2 , w3 onto u to obtain an OrthoSEFE of hG1 , G2 i. Note that the orthogonal constraint at u is satisfied since the triangle w1 , w2 , w3 ensures that the exclusive edges incident to u are embedded in the same face of G0∩ and hence in G∩ . Conversely, given an OrthoSEFE of hG1 , G2 i, due to the orthogonality constraints at u all exclusive edges incident to u are embedded in the same face of G∩ , and hence the replacement can be carried out locally without creating crossings. Note that, after the transformation, there are fewer endpoints of exclusive edges that have degree 3 in the shared graph. We iteratively apply this transformation to obtain the instance hG∗1 , G∗2 i. It remains to show that there always exists a suitable edge ux. Let T denote the SPQR-tree of the shared graph G∩ . Since u has degree 3, there is exactly one node µ of T whose skeleton contains u and where the degree of u in skel(µ) is 3. Note that µ is either a P-node or an R-node. First assume µ is an R-node and consider the position of v inside skel(µ), where it is either a vertex of skel(µ) or it is contained in a virtual edge εv of skel(µ). Since µ is an R-node, u and v (u and εv ) share at most two faces, both of which are incident to a virtual edge ε incident to u. We choose ux as the (unique) edge incident to u that is contained in the subgraph represented by ε. Second, assume µ is a P-node. If the other endpoint v of e is not a pole of µ, then v is contained in a virtual edge εv of skel(µ), and we can proceed as in the previous case; see Fig. 10. Now assume that v is the other vertex of skel(µ), i.e., e is an intra-pole edge. Since e cannot be isolated (due to the simplification step at the beginning), there exists an exclusive edge e0 in G2 incident to u or v. Since e 6= e0 , the edge e0 has an endpoint v 0 that is contained in a subgraph represented by a virtual edge ε of skel(µ). It follows that in every planar embedding of G2 , the edge e0 is embedded in a face incident to ε. By the orthogonality constraints 26 at the vertex shared by e and e0 , e also has to be embedded in a face incident to ε in any OrthoSEFE. We thus choose ux as the (unique) edge incident to u contained in the subgraph represented by ε. t u Lemma 5. Let hG1 , G2 i be an instance of OrthoSEFE-2 such that the shared graph G∩ is biconnected. Then hG1 , G2 i admits an OrthoSEFE if and only if all instances hGµ1 , Gµ2 i admit an OrthoSEFE. Proof. It is not hard to see that each hGµ1 , Gµ2 i can be obtained from hG1 , G2 i by removing some vertices and edges and suppressing subdivision vertices. Thus, if hG1 , G2 i admits an OrthoSEFE, so does each hGµ1 , Gµ2 i. Conversely, assume that each hGµ1 , Gµ2 i admits an OrthoSEFE. Recall that we have fixed a reference embedding for each skeleton of the SPQR-tree of the shared graph G∩ up to a flip. We fix the flips of all reference embeddings as follows. For each S-node µ and each neighbor ν, represented by a virtual edge ε in skel(µ), we consider the flips of the cycle Cε in the OrthoSEFE of hGµ1 , Gµ2 i with respect to the ordering Oε of the attachments of the subgraph represented by ε. If the reference embedding is used, we label the edge µν with label 1, otherwise we label it −1. Finally, we choose an arbitrary root µ0 of the augmented SPQR-tree for which we fix the reference embedding. For each skeleton skel(µ), µ 6= µ0 , we choose the reference embedding if and only if the product of the labels on the (unique) path from µ0 to µ is 1, and its flip otherwise. We denote the planar embedding of G∩ obtained in this way by E. It remains to determine the embeddings of G1 and G2 . After suitably flipping the given OrthoSEFEs, we can assume that their embeddings can be obtained from E by removing vertices and edges, and by contracting edges. We now determine the embeddings of G1 and G2 as follows. Recall that every vertex that is incident to exclusive edges has degree 2 in the shared graph. For each vertex v that is incident to exclusive edges of G1 (of G2 ), we consider the unique S-node µ whose skeleton contains v, and we choose the edge ordering as in the given OrthoSEFE of hGµ1 , Gµ2 i. We claim that this results in an OrthoSEFE hE1 , E2 i of hG1 , G2 i. Refer to Figs. 6(a) and 6(b). First observe that the orthogonality constraints are satisfied, since the edge ordering of each vertex is chosen according to one of the given OrthoSEFEs. It remains to show that the embeddings also satisfy the planarity constraints. Due to the construction of the embeddings, all the exclusive edges are embedded in faces of E; otherwise we would observe crossings in the skeletons of the (augmented) SPQR-tree. Consider two exclusive edges uv and u0 v 0 from the same graph that cross. Since uv and u0 v 0 cross, there exists a node µ of the (augmented) SPQR-tree such that for each of the two edges the endpoints are in different parts of skel(µ). If µ is a P-node or an R-node and all four parts containing these endpoints are distinct, then the parts containing the endpoints of these edges alternate around a face of skel(µ). This contradicts the planarity of the corresponding input graph G1 or G2 . Thus, in this case at least two attachments are contained in the same virtual edge ε of skel(µ). Let ν be the S-node of the augmented SPQR-tree corresponding to ε. Clearly, in skel(ν), the endpoints 27 of each of the two edges are distinct parts of skel(ν). It follows that the endpoints of the two edges alternate around the two faces of hGν1 , Gν2 i corresponding to the two faces of skel(µ). By construction of hE1 , E2 i this contradicts the assumption that the given drawing of hGν1 , Gν2 i is an OrthoSEFE. t u Lemma 6. hGµ1 , Gµ2 i admits an OrthoSEFE if and only if hGµ1 , Gµ2 i does. Proof. We simply show that, in terms of embeddings, the path Pε replacing Cε behaves the same as Cε . First, observe that the edge (a2 , x3 ) ensures that all exclusive edges of G1 incident to the clockwise uv-path of Ci are embedded on the same side of the path Pε . Similarly, (x2 , b1 ) ensures that all exclusive edges of G1 incident to the counterclockwise uv-path of Cε are embedded on the same side of the path Pε . Moreover, since the endpoints of the edges (a2 , x3 ) and (x2 , b1 ) alternate along Pε , they are embedded on different sides of Pε . Thus, the exclusive edges of G1 incident to the clockwise and counterclockwise uvpath of Cε cannot be embedded on the same side of Pε . Similarly, the exclusive edges (a1 , x4 ) and (x2 , b2 ) ensure that the exclusive edges of G2 incident to the clockwise uv-path are all on one side of Pε and the exclusive edges of G2 incident to the counterclockwise uv-path are on the other side of Pε . Finally, due to the alternation with (a2 , x3 ), the edges (x2 , x4 ) and (x2 , b1 ) must be embedded on the same side of Pε . By the orthogonality constraint at x2 , the edge (x2 , b2 ) must be also embedded on the same side as (x2 , x4 ). Thus, (a2 , x3 ) and (a1 , x3 ) are on the same side of Pε and likewise for (x2 , b1 ) and (x2 , b2 ). This ensures that the exclusive edges of G1 and G2 incident to the clockwise uv-path of Cε are embedded on the same side of Cε and likewise for those incident to the counterclockwise uv-path. t u Theorem 7. Let hG1 , . . . , Gk i be a positive instance of OrthoSEFE-k whose shared graph is biconnected. Then, there exists an OrthoSEFE hΓ1 , Γ2 , . . . , Γk i of hG1 , . . . , Gk i in which every edge has at most three bends. Proof. We assume that a cyclic order of the edges of the union graph around each vertex is given such that (a) it induces a planar embedding on each Gi , i = 1, . . . , k, and (b) we can assign the incident edges around a vertex to at most four ports such that at most one edge of each Gi is assigned to the same port. We adopt the method of Biedl and Kant [9]. First, we compute in linear time [15] an s-t-ordering on the shared graph, i.e., we label the vertices v1 , . . . , vn such that {v1 , vn } is an edge of the shared graph and, for each i = 2, . . . , n − 1, there are j < i < k such that {vj , vi } and {vi , vk } are edges of the shared graph. We choose the face to the left of (v1 , vn ) as the outer face of the union graph. We now draw the union graph by adding the vertices in the order in which they appear in the s-t-ordering while respecting the given order of the edges around each vertex. The edges will bend at most on y-coordinates near their incident vertices and are drawn vertically otherwise. We draw the edges around v1 as indicated in Fig. 11(a) where some of the incident edges might actually indicate several exclusive edges – at most one from each graph. 28 to vn or vi (a) around v1 vn vi v1 to v1 (b) around v2 , . . . , vn−1 (c) around vn Fig. 11. Constructing a drawing with at most three bends per edge For i = 2, . . . , n − 1, an edge may only leave vi to the bottom if it is incident to a neighbor with a lower index. Again, some of the ports might host several exclusive edges, even one to a vertex with a lower index and one to a vertex with a higher index. Special cases occur when the ordering around vi is such that four exclusive edges of two distinct graphs must be assigned to two consecutive ports. In particular, an edge leaving vi to a vertex with a lower index might bend twice around vi (see, e.g., the two small circles in Fig. 11(b)). Finally, the edges around vn are placed such that the edge {v1 , vn } enters it from the left. Thus, there are exactly three bends on {v1 , vn }; see Fig. 11(c). For any other edge, there is at most one bend around the endvertex with lower index and at most two bends around the endvertex with higher index. t u F Omitted or Sketched Proofs from Section 6 Lemma 7. Let hG1 , G2 i be an instance of OrthoSEFE-2 whose shared graph G∩ is a cycle. It is possible to construct in polynomial time an equivalent instance hG∗1 , G∗2 i of OrthoSEFE-2 such that (i) the shared graph G∗∩ is a cycle, (ii) graph G∗1 is outerplanar, and (iii) no two degree-4 vertices in G∗1 are adjacent to each other. Proof. The reduction works in two steps. In the first step, we construct an in+ stance hG+ 1 , G2 i satisfying properties (i) and (iii) that is equivalent to hG1 , G2 i; then, in the second step we construct the final instance hG∗1 , G∗2 i equivalent to + hG+ 1 , G2 i, which also satisfies property (ii). For the first step, we show how to construct an instance hG01 , G02 i of OrthoSEFE2 equivalent to hG1 , G2 i such that G0∩ is a cycle and the number of vertices with degree 4 in G01 not satisfying the condition of property (iii) is smaller than the number of vertices with degree 4 in G1 not satisfying this condition. Repeatedly + performing this transformation eventually yields the required instance hG+ 1 , G2 i. Consider a vertex v with degree 4 in G1 not satisfying the condition of property (iii). Let e = (u, v) and f = (v, w) be the two exclusive edges of G1 incident to v. Assume that u, v, and w appear in this order along G∩ , the other cases being analogous. Initialize G0∩ = G∩ ; refer to Fig. 12. Replace v in G0∩ by a path Pv composed of dummy vertices x1 , x2 , va , x3 , . . . , x8 , u0 , x9 , x10 , of vertex v, and of dummy vertices y1 , y2 , w0 , y3 , . . . , y8 , vb , y9 , y10 . Note that G0∩ contains all the vertices of 29 e0 e f u H1 v H2 w H1 u (a) e00 va f 00 v u0 w0 vb f0 H2 w (b) Fig. 12. Illustrations for the proof of Lemma 7 G∩ , plus a set of dummy vertices. We now describe the exclusive edges in E10 and E20 . Set Ei0 , with i = 1, 2, contains all the exclusive edges in Ei , except for e and f . Also, E10 contains edges e0 = (u, va ), e00 = (u0 , v), f 00 = (v, w0 ), and f 0 = (vb , w). Finally, E20 contains edges (x1 , va ), (x2 , x3 ), (va , x5 ), (x4 , x7 ), (x6 , u0 ), (x8 , x9 ), (u0 , x10 ), (y1 , w0 ), (y2 , y3 ), (w0 , y5 ), (y4 , y7 ), (y6 , vb ), (y8 , y9 ), and (vb , y10 ). We prove that hG01 , G02 i satisfies the required properties. First, graph G0∩ is a cycle by construction. Second, the number of vertices of degree 4 in G01 not satisfying the condition of property (iii) is smaller than the number of such vertices in G1 . In fact, any vertex x 6= v with degree 4 in G01 satisfies the required condition if and only if it satisfies the same condition in G1 . On the other hand, vertex v does not satisfy the condition in G1 , by hypothesis, but it satisfies the condition in G01 , since u0 and w0 have degree 3 in G01 and the path between them along G0∩ containing v only contains dummy vertices x9 , x10 , y1 , and y2 , which are not incident to any exclusive edge of G01 , by construction. We now prove that hG01 , G02 i is equivalent to hG1 , G2 i. Suppose that hG1 , G2 i admits an OrthoSEFE hΓ1 , Γ2 i, and let A be the corresponding assignment of the exclusive edges of E1 and of E2 to the two sides of G∩ , which exists by Theorem 2. We show how to construct an assignment A0 of the exclusive edges of E10 and of E20 to the two sides of G0∩ satisfying all the constraints. For each exclusive edge g ∈ E1 ∩ E10 , set A0 (g) = A(g). Also, set A0 (e0 ) = 0 00 A (e ) = A(e) and A0 (f 0 ) = A0 (f 00 ) = A(f ). For each exclusive edge g ∈ E2 ∩E20 , set A0 (g) = A(g). Also, set A0 (x1 , va ) = A0 (va , x5 ) = A0 (x6 , u0 ) = A0 (u0 , x10 ) = A0 (y1 , w0 ) = A0 (w0 , y5 ) = A0 (y6 , vb ) = A0 (vb , y10 ) = A(e) and set A0 (x2 , x3 ) = A0 (x4 , x7 ) = A0 (x8 , x9 ) = A0 (y2 , y3 ) = A0 (y4 , y7 ) = A0 (y8 , y9 ) = A(f ). We prove that the planarity constraints for the edges of G01 are satisfied by 0 A . Note that, by construction, edges e00 and f 00 do not alternate with any edge of G01 along G0∩ . Also, edges e0 and f 0 do not alternate with each other. Further, if edge e0 (edge f 0 ) alternates with an edge g ∈ G01 along G0∩ , then edge e (edge f ) alternates with g along G∩ . Finally, any two edges different from e0 , e00 , f 0 , f 00 that alternate along G0∩ also alternate along G∩ . In all the described cases, the planarity constraints are satisfied by A0 since they are satisfied by A. We prove that the planarity constraints for the edges of G02 are satisfied by 0 A . Note that, by construction, edges in E20 ∩ E2 do not alternate with any edge 30 incident to a dummy vertex along G0∩ , and alternate with each other along G0∩ if only if they alternate with each other along G∩ . Hence, the planarity constraints for these edges are satisfied by A0 since they are satisfied by A. On the other hand, it is easily verified that the planarity constraints are satisfied by A0 also for the edges incident to dummy vertices. We now prove that the orthogonality constraints are satisfied by A0 at every vertex in V 0 . For the non-dummy vertices in V 0 \ {u, v, w}, this is true since they are satisfied by A and since the edges incident to these vertices have the same assignment in A as in A0 . For vertex u, this is true since they are satisfied by A, since A0 (e0 ) = A(e), and since the other edges incident to u have the same assignment in A and in A0 . Analogously, for w this is true since they are satisfied by A, since A0 (f 0 ) = A(f ), and since the other edges have the same assignment in A and in A0 . For v, this is true since they are satisfied by A, since A0 (e00 ) = A(e), since A0 (f 00 ) = A(f ), and since the other edges have the same assignment in A and in A0 . For va , this is true since A0 (x1 , va ) = A0 (va , x5 ) = A0 (e0 ) = A(e). For u0 , this is true since A0 (x6 , u0 ) = A0 (u0 , x10 ) = A0 (e00 ) = A(e). For w0 , this is true since A0 (y1 , w0 ) = A0 (w0 , y5 ) = A0 (f 00 ) = A(f ). For vb , this is true since A0 (y6 , vb ) = A0 (vb , y10 ) = A0 (f 0 ) = A(f ). Since all the other dummy vertices have degree 2 in G01 , this concludes the proof that A0 satisfies the orthogonality constraints. Suppose that hG01 , G02 i admits OrthoSEFE hΓ10 , Γ20 i, and let A0 be the corresponding assignment of the exclusive edges of E10 and of E20 to the two sides of G0∩ . We show how to construct an assignment A of the exclusive edges of E1 and of E2 to the two sides of G∩ satisfying all the planarity and the orthogonality constraints. For each exclusive edge g ∈ E1 , set A(g) = A0 (g). Also, set A(e) = A0 (e0 ) and A(f ) = A0 (f 0 ). Finally, for each exclusive edge g ∈ E2 ∩ E20 , set A(g) = A0 (g). We prove that the planarity constraints for the edges of G1 are satisfied by A. Note that e and f do not alternate with each other since they are incident to the same vertex v. Also, if edge e (edge f ) alternates with an edge g ∈ G1 along G∩ , then edge e0 (edge f 0 ) alternates with g along G0∩ . Finally, any two edges different from e and f that alternate along G∩ also alternate along G0∩ . In all these cases, the planarity constraints are satisfied by A since they are satisfied by A0 . The planarity constraints for the edges of G2 are satisfied by A since any two of these edges alternate along G∩ if and only if they alternate along G0∩ , and since the planarity constraints are satisfied by A0 . We finally prove that the orthogonality constraints are satisfied by A at every vertex in V . For the vertices in V \ {v}, this is true since they are satisfied by A0 and since for every exclusive edge g ∈ E1 ∪ E2 incident to these vertices, we have g ∈ E10 ∪ E20 , by construction, and A(g) = A0 (g). To prove that this is true also for v, we first argue that A(e) = A0 (e00 ) and A(f ) = A0 (f 00 ): By planarity constraints, we get A0 (x1 , va ) = A0 (va , x5 ) = A0 (x6 , u0 ) = A0 (u0 , x10 ), since they belong to a sequence of alternating edges; hence, by orthogonality constraints at va and u0 , we get A0 (e0 ) = A0 (x1 , va ) = A0 (va , x5 ) = A0 (x6 , u0 ) = A0 (u0 , x10 ) = 31 A0 (e00 ); since A(e) = A0 (e0 ), by construction, we conclude A(e) = A0 (e00 ). The equality A(f ) = A0 (f 00 ) follows symmetrically. Hence, the orthogonality constraints at v are satisfied by A since they are satisfied at v by A0 . This concludes + the proof that hG+ 1 , G2 i, which satisfies properties (i) and (iii), is equivalent to hG1 , G2 i. + In order to construct an instance hG∗1 , G∗2 i equivalent to hG+ 1 , G2 i that also satisfies property (ii), we observe that the proof of Lemma 2 can be easily ex+ tended so that it can be applied to hG+ 1 , G2 i. This lemma, in fact, holds for instances hG1 , G2 i satisfying property (i) and a property that is stronger than (iii), namely that G1 has degree at most 3. This stronger condition, however, is only used to ensure that there exists no exclusive edge in E1 with an endpoint in H2 and the other one not in H2 ; refer to Fig. 8. In particular, it is used to ensure that there exists no edge connecting a vertex of H2 to any of u, v, w, z. However, it is possible to prove that property (iii) is already sufficient to ensure the absence of these edges. Namely, suppose that there exists an edge in E1 connecting a vertex x of H2 to vertex v, the other cases being analogous. This implies that v has degree 4 in G1 , since it is also adjacent to u. However, any path in cycle G∩ containing u, x, and v also contains either w or z, since e and f alternate along G∩ ; this is a contradiction to property (iii), since each of w and z is incident to an exclusive edge of G1 , namely f . This concludes the proof of the lemma. t u 32
8cs.DS
1 Throughput Maximization for UAV-Enabled Wireless Powered Communication Networks arXiv:1801.04545v3 [cs.IT] 22 Mar 2018 Lifeng Xie, Jie Xu, and Rui Zhang Abstract—This paper studies an unmanned aerial vehicle (UAV)-enabled wireless powered communication network (WPCN), in which a UAV is dispatched as a mobile access point (AP) to serve a set of ground users periodically. The UAV employs the radio frequency (RF) wireless power transfer (WPT) to charge the users in the downlink, and the users use the harvested RF energy to send independent information to the UAV in the uplink. Unlike the conventional WPCN with fixed APs, the UAV-enabled WPCN can exploit the mobility of the UAV via trajectory design, jointly with the wireless resource allocation optimization, to maximize the system throughput. In particular, we aim to maximize the uplink common (minimum) throughput among all ground users over a finite UAV’s flight period, subject to its maximum speed constraint and the users’ energy neutrality constraints. The resulted problem is non-convex and thus difficult to be solved optimally. To tackle this challenge, we first consider an ideal case without the UAV’s maximum speed constraint, and obtain the optimal solution to the relaxed problem. The optimal solution shows that the UAV should successively hover above a finite number of ground locations for downlink WPT, as well as above each of the ground users for uplink communication. Next, we consider the general problem with the UAV’s maximum speed constraint. Based on the above multi-location-hovering solution, we first propose an efficient successive hover-and-fly trajectory design, jointly with the downlink and uplink wireless resource allocation, and then propose a locally optimal solution by applying the techniques of alternating optimization and successive convex programming (SCP). Numerical results show that the proposed UAV-enabled WPCN achieves significant throughput gains over the conventional WPCN with fixed-location AP. Index Terms—Unmanned aerial vehicle (UAV), wireless powered communication network (WPCN), wireless power transfer (WPT), trajectory optimization, resource allocation. I. I NTRODUCTION R ADIO frequency (RF) wireless power transfer (WPT) has emerged as a promising solution to provide convenient and reliable energy supply to low-power Internet-of-things (IoT) devices such as sensors and RF identification (RFID) tags [2], [3]. Compared to the near-field WPT based on inductive coupling or magnetic resonant coupling, the far-field WPT via RF radiation is able to operate over a much longer range and charge multiple wireless devices (WDs) simultaneously even when they are moving and densely deployed, and yet with transceivers of significantly reduced form factor. Part of this paper will be presented at the IEEE Vehicular Technology Conference (VTC2018-spring), Porto, Portugal, June 3-6, 2018 [1]. L. Xie and J. Xu are with the School of Information Engineering, Guangdong University of Technology, Guangzhou, China (e-mail: [email protected], [email protected]). J. Xu is the corresponding author. R. Zhang is with the Department of Electrical and Computer Engineering, National University of Singapore, Singapore (e-mail: [email protected]). Fig. 1. Illustration of a UAV-enabled WPCN. There are in general two major applications of RF WPT in wireless communications, namely simultaneous wireless information and power transfer (SWIPT) and wireless powered communication network (WPCN), which unify the WPT and wireless information transfer (WIT) in a joint design framework over the same (downlink for both WPT and WIT) and opposite (downlink for WPT and uplink for WIT) transmission directions, respectively [4], [5]. In particular, WPCN enables dedicated wireless charging and information collection for massive IoT devices, thus significantly enhances the operation range and throughput of traditional backscattering-based wireless communications. However, in conventional WPCNs, access points (APs) are usually deployed at fixed locations, which cannot be changed once deployed [6], [7]. The conventional WPCN with fixed APs faces several challenges. First, due to the severe propagation loss of RF signals over distance, the end-to-end WPT efficiency is generally low when the distance from the AP to a WD becomes large. Next, the conventional WPCN suffers from the so-called “doubly near-far” problem [6], i.e., far-apart WDs from the AP receive lower RF energy in the downlink WPT, but they need to use higher transmit power in the uplink WIT to achieve the same rate as nearby WDs. The doubly near-far problem result in a severe user fairness issue among WDs when they are geographically distributed over a large area. To overcome the above issues, various approaches have been proposed in the literature, such as adaptive time and power allocation [6], multi-antenna beamforming [8]–[12], and user cooperation [13], [14]. However, all these prior works focused on the wireless resource allocation designs to enhance the WPT/WIT performances of WPCNs with fixed APs. By contrast, in this paper we propose an alternative solution based on a new unmanned aerial vehicle (UAV)-enabled WPCN architecture with UAVs employed as mobile APs. UAVs have found abundant applications such as for cargo 2 delivery, aerial surveillance, filming, and industrial IoT. Recently, UAVs-enabled/aided wireless communications have attracted substantial research interests, due to their advantages in flexible deployment, strong line-of-sight (LoS) channels with ground users, and controllable mobility [15]. For example, UAVs can be utilized as mobile relays to help information exchange between far-apart ground users [16], or as mobile base stations (BSs) to help enhance the wireless coverage and/or the network capacity for ground mobile users [17]–[25]. Furthermore, UAV-enabled WPT has been proposed in [26], [27], in which UAVs are used as mobile energy transmitters to charge low-power WDs on the ground. By exploiting its fully controllable mobility, the UAV can properly adjust its location over time (a.k.a. trajectory) to reduce the distances with target ground users, thus improving the efficiency for both WIT and WPT. Motivated by the UAV-enabled wireless communications as well as WPT, this paper pursues a unified study on both of them in a UAV-enabled WPCN as shown in Fig. 1. Specifically, a UAV following a pre-designed periodic trajectory is dispatched as a mobile AP to charge a set of ground users in the downlink via WPT, and the users use the harvested RF energy to send independent information to the UAV in the uplink. We investigate how to optimally exploit the UAV mobility via trajectory design, jointly with the wireless resource allocation, to maximize the uplink data throughput of the multiuser WPCN in a fair manner. To this end, we maximize the uplink common (minimum) throughput among all ground users over a given UAV’s flight period, by optimizing its trajectory, jointly with the downlink and uplink transmission resource allocations for WPT and WIT, respectively, subject to the UAV’s maximum speed and the users’ energy neutrality constraints. However, due to the complex data throughput and harvested energy functions in terms of coupled UAV trajectory and resource allocation design variables, our formulated problem is nonconvex and thus difficult to be solved optimally. To tackle this difficulty, we first consider an ideal case without considering the UAV’s maximum speed constraint. We show that the strong duality holds between this problem and its Lagrange dual problem, and thus it can be solved optimally via the Lagrange dual method. The optimal solution shows that the UAV should successively hover above a finite number of ground locations for downlink WPT, as well as above each of the ground users for uplink WIT, with the optimal hovering duration and wireless resource allocation for each location. Next, we address the general problem with the UAV’s maximum speed constraint considered. Based on the multi-location-hovering solution to the above relaxed problem, we first propose a heuristic successive hover-andfly trajectory design, jointly with the downlink and uplink resource allocations, to find an efficient suboptimal solution. The proposed solution is also shown to be asymptotically optimal as the UAV’s flight period becomes infinitely large. In addition, we further propose an alternating optimization based algorithm to obtain a locally optimal solution, which optimizes the wireless resource allocations and the UAV trajectory in an alternating manner, via convex optimization and successive convex programming (SCP) techniques, respectively. By employing the successive hover-and-fly trajectory as the UAV initial trajectory, the alternating-optimization-based algorithm iteratively refines the wireless resource allocations and the UAV trajectory to improve the uplink common throughput of all ground users until convergence. Finally, we present numerical results to validate the performance of our proposed UAV-enabled WPCN. It is shown that the joint trajectory and wireless resource allocation design significantly improves the uplink common throughput, as compared to the conventional WPCN with the AP at a fixed location. It is worth noting that there is another line of related research that employs ground moving vehicles to wirelessly charge and/or collect information from ground sensors, e.g., [29]–[31]. However, different from the UAV that can freely fly in the three-dimensional (3D) airspace, the ground vehicle can only move following a constrained path in the twodimensional (2D) plane. Furthermore, unlike UAVs that have strong LoS links with ground users, the wireless channels from ground vehicles to users usually suffer from severe fading, thus limiting the performance for both WIT and WPT. As such, the joint trajectory design and wireless resource allocation in the UAV-enabled WPCN is a new study different from that with ground moving vehicles, which has not been investigated in the literature to our best knowledge. The remainder of this paper is organized as follows. Section II presents the system model of the UAV-enabled WPCN, and formulates the uplink common throughput maximization problem of our interest. Section III considers an ideal case without the UAV’s maximum speed constraint and presents the optimal solution to the relaxed problem. Section IV and Section V present two efficient solutions to the general problem with the UAV’s maximum speed constraint considered. Section VI provides numerical results to validate the effectiveness of our proposed designs. Finally, Section VII concludes the paper. II. S YSTEM M ODEL AND P ROBLEM F ORMULATION As shown in Fig. 1, we consider a UAV-enabled WPCN, in which a UAV is dispatched to periodically charge a set K , {1, . . . , K} of ground users via WPT in the downlink, and each user k ∈ K uses its harvested energy to send independent information to the UAV in the uplink. Suppose that each user k ∈ K is at a fixed location (xk , yk , 0) on the ground in a 3D Cartesian coordinate system, where wk = (xk , yk ) is defined as the horizontal coordinate of user k. The users’ locations are assumed to be a-priori known by the UAV for its trajectory design and transmission resource allocation. We focus on one particular flight period of the UAV, denoted by T , (0, T ] with finite duration T in second (s), in which the UAV flies horizontally at a fixed altitude H > 0 in meter (m). At any given time instant t ∈ T , let q(t) = (x(t), y(t)) denote the location of the UAV projected on the horizontal plane. Accordingly, the distance between the UAV and each user k ∈ K is given by p (1) dk (q(t)) = kq(t) − wk k2 + H 2 , where k · k denotes the Euclidean norm of a vector. By denoting p the UAV’s maximum speed as Vmax in m/s, we have ẋ2 (t) + ẏ 2 (t) ≤ Vmax , ∀t ∈ T , where ẋ(t) and ẏ(t) 3 denote the first-derivatives of x(t) and y(t) with respect to t, respectively. Note that we assume the UAV can freely choose its initial location q(0) and final location q(T ) for performance optimization. We consider that the wireless channels between the UAV and the K ground users are dominated by LoS links. In this case, the free-space path loss model can be practically assumed, similarly as in [16], [26]. Accordingly, the channel power gain between the UAV and user k ∈ K at time instant t ∈ T is given by β0 hk (q(t)) = β0 d−2 , (2) k (q(t)) = kq(t) − wk k2 + H 2 where β0 denotes the channel power gain at a reference distance of d0 = 1 m. We consider a time-division multiple access (TDMA) transmission protocol, in which the downlink WPT for all users and the uplink WIT of different users are implemented in the same frequency band but over orthogonal time instants. At any time instant t ∈ T , we use the indicators ρk (t) ∈ {0, 1}, ∀k ∈ {0} ∪ K, to denote the transmission mode. We use ρ0 (t) = 1 and ρk (t) = 0, ∀k ∈ K, to indicate the downlink WPT mode, in which the UAV transmits RF energy to charge the K users simultaneously; while we use ρk (t) = 1, k ∈ K, and ρj (t) = 0, ∀j ∈ {0} ∪ K, j 6= k, to represent the uplink WIT mode for user k, in which user k sends its information to the UAV by using its harvested energy. As the TDMA protocol PK is employed, it follows that k=0 ρk (t) = 1, ∀t ∈ T . First, consider the downlink WPT mode at time instant t ∈ T , in which ρ0 (t) = 1, and ρk (t) = 0, ∀k ∈ K. Suppose that the UAV adopts a constant transmit power P in the downlink WPT mode. Accordingly, the harvested power at each user k ∈ K is given by Ek (ρ0 (t), q(t)) = ηP ρ0 (t)hk (q(t)) ηP β0 ρ0 (t) , (3) = kq(t) − wk k2 + H 2 where 0 < η ≤ 1 denotes the RF-to-direct current (DC) energy conversion efficiency at the energy harvester of each user.1 Therefore, the total harvested energy at user k over each period of duration T is given by Z T Êk ({ρ0 (t), q(t)}) = Ek (ρ0 (t), q(t))dt. (4) 0 Next, consider the WIT mode for user k ∈ K at time instant t ∈ T with ρk (t) = 1, and ρj (t) = 0, ∀j ∈ {0} ∪ K, j 6= k. Let Qk (t) denote the transmit power of user k for the uplink WIT to the UAV. Accordingly, the achievable data rate from user k to the UAV in bits/second/Hertz (bps/Hz) at time instant t ∈ T is given by   Qk (t)hk (q(t)) rk (ρk (t),q(t), Qk (t)) = ρk (t)log2 1 + σ2   Qk (t)γ , (5) = ρk (t)log2 1 + kq(t) − wk k2 + H 2 1 Note that in practice, the RF-to-DC energy conversion efficiency is generally non-linear and depends on the received RF power level [32] and signal waveform [33]. For the purpose of exposition, in (3) we consider a simplified constant RF-to-DC energy conversion efficiency by assuming that each receiver operates at the linear regime for RF-to-DC conversion. Nevertheless, the design principles in this paper are also extendable to the scenario with non-linear RF-to-DC energy conversion efficiency, which, however, is left for future work. where σ 2 denotes the noise power at the information receiver of the UAV, and γ , β0 /σ 2 is the reference signal-to-noise ratio (SNR). Therefore, the average achievable rate or throughput of user k over each period in bps/Hz is given by Z 1 T Rk ({ρk (t), q(t), Qk (t)}) = rk (ρk (t), q(t), Qk (t))dt. T 0 (6) Note that for the purpose of exposition, we consider that the energy consumption of each ground user is mainly due to the transmit power for its uplink WIT. In this case, the total energy consumption at user k ∈ K is Z T Q̂k ({ρk (t), Qk (t)}) = ρk (t)Qk (t)dt. (7) 0 In order to achieve the self-sustainable operation for the WPCN, we consider the energy neutrality constraint at each user k, such that the user’s energy consumption for uplink WIT (i.e., Q̂k ({ρk (t), Qk (t)}) in (7)) cannot exceed the energy harvested from the downlink WPT (i.e., Êk ({ρ0 (t), q(t)}) in (4)) in each period.2 As a result, we have the following energy neutrality constraints for the K users, Z T Z T ρk (t)Qk (t)dt ≤ Ek (ρ0 (t), q(t))dt, ∀k ∈ K. (8) 0 0 In this work, our objective is to maximize the uplink common throughput among all users (i.e., mink∈K Rk ({ρk (t), q(t), Qk (t)})) subject to the UAV’s maximum speed constraint and the K users’ energy neutrality constraints. The decision variables include the UAV trajectory {q(t)}, the transmission mode {ρk (t)}, and the transmit power {Qk (t)} for uplink WIT. As a result, the problem is formulated as (P1) : max min Rk ({ρk (t), q(t), Qk (t)}) {ρk (t),Qk (t),q(t)} k∈K Z T Z T s.t. ρk (t)Qk (t)dt ≤ Ek (ρ0 (t), q(t))dt, ∀k ∈ K (9) 0 0 Qk (t) ≥ 0, ∀k ∈ K, t ∈ T ρk (t) ∈ {0, 1}, ∀k ∈ {0} ∪ K, t ∈ T K X k=0 ρk (t) = 1, ∀t ∈ T (10) (11) (12) p ẋ2 (t) + ẏ 2 (t) ≤ Vmax , ∀t ∈ T , (13) where (13) denotes the UAV’s maximum speed constraint. It is observed that for problem (P1), the objective function is non-concave and constraints (9) and (11) are non-convex, due to the complicated rate and energy functions with respect to coupled variables ρk (t), q(t), and Qk (t), as well as the binary constraints on ρk (t)’s. Therefore, (P1) is a non-convex optimization problem. Furthermore, (P1) contains an infinite number of optimization variables over continuous time. For these reasons, problem (P1) is difficult to be solved optimally. To tackle this problem, in Section III we first consider an ideal 2 Note that we assume that at the beginning of each period, each user has sufficient energy in its storage and the storage has sufficiently large capacity. In this case, as long as the energy neutrality constraints are satisfied over each period, the more stringent energy causality constraints in energy harvesting wireless communications [34] will be automatically ensured, and thus each of the K users can sustain its operation without energy outage. 4 case by ignoring the UAV’s maximum speed constraint in (13), and solve the relaxed problem of (P1) as follows: (P2) : max min Rk ({ρk (t), q(t), Qk (t)}) {ρk (t),Qk (t),q(t)} k∈K s.t. (9), (10), (11) and (12). Note that problem (P2) also corresponds to the practical scenario when the UAV’s flight duration T is sufficiently large for any given finite Vmax such that the flying time of the UAV becomes negligible as compared to its hovering time (see Section III for details). In Section IV and Section V, we propose efficient algorithms to solve the general problem (P1) with the UAV’s maximum speed constraint based on the optimal solution obtained for the relaxed problem (P2). III. O PTIMAL S OLUTION TO P ROBLEM (P2) In this section, we consider problem (P2). By introducing an auxiliary variable R, problem (P2) can be equivalently expressed as (P2.1) : max R {ρk (t),Qk (t),q(t)},R Z T 1 rk (ρk (t), q(t), Qk (t))dt ≥ R, ∀k ∈ K (14) s.t. T 0 (9), (10), (11), and (12). Although problem (P2.1) is still non-convex, one can easily show that it satisfies the so-called time-sharing condition in [35]. Therefore, the strong duality holds between (P2.1) and its Lagrange dual problem. As a result, we can optimally solve (P2.1) by using the Lagrange dual method. Let λk ≥ 0 and µk ≥ 0, k ∈ K, denote the dual variables associated with the k-th constraints in (14) and (9), respectively. For notational convenience, we define λ , [λ1 , · · · , λK ] and µ , [µ1 , · · · , µK ]. The partial Lagrangian of (P2.1) is L({ρk (t), Qk (t), q(t)}, R, λ, µ) ! Z K K X X λk T λk R + = 1− rk (ρk (t), q(t), Qk (t))dt T 0 k=1 k=1 Z T  Z T K X + µk E(ρ0 (t), q(t))dt − ρk (t)Qk (t)dt . k=1 0 0 (15) The Lagrange dual function of (P2.1) is g(λ, µ) = max L({ρk (t), Qk (t), q(t)}, R, λ, µ) {q(t),Qk (t),ρk (t)},R s.t. (10), (11), and (12). (16) Lemma 3.1: In order for the dual function g(λ, µ) to be upperPbounded from above (i.e., g(λ, µ) < ∞), it must hold K that k=1 λk = 1. PK PK Proof: Suppose that k=1 λk > 1 (or k=1 λk < 1). Then by settingPR → −∞ (or R → ∞), we have g(λ, µ) → K ∞. Therefore, k=1 λk = 1 must hold in order for g(λ, µ) to be bounded from above, and this lemma is proved. Based on Lemma 3.1, the dual problem of problem (P2.1) is given by (D2.1) : min g(λ, µ) λ,µ s.t. K X k=1 λk = 1 λk ≥ 0, µk ≥ 0, ∀k ∈ K. For notational convenience, let X denote the set of λ and µ specified by the constraints in (D2.1). As the strong duality holds between (P2.1) and (D2.1), we can solve (P2.1) by equivalently solving (D2.1). In the following, we first obtain g(λ, µ) by solving problem (16) under any given (λ, µ) ∈ X , and then solve (D2.1) by finding the optimal λ and µ to minimize g(λ, µ). 1) Obtaining g(λ, µ) by Solving Problem (16) Under Given (λ, µ) ∈ X : First, consider problem (16) under any given (λ, µ) ∈ X . It is evident that problem (16) can be decomposed into the following subproblems.   K X (17) max 1− λk R. R k=1 max {Qk (t),ρk (t)},q(t) K X ρk (t)ϕk (q(t), Qk (t), λk , µk ) k=1 + ρ0 (t)φ(q(t), {µk }) Qk (t) ≥ 0, ∀k ∈ K s.t. ρk (t) ∈ {0, 1}, ∀k ∈ {0} ∪ K, (18) K X ρk (t) = 1, k=0 (19)   λk Qk (t)hk (q(t)) ϕk (q(t), Qk (t), λk , µk ) = log2 1 + T σ2 − µk Qk (t), k ∈ K, ∀t ∈ T , where φ(q(t), {µk }) = K X ηP µk hk (q(t)). k=1 Here, problem (18) consists of an infinite number of subproblems, each corresponding to one time instant t. NotePthat the optimal value of problem (17) is always zero K as 1− k=1 λk = 0 (see Lemma 3.1). In this case, the optimal solution R∗ to problem (17) can be chosen as any arbitrary real number. Therefore, we only need to focus on problem (18). As the subproblems in (18) are identical for different time instants t’s, we can drop the index t for notational convenience, and denote the optimal solution as {Q∗k }, q∗ , and {ρ∗k }. As for problem (18), there are a total of K + 1 feasible choices for {ρk } due to the constraints in (19). In the following, we solve problem (18) by first obtaining the maximum objective value (and the corresponding optimal {Qk } and q) under each of the K + 1 feasible {ρk }, and then comparing them to obtain the optimal {ρk }. First, consider that ρ0 = 1 and ρk = 0, ∀k ∈ K. In this case, problem (18) can be re-expressed as max φ(q, {µk }) {Qk (t)},q s.t. Qk ≥ 0, ∀k ∈ K, (20) for which the optimal solution is given as Qk = 0, ∀k ∈ K, (µ) and q = q̄ω , ω ∈ {1, . . . , Ω(µ) }, where Ω(µ) {q̄(µ) (21) ω }ω=1 = arg max φ(q, {µk }) q corresponds to the set of optimal hovering locations for downlink WPT, with Ω(µ) ≥ 1 denoting the number of optimal solutions to problem (21). Here, for the non-convex 5 problem (21), we solve it by using a 2D exhaustive search over the region [x, x] × [y, y], where x = mink∈K xk , x = maxk∈K xk , y = mink∈K yk , y = maxk∈K yk . Note that when the optimal solution to problem (21) is non-unique (or (µ) Ω(µ) > 1), we can arbitrarily choose any one of q̄ω ’s for obtaining the dual function g(λ, µ). Accordingly, the optimal (µ) value of problem (18) is given as φ(q̄ω , {µk }). Next, consider that ρk = 1 for any one k ∈ K and ρj = 0, ∀j ∈ {0} ∪ K, j 6= k. In this case, problem (18) can be re-expressed as max ϕk (q, Qk , λk , µk ) Qk ,q s.t. Qj ≥ 0, ∀j ∈ K. (22) Note that the objective function of problem (22) is concave with respect to Qk , and therefore, problem (22) is convex. By checking the Karush-Kuhn-Tucker (KKT) conditions, we (λk ,µk ) have the optimal ,  solution as q = wk , Qk = Q  λk T µk ln 2 + − H2 γ + , and Qj = 0, ∀j ∈ K, j 6= k, where (x) , max(x, 0). Therefore, the corresponding optimal value of problem (18) is ϕk (wk , Q(λk ,µk ) , λk , µk ). By comparing the K + 1 optimal values, i.e., (µ) φ({q̄ω }, {µk }) and ϕk (wk , Q(λk ,µk ) , λk , µk ), ∀k ∈ K, we have the following proposition, for which the proof is straightforward and thus omitted. Proposition 3.1: The optimal solution to problem (18) is obtained by considering following two cases. • (µ) If φ({q̄ω }, {µk }) ≥ ϕk (wk , Q(λk ,µk ) , λk , µk ), ∀k ∈ K, then the UAV operates in the downlink WPT mode, i.e., ρ∗0 = 1, ρ∗k = 0, Q∗k = 0, ∀k ∈ K, (µ) • (µ) q∗ ∈ {q̄1 , . . . , q̄Ω(µ) }, (23) where q∗ is generally non-unique when Ω(µ) > 1. Otherwise, we denote k∗ = arg maxk∈K ϕk (wk , Q(λk ,µk ) , λk , µk ). Then the UAV operates in the uplink WIT mode for user k ∗ , i.e., ρ∗0 = 0, ρ∗k∗ = 1, ρ∗j = 0, ∀j ∈ K, j 6= k ∗ , Q∗k∗ ∗ =Q (λk∗ ,µk∗ ) q = wk ∗ . , Q∗j ∗ = 0, ∀j ∈ K, j 6= k , (24) Note that if any two of the K + 1 optimal values (i.e., (µ) ϕk (wk , Q(λk ,µk ) , λk , µk ), ∀k ∈ K and φ({q̄ω }, {µk })) are equal, then the corresponding solutions in (23) and (24) are both optimal for problem (18). Based on Proposition 3.1, problem (18) is solved, and thus the function g(λ, µ) is obtained. 2) Finding Optimal λ and µ to Solve (D2.1): Next, we search over (λ, µ) to minimize g(λ, µ) for solving (D2.1). Since the dual problem (D2.1) is always convex but in general non-differentiable, we can use subgradient based methods, such as the ellipsoid method [37], to obtain the optimal λ and µ, denoted by λopt and µopt . Note that for the objective function g(λ, µ) in (D2.1), the subgradient with respect to (λ,  µ) is where R∗ = 0 is chosen for simplicity. 3) Constructing Optimal Primal Solution to (P2.1): With λopt and µopt at hand, it remains to construct the optimal opt opt primal solution to (P2.1), denoted by {ρopt k (t), Qk (t), q (t)} opt and R . Before proceeding, we have the following proposition. (µopt ) Proposition 3.2: It must hold that φ({q̄ω }, {µopt k }) = opt opt opt (µopt ) ϕk (wk , Q(λk ,µk ) , λopt , µ ), ∀k ∈ K, ω ∈ {1, . . . , Ω } at k k the optimal λopt and µopt . Proof: See Appendix A. By combining Propositions 3.1 and 3.2, it follows that under the optimal dual solution λopt and µopt to (D2.1), problem opt (18) has a total number of Ω(µ ) + K optimal solutions. (µopt ) Among them, the Ω optimal solutions are given in (23) for downlink WPT, and the other K solutions are given in (24) for uplink WIT (each for one user k). In this case, we need to time-share among these optimal solutions to construct the optimal primal solution to (P2.1). opt More specifically, notice that the Ω(µ ) solutions in (23) opt (µopt ) (µopt ) correspond to Ω(µ ) hovering locations q̄1 , . . . , q̄Ω(µopt ) for downlink WPT, at which only the UAV transmits at constant power P with ρ∗0 = 1; on the other hand, the kth solution in (24), k ∈ K, corresponds to that the UAV hovers above user k at location wk for uplink WIT, at which opt opt user k transmits with Q∗k = Q(λk ,µk ) and ρ∗k = 1. With time-sharing, let τω and ςk denote the hovering durations opt (µopt ) at the location q̄ω , ω ∈ {1, . . . , Ω(µ ) }, and wk , k ∈ {1, . . . , K}, respectively. In this case, we solve the following uplink common throughput maximization problem to obtain the optimal hovering durations τω ’s and ςk ’s for time sharing. (P2.2) : max R {ςk ≥0,τω ≥0},R ! opt opt ςk Q(λk ,µk ) hk (wk ) s.t. log2 1 + ≥ R, ∀k ∈ K (25) T σ2 opt ςk Q K X k=1 opt (λopt k ,µk ) ςk + ≤ Ω X (µ ) ΩX opt τω ηP hk (q̄(µ ω ω=1 τω = T. ) ), ∀k ∈ K (26) (27) ω=1 Note that problem (P2.2) is a linear program, which can be solved by standard convex optimization techniques in [37]. The optimal solution to (P2.2) is denoted as ςˆk , τ̂ω and R̂. Accordingly, we can divide the whole period T into opt (µopt ) Ω(µ ) + K sub-periods, sub-periods, Pω−1where Pωthe first Ω opt denoted by Tω = ( j=1 τ̂j , j=1 τ̂j ], ω ∈ {1, . . . , Ω(µ ) }, are for downlink WPT, and the next K sub-periods, denoted by Pk Pk−1 PΩ(µopt ) P (µopt ) TΩ(µopt ) +k = ( Ω ω=1 τ̂ω + j=1 ςˆj ], ω=1 τ̂ω + j=1 ςˆj , k ∈ K, are for uplink WIT of the K users. As a result, we have the following proposition, for which the proof is omitted for brevity. Proposition 3.3: The optimal solution to (P2.1) (and thus (P2)) is given as follows. During sub-period ω ∈ opt (µopt ) {1, . . . , Ω(µ ) }, the UAV hovers at the location q̄ω for r1 (ρ∗1 , q∗ , Q∗1 ), . . . , rK (ρ∗K , q∗ , Q∗K ), downlink WPT, i.e.,  opt opt opt (µopt ) opt ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ T E1 (ρ0 , q , P ) − T ρ1 Q1 , . . . , T EK (ρ0 , q , P ) − T ρK QK , q (t) = q̄ω , ρ0 (t) = 1, ρk (t) = 0, Qk (t) = 0, ∀k ∈ K, (28) 6 TABLE I A LGORITHM 1 FOR S OLVING P ROBLEM (P2) 6 5 (µ) 1) Obtain {q̄ω } to maximize φ(q, {µk }) in problem (21) via a 2D exhaustive search over the region [x, x] × [y, y]; 2) Obtain g(λ, µ) under given (λ, µ) by using Proposition 3.1; 3) Compute the subgradients of g(λ, µ), and then update λ and µ by using the ellipsoid method [37]. c) Until λ and µ converge with a prescribed accuracy. d) Set (λopt , µopt )← (λ, µ). e) Obtain the optimal solution to problem (P2.1) or (P2) based on Proposition 3.3. 4 H (m) a) Initialization: Given an ellipsoid ε((λ, µ), A) containing (λopt , µopt ), where (λ, µ) is the center point of ε((λ, µ), A) and the positive definite matrix A characterizes the size of ε((λ, µ), A). b) Repeat: Ground users' locations Optimal hovering locaitons for WIT Optimal hovering locations for WPT 3 2 1 0 -6 -4 -2 0 2 4 6 x (m) Fig. 2. The optimal hovering locations for WPT and WIT √ in a two-user WPCN, where D = 10 m and H = 5 m with D > 2H/ 3. 6 opt opt opt ρopt 0 (t) = 0, ρj (t) = 0, Qj (t) = 0, ∀j ∈ K, j 6= k, (29) ∀t ∈ TΩ(µopt ) +k , k ∈ K. The optimal uplink common throughput is given as Ropt = R̂ (with R̂ denoting the optimal solution obtained for (P2.2)). In summary, we present the overall algorithm for solving (P2) as Algorithm 1 in Table I, and we refer to such a solution as the multi-location-hovering solution. Remark 3.1: It is worth noting that similar multi-locationhovering solutions have been proposed in the UAV-enabled multiuser WPT system [26] and the UAV-enabled multiuser communication system with TDMA transmission [28], when the UAV’s flight period becomes sufficiently long. In [26], the UAV successively hovers above a given set of locations to maximize all users’ minimum received energy; while in [28], the UAV successively hovers above each user to maximize the minimum throughput of all users. As a matter of fact, our derived multi-location-hovering solution to problem (P2) in Proposition 3.3 unifies the results in [26] and [28], which opt consists of two sets of hovering locations: Ω(µ ) ones for WPT and the other K ones for WIT. Nevertheless, note that opt the Ω(µ ) hovering locations to problem (P2) for WPT are generally different from those in [26], as they are designed based on different objective functions (max-min communication throughput versus max-min harvested energy). Remark 3.2: To gain more insights, it is interesting to consider problem (P2) in the special case of K = 2 users. Without loss of generality, we assume that the two users are located at (−D/2, 0, 0) and (D/2, 0, 0) with x1 = −D/2, x2 = D/2, and y1 = y2 = 0, where D denotes the distance between the two users. Due to this symmetric setup, it can be shown that opt the optimal Ω(µ ) hovering locations for WPT to problem (P2) are actually identical to the optimal hovering locations to maximize the two users’ minimum harvested energy in the UAV-enabled WPT system [26], [27]. It follows from [26], [27] that the optimal hovering locations for downlink WPT are critically dependent on the UAV’s flying altitude H and 5 4 H (m) opt ∀t ∈ Tω , ω ∈ {1, . . . , Ω(µ ) }. During sub-period Ω(µ ) + k, k ∈ K, the UAV hovers above user k at wk , and user k sends information to the UAV in the uplink, i.e., opt opt opt opt qopt (t) = wk , ρk (t) = 1, Qk (t) = Q(λk ,µk ) , Ground users' locations Optimal hovering locations for WIT Optimal hovering location for WPT 3 2 1 0 -3 -2 -1 0 1 2 3 x (m) Fig. 3. The optimal hovering locations for WPT and WIT √ in a two-user WPCN, where D = 5 m and H = 5 m with D ≤ 2H/ 3. √ the users’ distance D. In particular, when D > 2H/ 3, there opt are Ω(µ ) = 2 hoveringqlocations at (−ǫ, 0, H) and (ǫ, 0, H) p for WPT, where ǫ , −(D2 /4 + H 2 ) + D4 /4 + H 2 D2 √ with limD→∞ ǫ = D/2; while when D ≤ 2H/ 3 there is opt only Ω(µ ) = 1 hovering location (0, 0, H) right above the middle point of two users. By combining the optimal hovering locations for √ WPT and those for WIT, it is evident that when D > 2H/ 3, the UAV needs to hover above four locations for efficient WPCN, √ as shown by the example in Fig. 2; while when D ≤ 2H/ 3, the UAV needs to hover above three locations, as shown by the example in Fig. 3. IV. P ROPOSED S OLUTION TO P ROBLEM (P1) WITH S UCCESSIVE H OVER - AND -F LY T RAJECTORY This section considers problem (P1) with the UAV’s maximum speed constraint considered. First, we present a successive hover-and-fly trajectory motivated by the multi-locationhovering solution to the relaxed problem (P2), in which the opt UAV sequentially visits the Ω(µ ) + K hovering locations for efficient WPT and WIT, respectively. Next, under such a flying trajectory, we design the duration at each hovering location and the transmission resource allocation for (P1) by discretizing the time period. Finally, we discuss the case when the UAV’s flight duration T is too small to visit all these hovering locations, for which the trajectory and transmission resource allocations are redesigned. 7 A. Successive Hover-and-Fly UAV Trajectory In the proposed successive hover-and-fly trajectory design, opt the UAV sequentially visits the Ω(µ ) + K optimal hovering (µopt ) (µopt ) , . . . , q̄Ω(µopt ) locations that are obtained for (P2), i.e., q̄1 for downlink WPT, and w1 , . . . , wK for uplink WIT. For opt notational convenience, we denote the Ω(µ ) + K hovering opt opt (µ ) opt locations as {qopt , k ∈ {1, . . . , Ω(µ ) }, k }, where qk = q̄k opt and qΩ(µopt ) +k = wk , k ∈ K. In order to maximize the time for efficient WPT and WIT, the UAV flies among these hovering locations by using the maximum speed Vmax , and the UAV aims to minimize the flying time by equivalently minimizing opt the traveling path among the Ω(µ ) + K locations. Towards this end, we define a set of binary variables opt {fj,k }, ∀j, k ∈ {1, . . . , Ω(µ ) + K}, j 6= k, where fj,k = 1 or fj,k = 0 indicates that the UAV flies or does not fly from the opt opt j-th hovering location qj to the k-th hovering location qk . Hence, the traveling path minimization problem becomes dePΩ(µopt ) +K PΩ(µopt ) +K termining {fj,k } to minimize j=1 k=1,k6=j fj,k dj,k , (µopt ) provided that each of the Ω + K locations is visited opt once, where dj,k = kqopt j − qk k denotes the distance between qopt and qopt j k . Note that as shown in [26], the flying distance minimization is similar to the well-established traveling salesman problem (TSP), with the following differences. The standard TSP requires the salesman (or the UAV in this paper) to return to the origin city (the initial hovering location) after visiting all the other cities (or hovering locations here), while the flying distance minimization problem of our interest does not have such a requirement since the initial and final hovering locations can be optimized. As shown in [36], we can transform our traveling distance minimization problem to the standard TSP as follows. First, we add a dummy hovering opt location, namely the (Ω(µ ) + K + 1)-th hovering location, opt whose distances to all the existing Ω(µ ) + K hovering locations are 0, i.e., dΩ(µopt ) +K+1,k = dk,Ω(µopt ) +K+1 = 0, opt ∀k ∈ {1, . . . , Ω(µ ) + K}. Note that this dummy hovering location is a virtual node that does not exist physically. Then, we obtain the desirable traveling path by solving the opt standard TSP problem for the Ω(µ ) + K + 1 hovering locations, and then removing the two edges associated with the dummy location. As a result, we use the permutation opt κ(·) over the set {1, . . . , Ω(µ ) + K} to denote the obtained traveling path, where the κ(1)-th hovering location is first visited, followed by the κ(2)-th, the κ(3)-th, etc., until the opt κ(Ω(µ ) +K)-th hovering location at last. We denote the traveling distance and traveling duration from the κ(i)-th hovering opt location qopt κ(i) to the κ(i + 1)-th hovering location qκ(i+1) as dκ(i),κ(i+1) and Tfly,i = dκ(i),κ(i+1) /Vmax , respectively, opt i ∈ {1, ..., Ω(µ ) + K − 1}. Hence, the total traveling distance PΩ(µopt ) +K−1 dκ(i),κ(i+1) , and duration are given as Dfly = i=1 and Tfly = Dfly /Vmax , respectively. We denote the obtained Tfly flying trajectory as {q̃(t)}t=0 . Note that in practice, the UAV’s flight duration T may not be exactly equal to the traveling distance Tfly . When T > Tfly , we further need to determine the hovering durations at opt the Ω(µ ) + K locations. When T < Tfly , the UAV does opt not have sufficient time to visit all the Ω(µ ) + K hovering locations, and thus we need to redesign the UAV trajectory in order to satisfy the duration requirements. We present the complete successive hover-and-fly trajectory for the above two cases in Sections IV-B and IV-C, respectively, together with the transmission resource allocations. B. Hovering Durations and Transmission Resource Allocation Optimization When T > Tfly Let τω denote theopttime duration for the UAV to hover at the opt (µ ) opt location qω = q̄ω for WPT, ω ∈ {1, . . . , Ω(µ ) }, and ςk denote the the time duration for the UAV to hover above user k at qopt = wk for WIT, k ∈ {1, . . . , K}. Furthermore, opt Ω(µ ) +k we define ( opt τκ(k) , if κ(k) ∈ {1, . . . , Ω(µ ) } opt opt υk , ςκ(k)−Ω(µopt ) , if κ(k) ∈ {Ω(µ ) + 1, . . . , Ω(µ ) + K}. Accordingly, we can divide the time duration opt into 2(Ω(µ ) + K) − 1 sub-periods, denoted by [T̂1 , T̂2 , . . . , T̂2(Ω(µopt ) +K)−1 ], which are defined as   Pk−1 Pk−1 T̂2k−1 , i=1 (υκ(i) + Tfly,i ), i=1 (υκ(i) + Tfly,i ) + υκ(k) opt for odd sub-periods with k ∈ {1, . . . , Ω(µ ) + K}, and  Pk Pk−1 T̂2k , i=1 (υκ(i) + Tfly,i ) i=1 (υκ(i) + Tfly,i ) + υκ(k) , opt for even sub-periods with k ∈ {1, . . . , Ω(µ ) + K − 1}. Therefore, within each odd sub-period 2k −1, the UAV should hover at the κ(k)-th hovering location (xκ(k) , yκ(k) , H), i.e., (µopt ) q(t) = qopt + K}. (30) κ(k) , ∀t ∈ T̂2k−1 , k ∈ {1, . . . , Ω During each even sub-period 2k, the UAV should fly from the κ(k)-th hovering location to the κ(k + 1)-th hovering location with the UAV’s maximum speed Vmax , whose time-varying location is   k X υκ(i) , (31) q(t) = q̃ t − i=1 opt ∀t ∈ T̂2k , k ∈ {1, . . . , Ω(µ ) + K − 1}. Therefore, the successive hover-and-fly trajectory is finally obtained, in which the hovering durations {υκ(k) }, or equivalently, {τω } and {ςk }, are optimization variables that will be determined next. Under the obtained successive hover-and-fly trajectory, we maximize the uplink common throughput of all users by optimizing the transmission mode {ρk (t)} and the power allocation {Qk (t)}, jointly with the hovering durations {τω } and {ςk }. Towards this end, we first discuss the transmission policy of the UAV-enabled WPCN during each sub-period. First, consider the (2k − 1)-th sub-period T̂2k−1 , k ∈ opt {1, . . . , Ω(µ ) + K}, in which the UAV hovers above qopt κ(k) as in (30). The transmission policy in this sub-period is based on the optimal solution to (P2) in Section III. In particular, if opt 1 ≤ κ(k) ≤ Ω(µ ) , we denote ω = κ(k) for convenience. (µopt ) In this case, the UAV hovers at q̄ω with duration τω , and the UAV works in the downlink WPT mode during this sub-period, by employing transmit power P . On the other opt opt hand, if Ω(µ ) + 1 ≤ κ(k) ≤ Ω(µ ) + K, we denote opt k̂ = κ(k) − Ω(µ ) . In this case, the UAV hovers above user k̂ at wk̂ with duration ςk̂ , and user k̂ works in the uplink WIT mode to transmit information to the UAV during the whole sub-period, by employing a certain transmit power, denoted 8 opt by Qhover . By combining the Ω(µ ) + K odd sub-periods, the k̂ uplink throughput and the harvested energy at each user k are respectively given as   hover fk (ςk , Qhover ) = ςk log2 1 + hk (wk )Qk R (32) k T σ2 opt fk ({τω }) = E (µ ) ΩX opt ηP hk (q̄(µ ω ) )τω . (33) ω=1 opt Next, consider even sub-periods T̂2k , ∀k ∈ {1, . . . , Ω(µ ) + K−1}, with total duration Tfly . We discretize these sub-periods into N slots each with duration δ = Tfly /N . For each slot n ∈ N , {1, . . . , N }, the location of the UAV is assumed to be constant and denoted as qfly [n] = q̃(nδ). Furthermore, in order to handle the binary transmission mode indicator {ρk (t)}, we consider that the (K + 1) transmission modes are time shared within one slot, by dividing each slot n into (K + 1) sub-slots without loss of generality. In the first sub-slot with duration τ0fly [n] ≥ 0, the UAV works in the downlink WPT mode with transmit power P , and in sub-slot k+1 with duration τkfly [n] ≥ 0, user k works in the uplink WIT mode byP using transmit K fly power Qfly [n], k ∈ K. Here, it follows that k k=0 τk [n] = δ, ∀n ∈ N . By combining the N slots over the even subperiods, the uplink throughput and the harvested energy at each user k are respectively given by fly fly fly Rk ({τk [n], Qk [n]}) N 1 X fly τ [n]log2 = T n=1 k Ekfly ({τ0fly [n]}) = N X 1+ hk (qfly [n])Qfly k [n] σ2 ! , (34) ηP hk (qfly [n])τ0fly [n]. (35) n=1 Based on (32), (33), (34), and (35), the uplink common throughput maximization problem is reformulated as follows, in which the optimization variables are {ςk }, {Qhover }, {τω }, k {τkfly [n]}, {Qfly [n]}, and R. k (P3) : max R fk (ςk , Qhover ) + Rfly ({τ fly [n], Qfly [n]}) ≥ R, ∀k ∈ K s.t. R k ςk Qhover + k k N X n=1 k k fly fk ({τω }) + E fly ({τ fly [n]}) τk [n]Qk [n] ≤ E 0 k ςk ≥ 0, ∀k ∈ K, τω ≥ 0, ∀ω ∈ {1, . . . , Ω Qhover k K X ≥ ςk + 0, Qfly k [n] (µopt ) ΩX ≥ 0, ∀k ∈ K, n ∈ N τω = T − T fly ω=1 k=1 fly τk [n] ≥ 0, ∀n ∈ N , k ∈ K X τkfly [n] = δ, ∀n ∈ N . k=0 K ∪ {0} (µopt ) , ∀k ∈ K } (36) (37) (38) (39) (40) Note that although problem (P3) is non-convex due to the coupling of some variables (e.g., ςk and Qhover ). Fortunately, via k a change of variables (e.g., by introducing Ekhover = ςk Qhover ), k we can transform problem (P3) into a convex optimization problem, which can thus be solved via standard convex optimization techniques. By combining the optimal solution to (P3) together with the successive hover-and-fly trajectory in (30) and (31), the solution to (P1) is finally found. Remark 4.1: It is worth noting that the proposed successive hover-and-fly trajectory design is asymptotically optimal for (P1) when T → ∞, as the total flying time Tfly becomes negligible as compared to the total hovering time T − Tfly . In this case, the obtained uplink common throughput approaches the optimal value of (P2), which serves as the upper bound for that of (P1). C. Trajectory Redesign and Transmission Resource Allocation When T < Tfly Next, we consider the case with T < Tfly , in which the Tfly UAV flying trajectory {q̃(t)}t=0 based on the TSP solution is no longer feasible since the duration T is not sufficient opt for the UAV to visit all the Ω(µ ) + K hovering locations. To overcome this problem, we first find the solution to (P1) when T is sufficiently small (i.e., T → 0) such that the UAV can only hover at one single location, and then reconstruct a modified UAV trajectory for the case of T < Tfly . First, when T → 0, the UAV should hover at one single fixed location. Let q denote the hovering location of the UAV. Then problem (P1) can be re-expressed as max min Rk ({ρk , q, Qk (t)}) {ρk ,Qk (t)},q k∈K Z T s.t. 0 ρk (t)Qk (t)dt ≤ Z T 0 Ek (ρ0 (t), q)dt, ∀k ∈ K Qk (t) ≥ 0, ∀k ∈ K, t ∈ T ρk (t) ∈ {0, 1}, ∀k ∈ {0} ∪ K, t ∈ T K X k=0 ρk (t) = 1, ∀t ∈ T . (41) Note that under given UAV hovering location q, it is easy to show that problem (41) is equivalent to the common throughput maximization problem in the conventional WPCN [6]. In this case, the transmission resource allocation can be obtained optimally by performing a TDMA protocol together with a joint time and power allocation, for which the details can be found in [6]. Therefore, the optimal q to problem (41) can be obtained via a 2D exhaustive search over [x, x] × [y, y], together with the joint time and power allocation under any given q. We denote the obtained optimal q to problem (41) as qfix . With qfix at hand, we then reconstruct the trajectory for problem (P1) as follows by down-scaling the previously obTfly tained traveling path {q̃(t)}t=0 for the case of T = Tfly linearly towards the center point (xfix , yfix , H), such that the resulting total flying distance equals Vmax T . Accordingly, we can zoom in the trajectory as follows: x∗∗ (t) = x̂(t/ν) + (1 − ν)(xfix − x̂(t/ν)), y ∗∗ (t) = ŷ(t/ν) + (1 − ν)(yfix − ŷ(t/ν)), ∀t ∈ [0, T ], (42) where ν = T /Tfly < 1 denotes the linear scaling factor. Note that when T → 0, we have ν → 0, and the above redesigned trajectory reduces to hovering at one single fixed location qfix ; when T → Tfly , we have ν → 1, and the above redesigned trajectory becomes identical to the TSPTfly based trajectory {q̃(t)}t=0 . 9 TABLE II A LGORITHM 2 FOR S OLVING P ROBLEM (P1) a) Solve problem (P2) by Algorithm 1 to find the Ω(µ opt b) c) d) e) (µopt ) opt ) + K optimal +K . hovering locations {qk }Ω k=1 opt Add a dummy hovering location, namely the (Ω(µ ) + K + 1)-th opt hovering location, and set its distances to all the existing Ω(µ ) + K hovering locations as 0. Tfly Obtain the desirable traveling path {q̃(t)}t=0 by solving the standard opt (µ ) TSP problem for the Ω + K + 1 hovering locations and then remove the two edges associated with the dummy location, where Tfly denotes the total flying time. If T ≥ Tfly , then find the optimal hovering time allocation, and transmission resource allocation by solving problem (P3); accordingly, obtain the corresponding trajectory as in Section IV-B. Otherwise, if T < Tfly , then obtain the trajectory based on (42), and accordingly find the optimal transmission resource allocation. (P3), we divide each slot n into (K + 1) sub-slots, with the first sub-slot with duration τ0 [n] ≥ 0 for downlink WPT, and the k-th sub-slot with duration τk [n] ≥ 0 for user k’s uplink transmit power Qk [n], k ∈ K. Here, we have PK WIT with e e k=0 τk [n] = δ, ∀n ∈ N . Accordingly, the uplink achievable data rate and the harvested energy at each user k ∈ K at slot e are respectively given as n∈N   hk (q[n])Qk [n] , rbk (τk [n], Qk [n], q[n]) =τk [n]log2 1 + σ2 (43) Ek (τ0 [n], q[n]) =ηP hk (q[n])τ0 [n]. (44) Accordingly, problem (P1) can be reformulated as e N 1 X (P4) : max min rbk (τk [n], Qk [n], q[n]) {τk [n],Qk [n],q[n]} k∈K T n=1 s.t. In this case, the successive hover-and-fly trajectory can be modified as {q∗∗ (t)} with q∗∗ (t) = (x∗∗ (t), y ∗∗ (t)), ∀t ∈ [0, T ]. Accordingly, the transmission resource allocation can be obtained similarly as for (P3) by discretizing the T into N time slots and employing TDMA within each slot for WPT and WIT, respectively. The details of the transmission resource allocation are omitted for convenience. In summary, we present the overall algorithm with the successive hover-and-fly trajectory for solving (P1) as Algorithm 2 in Table II, with both the cases of T ≥ Tfly and T < Tfly . Remark 4.2: It is worth further discussing the special case with K = 2 users to provide more design insights. As mentioned in Remark 3.2, at the optimal solution to (P2) in this case, there are in general three or four hovering locations, which are all located above the line connecting the two users. Therefore, the corresponding successive hoverand-fly trajectory always flies above the line between the two users, and the total flying duration for visiting all these hovering locations is Tfly = D/Vmax , with D denoting the distance between the two users. Due to the symmetry between the two users, it can be shown via contradiction that in the case of T ≥ Tfly , the successive hover-and-fly trajectory and the correspondingly obtained transmission resource allocation are indeed the globally optimal solution to problem (P1). Nevertheless, such a result does not hold in the general scenario with K > 2 users. V. A LTERNATING O PTIMIZATION BASED S OLUTION P ROBLEM (P1) TO In this section, we propose an alternative solution to problem (P1) based on the technique of alternating optimization, which optimizes the UAV trajectory and the transmission resource allocation in an alternating manner towards a locally optimal solution. Towards this end, we reformulate problem (P1) by discretizing the whole period T with duration T e time slots, each with duration into a finite number of N e e δ = T /N . Note that the duration δe is chosen to be sufficiently small, such that we can assume the UAV’s location is approximately unchanged during each slot n, which is denoted as e , {1, ..., N e }. Similarly as for problem (x[n], y[n], H), n ∈ N e N X n=1 τk [n]Qk [n] ≤ e N X n=1 Ek (τ0 [n], q[n]), ∀k ∈ K (45) 2 e2 e − 1} (46) δ , ∀n ∈ {1, ..., N kq[n + 1] − q[n]k2 ≤ Vmax K X k=0 e ∀n ∈ N e τk [n] ≤ δ, (47) where the constraints in (46) correspond to the discretized version of the UAV’s maximum speed constraint in (13). In the following, we optimize each of the UAV trajectory {q[n]} and the transmission resource allocation {τk [n], Qk [n]} for (P4), respectively, by assuming the other one is given. First, we optimize the UAV trajectory {q[n]} under any given {τk [n]} and {Qk [n]}. In this case, problem (P4) is not a convex optimization problem with respect to {q[n]}, as the rate function in the objective function and the energy function at the right-hand-side (RHS) of (45) are both non-concave. To tackle this issue, we propose an efficient algorithm by using the SCP technique, which updates the UAV trajectory {q[n]} in an iterative manner by transforming the non-convex problem into a convex approximate problem. Suppose that {q(0) [n]} = {(x(0) [n], y (0) [n])} denotes the initial UAV trajectory and {q(i) [n]} = {(x(i) [n], y (i) [n])} corresponds to the obtained UAV trajectory after iteration i ≥ 1. Then we have the lower bounds for Ek (τ0 [n], q[n]) and rbk (τk [n], Qk [n], q[n]) in the following lemma. Lemma 5.1: Under any given UAV trajectory {q(i) [n]}, it follows that (i) Ek (τ0 [n], q[n]) ≥ Ek (τ0 [n], q[n]), (48) rbk (τk [n], Qk [n], q[n]) ≥ rbk (i) (τk [n], Qk [n], q[n]) , ∀k ∈ K, n ∈ N , (49) where 2ηβ0 P τ0 [n] (i) Ek (τ0 [n], q[n]) , 2 H + kq(i) [n] − wk k2 ηβ0 P τ0 [n](H 2 + kq[n] − wk k2 ) − , (50) H 2 + kq(i) [n] − wk k2 ! Q [n]γ k rbk (i) (τk [n], Qk [n], q[n]) , τk [n]log2 1 + (i) H 2 + Sk [n] (i) − γQk [n]τk [n] log2 e(Sk [n] − Sk [n]) (i) (i) (H 2 + Sk [n])2 + γQk [n]τk [n](H 2 + Sk [n]) , (51) 10 (i) Ek (τ0 [n], q(i) [n]) = Ek (i) rbk (τk [n], Qk [n], q [n]) = rbk (i) (i) (τ0 [n], q(i) [n]), (52) (i) (τk [n], Qk [n], q [n]) , ∀k ∈ K, n ∈ N . (53) Proof: See Appendix B. Based on Lemma 5.1, at each iteration i + 1, we optimize over q[n] by replacing Ek (τ0 [n], q[n]) and rbk (τk [n], Qk [n], q[n]) in problem (P4) with their re(i) spective lower bounds Ek (τ0 [n], q[n]) in (50) and (i) rbk (τk [n], Qk [n], q[n]) in (51), respectively, with the obtained UAV trajectory {q(i) [n]} at the previous iteration i. More specifically, the UAV trajectory is updated as e N 1 X (i) q(i+1) [n] = arg max min rbk (τk [n], Qk [n], q[n]) {q[n]} k∈K T n=1 s.t. N X n=1 τk [n]Qk [n] ≤ N X n=1 Ek (i) (τ0 [n], q[n]), ∀k ∈ K (46). (54) Note that for problem (54), the function rbk (i) (τk [n], Qk [n], q[n]) is concave with respect to q[n], and the constraints are convex. As a result, problem (54) is a convex optimization problem, and thus can be optimally solved by standard convex optimization techniques such as the interior point method [37]. Furthermore, as shown in Lemma 4.1, the objective function in problem (54) serves as a lower bound for that in problem (P4). Therefore, after each iteration i, the objective function of problem (P4) achieved by {q[n]} monotonically increases. As problem (P4) has a finite optimal value, the SCP-based trajectory design in (54) will converge to a locally optimal solution. Next, we optimize the transmission resource allocation {τk [n]} and {Qk [n]} under given UAV trajectory {q[n]}. Although this problem is non-convex, similarly as for problem (P3), we can transform it into a convex optimization problem via change of variables, which can thus be solved via standard convex optimization techniques [37]. Finally, we optimize over the UAV trajectory {q[n]} via (54) based on the SCP technique and the transmission resource allocation via the convex optimization technique, in an alternating manner. Such an alternating optimization can ensure the objective function of (P4) to monotonically increase. As the optimal value of (P4) is bounded from above, the alternating optimization approach will eventually converge to a locally optimal solution to (P4). Note that the performance of the alternating-optimizationbased approach for solving (P4) critically depends on the initial point for iteration. In this paper, we choose the successive hover-and-fly trajectory based solution in Section IV as the initial point. In this case, the alternating-optimizationbased approach in this section can always achieve a common throughput no smaller than the successive hover-and-fly tra- 3.2 3.1 Common throughput (bps/Hz) with e defining the Euler’s number, Sk [n] , kq(i) [n]−wk k2 , (i) and Sk [n] , kq[n] − wk k2 . Notice that Ek (τ0 [n], q[n]) (i) and rbk (τk [n], Qk [n], q[n]) are both concave functions with respect to q[n]. The inequalities in (48) and (49) are tight for q[n] = q(i) [n], i.e., Successive hover-and-fly trajectory SCP-based trajectory Multi-location hovering for (P2) Static hovering 3 2.9 2.8 2.7 2.6 1 2 3 4 5 6 7 8 9 10 11 12 Duration, T (s) Fig. 4. Uplink common throughput versus the flight duration T for the case with K = 2 users and D = 10 m. jectory based design, as will be validated by our numerical results in the next section. VI. N UMERICAL R ESULTS In this section, we present numerical results to validate the performance of our proposed joint trajectory and transmission resource allocation design as compared to the following banchmark scheme. • Static hovering: The UAV hovers at a fixed location q over the whole period. In this case, the uplink common throughput maximization problem is reduced to problem (41), which can be solved efficiently via a 2D exhaustive search over q and solving for the transmission resource allocation under any given q. In the simulation, the UAV flies at a fixed altitude H = 5 m. The receiver noise power at the UAV is set as σ 2 = −80 dBm. The channel power gain at the reference distance d0 = 1 m is set as β0 = −30 dB. The energy harvesting efficiency is set as η = 50%. The maximum speed of the UAV is Vmax = 10 m/s. The transmit power at the UAV is P = 40 dBm. First, we consider the special case with K = 2 users. Fig. 4 shows the uplink common throughput in this case versus the flight duration T , in which the distance between the two users is set as D = 10 m. It is observed that the successive hover-and-fly trajectory has the same performance as the SCPbased trajectory. This is consistent with Remark 4.2, which shows that the successive hover-and-fly trajectory is indeed optimal to (P1) when T ≥ Tfly = D/Vmax = 1 s. As a result, in this case, the SCP-based trajectory cannot further improve the performance. Furthermore, it is observed that the proposed successive hover-and-fly trajectory and SCP-based trajectory achieve higher common throughput than the statichovering benchmark, and the performance gain becomes more substantial when T becomes larger. Last, the two proposed designs are observed to approach the performance upper bound by the multi-location-hovering solution for (P2). This is consistent with Remark 4.1. Next, we consider a system with K = 9 ground users that are randomly distributed within a 2D area of 20 × 20 m2 , as shown in Fig. 5. For illustration, Fig. 5 also shows the optimal hovering locations for problem (P2), the successive 11 10 0.85 8 0.8 6 4 y (m) 2 Common throughput (bps/Hz) SCP-based trajectory Successive hover-and-fly trajectory Users' location Optimal hovering locations for WPT 0 -2 -4 -6 Successive hover-and-fly trajectory SCP-based trajectory Multi-location hovering for (P2) Static hovering 0.75 0.7 0.65 0.6 -8 -10 -10 0.55 -8 -6 -4 -2 0 2 4 6 8 10 x (m) × 10-3 1.8 Successive hover-and-fly trajectory SCP-based trajectory Harvested energy (J) 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 1 2 3 4 5 6 7 8 9 User k Fig. 6. The harvested energy at the K users. 1 Successive hover-and-fly trajectory SCP-based trajectory Average achievable rate (bps/Hz) 0.9 2 3 4 5 6 7 8 9 10 11 12 Duration, T (s) Fig. 5. System setup for simulation and various trajectories obtained with T = 12 s. 2 1 0.8 0.7 0.6 0.5 0.4 Fig. 8. Uplink common throughput versus the flight duration T . the corresponding transmission resource allocation for (P1) can efficiently balance the communication rates among the K users, but generally may lead to unbalanced energy harvesting at these distributed users. This shows the difference between our design versus that in the UAV-enabled multiuser WPT system [26], in which the harvested energy of all users are designed to be identical. Fig. 8 shows the uplink common throughput of the K users versus the flight duration T . It is observed that the proposed successive hover-and-fly trajectory and SCP-based trajectory (jointly with the corresponding optimized transmission resource allocation) achieve higher common throughput than the static-hovering benchmark, and the performance gain becomes more substantial when T becomes larger. In particular, the SCP-based trajectory outperforms the successive hover-andfly trajectory, especially when the flight duration is small. Furthermore, with T being sufficiently large, the successive hover-and-fly trajectory and the SCP-based trajectory are both observed to approach the performance upper bound by the multi-location-hovering solution to (P2) with the UAV’s maximum speed constraint ignored. This is consistent with Remark 4.1. 0.3 VII. C ONCLUSION 0.2 0.1 0 1 2 3 4 5 6 7 8 9 User k Fig. 7. The average achievable rate at the K users. hover-and-fly trajectory as well as the SCP-based trajectory under the 9-user setup, where T = 12 s. It is observed that opt there are Ω(µ ) = 4 optimal hovering locations for WPT and opt a total of Ω(µ ) + K = 13 optimal hovering locations for problem (P2). Fig. 6 and Fig. 7 show the harvested energy of each ground user, in which T = 12 s. It is observed in Fig. 6 that in both proposed designs, the harvested energy values at different users are generally different. By contrast, it is observed in Fig. 7 that the achievable rates at these users are the same in order to maximize the common throughput. It can be inferred from Figs. 6 and 7 that the trajectory design and In this paper, we investigate the common throughput maximization problem in a new UAV-enabled WPCN setup, by jointly optimizing the UAV trajectory and the transmission resource allocation in both downlink WPT and uplink WIT, subject to the UAV’s maximum speed constraint and the users’ energy neutrality constraints. To solve this challenging problem, we first consider the ideal case without the UAV’s maximum speed constraint and solve the relaxed problem optimally. The optimal solution shows that the UAV should successively hover above two sets of optimal ground locations for downlink WPT and uplink WIT, respectively. Next, based on the optimal solution to the relaxed problem, we propose the successive hover-and-fly trajectory and the SCP-based trajectory to solve the problem with the UAV’s maximum speed constraint considered. Numerical results showed that the proposed UAV-enabled WPCN achieves near-optimal performance when the flight period is sufficiently large, and significantly enhances the uplink common throughput performance 12 over the conventional WPCN with a static AP located even at the optimal location, by effectively resolving the “doubly near-far” fairness issue. A PPENDIX A. Proof of Proposition 3.2 Suppose that these K terms are not identical at the optimal λopt and µopt , for which there are in general K + 1 cases discussed in the following. (µopt ) In the first case, suppose that φ({q̄ω }, {µopt k }) is smaller opt opt opt than any one of ϕk (wk , Q(λk ,µk ) , λopt , µ ). It follows from k k Proposition 3.1 that ρ0 (t) = 0, ∀t ∈ T , i.e., the WPCN does not work in the downlink WPT mode throughout the whole period with duration T . In this case, the K users cannot harvest any energy from the UAV, thus leading to a zero common throughput. This solution is thus not optimal. In the k-th one of the other K cases, suppose opt opt opt that ϕk (wk , Q(λk ,µk ) , λopt than any one k , µk ) is smaller (µopt ) of the other K terms (i.e., φ({q̄ω }, {µopt k }) and opt opt opt ϕj (wj , Q(λj ,µj ) , λopt , µ )’s, ∀j ∈ K, j = 6 k). It follows j j from Proposition 3.1 that ρk (t) = 0, ∀t ∈ T , i.e., the WPCN does not work in the uplink WIT mode for user k throughout the whole period with duration T . In this case, the throughput for user k is zero, thus resulting in a zero common throughput. This solution thus cannot be optimal as well. By combining all the K + 1 cases in the above, in order for the common throughput to be non-zero, it must hold opt opt (µopt ) opt opt opt that φ(q̄ω , {µk }) = ϕk (wk , Q(λk ,µk ) , λk , µk ), ∀k ∈ opt K, ω ∈ {1, . . . , Ω(µ ) }. Therefore, Proposition 3.2 is proved. B. Proof of Lemma 5.1 τ0 [n] and g2 (z) = Define function g1 (z) = ηPHβ20+z Qk [n]γ τk [n]log2 1 + H 2 +z , which are both convex with respect to z ≥ 0. As the first-order Taylor expansion of a convex function is a global under-estimator of the function values, for any given z0 ≥ 0, it follows that g(z) ≥ g(z0 )+g ′ (z0 )(z −z0 ), or equivalently, ηP β0 τ0 [n] ηP β0 τ0 [n] ηP β0 τ0 [n] − (z − z0 ), (55) ≥ H2 + z H 2 + z0 (H 2 + z0 )2     Qk [n]γ Qk [n]γ τk [n]log2 1 + 2 ≥ τk [n]log2 1 + 2 H +z H + z0 γQk [n]τk [n] log2 e (z − z0 ). (56) − (H 2 + z0 )2 + γQk [n]τk [n](H 2 + z0 ) For any given k ∈ K, n ∈ N , and i ≥ 0, by substituting z = kq[n]−wk k2 and z0 = kq(i) [n]−wk k2 into (55) and (56), then (50) and (51) follow, respectively. Furthermore, note that the equality holds for (55) and (56) for z = z0 , and therefore, the equality in (49) holds. Therefore, Lemma 5.1 is proved. R EFERENCES [1] L. Xie, J. Xu, and R. Zhang, “Throughput maximization for UAV-enabled wireless powered communication networks,” to appear in Proc. IEEE VTC Spring, 2018. [Online] Available: https://arxiv.org/abs/1801.04545 [2] X. Lu, P. Wang, D. Niyato, D. I. Kim, and Z. Han, “Wireless charging technologies: fundamentals, standards, and network applications,” IEEE Commun. Surveys Tuts., vol. 18, no. 2, pp. 1413–1452, Nov. 2015. [3] Y. Zeng, B. Clerckx, and R. Zhang, “Communications and signals design for wireless power transmission,” IEEE Trans. Commun., vol. 65, no. 5, pp. 2264–2290, May 2017. [4] S. Bi, C. K. Ho, and R. Zhang, “Wireless powered communication: opportunities and challenges,” IEEE Commun. Mag., vol. 53, no. 4, pp. 117–125, Apr. 2015. [5] S. Bi, Y. Zeng, and R. Zhang, “Wireless powered communication networks: an overview,” IEEE Wireless Commun., vol. 23, no. 4, pp. 10–18, Apr. 2016. [6] H. Ju and R. Zhang, “Throughput maximization in wireless powered communication networks,” IEEE Trans. Wireless Commun., vol. 13, no. 1, pp. 418–428, Jan. 2014. [7] S. Bi and R. Zhang, “Placement optimization of energy and information access points in wireless powered communication networks,” IEEE Trans. Wireless Commun., vol. 15, no. 3, pp. 2351–2364, Mar. 2016. [8] J. Xu and R. Zhang, “Energy beamforming with one-bit feedback,” IEEE Trans. Signal Process., vol. 62, no. 20, pp. 5370–5381, Oct. 2014. [9] L. Liu, R. Zhang, and K. C. Chua, “Multi-antenna wireless powered communication with energy beamforming,” IEEE Trans. Commun., vol. 62, no. 12, pp. 4349–4361, Dec. 2014. [10] J. Xu, L. Liu, and R. Zhang, “Multiuser MISO beamforming for simultaneous wireless information and power transfer,” IEEE Trans. Signal Process., vol. 62, no. 18, pp. 4798–4810, Sep. 2014. [11] J. Xu and R. Zhang, “A general design framework for MIMO wireless energy transfer with limited feedback,” IEEE Trans. Signal Process., vol. 64, no. 10, pp. 2475–2488, May 2016. [12] Y. Che, J. Xu, L. Duan, and R. Zhang, “Multiantenna wireless powered communication with co-channel energy and information transfer,” IEEE Commun. Lett., vol. 19, no. 12, pp. 2266–2269, Dec. 2015. [13] H. Ju and R. Zhang, “User cooperation in wireless powered communication networks,” in Proc. IEEE Globecom, 2014. [14] H. Chen, Y. Li, J. L. Rebelatto, B. F. U. Filho, and B. Vucetic, “Harvest-then-cooperate: wireless-powered cooperative communications,” IEEE Trans. Signal Process., vol. 63, no. 7, pp. 1700–1711, Apr. 2015. [15] Y. Zeng, R. Zhang, and T. J. Lim, “Wireless communications with unmanned aerial vehicles: opportunities and challenges,” IEEE Commun. Mag., vol. 54, no. 5, pp. 36–42, May 2016. [16] Y. Zeng, R. Zhang, and T. J. Lim, “Throughput maximization for UAVenabled mobile relaying systems,” IEEE Trans. Commun., vol. 64, no. 12, pp. 4983–4996, Dec. 2016. [17] J. Lyu, Y. Zeng, R. Zhang, and T. J. Lim, “Placement optimization of UAV-mounted mobile base stations,” IEEE Commun. Lett., vol. 21, no. 3, pp. 604–607, Mar. 2017. [18] C. Zhan, Y. Zeng, and R. Zhang, ”Energy-efficient data collection in UAV enabled wireless sensor network,” to appear in IEEE Wireless Commun. Lett., [Online] Available: https://arxiv.org/abs/1708.00221 [19] R. I. Bor-Yaliniz, A. El-Keyi, and H. Yanikomeroglu, “Efficient 3-D placement of an aerial base station in next generation cellular networks,” in Proc. IEEE ICC, 2016. [20] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Efficient deployment of multiple unmanned aerial vehicles for optimal wireless coverage,” IEEE Commun. Lett., vol. 20, no. 8, pp. 1647–1650, Aug. 2016. [21] Q. Wu, Y. Zeng, and R. Zhang, “Joint trajectory and communication design for multi-UAV enabled wireless networks,” IEEE Trans. Wireless Commun., vol. 17, no. 3, pp. 2109–2121, Mar. 2018. [22] Q. Wu, J. Xu, and R. Zhang, “Capacity characterization of UAVenabled two-user broadcast channel.” [Online] Available: https://arxiv.org/ abs/1801.00443 [23] Y. Wu, J. Xu, L. Qiu, and R. Zhang, “Capacity of UAV-enabled multicast channel: joint trajectory design and power allocation,” to appear in Proc. IEEE ICC, 2018. [Online] Available: https://arxiv.org/abs/1711.04387 [24] J. Chen, O. Esrafilian, D. Gesbert, and U. Mitra, “Efficient algorithms for air-to-ground channel reconstruction in UAV-aided communications,” in Proc. IEEE Globecom Workshop, 2017. [25] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Unmanned aerial vehicle with underlaid device-to-device communications: performance and tradeoffs,” IEEE Trans. Wireless Commun., vol. 15, no. 6, pp. 3949– 3963, Jun. 2016. [26] J. Xu. Y. Zeng, and R. Zhang, “UAV-enabled wireless power transfer: trajectory design and energy optimization,” submitted to IEEE Trans. Wireless Commun.. [Online] Available: https://arxiv.org/abs/1706.07010 [27] J. Xu, Y. Zeng, and R. Zhang, “UAV-enabled wireless power transfer: trajectory design and energy region characterization,” in Proc. IEEE Globecom Workshop, 2017. [28] Q. Wu, Y. Zeng, and R. Zhang, “Joint trajectory and communication design for UAV-enabled multiple access,” in Proc. IEEE Globecom, 2017. 13 [29] L. Xie, Y. Shi, Y. T. Hou, and H. D. Sherali, “Making sensor networks immortal: an energy-renewal approach with wireless power transfer,” IEEE/ACM Trans. Netw., vol. 20, no. 6, pp. 1748–1761, Dec. 2012. [30] Y. Shu, H. Yousefi, P. Cheng, J. Chen, Y. Gu, T. He, and K. G. Shin, “Near-optimal velocity control for mobile charging in wireless rechargeable sensor networks,” IEEE Trans. Mobile Comput., vol. 15, no. 7, pp. 1699–1713, Jul. 2016. [31] Y. Zhang, S. He, and J. Chen, “Near optimal data gathering in rechargeable sensor networks with a mobile sink,” IEEE Trans. Mobile Comput., vol. 16, no. 6, pp. 1718–1729, Jun. 2017. [32] E. Boshkovska, D. W. K. Ng, N. Zlatanov, and R. Schober, “Practical non-linear energy harvesting model and resource allocation for SWIPT systems,” IEEE Commun. Lett., vol. 19, no. 12, pp. 2082–2085, Dec. 2015. [33] B. Clerckx and E. Bayguzina, “Waveform design for wireless power transfer,” IEEE Trans. Signal Process., vol. 64, no. 23, pp. 6313–6328, Dec. 2016. [34] H. Li, J. Xu, R. Zhang, and S. Cui, “A general utility optimization framework for energy harvesting based wireless communications,” IEEE Commun. Mag., vol. 53, no. 4, pp.79–85, Apr. 2015. [35] W. Yu and R. Lui. “Dual methods for nonconvex spectrum optimization of multicarrier systems.” IEEE Trans. Commun., vol. 54, no. 7, pp. 1310– 1322, Jul. 2006. [36] E. L. Lawler, J. K. Lenstra, A. H. G. R. Kan, and D. B. Shmoys, The Traveling Salesman Problem: A Guided Tour of Combinatorial Optimization, 1st ed. Wiley, 1985. [37] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge, U.K.: Cambridge Univ. Press, Mar. 2004.
7cs.IT
Automatic Segmentation of Spine using Convolutional Neural Networks via Redundant Generation of Class Labels arXiv:1712.01640v1 [cs.CV] 29 Nov 2017 Malinda Vaniaa,b,1 , Dawit Murejac,1 , Deukhee Leea,b,∗ a Center for Bionics, Korea Institute of Science and Technology, Seoul 02792, Republic of Korea b Division of Bio-Medical Science & Technology, KIST School, Korea University of Science and Technology, Seoul 02792, Republic of Korea c Korea Advanced Institute of Science and Technology, Daejeon 34141, Republic of Korea Abstract There has been a significant increase from 2010 to 2016 in the number of people suffering from spine problems. The automatic image segmentation of the spine obtained from a computed tomography (CT) image is important for diagnosing spine conditions and for performing surgery with computer-assisted surgery systems. The spine has a complex anatomy that consists of 33 vertebrae, 23 intervertebral disks, the spinal cord, and connecting ribs. As a result, the spinal surgeon is faced with the challenge of needing a robust algorithm to segment and create a model of the spine. In this study, we developed an automatic segmentation method to segment the spine, and we compared our segmentation results with reference segmentations obtained by experts. We developed a fully automatic approach for spine segmentation from CT based on a hybrid method. This method combines the convolutional neural network (CNN) and fully convolutional network (FCN), and utilizes class redundancy as a soft constraint to greatly improve the segmentation results. The proposed method was found to significantly enhance the accuracy of the segmentation results and the system processing time. Our comparison was based on 12 measurements: the Dice coefficient (94%), Jaccard index (93%), volumetric similarity (96%), sensitivity (97%), specificity (99%), precision (over segmentation; 8.3 and under segmentation 2.6), accuracy (99%), Matthews correlation coefficient (0.93), mean surface distance (0.16 mm), Hausdorff distance (7.4 mm), and global consistency error (0.02). We experimented with CT images from 32 patients, and the experimental results demonstrated the efficiency of the proposed method. Keywords: Automatic Segmentation, Computed Tomography, Spine Segmentation, CNN, FCN. ∗ Corresponding author Email address: [email protected] (Deukhee Lee) 1 Both authors contributed equally to this work. Preprint submitted to Journal of Compuational Design and Engineering December 6, 2017 1. Introduction Computer-based technology plays an important role in defining how surgery is performed today [1]. In computer-assisted surgery (CAS), the surgeon uses a surgical navigation system to navigate an instrument in relation to the anatomy of the patient. A computer-based system uses medical images, such as computed tomography (CT) images, of the patients to extract relevant information and create a 3D model of the patient. This model can be manipulated easily by the surgeon to provide views from any angle and at any depth within the volume. Thus, the surgeon can more thoroughly assess the situation and establish a more accurate diagnosis; such an approach is utilized in computer-assisted spinal diagnosis and therapy support systems [2]. The most important element of CAS is image segmentation. This process is used to construct an accurate model of the patient. Image segmentation is important for extracting information from an image. The segmentation process subdivides an image into its constituent parts or objects, depending on the problem to be solved. Segmentation is stopped when the region of interest in a specific application has been isolated. One of the most difficult tasks in this process is the autonomous segmentation method. This step determines the eventual success or failure of the image analysis in which organ visualization is a critical aspect [3, 4]. Today, medical imaging modalities generate high resolutions and large number of images that cannot be examined manually. This drives the development of more efficient and robust problem-tailored image analysis methods for medical imaging. Automated image segmentation could increase precision by eliminating the subjectivity of the clinician. It also saves tremendous time and effort by eliminating an exhaustive process, where the results are hardly repeatable [5]. The automatic spine segmentation process used to generate anatomically correct 3D models has challenges associated with it use. Some of these challenges are attributed to the anatomic complexity of the spine (33 vertebrae, 23 intervertebral disks, spinal cord, connecting ribs, etc.), image noise (all real-life data and CT images contain noise), low intensity (in spongy bones and softer bones), and the partial volume effect. Many methods have been proposed to alleviate these challenges in recent years [3, 6, 7, 8, 9]. Recent spine segmentation research can be categorized into two main approaches: free estimation methods and trainable methods. Free-estimation methods do not require an explicit model for segmentations and include the following: classical region growing, watershed, active contours, and graph-cut methods [3, 7]. However, the trainable methods have a central assumption that the structures of interest/organs have repetitive geometry. Therefore, we can utilizethe repetitive geometry into a probabilistic representation aimed toward explaining the variation in the shape of the organ and then when segmenting an image uses this information. 2 Kang et al. [10] utilized adaptive thresholding combined with region growing to conduct 3D bone segmentation. Mastmeyer et al. [11] used a region growing method that was capable of detecting the disks between the vertebrae. Sambucetti et al. [12] proposed a 2D active contour segmentation on a per-slice basis to construct 3D bone volume. All of these methods require expert human intervention and the manual adjustment of the parameter settings at several distinct steps. Several automatic methods have been proposed for vertebral column segmentation from CT images [13, 14, 15, 16, 17]. Most of these methods consist of two steps: identification of the spine and separate individual segmentation of the spine vertebrae. Yao et al. [13] used watershed segmentation and directed graph search methods to locate the vertebral body surfaces. The method was performed for several datasets and leakages occurred in 14 cases. Furthermore, identification and segmentation of the vertebrae were not carried out. Klinder et al. [14] used a 3D deformable model through mesh adaptation. The disadvantages of this method lie in its dependency on tremendous parameter setup. Recent advances in medical image segmentation techniques employ machine learning techniques to increase segmentation accuracy and gradually reduce human intervention. Huang et al. [15] constructed vertebrae detectors by using Adaboost. Ma and Liu [16] used low-level edge descriptors for vertebrae detection. Glocker et al. [17] detected vertebrae shapes and labeled them using a model trained with supervised classification forest; however, this method required selecting an appropriate feature and relying on a priori knowledge of the spine shape. Therefore, this method is less applicable for general and varying image data. In this study, we propose the utilization of class redundancy combined with an improved hybrid of the convolutional neural network (CNN) and fully convolutional network (FCN) methods to overcome the drawbacks of previous methods and provide a practical solution. We present a fully automatic approach for spine segmentation from CT based on a hybrid method of CNN and FCN with the following main contributions: 1. We propose an efficient hybrid training scheme by utilizing a mask on sampled image segments and analyze its behavior in adapting to the class imbalance of the segmentation problem at hand. 2. We demonstrate the capabilities of our system using class redundancy as soft constraints that greatly improves segmentation results. The efficiency of the proposed method will be demonstrated through the experimental results. This study is organized as follows. In section 2, we introduce histogrambased segmentation. Next, we explain the CNN and describe the proposed method in detail. Section 3 presents the experimental results and analysis. We evaluate the method on a database of 32 cases and compare the results with other automatic methods in order to draw reliable conclusions. Section 4 concludes the study and discusses results and future work. 3 2. Methodology 2.1. Histogram-based level set segmentation Let the given CT data be represented as the following function, f : Ω × {1, ....., m} ⊂ R2 × N → R, where m is the number of sliced images in a volume data. Before segmenting, the CT data are preprocessed. This includes morphological image processing, region-based image processing, and contrast adjustment. After data preprocessing, histogram-based multiphase segmentation is performed. This method was first introduced by [18] and has been used in many previous works [19, 20]. The segmentation model in [18] deals with a 1D histogram of the preprocessed data, and uses adaptive global maximum clustering (AGMC) to automatically obtain the number of significant local maxima of the histogram using 2-means clustering. This procedure provides n number of distinct regions and their corresponding subdomains. Then, a region-based segmentation is performed using a level-set function to label the different regions in the data. Once n-labeled data are obtained, the region of interest can be chosen differently by assigning a different label in accordance with the property of the desirable objects. Then, a detailed segmentation of the desired region is carried out using an active contour model, as suggested in [19]. This model performs a segmentation based on a variable-weighted combination of local and global intensity. This will enable it to divide an object surrounded by both and weak boundaries and to distinguish very adjacent objects with those boundaries. This method generally yields a good result, but it has several shortcomings. First, it is tedious. The original CT data have to pass through various segmentation phases in order to obtain the final desired result. Second, setting the optimal value of the parameters used in the segmentation models ([18, 19]) is not an easy task. Third, it is dependent on the specific CT dataset. Different spine CT datasets will have different numbers of distinct regions, based on the histogram-based multiphase segmentation. Hence, the manual selection of labels is necessary for different datasets. Some parameters are also different for different CT datasets. As a result, a segmentation model that would potentially eliminate the downsides of the current method is necessary. 2.2. Convolutional neural networks A CNN [21] is a type of feed-forward network that utilizes grid-like topology to analyze data. Through the use of local receptive fields, weight sharing and subsampling mechanisms, CNNs have proved themselves to be successful in various supervised tasks, such as image classification, object recognition [22], and image segmentation [23, 24]. In such tasks, models using CNNs are trained using an image dataset that is associated with a certain class label. In this regard, to perform spine segmentation using these networks, we first have to transform the CT data into an image dataset that can be analyzed by the networks. 4 (a) Training frame (b) Ground truth Figure 1: When preparing the dataset, we take a pixel inside the box of a given training frame and we will form a patch size of n × n around it. If the same pixel is part of the spine in the ground truth, then we label the patch as a spine (label 1). Otherwise, it will be labelled as background (label 0). We repeat the same procedure for all pixels in the box by taking a sliding interval of s. The test data is also prepared in a similar manner. This is called 2D patching. 2.2.1. Preparing training and testing data Because spine CT data are volumetric data, they are processed frame by frame. However, each frame of the CT data is not necessary for training the network, as the structure of the spine is symmetric across some interval of frames (the inter-vertebral disk). In order to prepare a label for the training data, we first segmented the training frames using the histogram-based level set segmentation method discussed in section 2.1. This can also be performed manually by using software programs such as 3D slicer. Once the ground truth was obtained, training data were prepared using 2D patching. Because the task involves only segmenting the spine that is located in a certain area of a frame, there is no need to process the entire frame. For computational simplicity, we formed a rectangular box around the area of the spine in each training frame. Note that the box size should be big enough to tolerate the spatial variation of the spine structure across the frames of the CT data. In 2D patching, we take a pixel inside the rectangular box of a given frame and form an n × n sized patch around it. By juxtaposition, if the same pixel in the ground truth is a part of the spine, then the patch will be labeled as 1 (class 1). Otherwise, it will be labeled as 0 (class 2). By taking a certain stride size of s, we repeated the same procedure across all pixels inside the box to prepare the training images (patches). Padding was not required to form the patches of the boundary pixels in the box because the respective distances between the edges of a given frame and the boundaries of the box were longer than n/2. Now, we have two classes of training data (0 and 1), and each class contains training 5 patches obtained from the training frames. Testing data were also prepared in the same manner. We used the same rectangular box in the testing frames and then prepared the test images (patches) by using a 2D patching method. During testing, the model is expected to classify if a certain patch is part of the spine or not. Based on the result, we reconstruct the ground truth of each frame of the spine; in other words, we segment the spine (Fig. 1). In the above scenario, the segmentation of the spine is treated as an image classification task with two classes. The results of this method are shown in the experimental section. Although the segmentation result was encouraging, it was not very accurate, particularly at the boundaries of the spine, as shown in the qualitative evaluation in section 3.3. In our work, we found out that using only two classes has certain downsides. First, there is an imbalance in the number of training images in the two classes prepared from the 2D patching. Because the spatial area of the spine in Figure 2: A testing frame a given frame is small compared to the area of the containing a rib. background, most of the training patches are labeled as 0. Hence, the model eventually learns more about the background than the spine itself. The other problem is related to the pixel intensity. During testing, if the frame contains some parts of the rib (see Fig. 2), the model will most likely classify it as part of the spine (label 1) because the pixel intensity of the two is very similar in the CT data. Moreover, because we only have two classes, if the trained model is uncertain about the class a certain patch belongs to, there is a 50% chance it will commit an error. 2.2.2. Proposed method: redundant generation of class labels In order to address the above problems, we propose a new way of preparing the training data. The proposed method involves generating redundant class labels by masking the spine structure in each training frame. The ground truth of the training frames in Fig. 1 has a pixel value of 1 for the spine and a pixel value of 0 for the background. We generate the first redundant class by masking the area of the spine with a different pixel value (2, for instance). This class is important for training the model because it is used as a mechanism for punishing the model to accurately distinguish between the spine and its surrounding environment. In a similar manner, we generate more classes by continuing the masking with different pixel values, each associated with different class label (see Fig. 3). This enables us to obtain a proportional area of different classes in the ground truth. In general, the proposed method has the following advantages: • A proportional amount of training patches can be prepared for each class. • The model properly learns to segment the boundaries of the spine within a given frame because the masking is applied along the spine boundary. 6 Class 1 Class 2 Masking Class 3 Class 4 Figure 3: Masking. We generate redundant classes by masking the spine in the ground truth of a given training frame. In this work, we used four classes for classification. Class 4 represents the spine. Class 1 represents the background. The other two classes are redundant. The first redundant class (class 3) was generated by masking the ground truth with a pixel value of 2. The second masking (class 2) was done with a pixel value of 3. The masking thickness was chosen so as to balance the data in each distinct class. • Anything inside the rectangular box that is not part of the spine will now have a specific label. For example, if the testing frame contains parts of the rib, the model will be able to identify those parts by the labels. This will avoid segmenting the rib along with the spine, if trained properly. • The probability that the model makes errors will be reduced because we will have more classes. 2.2.3. Model architecture In this work, we used the architecture shown in Fig. 4. We implemented a simple neural architecture (2 CNN layers and 3 Fully Connected layers) compared to other deep networks used for medical image segmentation such as U-net (23 CNN layers) [24]. Here is how our network works. • The input image (patch) has a size of 32×32. The first 2D convolution has a local receptive field size of 5 × 5 and it outputs a convolutional feature map with 32 layers (neurons). A zero-padding is used for boundary pixels while doing the convolution, hence, each layer in the feature map has the same size as the input image. A rectified linear unit (ReLU) activation layer that computes element wise non-linearity follows this layer. Later, a max pooling layer of size 2 × 2 subsamples the spatial size of the input patch. The output of this stage has a size of 16 × 16 × 32. • The second 2D convolution also has a local receptive field size of 5 × 5. However, it outputs a convolutional feature map of 64 layers (neurons). This layer is followed by a ReLU activation layer and a max pooling layer of size 2x2. The output of this stage has a size of 8 × 8 × 64. 7 Figure 4: Convolutional network architecture • The output of the second stage is flattened and fed into a fully connected layer with 1024 neurons. This layer has a ReLU activation layer. Following this layer, a dropout is applied. Every time before the input is presented, the neural network drops out different set of neurons with probability of 0.5. This reduces the network’s dependency on the presence of particular neurons. A second fully connected layer with 2048 neurons uses the output of the previous stage as an input. This layer also has a ReLU activation layer. A dropout is also implemented in a similar manner as the previous layer. • The final layer is a softmax layer with 4 neurons. This layer outputs a probabilistic prediction array for the 4 classes we used. Note that, even though we treated spine segmentation as a 4-class classification problem, only one class is important for labeling the spine. The other three classes are irrelevant (redundant). 3. Experimental design and results There are several factors that affected the diversity of the medical image data, such as the imaging modalities, the type of CT-scan machine used, the radiation dose, the scan time, and the patients. In order to prove the robustness of our method, we tested our algorithm for a diverse array of patient datasets. 8 3.1. Data sources We tested our method on several public datasets that were obtained from the Spineweb website and Gangnam Severance Hospital. The datasets were CT scan data with 512 × 512 pixel images for 32 patients. To validate the segmentation result, we used the ground truth from Gangnam Severance Hospital. The dataset contained 512 × 512 pixel CT images with 0.3125 mm thicknesses, which act as the gold standard in this project. The reason we used data from Spineweb is to ensure there is diversity in the real clinical data with respect to patients and the CT-Scan imaging machine. 3.2. Parameter selection and training computational time The patch size was set to 32 × 32 with a sliding interval of 1 pixel. Our network was implemented in Python and Tensorflow. All experiments were performed on a PC (GeForce GTX 1080 Ti. and 32 GB RAM). The training stage is time-consuming for CPU implementation. We compared our network with U-Net [24]. U-Net took approximately 25 h for training and approximately 6.5 min (385 s) to process one image. However, these stages were completed faster in our network compared to U-Net where training the network took approximately 13 h and segmentation processing time took approximately 3 min (170 s) for each image. The training process was performed on patches that were obtained from the images (frames in the CT data). To obtain a large training dataset, patches were used instead of the entire image to train the network. In addition, patches can represent the local structures at a higher quality level. A corresponding label and 1.388.800 patches of spine data were selected randomly from the training dataset. To ensure fairness of the analysis results, the testing dataset that we used was different from the dataset that was used in the training set. The fairness in here is necessary in order to show that even if the system had never seen the input data, it can produce good segmentation results based on previous training data. A number of measurement metrics used in the evaluation of segmentation were computed for each segmentation result [25, 26]. The metrics chosen for quantitative analysis were divided into similarity, classic, and distance measurements. The similarity metrics included the Dice coefficient (DC), Jaccard index, and volumetric similarity (VS). The classic measurements used sensitivity, specificity, precision (over segmentation (OS) and under segmentation (US)) and accuracy. The distance measurement used the mean surface distance (MSD), Hausdorff distance (HD), and the global consistency error (GCE). We also used the Matthews correlation coefficient (MCC). In this work , we prepared the training and testing dataset using the method discussed in section 2.2.1. Then we used a simple CNN model discussed in section 2.2.2 to do the spine segmentation. We also used the prepared data with a rather deep neural network (U-net) [24] for comparison purpose. In addition, two different methods (level-set and CNN with two classes) were compared against our method. 9 (a) Our method (b) 2-Class (c) U-net (d) Level set (e) Our method (f) 2-Class (g) U-net (h) Level set (i) Our method (j) 2-Class (k) U-net (l) Level set (m) Our method (n) 2-Class (o) U-net (p) Level set Figure 5: Qualitative comparison between our method, 2-Class CNN, Level set and U-net model. 10 3.3. Experimental results 3.3.1. Qualitative evaluation We selected a few representative slices from the results of the testing set. Fig. 5 presents the results obtained using the different methods. The results demonstrate that the proposed method segments more accurately than the other methods. 3.3.2. Quantitative evaluation The manually labeled images of each subject are used as gold standards and the results of each segmentation method are converted into binary images with the same voxel resolution and image dimensions as the query image. In the following description of the measures, the segmentation result is indicated by S and the gold standard by GT. (A) Similarity metrics (i) Dice coefficient (DC) DC measures the extent of the spatial overlap between two binary images. DC values range between 0 (no overlap) and 1 (perfect agreement). In this study, the DC values are obtained using Equation (1). DC = 2 |S ∩ GT | |S| + |GT | (1) (ii) The Jaccard index The Jaccard index or Jaccard coefficient (J) is used to measure the spatial overlap of the intersection divided by the size of the union of two label sets. It can be expressed as shown in Equation (2), and can be obtained from the Dice measure by Equation (3). J= |S ∩ GT | |S ∪ GT | (2) J= DC (2 − DC) (3) (iii) Volumetric Similarity (VS) VS is defined as the absolute volume difference divided by the sum of the compared volumes that are obtained by Equation (4). VS =1−| |S| − |GT | | |S| + |GT | (4) As demonstrated by Table 1, our method shows a greater improvement in the segmentation results compared to other methods. Based on the DC score, we obtain 94%97% similarity with the ground truth. This result is also supported by the improvement in the volumetric similarity, which is 96% 98%. 11 Table 1: Similarity measurements of the four different segmentation algorithms Method Our Method 2-Class CNN U-Net Level Set Mean 0.942861 0.869014 0.959566 0.802039 DC SD 0.032463 0.048805 0.000408 0.068426 Jaccard Mean 0.933569 0.771535 0.956901 0.869837 Index SD 0.057144 0.077090 0.000164 0.012317 Volumetric Mean 0.967046 0.926983 0.973044 0.909029 (B) Classic measurements We utilize the confusion matrix to perform classic measurements by utilizing four variables: true positive (TP), false positive (FP), true negative (TN), and false negative (FN). • TP: pixels correctly segmented as the spine in the ground truth and algorithm • FP: pixels not classified as the spine in the ground truth, but are classified as the spine by algorithm (falsely segmented) • TN: pixels not classified as the spine in the ground truth and by algorithm (correctly detected as background) • FN: pixels classified as the spine in the ground truth, but not classified as the spine by the algorithm (falsely detected as background) (i) Sensitivity Sensitivity measures the portion of positive pixels in the ground truth that are also identified as positive by the algorithm being evaluated. It is used to check algorithm sensitivity for detecting proper spine pixels. Sensitivity can be obtained by Equation (5) Sensitivity = TP TP + FN (5) (ii) Specificity Specificity measures the portion of negative pixels in the ground truth that are also identified as negative by the algorithm being evaluated. It checks how sensitive the algorithm is to the detection of correct background pixels. This metric is obtained by Equation (6) Specif icity = TN TN + FP (6) (iii) Over segmentation (OS) and under segmentation (US) To further characterize the segmentation result, OS and US are obtained using Equations (7) and (8).. GT − and S − are the compliments of the gold standard and the segmentation results respectively. OS = 2 S ∩ GT − × 100 |S| + |GT | 12 (7) Similarity SD 0.026951 0.047280 0.002919 0.089093 Table 2: Classic measurements of the four different segmentation algorithms Method Our Method 2-Class CNN U-Net Level Set Sensitivity Mean SD 0.972481 0.026948 0.913395 0.077838 0.972806 0.003227 0.779723 0.123521 Specificity Mean SD 0.991616 0.006239 0.973564 0.019650 0.998899 0.000166 0.987453 0.009751 Over Segmentation Mean SD 8.349675 5.186457 16.26140 8.085861 1.356735 0.248874 15.15807 9.973117 Method Our Method 2-Class CNN U-Net Level Set Under Segmentation Mean SD 2.643167 2.667397 8.944368 9.749133 2.758809 0.343256 27.792356 22.945279 Accuracy Mean SD 99.01979 0.575767 96.76865 1.907359 99.69015 0.016499 96. 98377 1.231796 MCC Mean SD 0.938438 0.033661 0.852528 0.047596 0.977921 0.000479 0.792400 0.066891 2 S − ∩ GT × 100 |S| + |GT | US = (8) (iv) Accuracy Accuracy is defined by Equation (9). Accuracy = TP + TN TP + TN + FP + FN (9) (v) Matthew correlations coefficient (MCC) MCC, introduced by B.W. Matthews, gives a summary of the performance of the segmentation algorithm. The MCC analyzes the segmentation result and the ground truth as two sets and takes into account TP and FN to compute a correlation coefficient that ranges between -1 (complete disagreement) and 1 (complete agreement). A value of zero shows that the segmentation was not correlated with the ground truth. The MCC can be defined as shown in Equation (10) M CC = p TP × TN − FP × FN (T P + F N ) × (T P + F P ) × (T N + F P ) × (T N + F N ) (10) Our method shows the greatest improvement in the sensitivity and specificity categories. Sensitivity is 97%, specificity is 99%, and the accuracy is 99%, which is similar to the U-Net method. The results also showed a significant improvement in the OS and US categories compared to the classic CNN method. These results are also supported by the improvement in the MCC, which showed that our segmentation results achieved very good agreement with the ground truth. (C) Distance measurements 13 Table 3: Distance measurements of the four different segmentation algorithms Method Our Method 2-Class CNN U-Net Level Set MSD Mean SD 0.167692 0.229274 0.784873 0.437554 0.099893 0.005112 0.342764 0.352889 Mean 7.429119 9.502033 6.289700 8.788802 HD SD 1.635649 1.803488 0.491765 4.781518 GCE Mean SD 0.020012 0.010673 0.064253 0.039871 0.030092 0.001779 0.054126 0.021791 (i) Mean surface distance (MSD) MSD is the mean of the absolute values of the surface distance for all ns surface voxels. This value is obtained by Equation (11). M SDsg = ns 1 X dsg ns i=1 l (11) This metrics attempts to estimate the error between the surfaces of S and GT using distances between their surface voxels. We define the surface distance of the ith surface voxel on S, dsg l as the distance to the closest voxel on GT. The surface distance values were calculated as described in Gerig et al. [27]. In the distance transform image, the value of each voxel is the Euclidean distance in millimeters to the nearest surface voxel. The values on the surface voxels are 0. (ii) Hausdorff distance (HD) HD measures the distance between the ground truth surface and the segmented surface. To compute the HD, we use the same surface models that were generated to compute the MSD. A smaller HD indicates better segmentation accuracy. This metric can be defined by Equation (12) HD = max(h(S, GT ), h(GT, S)) (12) where, h(S, GT ) = max min ka − bk a∈S b∈GT (iii) Global consistency Error (GCE) GCE is defined as an error measurement between two segmentations (the error averaged over all voxels) and is given by Equation (13)   n n X  X 1 GCE(S, GT ) = min E(S, GT, xi ), E(GT, S, xi ) (13)   N i i where, N is the total voxel, n is the set difference and the error E at |(S,x)n(GT,x)| voxel x is defined as E(S, GT, x) = (S,x) Based on the distance metrics, our methods obtained better results than the classic CNN and the level set methods. We showed improvement and 14 achieved good results based on the GCE and both distance metrics. The GCE distribution of the proposed method is close to zero, which indicates a very low error segmentation result. 4. Conclusion and discussion In this study, a new approach for medical image segmentation was proposed. This approach uses class redundancy as a soft constraint in the CNN architecture. The proposed method achieved a competitive result compared to several widely used medical image segmentation methods. The proposed algorithm was tested on real medical data and evaluated through similarity metrics such as the Dice coefficient, Jaccard index, volumetric similarity, sensitivity, specificity, precision (over segmentation and under segmentation), accuracy, Matthews correlation coefficient, mean surface distance, Hausdorff distance, and global consistency error. These similarity metrics generated results of 94%, 93%, 96%, 97%, 99%, 8.3, 2.6, 99%, 0.93, 0.16, 7.4, and 0.02, respectively. These results demonstrate the effectiveness of our method. In conclusion, the main contribution of this study is the presentation of a new approach to prepare data and its use in a simple CNN model to improve the accuracy of segmentation results. The experimental results quantitatively and qualitatively showed that our proposed method improves accuracy, corrects error segmentation, and exhibits better segmentation performance compared to other conventional methods. The results also showed high specificity and sensitivity, and high overlap and low distance between the manual annotation and the proposed method. In future, we intend to assess our method using a deeper network and broad training data, such as low contrast data, to improve the performance. The hope is to eventually be capable of handling a broader range of medical data. Another possible direction is to investigate improvements to computation time during the training stage through optimization. Currently, for 1.388.800 patches of spine data, it takes 13 h to train using our CPU, and processing time takes 3 min for each image. 5. Acknowledgment This research was supported by WC300 Project (2E26210) funded by the Ministry of SMEs and Startups (MSS, Korea) [Project NO. S2482672]. References References [1] G. Kumar, Analysis of medical image processing and its applications in healthcare industry, International Journal of Computer Technology & Applications 5 (3) (2014) 851–860. 15 [2] M. Shoham, I. Lieberman, E. Benzel, E. Zehavi, B. Zilberstein, M. Roffman, A. Bruskin, A. Fridlander, L. Joskowicz, S. Brink-Danan, N. Knoller, Robotic assisted spinal surgery: from concept to clinical practice, Computer Aided Surgery, 12 (2007) 105–115. [3] D. L. Pham, C. Xu, J. Prince, Current methods in medical image segmentation, Annual Review of Biomedical Engineering, 2 (2000) 315–337. [4] Elsevier, I. Bankman, Handbook of medical image processing and analysis, 2nd ed. (2009) 71–257. [5] G. Harris, N. C. Andreasen, J. M. B. T. Cizadlo, H. J. Bockholt, V. A. Magnotta, S. Arndt, Improving tissue classification in MRI: A threedimensional multispectral discriminant analysis method with automated training class selection, Journal of Computer Assisted Tomography, 23 (1) (1999) 144–154. [6] T. Klinder, J. Ostermann, M. Ehm, A. Franz, R. Kneser, C. Lorenz, Automated model-based vertebra detection identification and segmentation in ct images, Medical Image Analysis, 13 (3) (2009) 471–482. [7] M. Krcah, G. Szekely, R. Blanc, Fully automatic and fast segmentation of the femur bone from 3D–CT images with no shape prior, Proceedings of IEEE International Symposium of Biomedical Imaging, (2011) 2087–2090. [8] M. Hardisty, L. Gordon, P. Agarwal, T. Skrinskas, C. Whyne, Quantitative characterization of metastatic disease in the spine. Part I. Semiautomated segmentation using atlas-based deformable registration and the level set method, Medical Physics, 34 (8) (2007) 3127–3134. [9] L. Gordon, M. Hardisty, T. Skrinskas, F. Wu, C. Whyne, Automated atlasbased 3D segmentation of the metastatic spine, Journal of Bone & Joint Surgery, 90 (2008) 128. [10] Y. Kang, K. Engelke, W. A. Kalender, A new accurate and precise 3D segmentation method for skeletal structures in volumetric CT data, IEEE Transactions on Medical Imaging, 22 (2003) 586–598. [11] A. Mastmeyer, K. Engelke, C. Fuchs, W. A. Kalender, A hierarchical 3D segmentation method and the definition of vertebral body coordinate systems for QCT of the lumbar spine, Medical Image Analysis, 10 (2006) 560–577. [12] G. Sambuceti, M. Brignone, a. M. M. Marini, C., F. Fiz, S. Morbelli, A. Buschiazzo, C. Campi, R. Piva, A. Massone, M. Piana, F. Frassoni, Estimating the whole bone-marrow asset in humans by a computational approach to integrated PET/CT imaging, European Journal of Nuclear Medicine, Molecular Imaging, 39 (8) (2012) 1326–1338. 16 [13] J. Yao, S. OConnor, R. Summers, Automated spinal column extraction and partitioning, Proceedings of IEEE International Symposium of Biomedical Imaging, (2006) 390–393. [14] T. Klinder, J. Ostermann, M. Ehm, A. Franz, R. Kneser, C. Lorenz, Automated model-based vertebra detection, identification, and segmentation in CT images, Medical Image Analysis, 13 (3) (2009) 471–482. [15] S. Huang, Y. Chu, S. Lai, C. Novak, Learning-based vertebra detection and iterative normalized-cut segmentation for spinal MRI, IEEE Transactions on Medical Imaging, 28 (10) (2009) 1596–1605. [16] J. Ma, L. Lu, Hierarchical segmentation and identification of thoracic vertebra using learning-based edge detection and coarsetofine deformable model, Computer Vision and Image Understanding, 117 (9) (2013) 1072–1083. [17] B. Glocker, D. Zikic, E. Konukoglu, D. Haynor, A. Criminisi, Vertebrae localization in pathological spine CT via dense classification from sparse annotations, Proceedings of International Conference on Medical Computing and Computer Assisted Intervention, (2013) 262–270. [18] S. Kim, M. Kang, Multiple-region segmentation without supervision by adaptive global maximum clustering, IEEE Transactions on Image Processing, 21 (4) (2012) 1600–1612. [19] S. Kim, Y. Kim, S. Park, D. Lee, Automatic segmentation of leg bones by using active contours, Conf Proc IEEE Eng Med Biol Soc., (2014) 4695– 4698. [20] M. Vania, S. Kim, D. Lee, Automatic multisegmentation of abdominal organs by level set with weighted global and local forces, Student Conference (ISC), 2016 IEEE EMBS International. [21] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, L. D. Jackel, Backpropagation applied to handwritten zip code recognition, Neural Computation, 1 (4) (1989) 541–551. [22] A. Krizhevsky, S. Ilya, G. E. Hinton, Imagenet classification with deep convolutional neural networks, Advances in Neural Information Processing Systems 25 (2012) 1097–1105. [23] J. Long, E. Shelhamer, T. Darrell, Fully convolutional networks for semantic segmentation, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, (2015) 3431–3440. [24] O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image segmentation, Medical Image Computing and ComputerAssisted Intervention MICCAI 2015, (2015) 234–241. 17 [25] A. A. Taha, A. Hanbury, Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool. bmc medical imaging, BMC Medical Imaging (2015) 15–29. [26] J. Udupa, V. Leblanc, Y. Zhuge, C. Imielinska, H. Schmidt, L. Currie, B. Hirsch, J. Woodburn, A framework for evaluating image segmentation algorithms, Computerized Medical Imaging and Graphics, 30 (2) (2006) 75–87. [27] G. Gerig, M. Jomier, M. Chakos, Valmet: a new validation tool for assessing and improving 3d object segmentation, Proceedings of the 4th International Conference on Medical Image Computing and Computer-Assisted Intervention (2001) 516–523. [28] S. Lashari, R. A. Ibrahim, A framework for medical images classification using soft set, Procedia Technology, 11 (2013) 548–556. [29] D. Yang, J. Zheng, A. Nofal, J. Deasy, I. M. E. Naqa, Techniques and software tool for 3D multimodality medical image segmentation, Journal of Radiation Oncology Informatics, 1 (1). [30] I. Goodfellow, Y. Bengio, A. Courville, Deep learning, MIT press (2016) 321–363. 18
1cs.CV
Experience Report: Developing the Servo Web Browser Engine using Rust arXiv:1505.07383v1 [cs.PL] 26 May 2015 Brian Anderson Lars Bergstrom David Herman Josh Matthews Keegan McAllister Jack Moffitt Simon Sapin Manish Goregaokar Indian Institute of Technology Bombay [email protected] Mozilla Research {banderson,larsberg,dherman,jdm, kmcallister,jack,ssapin}@mozilla.com Abstract All modern web browsers — Internet Explorer, Firefox, Chrome, Opera, and Safari — have a core rendering engine written in C++. This language choice was made because it affords the systems programmer complete control of the underlying hardware features and memory in use, and it provides a transparent compilation model. Servo is a project started at Mozilla Research to build a new web browser engine that preserves the capabilities of these other browser engines but also both takes advantage of the recent trends in parallel hardware and is more memory-safe. We use a new language, Rust, that provides us a similar level of control of the underlying system to C++ but which builds on many concepts familiar to the functional programming community, forming a novelty — a useful, safe systems programming language. In this paper, we show how a language with an affine type system, regions, and many syntactic features familiar to functional language programmers can be successfully used to build state-ofthe-art systems software. We also outline several pitfalls encountered along the way and describe some potential areas for future research. 1. Introduction The heart of a modern web browser is the browser engine, which is the code responsible for loading, processing, evaluating, and rendering web content. There are three major browser engine families: 1. Trident/Spartan, the engine in Internet Explorer [IE] 2. Webkit[WEB]/Blink, the engine Chrome [CHR], and Opera [OPE] in Safari [SAF], 3. Gecko, the engine in Firefox [FIR] All of these engines have at their core many millions of lines of C++ code. While the use of C++ has enabled all of these browsers to achieve excellent sequential performance on a single web page, on mobile devices with lower processor speed but many more processors, these browsers do not provide the same level of interactivity that they do on desktop processors [MTK+ 12, CFMO+ 13]. Further, in an informal inspection of the critical security bugs in Gecko, we determined that roughly 50% of the bugs are use after free, out of range access, or related to integer overflow. The other 50% are split between errors in tracing values from the JavaScript heap in the C++ code and errors related to dynamically compiled code. Servo [SER] is a new web browser engine designed to address the major environment and architectural changes over the last decade. The goal of the Servo project is to produce a browser that enables new applications to be authored against the web platform that run with more safety, better performance, and better power usage than in current browsers. To address memory-related safety issues, we are using a new systems programming language, Rust [RUS]. For parallelism and power, we scale across a wide variety of hardware by building either data- or task-parallelism, as appropriate, into each part of the web platform. Additionally, we are improving concurrency by reducing the simultaneous access to data structures and using a message-passing architecture between components such as the JavaScript engine and the rendering engine that paints graphics to the screen. Servo is currently over 400k lines of Rust code and implements enough of the web platform to render and process many pages, though it is still a far cry from the over 7 million lines of code in the Mozilla Firefox browser and its associated libraries. However, we believe that we have implemented enough of the web platform to provide an early report on the successes, failures, and open problems remaining in Servo, from the point of view of programming languages and runtime research. In this experience report, we discuss the design and architecture of a modern web browser engine, show how modern programming language techniques — many of which originated in the functional programming community — address these design constraints, and also touch on ongoing challenges and areas of research where we would welcome additional community input. 2. Browsers Modern web browsers do not just load static pages, but can also handle pages that have similar complexity to native applications. From application suites such as the Google Apps1 to games based [Copyright notice will appear here once ’preprint’ option is removed.] 1 https://apps.google.com HTML CSS JS Parsing DOM Styling Flow Tree Display Lists Layout Rendering Layers Compositing Final Output Script Script Figure 1. Processing stages and intermediate representations in a browser engine. on the Unreal Engine,2 modern browsers have the ability to handle much more than simple static pages. Figure 1 shows the steps in processing a site. While the naming is specific to the Servo browser, similar steps are used in all modern browsers.3 2.1 2.3 Parsing HTML and CSS A URL identifies a resource to load. This resource usually consists of HTML, which is then parsed and typically turned into a Document Object Model (DOM) tree. From a programming languages standpoint, there are several interesting aspects of the parser design for HTML. First, though the specification allows the browser to abort on a parse error,4 in practice browsers follow the recovery algorithms described in that specification precisely so that even illformed HTML will be handled in an interoperable way across all browsers. Second, due to the presence of the <script> tag, the token stream can be modified during operation. For example, the below example that injects an open tag for the header and comment blocks works in all modern browsers. 2.4 Rendering Once all of the elements to appear on screen have been computed, these elements are rendered, or painted, into memory buffers or directly to graphics surfaces. 2.5 Compositing The set of memory buffers or graphical surfaces, called layers, are then transformed and composited together to form a final image for presentation. Layerization is used to optimize interactive transformations like scrolling and certain animations. 2.6 <html> <script> document.write("<h"); </script>1> This is a h1 title Scripting Whether through timers, <script> blocks in the HTML, user interactions, or other event handlers, JavaScript code may execute at any point during parsing, layout, and painting or afterwards during display. These scripts can modify the DOM tree, which may require rerunning the layout and painting passes in order to update the output. Most modern browsers use some form of dirty bit marking to attempt to minimize the recalculations during this process. <script> document.write("<!-"); </script>This is commented --> </html> 3. Rust Rust is a statically typed systems programming language most heavily inspired by the C and ML families of languages [RUS]. Like the C family of languages, it provides the developer fine control over memory layout and predictable performance. Unlike C programs, Rust programs are memory safe by default, only allowing unsafe operations in specially-delineated blocks. Rust features an ownership-based type system inspired by the region systems work in the MLKit project [TB98] and especially as implemented in the Cyclone language [GMJ+ 02]. Unlike the related ownership system in Singularity OS [HLA+ 05], Rust allows programs to not only transfer ownership but also to temporarily borrow owned values, significantly reducing the number of region and ownership annotations required in large programs. The ownership model encourages immutability by default while allowing for controlled mutation of owned or uniquely-borrowed values. Complete documentation and a language reference for Rust are available at: http://doc.rust-lang.org/. This requires parsing to pause until JavaScript code has run to completion. But, since resource loading is such a large factor in the latency of loading many webpages (particularly on mobile), all modern parsers also perform speculative token stream scanning and prefetch of resources likely to be required [WLZC11]. 2.2 Layout The flow tree is then processed to produce a set of display list items. These list items are the actual graphical elements, text runs, etc. in their final on-screen positions. The order in which these elements are displayed is well-defined by the standard.5 Styling After constructing the DOM, the browser uses the styling information in linked CSS files and in the HTML to compute a styled tree of flows and fragments, called the flow tree in Servo. This process can create many more flows than previously existed in the DOM — for example, when a list item is styled to have an associated counter glyph. 2 https://www.unrealengine.com/ 3 http://www.html5rocks.com/en/tutorials/internals/ howbrowserswork/ 4 https://html.spec.whatwg.org/multipage/#parsing 5 http://www.w3.org/TR/CSS21/zindex.html# painting-order 2 2015/5/28 Site Reddit CNN fn main() { // An owned pointer to a heap-allocated // integer let mut data = Box::new(0); Servo 1 thread 100 50 Servo 4 threads 55 35 Table 1. Performance of Servo against Mozilla’s Gecko rendering engine on the layout portion of some common sites. Times are in milliseconds, where lower numbers are better. Thread::spawn(move || { *data = *data + 1; }); // error: accessing moved value print!("{}", data); With relatively few simple rules, ownership in Rust enables foolproof task parallelism, but also data parallelism, by partitioning vectors and lending mutable references into properly scoped threads. Rust’s concurrency abstractions are entirely implemented in libraries, and though many advanced concurrent patterns such as work-stealing [ABP98] cannot be implemented in safe Rust, they can usually be encapsulated in a memory-safe interface. } Figure 2. Code that will not compile because it attempts to access mutable state from two threads. fn main() { // An immutable borrowed pointer to a // stack-allocated integer let data = &1; 4. Servo A crucial test of Servo is performance — Servo must be at least as fast as other browsers at similar tasks to succeed, even if it provides additional memory safety. Table 1 shows a preliminary comparison of the performance of the layout stage (described in Section 2) of rendering several web sites in Mozilla Firefox’s Gecko engine compared to Servo, taken on a modern MacBook Pro. In the remainder of this section, we cover specific areas of Servo’s design or implementation that make use of Rust and the impacts and limitations of these features. Thread::scoped(|| { println!("{}", data); }); print!("{}", data); } Figure 3. Safely reading immutable state from two threads. 4.1 Rust’s syntax Rust has struct and enum types (similar to Standard ML’s record types and datatypes [MTHM97]) as well as pattern matching. These types and associated language features provide two large benefits to Servo over traditional browsers written in C++. First, creating new abstractions and intermediate representations is syntactically easy, so there is very little pressure to tack additional fields into classes simply to avoid creating a large number of new header and implementation files. More importantly, pattern matching with static dispatch is typically faster than a virtual function call on a class hierarchy. Virtual functions can both have an inmemory storage cost associated with the virtual function tables (sometimes many thousands of bytes6 ) but more importantly incur indirect function call costs. All C++ browser implementations transform performance-critical code to either use the final specifier wherever possible or specialize the code in some other way to avoid this cost. Rust also attempted to stay close to familiar syntax, but did not require full fidelity or easy porting of programs from languages such as C++. This approach has worked well for Rust because it has prevented some of the complexity that arose in Cyclone [GMJ+ 02] with their attempts to build a safe language that required minimal porting effort for even complicated C code. fn main() { // A heap allocated integer protected by an // atomically-reference-counted mutex let data = Arc::new(Mutex::new(0)); let data2 = data.clone(); Thread::scoped(move || { *data2.lock().unwrap() = 1; }); print!("{}", *data.lock().unwrap()); } Figure 4. Safely mutating state from two threads. 3.1 Gecko 250 105 Ownership and concurrency Because the Rust type system provides very strong guarantees about memory aliasing, Rust code is memory safe even in concurrent and multithreaded environments, but beyond that Rust also ensures data-race freedom. In concurrent programs, the data operated on by distinct threads is also itself distinct: under Rust’s ownership model, data cannot be owned by two threads at the same time. For example, the code in Figure 2 generates a static error from the compiler because after the first thread is spawned, the ownership of data has been transferred into the closure associated with that thread and is no longer available in the original thread. On the other hand, the immutable value in Figure 3 can be borrowed and shared between multiple threads as long as those threads don’t outlive the scope of the data, and even mutable values can be shared as long as they are owned by a type that preserves the invariant that mutable memory is unaliased, as with the mutex in Figure 4. 4.2 Compilation strategy Many statically typed implementations of polymorphic languages such as Standard ML of New Jersey [AM91] and OC AML [Ler00] have used a compilation strategy that optimizes representations of data types when polymorphic code is monomorphically used, but defaults to a less efficient style otherwise, in order to share code [Ler90]. This strategy reduces code size, but leads to unpredictable performance and code, as changes to a codebase that either add a new instantiation of a polymorphic function at a given type or, 6 https://chromium.googlesource.com/chromium/blink/ +/c048c5c7c2578274d82faf96e9ebda4c55e428da 3 2015/5/28 Language C or C++ Rust HTML or JavaScript in a modular compilation setting, that expose a polymorphic function externally, can change the performance of code that is not local to the change being made. Monomorphization, as in MLton [CFJW], instead instantiates each polymorphic code block at each of the types it is applied against, providing predictable output code to developers at the cost of some code duplication. This strategy is used by virtually all C++ compilers to implement templates, so it is proven and well-known within systems programming. Rust also follows this approach, although it improves on the ergonomics of C++ templates by embedding serialized generic function ASTs within “compiled” binaries. Rust also chooses a fairly large default compilation unit size. A Rust crate is subject to whole-program compilation [Wee06], and is optimized as a unit. A crate may comprise hundreds of modules, which provide namespacing and abstraction. Module dependencies within a crate are allowed to be cyclic. The large compilation unit size slows down compilation and especially diminishes the ability to build code in parallel. However, it has enabled us to write Rust code that easily matches the sequential speed of its C++ analog, without requiring the Servo developers to become compiler experts. Servo contains about 600 modules within 20 crates. 4.3 Table 2. Lines of code in Servo against C++ code. In both cases, Servo uses C wrapper code to call into the code that Rust cannot directly reach. While this approach is not a large problem for varargs-style functions, it defeats many of the places where the C++ code has been crafted to ensure the code is inlined into the caller, resulting in degraded performance. We intend to fix this through cross-language inlining, taking advantage of the fact that both rustc and clang++ can produce output in the LLVM intermediate representation [LLV], which is subject to link-time optimization. We have demonstrated this capability at small scale, but have not yet deployed it within Servo. 4.5 Libraries and abstractions Many high-level languages provide abstractions over I/O, threading, parallelism, and concurrency. Rust provides functionality that addresses each of these concerns, but they are designed as thin wrappers over the underlying services, in order to provide a predictable, fast implementation that works across all platforms. Therefore, much like other modern browsers, Servo contains many of its own specialized implementations of library functions that are tuned for the specific cases of web browsers. For example, we have special “small” vectors that allow instantiation with a default inline size, as there are use cases where we create many thousands of vectors, nearly none of which have more than 4 elements. In that case, removing the extra pointer indirection — particularly if the values are of less than pointer size — can be a significant space savings. We also have our own work-stealing library that has been tuned to work on the DOM and flow trees during the process of styling and layout, as described in Section 2. It is our hope that this code might be useful to other projects as well, though it is fairly browser-specific today. Concurrency is available in Rust in the form of CML-style channels [Rep91], but with a separation between the reader and writer ends of the channel. This separation allows Rust to enforce a multiple-writer, single-reader constraint, both simplifying and improving the performance of the implementation over one that supports multiple readers. We have structured the entire Servo browser engine as a series of threads that communicate over channels, avoiding unsafe explicitly shared global memory for all but a single case (reading properties in the flow tree from script, an operation whose performance is crucially tested in many browser benchmarks). The major challenge we have encountered with this approach is the same one we have heard from other designers of large concurrent systems — reasoning about whether protocols make progress or threads eventually terminate is manual and quite challenging, particularly in the presence of arbitrary thread failures. Memory management As described in Section 3, Rust has an affine type system that ensures every value is used at most once. One result of this fact is that in the more than two years since Servo has been under development, we have encountered zero use-after-free memory bugs in safe Rust code. Given that these bugs make up such a large portion of the security vulnerabilities in modern browsers, we believe that even the additional work required to get Rust code to pass the type checker initially is justified. One area for future improvement is related to allocations that are not owned by Rust itself. Today, we simply wrap raw C pointers in unsafe blocks when we need to use a custom memory allocator or interoperate with the SpiderMonkey JavaScript engine from Gecko. We have implemented wrapper types and compiler plugins that restrict incorrect uses of these foreign values, but they are still a source of bugs and one of our largest areas of unsafe code. Additionally, Rust’s ownership model assumes that there is a single owner for each piece of data. However, many data structures do not follow that model, in order to provide multiple traversal APIs without favoring the performance of one over the other. For example, a doubly-linked list contains a back pointer to the previous element to aid in traversals in the opposite direction. Many optimized hashtable implementations also have both hash-based access to items and a linked list of all of the keys or values. In Servo, we have had to use unsafe code to implement data structures with this form, though we are typically able to provide a safe interface to users. 4.4 Lines of Code 1,678,427 410,817 217,827 Language interoperability Rust has nearly complete interoperability with C code, both exposing code to and using code from C programs. This support has allowed us to smoothly integrate with many browser libraries, which has been critical for bootstrapping a browser without rewriting all of the lower-level libraries immediately, such as graphics rendering code, the JavaScript engine, font selection code, etc. Table 2 shows the breakdown between current lines of Rust code (including generated code that handles interfacing with the JavaScript engine) and C code. This table also includes test code, though the majority of that code is in HTML and JavaScript. There are two limitations in the language interoperability that pose challenges for Servo today. First, Rust cannot currently expose varargs-style functions to C code. Second, Rust cannot compile 4.6 Macros Rust provides a hygienic macro system. Macros are defined using a declarative, pattern-based syntax [KW87]. The macro system has proven invaluable; we have defined more than one hundred macros throughout Servo and its associated libraries. For example, our HTML tokenizer rules, such as those shown in Figure 5, are written in a macro-based domain specific language that closely matches the format of the HTML specification.7 An7 https://html.spec.whatwg.org/multipage/syntax. html#tokenization 4 2015/5/28 of the Rust community can share lint checks that they have found useful.10 Future plans include refining our safety checks for garbage collected values, such as flagging invalid ownership transference, and introducing compile-time checks for constructs that are nonoptimal in terms of performance or memory usage. match self.state { states::Data => loop { match get_char!(self) { '&' => go!(self: consume_char_ref), '<' => go!(self: to TagOpen), '\0' => go!(self: error; emit '\0'), c => go!(self: emit c), 5. Figure 5. Incremental HTML tokenizer rules, written in a succinct form using macros. Macro invocations are of the form identifier!(...). Just-in-time code. JavaScript engines dynamically produce native code that is intended to execute more efficiently than an interpreted strategy. Unfortunately, this area is a large source of security bugs. These bugs come from two sources. First, there are potential correctness issues. Many of these optimizations are only valid when certain conditions of the calling code and environment hold, and ensuring the specialized code is called only when those conditions hold is non-trivial. Second, dynamically producing and compiling native code and patching it into memory while respecting all of the invariants required by the JavaScript runtime (e.g., the garbage collector’s read/write barriers or free vs. in-use registers) is also a challenge. other macro handles incremental tokenization, so that the state machine can pause at any time to await further input. If no next character is available, the get_char! macro will cause an early return from the function that contains the macro invocation. This careful use of non-local control flow, together with the overall expressionoriented style, makes Servo’s HTML tokenizer unusually succinct and comprehensible. The Rust compiler can also load compiler plugins written in Rust. These can perform syntactic transformations beyond the capabilities of the hygienic pattern-based macros. Compiler plugins use unstable internal APIs, so the maintenance burden is high compared to pattern-based macros. Nevertheless, Servo uses procedural macros for a number of purposes, including building perfect hash maps at compile time,8 interning string literals, and auto-generating GC trace hooks. Despite the exposure to internal compiler APIs, the deep integration with tooling makes procedural macros an attractive alternative to the traditional systems metaprogramming tools of preprocessors and code generators. 4.7 Open problems While this work has discussed many challenges in browser design and our current progress, there are many other interesting open problems. Integer overflow/underflow. It is still an open problem to provide optimized code that checks for overflow or underflow without incurring significant performance penalties. The current plan for Rust is to have debug-only checking of integer ranges and for Servo to run debug builds against a test suite, but that may miss scenarios that only occur in optimized builds or that are not represented by the test suite. Unsafe code correctness. Today, when we write unsafe code in Rust there is limited validation of memory lifetimes or type safety within that code block. However, many of our uses of unsafe code are well-behaved translations of either pointer lifetimes or data representations that cannot be annotated or inferred in Rust. We are very interested in additional annotations that would help us prove basic properties about our unsafe code, even if these annotations require a theorem prover or ILP solver to check. Project-specific static analysis Compiler plugins can also provide “lint” checks9 that use the same infrastructure as the compiler’s built-in warnings. This allows project-specific safety or style checks to integrate deeply with the compiler. Lint plugins traverse a typechecked abstract syntax tree (AST), and they can be enabled/disabled within any lexical scope, the same way as built-in warnings. Lint plugins provide some essential guarantees within Servo. Because our DOM objects are managed by the JavaScript garbage collector, we must add GC roots for any DOM object we wish to access from Rust code. Interaction with a third-party GC written in C++ is well outside the scope of Rust’s built-in guarantees, so we bridge the gap with lint plugins. These capabilities enable a safer and more correct interface to the SpiderMonkey garbage collector. For example, we can enforce at compile time that, during the tracing phase of garbage collection, all Rust objects visible to the GC will report all contained pointers to other GC values, avoiding the threat of incorrectly collecting reachable values. Furthermore, we restrict the ways our wrappers around SpiderMonkey pointers can be manipulated, thus turning potential runtime memory leaks and ownership semantic API mismatches into static compiler errors instead. As the “lint” / “warning” terminology suggests, these checks may not catch all possible mistakes. Ad-hoc extensions to a type system cannot easily guarantee soundness. Rather, lint plugins are a lightweight way to catch mistakes deemed particularly common or damaging in practice. As plugins are ordinary libraries, members Incremental computation. As mentioned in Section 2, all modern browsers use some combination of dirty bit marking and incremental recomputation heuristics to avoid reprocessing the full page when a mutatation is performed. Unfortunately, these heuristics are not only frequently the source of performance differences between browsers, but they are also a source of correctness bugs. A library that provided a form of self adjusting computation suited to incremental recomputation of only the visible part of the page, perhaps based on the Adapton [HPHF14] approach, seems promising. 6. Related browser research The ZOOMM browser was an effort at Qualcomm Research to build a parallel browser, also focused on multicore mobile devices [CFMO+ 13]. This browser includes many features that we have not yet implemented in Servo, particularly around resource fetching. They also wrote their own JavaScript engine, which we have not done. Servo and ZOOMM share an extremely concurrent architecture — both have script, layout, rendering, and the user interface operating concurrently from one another in order to maximize interactivity for the user. Parallel layout is one major area that was not investigated in the ZOOMM browser, but is a focus of Servo. The other major difference is that Servo is implemented in 8 https://github.com/sfackler/rust-phf 9 http://doc.rust-lang.org/book/plugins.html# 10 https://github.com/Manishearth/rust-clippy lint-plugins 5 2015/5/28 Rust, whereas ZOOMM is written in C++, similarly to most modern browser engines. Ras Bodik’s group at the University of California Berkeley worked on a parallel browsing project (funded in part by Mozilla Research) that focused on improving the parallelism of layout [MB10]. Instead of our approach to parallel layout, which focuses on multiple parallel tree traversals, they modeled a subset of CSS using attribute grammars. They showed significant speedups with their system over a reimplementation of Safari’s algorithms, but we have not used this approach due to questions of whether it is possible to use attribute grammars to both accurately model the web as it is implemented today and to support new features as they are added. Servo uses a very similar CSS selector matching algorithm to theirs. [Ler90] Leroy, X. Efficient data representation in polymorphic languages. In P. Deransart and J. Małuszyński (eds.), Programming Language Implementation and Logic Programming 90, vol. 456 of Lecture Notes in Computer Science. Springer, 1990. [Ler00] Leroy, X. The Objective Caml System (release 3.00), April 2000. Available from http://caml.inria.fr. [LLV] The LLVM compiler infrastructure. http://llvm.org. [MB10] Meyerovich, L. A. and R. Bodik. Fast and Parallel Webpage Layout. In Proceedings of the 19th International Conference on World Wide Web, WWW ’10, Raleigh, North Carolina, USA, 2010. ACM, pp. 711–720. [MTHM97] Milner, R., M. Tofte, R. Harper, and D. MacQueen. The Definition of Standard ML (Revised). The MIT Press, Cambridge, MA, 1997. [MTK+ 12] Mai, H., S. Tang, S. T. King, C. Cascaval, and P. Montesinos. A Case for Parallelizing Web Pages. In Proceedings of the 4th USENIX Conference on Hot Topics in Parallelism, HotPar’12, Berkeley, CA, 2012. USENIX Association. [OPE] The Opera web browser. http://www.opera.com/. [Rep91] Reppy, J. H. CML: A higher-order concurrent language. In PLDI ’91. ACM, June 1991, pp. 293–305. Acknowledgments The Servo project was started at Mozilla Research, but has benefitted from contributions from Samsung, Igalia, and hundreds of volunteer community members. These contributions amount to more than half of the commits to the project, and we are grateful for both our partners’ and volunteers’ efforts to make this project a success. [RUS] The Rust language. http://www.rust-lang.org/. [SAF] The Apple Safari web browser. http://www.apple. com/safari. [SER] The Servo web browser engine. https://github.com/ servo/servo. [TB98] Tofte, M. and L. Birkedal. A region inference algorithm. ACM Trans. Program. Lang. Syst., 20(4), July 1998, pp. 724– 767. [WEB] The WebKit open source project. http://www.webkit. org. [Wee06] Weeks, S. Whole program compilation in MLton. Invited talk at ML ’06 Workshop, September 2006. [WLZC11] Wang, Z., F. X. Lin, L. Zhong, and M. Chishtie. Why Are Web Browsers Slow on Smartphones? In Proceedings of the 12th Workshop on Mobile Computing Systems and Applications, HotMobile ’11, Phoenix, Arizona, 2011. ACM. References [ABP98] Arora, N. S., R. D. Blumofe, and C. G. Plaxton. Thread scheduling for multiprogrammed multiprocessors, 1998. [AM91] Appel, A. W. and D. B. MacQueen. Standard ML of New Jersey. In PLIP ’91, vol. 528 of LNCS. Springer-Verlag, New York, NY, August 1991, pp. 1–26. [CFJW] Cejtin, H., M. Fluet, S. Jagannathan, and S. Weeks. The MLton Standard ML compiler. Available at http://mlton. org. [CFMO+ 13] Cascaval, C., S. Fowler, P. Montesinos-Ortego, W. Piekarski, M. Reshadi, B. Robatmili, M. Weber, and V. Bhavsar. ZOOMM: A parallel web browser engine for multicore mobile devices. In Proceedings of the 18th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, PPoPP ’13, Shenzhen, China, 2013. ACM, pp. 271–280. [CHR] The Google Chrome web browser. google.com/chrome. http://www. [FIR] The Mozilla Firefox web browser. firefox.com/. http://www. [GMJ+ 02] Grossman, D., G. Morrisett, T. Jim, M. Hicks, Y. Wang, and J. Cheney. Region-based memory management in cyclone. In Proceedings of the ACM SIGPLAN 2002 Conference on Programming Language Design and Implementation, PLDI ’02, Berlin, Germany, 2002. ACM, pp. 282–293. [HLA+ 05] Hunt, G., J. R. Larus, M. Abadi, M. Aiken, P. Barham, M. Fahndrich, C. Hawblitzel, O. Hodson, S. Levi, N. Murphy, B. Steensgaard, D. Tarditi, T. Wobber, and B. D. Zill. An Overview of the Singularity Project. Technical Report MSR-TR-2005-135, Microsoft Research, October 2005. [HPHF14] Hammer, M. A., K. Y. Phang, M. Hicks, and J. S. Foster. Adapton: Composable, demand-driven incremental computation. In Proceedings of the 35th ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI ’14, Edinburgh, United Kingdom, 2014. ACM, pp. 156–166. [IE] The Microsoft Internet Explorer web browser. http:// www.microsoft.com/ie. [KW87] Kohlbecker, E. E. and M. Wand. Macro-by-example: Deriving syntactic transformations from their specifications. In Proceedings of the 14th ACM SIGACT-SIGPLAN Symposium on Principles of Programming Languages, POPL ’87, Munich, West Germany, 1987. ACM, pp. 77–84. 6 2015/5/28
6cs.PL
POSITIVITY OF DENOMINATOR VECTORS OF CLUSTER ALGEBRAS arXiv:1712.06975v1 [math.RT] 15 Dec 2017 PEIGEN CAO FANG LI Abstract. In this paper, we prove that positivity of denominator vectors holds for any skewsymmetric cluster algebra. 1. introduction Cluster algebras were introduced by Fomin and Zelevinsky in [6]. The motivation was to create a common framework for phenomena occurring in connection with total positivity and canonical bases. A cluster algebra A(S) of rank n is a subalgebra of an ambient field F generated by certain combinatorially defined generators (i.e., cluster variables), which are grouped into overlapping clusters. One of the important features of cluster algebras is their Laurent phenomenon. Thus the denominator vectors of cluster variables can be defined following this phenomenon in Section 2. Fomin and Zelevinsky conjectured that Conjecture 1.1. (Conjecture 7.4 (i) in [7]) Positivity of denominator vectors) For any given a / xt0 , the denominator vector dt0 (x) cluster xt0 of a cluster algebra A(S), and any cluster variable x ∈ of x with respect to xt0 is in Nn . In the sequel, we always write denominator vectors briefly as d-vectors. Positivity of d-vectors has been affirmed in the following cases: (i) cluster algebras of rank 2 (Theorem 6.1 of [6]), (ii) cluster algebras arising from surfaces (see [5]), and (iii) cluster algebras of finite type (see [1, 4]). Caldero and Keller proved a weak version of Conjecture 1.1 in [2], that is, for a skew-symmetric cluster algebra A(S), if its seed (xt0 , yt0 , Bt0 ) is acyclic, then the d-vector dt0 (x) of any cluster variable x ∈ / xt0 is in Nn . In this paper, we will give a positive answer to Conjecture 1.1 for skew-symmetric cluster algebras. Theorem 1.2. Positivity of d-vectors holds for any skew-symmetric cluster algebra. This paper is organized as follows. In Section 2, some basic definitions and notations are introduced. In Section 3, the proof of Theorem 1.2 is given. 2. Preliminaries Recall that (P, ⊕, ·) is a semifield if (P, ·) is an abelian multiplicative group endowed with a binary operation of auxiliary addition ⊕ which is commutative, associative, and distributive with respect to the multiplication · in P. Let T rop(ui : i ∈ I) be a free abelian group generated by {ui : i ∈ I} for a finite set of index I. We Q min(ai ,bi ) Q Q , then (T rop(ui : i ∈ I), ⊕) define the addition ⊕ in T rop(ui : i ∈ I) by uai i ⊕ ubi i = ui i is a semifield, which is called a tropical semifield. i i Mathematics Subject Classification(2010): 13F60, 05E40 Keywords: skew-symmetric cluster algebra, denominator vector, positivity. Date: version of December 20, 2017. 1 2 PEIGEN CAO FANG LI The multiplicative group of any semifield P is torsion-free for multiplication [6], hence its group ring ZP is a domain. We take an ambient field F to be the field of rational functions in n independent variables with coefficients in ZP. An n × n integer matrix B = (bij ) is called skew-symmetric if bij = −bji for any i, j = 1, · · · , n; and in general, B is called skew-symmetrizable if there exists a positive integer diagonal matrix S such that SB is skew-symmetric. For a skew-symmetrizable matrix B, we can encode the sign pattern of matrix entries of B by the directed graph Γ(B) with the vertices 1, 2, · · · , n and the directed edges (i, j) for bij > 0. A skew-symmetrizable matrix matrix B is called acyclic if Γ(B) has no oriented cycles. Definition 2.1. A seed in F is a triplet (x, y, B) such that (i) x = {x1 , · · · , xn }, called a cluster, is a free generating set of F , where x1 · · · , xn are called cluster variables. (ii) y = {y1 , · · · , yn } is a subset of P, where y1 , · · · , yn are called coefficients. (iii) B = (bij ) is a n × n skew-symmetrizable integer matrix, called an exchange matrix. The seed (x, y, B) is called acyclic if B is acyclic. Denote by [a]+ = max{a, 0} for any a ∈ R. Definition 2.2. Let (x, y, B) be a seed in F . Define the mutation of the seed (x, y, B) at k ∈ {1, · · · , n} as a new triple µk (x, y, B) = (x̄, ȳ, B̄) given by   if i 6= k xi , n n Q Q [bjk ]+ [−bjk ]+ x̄i = (1) yk x x +   j=1 j L j=1 j , if i = k. (1 yk )xk  y −1 , i=k k (2) ȳi = L [b ] yi y ki + (1 yk )−bki , otherwise. k  −b , i = k or j = k; ij (3) . b̄ij = bij + bik [−bkj ]+ + [bik ]+ bkj , otherwise. It can be seen that µk (x, y, B) is also a seed and µk (µk (x, y, B)) = (x, y, B). Definition 2.3. A cluster pattern S in F is an assignment of a seed (xt , yt , Bt ) to every vertex t of the n-regular tree Tn , such that for any edge t k t′ , (xt′ , yt′ , Bt′ ) = µk (xt , yt , Bt ). We always denote by xt = {x1;t , · · · , xn;t }, yt = {y1;t , · · · , yn;t }, Bt = (btij ). Definition 2.4. Let S be a cluster pattern, the cluster algebra A(S) associated with the given cluster pattern S is the ZP-subalgebra of the field F generated by all cluster variables of S. • If S is cluster pattern with coefficients in T rop(y1 , · · · , ym ), the corresponding cluster algebra A(S) is said to be a cluster algebra of geometric type. • If S is cluster pattern with coefficients in T rop(y1 , · · · , yn ) and there exists a seed (xt0 , yt0 , Bt0 ) such that yi;t0 = yi for i = 1, · · · , n, then the corresponding cluster algebra A(S) is called a cluster algebra with principal coefficients at t0 . Theorem 2.5. Let A(S) be a skew-symmetrizable cluster algebra, and (xt0 , yt0 , Bt0 ) be a seed of A(S). Then (i).(Theorem 3.1 in [6], Laurent phenomenon) any cluster variable xi;t of A(S) is a sum of Laurent monomials in xt0 with coefficients in ZP. POSITIVITY OF DENOMINATOR VECTORS OF CLUSTER ALGEBRAS 3 (ii). ([9], positive Laurent phenomenon) any cluster variable xi;t of A(S) is a sum of Laurent monomials in xt0 with coefficients in NP. Denote by xat0 := n Q i=1 xai;ti 0 for a ∈ Zn . Let A(S) be a skew-symmetrizable cluster algebra, and (xt0 , yt0 , Bt0 ) be a seed of A(S), by Laurent phenomenon, any cluster variable x of A(S) has the P cv xvt0 , where V is a subset of Zn , 0 6= cv ∈ ZP. Let −dj be the minimal exponent form of x = v∈V P cv xvt0 . Then x has the form of of xj;t0 appearing in the expansion x = v∈V x= (4) f (x1;t0 , · · · , xn,t0 ) , 1 n xd1;t · · · xdn;t 0 0 where f ∈ ZP[x1;t0 , · · · , xn;t0 ] with xj;t0 ∤ f for j = 1, · · · , n. The vector dt0 (x) = (d1 , · · · , dn )⊤ is called the denominator vector (briefly, d-vector) of the cluster variable x with respect to xt0 . Proposition 2.6. ((7.7) in [7]) Denote the standard basis vectors in Zn by e1 , · · · , en . The vectors dt0 (xl;t ) are uniquely determined by the initial conditions dt0 (xl;t0 ) = −el together with the recurrence relations t0 d (xl;t′ ) = for t k t′ .  t0   d (xl;t ) if l 6= k; t0   −d (xk;t ) + max P btik >0 btik dt0 (xi;t ), P btik <0 ! −btik dt0 (xi;t ) if l = k From the above proposition, we know that the notion of d-vectors is independent of the choice of coefficient system. So when studying the d-vector dt0 (x) of a cluster variable x, we can focus on the cluster algebras of geometric type. Proposition 2.7. (Proposition 2.5 in [11]) Let A(S) be a cluster algebra, x be a cluster variable, and (xt , yt , Bt ), (xt′ , yt′ , Bt′ ) be two seeds of A(S) with (xt′ , yt′ , Bt′ ) = µk (xt , yt , Bt ). Suppose that ′ dt (x) = (d1 , · · · , dn )⊤ , dt (x) = (d′1 , · · · , d′n ) are the d-vectors of x with respective to xt and xt′ respectively, then di = d′i for i 6= k. Proof. For the convenience of the readers, we give the sketch of the proof here. By (xt′ , yt′ , Bt′ ) = Q btik Q −btik yk;t yk;t xi;t + 1⊕y xi;t . Thus µk (xt , yt , Bt ), we know that xk;t xk;t′ = 1⊕y k;t k;t btik >0 (5) xk;t = x−1 k;t′ ( btik <0 Y bt Y −bt yk;t yk;t xi;tik + xi;t ik ). 1 ⊕ yk;t t 1 ⊕ yk;t t bik >0 bik <0 By Laurent phenomenon, the expansion of x with respect to xt has the following form (6) x= f (x1;t , · · · , xn,t ) , n 1 · · · xdn;t xd1;t where f ∈ ZP[x1;t , · · · , xn;t ] with xj;t ∤ f for j = 1, · · · , n. Then replacing the xk;t in the equation (6) by the one in the equation (5), we get the expansion of x with respect to xt′ . And it can be seen that di = d′i for i 6= k.  4 PEIGEN CAO FANG LI 3. The proof of Theorem 1.2 Let A(S) be a cluster algebra and z be a cluster variable of A(S). Denote by I(z) the set of clusters xt0 of A(S) such that z is a cluster variable in xt0 . For two vertices t1 , t2 of Tn , let l(t1 , t2 ) be the distance between t1 and t2 in the n-regular tree Tn . For a cluster xt of A(S), define the distance between z and xt as dist(z, xt ) := min{l(t0, t)| t0 ∈ I(z)}. The following theorem is from [10]: Theorem 3.1. (Theorem 4.1, Proposition 5.1 and 5.5 in [10]) Let A(S) be a skew-symmetric cluster algebra of geometric type with a cluster xt . Let z be a cluster variable of A(S), and xt0 be a cluster containing z such that dist(z, xt ) = l(t0 , t). Let σ be the unique sequence in Tn relating the seeds at t0 and t in which p, q denote the last two directions (clearly, p 6= q): σ : t0 k1 k2 t1 p=km−1 · · · tm−2 tm−1 = u q=km tm = t. For e 6= p, q, let u e v and t e w be the edges of Tn in the same direction e. Then, ±1 (i) z can be written as z = Pt + Qt , with Pt ∈ L1 := NP[xq;u , x±1 p;t ; (xi;t )i6=p,q ], and Qt ∈ L2 := ±1 ±1 NP[xq;t , xp;t ; (xi;t )i6=p,q ]. Moreover, Pt and Qt are unique up to L1 ∩ L2 . (ii) There exists a Laurent monomial F appearing in the expansion z = Pt + Qt such that the variable xe;t , has nonnegative exponent in F . (iii) There exist P1;t ∈ L3 , P2;t ∈ L4 , Q1;t ∈ L5 , Q2;t ∈ L6 such that P1;t + P2;t = Pt and Q1;t + Q2;t = Qt , where L3 := ±1 NP[xq;u , xe;v ; (x±1 i;t )i6=p,q ], L4 := NP[xq;u , xe;u ; (xi;t )i6=p,q ], L5 := ±1 NP[xq;t , xe;t ; (x±1 i;t )i6=p,q ], L6 := NP[xq;t , xe;w ; (xi;t )i6=p,q ]. Thus z has the form of z = P1;t + P2;t + Q1;t + Q2;t , where P1;t , P2;t , Q1;t , Q2;t are unique up to L3 ∩ L4 ∩ L5 ∩ L6 . Lemma 3.2. Keep the notations in Theorem 3.1. There exists a Laurent monomial in Laurent expansion of z with respect xt in which xe;t appears with nonnegative exponent for e 6= p, q. Proof. By Theorem 3.1 (ii), there exists a Laurent monomial F appearing in the expansion z = Pt +Qt such that the variable xe;t has non-negative exponent in F . We know n n Q Q [−bt ]+ [bt ]+ xj;t jk yq;t xj;tjq + j=1 j=1 L . xq;u = (1 yq;t )xq;t Substituting the above equality into z = Pt + Qt , then we obtain the Laurent expansion of z with respect to xt . It is easy to see that in this Laurent expansion there exists a Laurent monomial such that xe;t appear in this Laurent monomial with nonnegative exponent.  Now we give the proof of Theorem 1.2. Proof. By Proposition 2.6, we know the notion of d-vectors is independent of the choice of coefficient system. Hence, we can assume that A(S) is a skew-symmetric cluster algebra of geometric type. / xt′ , then dist(z, xt′ ) =: m + 1 > 0. Let xt′ be a cluster of A(S), z be a cluster variable with z ∈ t′ We will show the d-vector d (z) of z with respect to xt′ is in Nn . Let xt0 ∈ I(z) such that l(t0 , t′ ) = dist(z, xt′ ) = m + 1 > 0. Since Tn is a tree, there is an unique sequence σ linking t0 and t′ in Tn : t0 k1 t1 k2 p=km−1 · · · tm−2 tm−1 = u q=km tm = t km+1 (6=q) tm+1 = t′ . POSITIVITY OF DENOMINATOR VECTORS OF CLUSTER ALGEBRAS 5 By the choice of xt0 , we know dist(z, xtj ) = l(t0 , tj ) = j for j = 1, 2, · · · , m + 1. ′ We prove dt (z) ∈ Nn by induction on dist(z, xt′ ) = m + 1 > 0. ′ Clearly, dt (z) ∈ Nn if dist(z, xt′ ) = 1, 2. ⋆ Inductive Assumption: Assume that dt (z) ∈ Nn for any t⋆ such that 1 ≤ dist(z, xt⋆ ) < dist(z, xt′ ). Because 1 ≤ dist(z, xt ) < dist(z, xt′ ), we know dt (z) ∈ Nn , by inductive assumption. By Proposition 2.7 and xt′ = µkm+1 (xt ), we get that the i-th component of dt (z) and the i-th component of ′ ′ dt (z) are equal for i 6= km+1 . Thus, in order to show dt (z) ∈ Nn , we only need to show that the ′ km+1 -th component of dt (z) is nonnegative. We prove this in two cases. Case (I): km+1 6= p and Case (II): km+1 = p. The proof in Case (I): By positive Laurent phenomenon, the expansion of z with respect to xt′ has the form of z = P ′ cv xvt′ , where V is a subset of Zn , and 0 6= cv ∈ NP. If km+1 -th component of dt (z) is negative v∈V (here km+1 6= p, q), then the exponent of xkm+1 ;t′ in each xvt′ must be positive, v ∈ V . We know ′ ′ n [−btjk ]+ [btjk ]+ Q xj;t′ m+1 xj;t′ m+1 + j=1 j=1 L xkm+1 ;t′ = . (1 ykm+1 ;t′ )xkm+1 ;t P cv xvt′ , we can obtain the Laurent expansion of z with Substituting the above equation into z = ykm+1 ;t′ n Q v∈V respect to xt . And the exponent of xkm+1 ;t in each Laurent monomial appearing in the obtained Laurent expansion is negative. This contradicts Lemma 3.2. Thus if km+1 6= p, q, then km+1 -th ′ ′ component of dt (z) is nonnegative and we have dt (z) is in Nn . The preparation for the proof in Case (II): Consider the maximal rank two mutation subsequence in directions p, q at the end of σ. This sequence connects t′ to a vertex tr . Thus we have the following two cases (7) kr (6=q) t0 · · · · tr−1 tr p q tr+1 p p tr+2 · · · tm−2 q tm−1 tm p tm+1 = t′ . p tm+1 = t′ . or (8) kr (6=p) t0 · · · · tr−1 tr q p tr+1 q q p tr+2 · · · tm−2 q tm−1 tm p Let tr ta in the first case and let tr tb in the second case. It is easy to see that r ≤ m− 2 in the first case and r ≤ m − 3 in the second case. q ta , the p-th components of d-vectors of xp;ta , xq;ta , xp;tr , xp;tr+1 , Lemma 3.3. (i) In the first case tr xq;tr+1 , xq;tr+2 with respect to xt′ are nonnegative. p (ii) In the second case tr tb , the p-th components of d-vectors of xp;tb , xq;tb , xq;tr , xp;tr+1 , xq;tr+1 , xp;tr+2 with respect to xt′ are nonnegative. ′ Proof. (i) If the p-th components of d-vector dt (x) of some cluster variable x ∈ {xp;ta , xq;ta , xp;tr , xp;tr+1 , xq;tr+1 , xq;tr+2 } with respect to xt′ is negative. Say x ∈ xtE , E ∈ {a, r, r + 1, r + 2}. Since the clusters xta , xtr , xtr+1 , xtr+2 can be obtained from the cluster xt′ by sequences of mutations using only two directions p and q. Then by Theorem 6.1 in [6], we know x = xp;t′ . Hence, xtE and xt′ have at least n − 1 common cluster variables. By Theorem 5 of [8] or Theorem 4.23 (c) of [3], we know xt′ = xtE or xt′ = µp (xtE ) or xt′ = µq (xtE ). Thus m + 1 = dist(z, xt′ ) ≤ max{dist(z, xtE ), dist(z, µp (xtE ), dist(z, µq (xtE )}. On the other hand, by the sequence (7) and tr q ta , we know max{dist(z, xtE ), dist(z, µp (xtE ), dist(z, µq (xtE )} ≤ r + 3 ≤ m + 1. 6 PEIGEN CAO FANG LI Thus we obtain max{dist(z, xtE ), dist(z, µp (xtE ), dist(z, µq (xtE )} = m + 1, and this will result in that r = m − 2 and E = m, i.e. xtE = xtm . But this contradicts that xp;t′ = x ∈ xtE = xtm . So the p-th components of d-vectors of xp;ta , xq;ta , xp;tr , xp;tr+1 , xq;tr+1 , xq;tr+2 with respect to x′t are nonnegative. (ii) By the same argument with (i), we have (9) m + 1 = dist(z, xt′ ) ≤ max{dist(z, xtE ), dist(z, µp (xtE ), dist(z, µq (xtE )}, and, on the other hand, max{dist(z, xtE ), dist(z, µp (xtE ), dist(z, µq (xtE )} ≤ r + 3 ≤ m, which contradicts to (9). So, the p-th components of d-vectors of xp;tb , xq;tb , xq;tr , xp;tr+1 , xq;tr+1 , xp;tr+2  with respect to xt′ are nonnegative. Return to the proof of Theorem 1.2. The proof in case (II): Applying Theorem 3.1 (iii) at the vertex tr+1 with respect to the directions q (in case of (7)) and p (in case of (8)), we get that either z ∈ NP[xp;ta , xq;ta , xp;tr , xp;tr+1 , xq;tr+1 , xq;tr+2 ; (x±1 i;tr+1 )i6=p,q ], or z ∈ NP[xp;tb , xq;tb , xq;tr , xp;tr+1 , xq;tr+1 , xp;tr+2 ; (x±1 i;tr+1 )i6=p,q ]. ′ ′ Then by Lemma 3.3, the p-th component of dt (z) is nonnegative and so, dt (z) is in Nn .  Acknowledgements: This project is supported by the National Natural Science Foundation of China (No.11671350 and No.11571173). References [1] P. Caldero, F. Chapoton, and R. Schiffler, Quivers with relations arising from clusters (An case), Trans. Amer. Math. Soc. 358 (2006), no. 3, 1347-1364. [2] P. Caldero and B. Keller, From triangulated categories to cluster algebras II, Annales Sci. de l’École Norm. Sup. (4) 39 (2006), 983-1009. [3] P. Cao and F. Li, Some conjectures on generalized cluster algebras via the cluster formula and D-matrix pattern, J. Algebra 493 (2018), 57C78. [4] C. Ceballos and V. Pilaud, Denominator vectors and compatibility degrees in cluster algebras of finite type, Trans. Amer. Math. Soc. 367 (2015), no. 2, 1421C1439. [5] S. Fomin, M. Shapiro, and D. Thurston, Cluster algebras and triangulated surfaces. I. Cluster complexes. Acta Math., 201(1):83C146, 2008. [6] S. Fomin and A. Zelevinsky, Cluster algebras. I. Foundations, J. Amer. Math. Soc. 15 (2002), no. 2, 497-529. [7] S. Fomin and A. Zelevinsky. Cluster algebras. IV. Coefficients. Compos. Math. 143(2007), 112-164. [8] M. Gekhtman, M. Shapiro, and A. Vainshtein, On the properties of the exchange graph of a cluster algebra, Math. Res. Lett. 15 (2008), 321-330. [9] M. Gross, P. Hacking, S. Keel, M. Kontsevich, Canonical bases for cluster algebras, arXiv:1411.1394 [math.AG]. [10] K. Lee and R. Schiffler, Positivity for cluster algebras, Annals of Mathematics 182 (2015), 73C125. [11] N. Reading and S. Stella, Initial-seed recursions and dualities for d-vectors, Pacific J. Math. 293 (2018), no. 1, 179-206. Peigen Cao Department of Mathematics, Zhejiang University (Yuquan Campus), Hangzhou, Zhejiang 310027, P.R.China E-mail address: [email protected] POSITIVITY OF DENOMINATOR VECTORS OF CLUSTER ALGEBRAS 7 Fang Li Department of Mathematics, Zhejiang University (Yuquan Campus), Hangzhou, Zhejiang 310027, P.R.China E-mail address: [email protected]
0math.AC
K. Horie OIMC 07P03556 An Upper Limit of AC Huffman Code Length in JPEG Compression Kenichi Horie* R&D Division, Olympus Imaging Corporation, 2951 Ishikawa, Hachioji, 192-8507 Tokyo, Japan E-mail: [email protected] Preprint: OIMC 07P03556 Revised version v1.3. Abstract. A strategy for computing upper code-length limits of AC Huffman codes for an 8x8 block in JPEG Baseline coding is developed. The method is based on a geometric interpretation of the DCT, and the calculated limits are as close as 14% to the maximum code-lengths. The proposed strategy can be adapted to other transform coding methods, e.g., MPEG 2 and 4 video compressions, to calculate close upper code length limits for the respective processing blocks. Keywords: bit rate requirement, discrete cosine transforms, Huffman codes, image coding, JPEG, upper bound bit length 1 of 27 09/01/20 K. Horie OIMC 07P03556 I. INTRODUCTION JPEG Baseline coding [1] is by far the most widely implemented method for coding still images. Statistical behavior of the DCT coefficients and the correlations between the runs of zeros and the sizes of coefficients have been studied, which resulted in the JPEG 2-dimensional Huffman coding [2], [3]. Furthermore, the distribution of the DCT coefficients has also been investigated in detail [3]-[5]. While these statistical behaviors of JPEG are well known, little is known about the mathematical bounds on the code-lengths. Of particular interest are the AC Huffman codes, which exhibit large variations in code-length and thereby determine the most important part of the total code-length of a JPEG file. The mathematical study of upper limits of AC Huffman code-length is not only of academic interest, but stems from needs in consumer electronics. Most digital cameras employ the JPEG Baseline sequential coding, and the knowledge of a close upper code-length limits for an 8x8 block (and the JPEG image file) is important in assigning economic buffer memory spaces at various stages of image encoding in a digital camera. Furthermore, a digital camera usually displays the number of images fitting into the rest of the memory space, and to ensure that this number of JPEG images can indeed be stored, the quantization tables for each JPEG image are optimized to keep the file size close to or below a targeted value. However, such a rate control is achieved by iterating the quantization, the Huffman encoding, and the bit counting of coded data length [6] or by other, sophisticated file size prediction schemes [7], both of which require extensive processing. Alternatively, rate control can be avoided by calculating the number of images based on an average JPEG file size and its standard variation [8]. However, such a statistical estimation method is not reliable, especially in a repetitive shooting mode wherein successive images are often statistically correlated with each other. With the knowledge of a close upper limit of the JPEG file size, a rate control scheme is not necessary anymore, or at least its use greatly reduced. Additionally, the reliability of statistical methods for calculating the number of images can be significantly improved by taking the upper limit into account. A JPEG file may be formatted according to JFIF [9] or Exif [10] and consists of AC Huffman 2 of 27 09/01/20 K. Horie OIMC 07P03556 codes, the DC Huffman codes, the Huffman and the quantization tables, and other portions including a header. Close upper code-length limits or even the maximum code-lengths of those portions except the AC Huffman codes can be easily calculated. On the other hand, it is extremely difficult to calculate a close limit of the AC Huffman code-length. One reason for this difficulty is that the AC Huffman codes are assigned to the runlength-size combination, which extends over several DCT coefficients rather than to individual coefficients. A further, subtle reason is that little is known about the space of the DCT coefficients and the functional behavior of the code-length in this space. In this work we clarify the geometrical aspect of the DCT coefficients to derive a computational method for calculating a close upper limit on the AC Huffman code-length for an 8x8 block. The calculation method is derived in dependence on the scaling of the example quantization tables in annex K of the JPEG specification [1]. Although only the JPEG case is described in this work, the general principles of our method are applicable to other transform coding methods. In particular, these approaches can be adapted to calculate close upper limits of the VLC code-lengths in MPEG 2 and MPEG 4. Some of the proposed ideas might be applicable even for H.264/AVC, VC-1, and JPEG XR video/still image compressions. II. JPEG BASELINE CODING AND THE GEOMETRY OF THE DCT COEFFICIENTS We introduce terms and definitions used in this work and thereby briefly review some aspects of the JPEG Baseline encoding process [1]-[3]. We also add mathematical comments and observations which are necessary in introducing and developing our calculation strategy. In the JPEG Baseline coding a (color) image consists of, for example, one luminance (Y) and two chrominance (Cb, Cr) matrices of 8-bit pixel data. Each of these matrices is sectioned into 8x8 blocks. In each pixel block, its pixel values are shifted from unsigned integers in the range [0,28 − 1] to signed integers in the range [ −2 7 ,2 7 − 1] . Let f xy ( x, y = 0,...,7 ) denote the shifted pixel values of an 3 of 27 09/01/20 K. Horie OIMC 07P03556 8x8 pixel block, which can be viewed as a vector in a 64-dimensional space. The set of all possible vectors furnishes a cube I = [−2 7 ,27 − 1]64 , which is off-centered from the coordinate origin. This cube in turn is contained in a ball B0 around the coordinate origin with radius R = 27 × 64 = 210 . +1) uπ Let K xu = 12 C (u ) cos ( 2 x16 be the Discrete Cosine Transform (DCT) matrix, where x, u = 0,...,7 , and C (u ) = 1 / 2 for u = 0 and C (u ) = 1 for u = 1,...,7 . The DCT coefficients of the pixel block are given by the expression Fuv = ∑ x , y =0 K xu f xy K yv . In matrix notation this reads x , y =7 F = K T fK , where the superscript T denotes the transpose of a matrix. Among the DCT coefficients Fuv , the upper-left coefficient F00 is 8 times the average value of all pixel data within the 8x8 pixel block and is called the DC coefficient. The remaining 63 coefficients are called AC coefficients. We call Fuv or F a (DCT) configuration, and f xy or f simply a vector. Although the DCT is introduced in literature as a mapping from the image domain into the spatial frequency domain, we can interpret the DCT simply as an orthogonal mapping acting upon vectors within the 64-dimensional space. To see this, we use the fact that the DCT matrix is an orthogonal matrix, i.e. KK T = K T K = I . This can be derived naturally by noting that these column vectors are eigenvectors of a real symmetric second difference matrix, which automatically ensures that the column vectors are orthogonal [11]. In order to verify the orthogonal feature of the DCT as a 64-dimensional transformation, we have to show that the natural inner product in the 64-dimensional space is preserved under the DCT. The inner product between two general 64-dimensional vectors v and w is given by the trace Tr (v ⋅ wT ) , wherein the vectors are assumed to take the form of an 8x8 matrix. Let V = K T vK and W = K T wK be the image vectors under the DCT. Using the orthogonality of the DCT matrix and the invariance of the trace under cyclic permutation of the matrices, it follows Tr (VW T ) = Tr ( K T vKK T wT K ) = Tr (vwT ) . Since the DCT is orthogonal, it is a 64-dimensional rotation which rotates the cube I to another cube J , but leaves the ball B0 unchanged. We call J the configuration space, which by definition contains all possible DCT configurations and is a subset of B0 . Since the DCT configuration Fuv lies in the ball B0 , the coefficients satisfy the relation ∑ (F ) 2 u ,v uv ≤ 2 20 . In particular, all coefficients must lie in the range of [−210 ,210 ] . As we will be focusing on the AC coefficients in the subsequent sec- 4 of 27 09/01/20 K. Horie OIMC 07P03556 tions, we need further characterizations of the AC coefficients with respect to the ball. We have: Proposition 1. Regardless of the DC coefficient value, the AC coefficients fulfill the following strict inequality: ∑ (F ) 2 ( u ,v ) ≠( 0.0 ) uv < 2 20 (AC Ball condition). Proof. If this is wrong, a DCT configuration F would exist with the property (1) ∑ ( u ,v ) ≠( 0 , 0 ) (Fuv )2 = 220 . Then the DC coefficient F00 must be necessarily zero, and F lies on the sphere surface S (with radius R ) of the ball B0 . Since the DCT is a rotation, the inverse image vector f must also lie on the sphere S . However, due to the off-centered nature of the cube I , the sum of the squared amplitudes of the pixel data f xy cannot become the square of the radius R unless all pixel data are equal to − 27 . This implies that the image vector f is flat, with no spatial frequency, so that under the DCT, all AC coefficients would become zero, and the DC coefficient would have the value − 210 , contradicting the assumption and proving the assertion. The inequality (1) means geometrically that DCT configurations cannot lie on the 62-dimensional equator E of the sphere S , which is defined by the set of the two equations F00 = 0 and ∑ ( u ,v ) ≠ ( 0 , 0 ) (Fuv )2 = 2 20 . Thus the DCT configurations lie in the “half-open” ball B = B0 \ E , in which this equator has been taken away from the ball B0 . Henceforth we call this half-open ball B simply the Ball. Expressing the condition that a configuration Fuv lies in the configuration space J is not an easy task because the rotated cube J is not symmetric with respect to the coordinate axes. Instead, it is easier to apply the inverse DCT to Fuv and to express that the inverse vector lies in the cube I . More precisely, the inverse vector must lie on the integer coordinate lattice within the cube I : − 2 7 ≤ ∑u ,v=0 K xu Fuv K yv ≤ 2 7 for all x, y = 0,...,7 (Cube Condition) u ,v = 7 5 of 27 (2) 09/01/20 K. Horie OIMC 07P03556 ∑ u ,v =7 u ,v =0 K xu Fuv K yv is integer for all x, y = 0,...,7 (Integer Condition) (3) The DCT coefficients are quantized with an integer-valued quantization table Quv (≥ 1) . The quantized DCT coefficients can then be calculated as Duv = ΙΝΤ( Fuv / Quv ) , where ΙΝΤ denotes the integer portion (one may also use the rounding function instead). Due to the AC Ball condition (1), the AC coefficients cannot become − 210 nor 210 , and the quantized AC coefficient values can be constrained to lie in the interval [−210 + 1,210 − 1] . 16 12 14 14 18 24 49 72 11 12 13 17 22 35 64 92 10 14 16 22 37 55 78 95 16 19 24 29 56 64 87 98 24 26 40 51 68 81 103 112 40 58 57 87 109 104 121 100 51 60 69 80 103 113 120 103 61 55 56 62 77 92 101 99 Table K.1 – Luminance quantization table 17 18 24 47 99 99 99 99 18 21 26 66 99 99 99 99 24 26 56 99 99 99 99 99 47 66 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 99 Table K.2 – Chrominance quantization table The JPEG Baseline coding allows control over the file size by varying the quantization table. A frequently employed method consists in multiplying a scale factor SF to the example quantization tables K.1 and K.2 of the JPEG specification to obtain scaled quantization tables: Quv = max(ΙΝΤ( SF × Q0 uv ),1) . (4) The smaller the scale factor, the finer the quantization and the larger the JPEG file size. When SF = 1 / 64 , all quantization factors become one, whereas when SF = 1 , we have the original example quantization tables K.1 and K.2, whose largest quantization factor is 121 . In this work we consider scale factors between 1 and 1 / 64 . After the quantization the quantized AC coefficients are ordered in a zigzag scan order and are represented by runlength, size, and amplitude. Let D (k ) denote the quantized DCT coefficients in the zigzag scan order of the JPEG specification, where k = 0,...,63 is the index in the zigzag ordering. The DC coefficient is D (0) . In what follows, we also use the notations F (k ) := Fuv and Q (k ) := Quv for 6 of 27 09/01/20 K. Horie OIMC 07P03556 the unquantized coefficients and the quantization values, respectively, in the zigzag ordering. For each nonzero quantized AC coefficient D (k ) , k = 1,...,63 , the runlength (consecutive number, possibly zero) of zero-valued AC coefficients which precede it in the zigzag sequence is determined. Each such runlength/nonzero-coefficient combination is represented by a duplet of integer numbers (r, s) (symbol) together with the amplitude a of the nonzero-coefficient. Here r is the runlength 0 to 15 , and the size s is the number of bits used to encode the amplitude in the signed-integer encoding method of JPEG, so that 2 s−1 ≤ a < 2 s and s = 1,...,10 . For zero-runs greater than 15 , the extension duplet (15,0) is interpreted as a succession of 16 zeros. Up to three consecutive (15,0) extensions may precede a duplet (r , s ) with the total runlength 16 × 3 + r . If the last run of zeros includes the last (63rd) AC coefficient, a special duplet (0,0) is assigned to this run of zeros to indicate the EOB (end of block), which terminates the 8x8 block. If the last AC coefficient is not zero, the EOB symbol is not assigned. In this work we re-define the size s to be the smallest integer satisfying a < 2 s . This definition is the same as the usual definition for s = 1,...,10 . Additionally, it introduces the zero size s = 0 , which implies that the quantized coefficient D (k ) is in fact zero. In the JPEG encoding process, the sequence D (k ) of quantized AC coefficients are represented by a sequence of pairs ( r , s ) a of symbols (r , s ) and amplitudes a in ascending order of the zigzag scan: (r1 , s1 )a1 , (r2 , s2 )a2 , (r3 , s3 )a3 ,... . Huffman codes Huff (r , s ) of length up to 16 bits are assigned to each symbol (r , s ) , whereas each amplitude a is encoded with s bits. The EOB is 4 (2) bits long for luminance (chrominance) AC coefficients. Throughout this work we use the typical Huffman tables in annex K of the JPEG specification [1]. Let len denote the code-length in bits of a (sub-)sequence of AC coefficients, represented as pairs of symbols (r , s ) and amplitudes a in the ascending zigzag scan order, or of Huffman codes Huff ( r , s ) themselves. Since each symbol ( r , s ) already determines the code-length of the pair ( r , s ) a , we define len( r , s ) to be the code-length of such a pair. For a coefficient D (k ) = a with size s > 0 and preceded by r zeros, the relation len(0,0,...,0, D ( k )) = len( r , s ) = len( Huff ( r , s )) + s holds. For example, the zero runlength sequence 0,0,0,0,−1 of chrominance AC coefficients results in len(0,0,0,0,−1) = len( 4,1) = len( Huff ( 4,1)) + 1 = len("111010" ) + 1 = 7 , see Table I. Since the code-length 7 of 27 09/01/20 K. Horie OIMC 07P03556 depends on the amplitudes only through the respective sizes, we also use a simplified notation len( s1 , s2 ,..., sl ) for a sequence of AC coefficients having coefficient sizes s1 , s2 ,..., sl (some of which might be zero). Furthermore, len(F ) denotes the (luminance or chrominance) AC code-length of a configuration F . The typical Huffman tables are structured such that len(r , s) > len(r ' , s' ) for r ≥ r '≥ 0 and for s > s ' > 0 , see Tables I and II below. Therefore, the AC code-length len(F ) of a configuration F increases when the quantized AC amplitudes increase and/or the runlengths increase. In particular, one might expect that, at least in general, the code-length increases when the unquantized AC amplitudes increase, so that a configuration with the maximum code-length would lie on the boundary of the configuration space J . More precisely, we have the following fact: Proposition 2. For any maximum code-length configuration, there exists a configuration on the boundary having the same code-length. Furthermore, a maximum code-length configuration lies always on the boundary of J or at most in such a distance away from it where the code-length does not change. Proof. Assume that F is a maximum code-length configuration. Certainly then, not all AC coefficients of F can become zero after quantization, so that at least one coefficient Fuv remains non-zero after quantization. If F is not already on the boundary of J , we can gradually enlarge the amplitude of Fuv until we hit the boundary and thereby obtain a configuration F ' on the boundary. During this process the code-length len(F ) cannot become smaller due to the above mentioned monotony structure of the code-lengths len( r , s ) , but also cannot become larger, since F is assumed to be a maximum code-length configuration. Thus F has the same code-length as F ' , which proves the assertion. Since the arguments do not depend on the specific geometry of J , the same statements are true if we consider configurations in the Ball B . Having discussed the coding of AC coefficients, we briefly review the DC coefficients. The DC coefficients are encoded separately from the AC coefficients and take values in the range [−210 ,210 ] . The quantized DC coefficient is encoded as the difference from the DC term of the previous block in the 8 of 27 09/01/20 K. Horie OIMC 07P03556 raster-scan order. The difference is represented by the amplitude a and its size s = 1,...,11 , 2 s−1 ≤ a < 2 s . The size is Huffman coded, and the amplitude is encoded with s bits. The maximum amplitude of the difference between two quantized DC coefficients is given by ΙΝΤ(2 × 210 / Q00 ) , where Q00 is the DC quantization value. III. GENERAL STRATEGY The AC code-length of a DCT configuration can certainly not exceed 63 times the maximum code-length in the Huffman table plus the length of the EOB code. The maximum length of Huffman codes is 16 bits, the maximum size is 10 , and the EOB is 4 or 2 bits long. Thus a crude upper bound in bits for AC code-length is given by l0 = 63 × (16 + 10) + 4 = 1642 . (5) Besides being obviously too high and therefore useless, this number was derived with no reference to the quantization table. Contrary to the case of the DC coefficients, the AC coefficients are encoded in context with each other in combination of runlength and size, and the underlying geometry of the configuration space is rather complicated. These make it very difficult to find a close upper code-length limit for the AC codes, let alone to find a maximum AC code-length configuration. In order to develop a strategy for deriving close upper limits of the AC code-lengths, we simplify the problem. We note that the AC coefficients cannot be disentangled from the DC coefficient in the Cube Condition (2) and the Integer Condition (3). In contrast, the AC Ball condition (1) makes no reference to the DC coefficient and is certainly much simpler than the system of 64 inequalities (2). For these reasons, we drop the Cube Condition (2) and the Integer Condition (3) and work with the AC Ball condition (1) alone. Thus instead of considering the configuration space J , we allow for configurations within the entire Ball B . 9 of 27 09/01/20 K. Horie OIMC 07P03556 We only need to calculate an upper limit of the AC code-length for the entire Ball B , since such an upper limit is certainly also an upper limit of the code-length for any subset of B , in particular for the original configuration space J and the sub-lattice defined by the Integer Condition. In fact, we only need to calculate an upper limit for configurations on the surface S of the Ball B , since we have shown in the previous section that the code-length len(F ) takes its maximum value on the surface S . For this reason, we henceforth consider only configurations on the sphere S . In what follows, we characterize a configuration F on the sphere S by its AC coefficients in the zigzag scan order, i.e. F (k ) , k = 1,...,63 . In view of the AC Ball condition (1), the DC coefficient is fixed up to sign by the square root of 2 20 − ∑( u ,v )≠( 0, 0 ) (Fuv ) > 0 . 2 Let F be an arbitrary configuration on S . Let Q (k ) , k = 1,...,63 , be the AC quantization values in the zigzag scan order. The code-length len(F ) of F depends on each coefficient F (k ) only through the respective size sk of its quantized value D (k ) = ΙΝΤ( F (k ) / Q (k )) , where sk is the smallest integer with the property D(k ) < 2 sk . Therefore, each coefficient F (k ) of F can be replaced by the smallest possible non-negative value Q(k ) × 2 sk −1 when sk > 0 or by zero when sk = 0 without changing the code-length. Through these replacements the Huffman codes remain the same, and only the fixed-length codes for the amplitudes change. Each replaced coefficient is smaller than or equal to the original coefficient amplitude, so that the configuration F still satisfies the AC Ball condition (1), which now reads as follows: ∑ Q(k ) 2 × 2 2 sk − 2 < 2 20 . (6) k ( sk >0) Note that the configuration F still lies on the sphere S , although the DC coefficient may have changed. In order to calculate an upper code-length limit for configurations on S , we need to calculate an upper code-length limit only for such configurations F having smallest non-negative coefficients. These configurations on S shall be called reduced configurations. Unless specified otherwise, in the following all configurations are reduced configurations. 10 of 27 09/01/20 K. Horie OIMC 07P03556 We now outline our strategy for calculating an upper code-length limit. We first choose a reference configuration FR on the sphere S . The reference FR cannot be just any configuration, but should have large coefficient sizes. We then compare FR with an arbitrary configuration F on S by introducing a set of operations which replace the coefficient sizes of FR by those of F . These operations generally correspond to the JPEG symbols (r, s) , so that a single operation may replace r + 1 coefficients of FR to produce a run of r zeros followed by a coefficient of (quantized) size s in F . Rather than comparing FR and F in terms of individual coefficients, we compare them in terms of symbols. This method enables a direct calculation of code-length gain or loss induced by each operation, since the Huffman codes are assigned to the symbols. Depending on F , some operations (“positive operations”) enlarge certain coefficient sizes which generally lead to a longer code-length, whereas other operations (“negative operations”) reduce some coefficient sizes or even introduce runs of zeros, which generally lessen the code-length. Since FR has already large coefficient sizes, there are only a limited number of positive operations possible at all. In general, not all operations can be positive, since otherwise the AC Ball condition (6) would be violated. Instead, positive operations must be accompanied by negative operations in order to fufill the AC Ball condition. We will derive an interdependence rule between the positive and negative operations from the AC Ball condition. By carefully choosing the reference FR (and introducing a few more conditions), this interdependence rule can be made very restrictive, such that only a few combinations of positive and negative operations are allowed. One of the key features of our method is the fact that for these limited number of combinations, an upper bound on the code-length gain can be calculated by a simple table look-up method. Finally, this upper bound is added to the code-length of the reference configuration FR to yield the upper limit on the AC code-length of configurations on the sphere S . In order to specify the outlined strategy, it is helpful to straight away introduce an appropriate reference configuration FR and together with other necessary conditions which will lead to the restrictive interdependence rule. To this end, consider the AC Ball condition (6) and suppose for a moment that all quantization factors are powers of 2. Then, all AC coefficients of the configuration F are either zero or powers of 2 since F is a reduced configuration. We introduce the (unquantized) size S (k ) of 11 of 27 09/01/20 K. Horie OIMC 07P03556 each coefficient by defining F (k ) = 2 S ( k )−1 ( S (k ) = 1,...,10 ) if F (k ) > 0 or by setting S (k ) = 0 if F (k ) = 0 . For indices i = 1,2,..., a , let pi be positions in the zigzag order with the property F ( pi ) = 28 , i.e., S ( pi ) = 9 . Likewise, for j = 1,2,..., b , let q j be positions such that F (q j ) = 29 ( S (q j ) = 10 ). The AC Ball condition (6) can now be converted into the following inequality: ∑2 2 S ( k )−2 ⇔ < 2 20 k ( S ( k )>0 ) a ∑2 b 16 i =1 + ∑ 218 + j =1 ∑2 2 S ( k )−2 ∑2 2S (k )−2 < 2 20 ⇔ k ≠ pi , k ≠ q j , S ( k ) > 0 < (64 − 4a − 16b) ⋅ 214 . (7) k ≠ pi ,k ≠ q j ,S ( k ) >0 This inequality implies the constraint a + 4b < 16 , which restricts the numbers of size 9 and size 10 indices to just 40 combinations of the pair (a, b) . In particular, 0 ≤ a ≤ 15 and 0 ≤ b ≤ 3 . The right hand side is a sum of 64 − 4a − 16b copies of 214 , whereas the left hand side sums over at most 63 − a − b indices k . Hence we conclude that at least 3a + 15b terms on the left hand side must be smaller than 214 . These terms have respective sizes S ( k ) less than 8 . The maximum value for the expression 3a + 15b is 54 . These observations suggest a natural choice for the reference configuration FR , namely to set each of its coefficients to the same value FR (k ) = 27 with size 8 . If a ( b ) coefficients of FR are replaced by size 9 ( 10 ) coefficients (“positive operations”), at least 3a ( 15b ) coefficients of FR must be replaced by smaller sized coefficients (“negative operations”). Altogether, no more than 15 coefficients can be replaced by size 9 coefficients (with amplitude 28 ) and no more than 3 coefficients can be replaced by size 10 coefficients, since otherwise the AC Ball condition is violated no matter how small the other coefficients might be. The necessary condition for this restrictive interdependence rule is that the quantization values be powers of 2. In order to derive the same rule for an arbitrary quantization table Q(k ) , we replace it by the “power-of-2 quantization table” defined as follows: 12 of 27 09/01/20 K. Horie OIMC 07P03556 Q2 ( k ) := 2C ( k ) , where C (k ) = ΙΝΤ(log 2 Q (k )) (“Power-of-2 Q-Table”). (8) We are allowed to make this replacement, since we only need to find an upper limit of AC code-length for this power-of-2 quantization table Q2 (k ) rather than for the original quantization table Q(k ) . This is true in view of the following proposition: Proposition 3. On the sphere S , the maximum code-length for Q2 (k ) is larger than or equal to the maximum code-length for the original quantization table Q(k ) . Proof. To prove this assertion, let F be a maximum code-length configuration on S (based on the quantization table Q(k ) ) satisfying the AC Ball condition (6). Define a new configuration F ' on S by F ' (k ) := Q2 ( k ) / Q (k ) × F ( k ) . Since Q2 ( k ) ≤ Q ( k ) , the new configuration F ' still fulfills the AC Ball condition (6), wherein each Q(k ) is replaced by Q2 (k ) . By construction, quantizing F ( k ) with Q(k ) yields the same result as when quantizing F ' ( k ) with Q2 (k ) , so that the code-length of F is the same as the code-length of F ' based on Q2 (k ) . By the very definition of the maximum code-length for Q2 (k ) , it is greater than or equal to this code-length of F ' , which proves the assertion. This assertion is also true for the entire Ball B , but might not be true for the rotated cube J . The difficulties involved in the above reasoning can be visualized without going into the details of the inequalities (2). Let F be a configuration at or near a vertex of the cube. If the cube was not rotated, reducing coefficient amplitudes of F , e.g. in order to define F ' , will move F entirely within the cube. However, for a rotated cube like J , reductions of some of the amplitudes might move F along a coordinate direction outside the cube, so that the new configuration F ' may not be well-defined. Although this does not disprove the assertion, we see that it is not straightforward to introduce the concept of the power-of-2 quantization table (8) without enlarging the configuration space J to the Ball B . In summary, the arguments set forth in this section and Proposition 3 simplifies the problem of finding an upper AC code-length limit in the following way: 13 of 27 09/01/20 K. Horie OIMC 07P03556 Theorem 1. For a given quantization table Q(k ) , the AC code-length of any DCT configuration in the configuration space J is less than or equal to the maximum code-length of reduced configurations on the sphere S using the power-of-2 quantization table Q2 (k ) based on Q(k ) . Therefore, an upper code-length limit for such reduced configurations is also an upper code-length limit of all DCT configurations. The introduction of a power-of-2 quantization table via (8) is not mandatory in deriving a restrictive interdependence rule. However, for general quantization factors Q(k ) , it seems that much more work is needed to derive from the AC Ball condition (6) a convenient interdependence rule. For example, the quantized coefficient sizes sk of the reference may be defined as sk = ΙΝΤ(8 − log 2 Q( k )) , so that each square term in (6) comes close to 214 . However, since these reference values may not be constant over the positions k , it is very difficult to derive a simple interdependence rule in the manner shown above. In the rest of this work we restrict ourselves to the power-of-2 quantization table restriction (8). We leave the study of the general case to future work. The unquantized sizes S ( k ) determine the coefficients F ( k ) of a configuration F . The quantization with a power-of-2 quantization table amounts to reducing these sizes by the exponents C (k ) of the quantization factors. The quantized sizes sk are given by sk = max( S (k ) − C (k ),0) . Since the coefficients F (k ) are chosen to be minimal, S (k ) is zero if sk is zero. IV. DETAILED STRATEGY FOR CALCULATING THE UPPER LIMIT Following the outline of the strategy, let FR be the reference configuration on S having coefficients with the constant value FR (k ) = 27 and size S R (k ) = 8 . The quantized sizes are given by sk = max( S R (k ) − C (k ),0) . Since the largest quantization factor in this work is 121 , the largest power-of-2 quantization factor is 64 = 26 < 121 . Thus the quantized sizes of FR are always larger than 14 of 27 09/01/20 K. Horie OIMC 07P03556 or equal to 2 , sk ≥ 2 . In particular, all coefficients of FR are non-zero after quantization. Let F be an arbitrary configuration on S having unquantized sizes S (k ) and quantized sizes sk = max( S (k ) − C (k ),0) . We now define the following operations OP1 to OP6, which generally correspond to symbols (r, s) . Depending on the size S ( p) of the target configuration F , a different operation replaces the size S R ( p ) of the reference configuration FR by the size S ( p) of F : OP1: If 0 < S ( p ) < S R ( p ) and S ( p − 1) > 0 or p = 1 , replace S R ( p ) 6 S ( p ) . OP2: If 0 < S ( p ) < S R ( p ) , S ( p − 1) = S ( p − 2) = ... = S ( p − r ) = 0 for r > 0 , and S ( p − r − 1) > 0 p − r =1 , or replace simultaneously r +1 sizes S R ( p − r ),..., S R ( p ) 6 S ( p − r ),..., S ( p − 1), S ( p ) . OP3: If S ( p ) ≥ S R ( p ) , S ( p − 1) = S ( p − 2) = ... = S ( p − r ) = 0 for r > 0 , and S ( p − r − 1) > 0 p − r =1 , or replace simultaneously r sizes S R ( p − r ),..., S R ( p − 1) 6 S ( p − r ),..., S ( p − 1) . OP4: If S ( p ) > 0 and S ( q ) = 0 for all q > p , replace S R ( q ) 6 S ( q ) for all q > p simultaneously. OP5: If S ( p ) > S R ( p ) and S ( p − 1) > 0 or p = 1 , replace S R ( p ) 6 S ( p ) . OP6: If S ( p ) > S R ( p ) , S ( p − 1) = S ( p − 2) = ... = S ( p − r ) = 0 for r > 0 , and S ( p − r − 1) > 0 or p − r = 1 , then, after performing OP3, replace S R ( p) 6 S ( p) . Proposition 4. Given a reduced configuration F for a power-of-2 quantization table, the set of opera- tions OP1 to PO6 is uniquely defined. Proof. It suffices to show that each coefficient FR ( p) can be replaced by the corresponding coefficient F ( p ) through exactly one of these operations if FR ( p) ≠ F ( p) . We consider several cases of F ( p ) . The coefficient size S ( p ) can be either zero or non-zero. If S ( p ) is non-zero and not preceded by zeros, three cases may arise: if S ( p ) = S R ( p ) , we see that none of the operations OP1 to OP6 can be applied at all; if S ( p ) < S R ( p ) , then S R ( p ) is replaced by S ( p ) only through OP1; if S ( p ) > S R ( p ) , 15 of 27 09/01/20 K. Horie OIMC 07P03556 then S R ( p ) is replaced by S ( p ) only through OP5. On the other hand, if the non-zero S ( p ) is preceded by a run of r zeros, again three cases may arise: if S ( p ) = S R ( p ) , then r preceeding coefficient sizes S R ( p − r ),..., S R ( p − 1) are replaced by zeros through OP3, but OP3 does not change S R ( p ) itself; if S ( p ) < S R ( p ) , then r + 1 coefficient sizes S R ( p − r ),..., S R ( p ) are replaced by the respective sizes of F through OP2; if S ( p ) > S R ( p ) , then r coefficient sizes S R ( p − r ),..., S R ( p − 1) are replaced by zeros through OP3, whereupon the coefficient S R ( p ) is replaced by S ( p ) through OP6. Now assume that S ( p ) is zero. Then S ( p ) is either a member of a zero run followed by a non-zero coefficient or it is a member of a zero run including the last zigzag coefficient. In the first case, S R ( p ) will be replaced by zero through either OP2 or OP3, whereas in the latter case, S R ( p ) will be replaced by zero through OP4. In summary, each coefficient FR ( p ) of FR is correctly replaced by the respective coefficient F ( p) of F by one and only one operation from OP1 to OP6. Following the outline of our strategy, we examine the changes in the AC code-length induced by each of the operations. To this end, we introduce the notion of local (or individual) code-length dif- ference δ , which is defined to be the change in code-length per affected position: if an operation replaces sizes at m indices (or positions) pi and leaves sizes at n other indices unaffected, then the local code-length change is defined to be δ = δ (ti , u j ) = (len(ti , u j ) − len(ti , u j )) / m , where ti are the quantized sizes at positions pi , ti are the corresponding replaced quantized sizes, and u j are the unaffected quantized sizes at n positions. We assign to each of the m affected positions pi the code-length difference δ . The differences at these m positions together yield the total change in code-length through the operation. For example, OP2 replaces r + 1 coefficients simultaneously, so that m = r + 1 and n = 0 . As for OP3, it replaces r coefficients, so that m = r . Since it leave the size at p unaffected, we have n = 1 . For OP6, we have m = 1 and n = r . Proposition 5. The operations OP1 to OP4 reduce the code-length (“negative operations”), i.e., the re- spective local code-length differences are negative or zero. The operations OP5 and OP6 increase the code-length (“positive operations”), whereby the local code-length differences are strictly positive. 16 of 27 09/01/20 K. Horie OIMC 07P03556 Proof. We consider each operation in succession. OP1 reduces the unquantized size S R ( p ) = 8 to a lower size S ( p ) and so induces a change of the quantized size s p to a smaller size s p . The size s p is non-zero, since otherwise S ( p ) must be zero, so that OP1 would not have been applied. As noted earlier, for any runlength r ≥ 0 and for any sizes s > s ' > 0 , the relation len(r , s) > len(r , s' ) holds. The local code-length difference is given by − δ 1p ,s = len(0, s ) − len(0, s p ) < 0 , where we have set s = s p to simplify the notation. Thus OP1 is a negative operation, which reduces the AC code-length. As for OP2, we first note that for the example quantization tables K.1 and K.2 of the JPEG specification, preceding quantization factors in the zigzag scan order are always smaller than twice the current quantization factor, i.e., Q0 (l ) < 2Q0 (k ) if l < k . After scaling the quantization factors according to (4), we still have a weaker relation Q(l ) ≤ 2Q(k ) + 1 . For the exponents of the power-of-2 quantization factors defined according to (8), we can deduce the relation C (l ) ≤ C ( k ) + 1 by using the fact that the quantization factors are integers. For the quantized sizes in the reference configuration FR this relation implies s p −r ,..., s p −1 ≥ s p − 1 > 0 . It follows that len ( s p −r ," s p −1 , s p ) = len(0, s p−r ) + " + len(0, s p ) ≥ r ⋅ len(0, s p − 1) + len (0, s p ) . Tables I and II below reveal that the right hand side expression is always greater than or equal to len ( r , s p ) , which is greater than len(r , s p ) since s p > s p . In summary, len( s p −r ," s p −1 , s p ) > len(r , s p ) , which was to be shown. Let − δ p2,r ,s = (len (r , s ) − len ( s p −r ," s p −1 , s p )) /( r + 1) < 0 be the local code-length difference assigned to each of the r + 1 positions p − r to p , where we have set s = s p . OP3 differs from OP2 in that the size at position p is not changed. Repeating the arguments set forth for OP2, we see that the code-length is reduced or remains the same. Let − δ p3,r = (len(r , s p ) − len( s p−r ," s p−1 , s p )) / r ≤ 0 denote the local code-length difference assigned to each of the r positions p − r to p − 1 . As for OP4, we consider the inequality len ( s p +1 ,..., s63 ) ≥ (63 − p ) ⋅ len (0,2) , which is true since quantized sizes of the reference configuration are greater than or equal to 2 . For luminance AC coeffi- 17 of 27 09/01/20 K. Horie OIMC 07P03556 cients, we have the equality len(0,2) = 4 = len(ΕΟΒ) , whereas len(0,2) = 5 > 2 = len(ΕΟΒ) for chrominance AC coefficients. Thus OP3 always reduces the code-length except when p = 62 and s63 = 2 for luminance AC coefficients, in which case there is no change in the code-length. Let − δ p4 = (len(ΕΟΒ) − len( s p+1 ," , s63 )) /(63 − p) ≤ 0 be the local code-length difference assigned to each of the affected 63 − p positions. OP5 enlarges the size s p to either s p = s p + 1 or to s p = s p + 2 . These are the only possible cases, since for the unquantized size S ( p ) of the target configuration F only two sizes 9 and 10 are available above the unquantized size S R ( p ) = 8 . Let α 1p = len (0, s p + 1) − len (0, s p ) > 0 be the local code-length gain for the case s p = s p + 1 , and similarly let β 1p = len (0, s p + 2) − len (0, s p ) > 0 be the local code-length gain for the case s p = s p + 2 . OP6 induces a code-length gain of len(r , s p ) − len(r , s p ) > 0 . Similar to the case of OP5, only the two cases s p = s p + 1 or s p = s p + 2 may occur. Let α p2 , r = len ( r , s p + 1) − len ( r , s p ) > 0 be the local code-length gain for s p = s p + 1 , and similarly let β p2, r = len ( r , s p + 2) − len ( r , s p ) > 0 be the local code-length gain for s p = s p + 2 . In the above proof, we have introduced notations for the local code-length losses and gains: − δ 1p ,s , − δ p2,r ,s , − δ p3,r , − δ p4 , α 1p , β 1p , α p2 ,r , and β p2,r . Note that the indices assigned for each code-length change uniquely determine the respective operation. For example, − δ 1p ,s implies that OP1 is applied to a coefficient at zigzag scan position p where the quantized size of the configuration F is given by s . OP3 and OP4 do not depend on the quantized size, so that their code-length losses are indexed without the size. The code-length of F can be represented as a sum of the code-length of FR and the local code-length differences for all positions affected by the operations OP1 to OP6. With the help of the indices for the local code-length differences, we can express the code-length of F as follows: 18 of 27 len( F ) = len( FR ) + A + B − Δ , wherein (9) A = α q11 + ... + α q1a′ + α q2a′+1 , ra′+1 + ... + α q2a , ra , (10) 09/01/20 K. Horie OIMC 07P03556 B = β q11 + ... + β q1b′ + β q2b′+1 , rb′+1 + ... + β q2b , rb , and (11) Δ = 1 ⋅ δ 1p1 , s1 + 1 ⋅ δ 1p2 , s2 + ... + (r1′ + 1) ⋅ δ p21′ ,r1′,s1′ + (r2′ + 1) ⋅ δ p2′2 ,r2′ ,s′2 + ... (12) + (r ' '1 ) ⋅ δ p3''1 , r ''1 , s ''1 + (r ' ' 2 ) ⋅ δ p3''2 , r ''2 , s ''2 + ... + (63 − p) ⋅ δ p4 . The sum A counts the local code-length gains induced by the positive operations OP5 and OP6 for the case of unquantized size 9 , whereas the sum B counts the local code-length gains for the case of unquantized size 10 . As indicated by the indices, we assume that there are a terms in A and b terms in B . Thus the configuration F is assumed to posses a positions having unquantized size 9 and b positions with unquantized size 10 . Due to the AC Ball condition (7), there must exist at least 3a + 15b positions in F with unquantized size less than 8 . Each such position must have been created by one of the negative operations OP1 to OP4, i.e. it is either a position within a zero run or a position having non-zero size smaller than 8 . The local code-length losses of these 3a + 15b or more positions are summarized in the sum Δ . It has contributions from four lines on the right-hand side of equation (12), which are the sums of code-length losses generated by the negative operations OP1, OP2, OP3, and OP4, respectively. We now estimate an upper code-length limit for the expression (9). To this end, we observe that the sum Δ of local code-length losses in (12) has been generated by operations OP1 to OP4 uniquely defined for the particular configuration F . Therefore, the loss Δ is certainly greater than or equal to the sum Δ (3a + 15b) , which is defined to be the sum of only the 3a + 15b smallest local code-length losses among any and all possible local code-length losses created by any and all possible operations OP1 to OP4 for any and all target configurations. More precisely, let Δ{} be the set of all local code-length losses δ 1p ,s for all possible pairs ( p, s ) , r + 1 copies of δ p2,r ,s for all possible triplets ( p, r , s ) , r copies of δ p3,r for all possible pairs ( p, r ) , and (63 − p ) copies of δ p4 for all positions p . Define the loss function Δ (n) to be the sum of the n smallest values in Δ{} . Then the relation Δ ≥ Δ(3a + 15b) holds. 19 of 27 09/01/20 K. Horie OIMC 07P03556 In a similar fashion, let A{} be the set of all local code-length gains α 1p for all positions p and all α p2 ,r for all possible pairs ( p, r ) for any and all target configurations. Likewise, let B{} be the set of all local code-length gains β 1p for all positions p and all β p2,r for all possible pairs ( p, r ) for any and all target configurations. Define the gain function A( n) to be the sum of the n largest values in the set A{} , and the gain function B ( n) to be the sum of the n largest values in B{} . Clearly then, the relations A ≤ A( a ) and B ≤ B (b) hold. In summary we obtain the inequality len( F ) ≤ len( FR ) + ( A( a ) + B (b) − Δ (3a + 15b) ) . (13) The second term on the right-hand side is certainly not greater than the maximum of the same expression A( a ) + B (b) − Δ (3a + 15b) among all 40 possible combinations of the pair ( a, b) . Thus, we have just proved the following: Theorem 2. Given an arbitrary reduced configuration F , its code-length satisfies the inequality len( F ) ≤ Λ , where (14) Λ = len ( FR ) + max{A( a ) + B (b) − Δ (3a + 15b)} . (15) ( a ,b ) The formula (15) represents an upper limit of the AC code-length for reduced configurations on the sphere S when using a power-of-2 quantization table. By Theorem 1, Λ is an upper code-length limit for all DCT configurations. The code-length len ( FR ) can be readily calculated by using the power-of-2 quantization values Q2 ( k ) in (8). In order to calculate the maximum term in (15), we have to determine the 15 largest local code-length gains in the set A{} , the 3 largest local code-length gains in the set B{} , and the 54 smallest local code-length losses in the set Δ{} . In principle, these local code-length changes can be determined by applying each of the operations OP1 to OP6 for all possible parameter values of r 20 of 27 09/01/20 K. Horie OIMC 07P03556 and s at all possible positions p in FR and calculating the induced code-length changes: for the code-length loss − δ 1p ,s , there are 63 possible positions p and 7 possible non-zero sizes s < 8 , so that we may have to calculate the loss values for up to 63 × 7 cases; for − δ p2,r ,s , there are at most (63 × 62) / 2 × 7 cases to calculate, where we accounted for the fact that r < p ; for − δ p3,r , there are at most (63 × 62) / 2 cases; for − δ p4 , there are at most 62 cases; for α 1p and β 1p , there are at most 63 cases; for α p2 , r and β p2, r , there are at most (63 × 62) / 2 cases. By adding all these numbers, we obtain 20159 calculations, which can be easily done by a computer. This relatively small number of calcula- tions is sufficient to completely determine the loss and gain functions. On the other hand, if we had to consider the combinations of the operations for all possible configurations F , the calculations of the code-length changes would have been impossible. It turns out that only a few combinations of the parameter values need to be considered to determine the loss and gain functions in the required ranges. Tables I and II reveal that len(r , s) > len(r , s' ) for r ≥ 0 and s > s ' > 0 and len(r , s) = len(r ' , s) for large values of r and r ' . Using these relations, calculations for large runlength values and large/small sizes may be skipped. In this way it is possible to calculate the upper limit value by hand, see the calculation example in the next section. We can easily improve the upper limit Λ . To see how, let us assume that the quantized coefficient sizes s p of FR have the same value when p = 60,61,62,63 . Since the sizes are the same, the local code-length losses δ 612 ,1,1 , δ 622 ,1,1 , and δ 632 ,1,1 in the set Δ{} have the same value, say δ . There are altogether 6 copies of δ in Δ{} , since each of the underlying operations OP2 affects two positions. If δ was among the smallest values in Δ{} , the loss function Δ (n) would add δ up to 6 times. On the other hand, there are only 4 positions p = 60,61,62,63 available for assigning δ , so that 2 copies of δ are superfluous in Δ{} . In other words, no configuration can have more than 4 times the local loss ~ δ . By eliminating such superfluous copies in the set Δ{} , we obtain a subset Δ{} . In practical application, it is difficult to find and eliminate all superfluous copies, so that we may delete some of such cop- ~ ies to obtain the subset Δ{} . The term Δ summarizes the local code-length losses for a real configuration F , so that it must be still greater than or equal to the sum of the 3a + 15b smallest local ~ code-length losses within this subset Δ{} , wherein superfluous copies have been eliminated. Since 21 of 27 09/01/20 K. Horie OIMC 07P03556 ~ ~ ~ Δ{} is a subset of Δ{} , the sum Δ(n) of the n smallest values in Δ{} is greater than or equal to ~ Δ (n) , so that we obtain the relation Δ ≥ Δ (3a + 15b) ≥ Δ (3a + 15b) . Arguing in a similar fashion, we can eliminate (at least some) superfluous counts of local code-length gains in the sets A{} and B{} to ~ ~ ~ ~ obtain respective subsets A{} and B{} . Let A ( n) and B ( n) be the sums of the n smallest local ~ ~ ~ code-length gains in A{} and B{} , respectively, for which the relations A ≤ A( a ) ≤ A( a ) and ~ B ≤ B (b) ≤ B (b) are valid. In summary, we obtain a new, possibly lower upper limit { } ~ ~ ~ ~ Λ = len( FR ) + max A(a) + B (b) − Δ(3a + 15b) ≤ Λ . ( a ,b ) (16) To further improve this upper limit, we note that the upper limit needs to be calculated for maximum code-length configurations only. Obviously, the sum Δ of a maximum code-length configuration must be greater than or equal to the sum of 3a + 15b smallest local code-length losses for all pos- ~ sible maximum code-length configurations. Therefore we can eliminate from the subset Δ{} those losses which do not arise for a maximum code-length configuration, and obtain an improved loss func- ~ ~ ~ ~ ~ ~ tion Δ(n) ≥ Δ(n) with the property Δ ≥ Δ (3a + 15b ) ≥ Δ (3a + 15b ) . Similarly, by eliminating (at least ~ some) gains which can not occur for a maximum code-length configuration from the sets A{} and ~ ~ ~ ~ ~ ~ ~ ~ B{} , we obtain smaller gain functions A ( n ) and B ( n ) with the properties A ≤ A(a ) ≤ A(a ) and ~ ~ ~ ~ ~ ~ B ≤ B (b) ≤ B (b) . In summary, we obtain a new, possibly lower upper limit Λ ≤ Λ . To find those local code-length changes which do not occur in a maximum code-length configuration, consider the operations OP2, OP3, and OP6. These operations generate runs of zeros in F , say 0,...,0, S ( p ) . Let (r , s p ) be the symbol after quantization, len(r , s p ) its code-length, and assume s p > 2 . We compare the sub-sequence 0,...,0, S ( p ) with another possible sub-sequence of r + 1 coef- ficients at the same positions, wherein S ( p ) is replaced by the value S ( p ) − 1 and at most 3 preceding zeros are replaced by the same non-zero value S ( p ) − 1 . Such replacements within the configuration F are allowed, since the replaced configuration still satisfies the AC Ball condition (5) due to the identity 2 2 S ( p )−2 = 4 ⋅ 2 2 ( S ( p )−1)−2 . In view of the relation C (l ) ≤ C ( k ) + 1 ( l < k ) for the exponents of the power-of-2 quantization factors, the new sub-sequence has up to 3 non-zero quantized sizes greater than 22 of 27 09/01/20 K. Horie OIMC 07P03556 or equal to s p − 2 , followed by s p − 1 . If the code-length len ( r , s p ) is smaller than the code-length of this replaced sub-sequence, the original sequence 0,...,0, S ( p ) cannot occur in a maximum code-length configuration, since it is replaceable with a sub-sequence with a longer code-length. Tables I and II summarize the code-lengths len(r , s ) of zero run sub-sequences. Runlength/size combinations of shaded cells in Tables I and II cannot appear in a maximum code-length configuration. Table I: Code-lengths (Huffman + size) for chrominance AC coefficients Size 0 1 2 3 4 5 6 7 8 9 10 0 2 3 5 7 9 10 12 14 17 19 22 1 2 3 4 5 6 Runlength 7 8 9 10 11 12 13 14 5 8 11 13 16 18 23 24 25 26 6 10 13 16 20 22 23 24 25 26 6 10 13 16 21 22 23 24 25 26 7 11 19 20 21 22 23 24 25 26 7 12 19 20 21 22 23 24 25 26 8 13 19 20 21 22 23 24 25 26 8 13 19 20 21 22 23 24 25 26 10 18 19 20 21 22 23 24 25 26 10 18 19 20 21 22 23 24 25 26 10 18 19 20 21 22 23 24 25 26 10 18 19 20 21 22 23 24 25 26 12 18 19 20 21 22 23 24 25 26 15 18 19 20 21 22 23 24 25 26 9 18 19 20 21 22 23 24 25 26 15 10 16 18 19 20 21 22 23 24 25 26 Table II: Code-lengths (Huffman + size) for luminance AC coefficients Size 0 1 2 3 4 5 6 7 8 9 10 0 4 3 4 6 8 10 13 15 18 25 26 1 2 3 4 5 6 Runlength 7 8 9 10 11 12 13 14 5 7 10 13 16 22 23 24 25 26 6 10 13 16 21 22 23 24 25 26 7 11 15 20 21 22 23 24 25 26 7 12 19 20 21 22 23 24 25 26 8 13 19 20 21 22 23 24 25 26 8 14 19 20 21 22 23 24 25 26 9 14 19 20 21 22 23 24 25 26 10 18 19 20 21 22 23 24 25 26 10 18 19 20 21 22 23 24 25 26 11 18 19 20 21 22 23 24 25 26 11 18 19 20 21 22 23 24 25 26 12 18 19 20 21 22 23 24 25 26 17 18 19 20 21 22 23 24 25 26 V. 10 17 19 20 21 22 23 24 25 26 15 11 17 18 19 20 21 22 23 24 25 26 CALCULATION AND RESULTS We demonstrate the calculation of the upper limit Λ (15) for chrominance AC coefficients when using the unscaled ( SF = 1 ) quantization Table K.1. Its values are given in zigzag scan order: Q (k ) = 18, 18, 24, 21, 24, 47, 26, 26, 47, 99, 66, 56, 66, 99,..., 99 for k = 1,...,63 . 1. The power-of-2 quantization table is calculated according to (8). Its exponents are given by: 23 of 27 09/01/20 K. Horie OIMC 07P03556 C ( k ) = 4 for k = 1,2,3,4,5,7,8 , C ( k ) = 5 for k = 6,9,12 , and C (k ) = 6 for all other positions. 2. The quantized sizes of the reference FR are given by sk = 4 for k = 1,2,3,4,5,7,8 , sk = 3 for k = 6,9,12 , and sk = 2 for the remaining 53 positions. With the help of Table I, we obtain len( FR ) = 7 × 9 + 3 × 7 + 53 × 5 = 349 . 3. OP1: if we set s := s p − 1 , then the local code-length losses at any of the 63 positions p have the same value δ 1p ,s = len(0, s p ) − len(0, s p − 1) = 2 . Any smaller choice of s leads to a greater loss δ 1p ,s > 2 . Since we know already that the set Δ{} contains more than 54 copies of 2 , losses greater than or equal to 2 need not be considered anymore. 4. OP2: it is always δ p2,r ,s ≥ 2 . For example, the replacement of the quantized sizes 2,2 → 0,1 yields a local loss of δ p2,1,1 = (len(0,2) + len(0,2) − len(1,1)) / 2 = ((5 + 5) − 5) / 2 = 2.5 . Another example is δ 72,1,3 = (len(0,3) + len(0,4) − len(1,3)) / 2 = ((7 + 9) − 11) / 2 = 2.5 at position p = 7 . 5. OP3: at position p = 12 , we have δ123 ,1 = (len(0,2) + len (0,3) − len(1,3)) / 1 = 1 . For any other choices of p and r , a simple lookup of Table I yields the result δ p3,r > 2 . 6. OP4: if s63 = 2 is replaced by zero, we obtain δ 624 = len(0,2) − len(ΕΟΒ) = 3 > 2 . The more non-zero coefficients we replace by zeros, the larger the local code-length losses. 1 7. OP5: for all positions p , we have α 1p ≤ 2 . For example, α 45 = len(0,3) − len(0,2) = 2 . Similarly, we obtain the result β 1p ≤ 4 for all p . 8. OP6: for positions p = 10,11,13,14,...,63 and runlength r = 1,2 , we obtain α p2 ,r = 3 . For example, α162 , 2 = len(2,3) − len(2,2) = 13 − 10 = 3 . For all other combinations of p and r , α p2 ,r ≤ 3 . Likewise, β p2, 2 = 6 for p = 10,11,13,14,...,63 , and β p2,r < 6 for all other p and r . 9. Summary: the 54 smallest local code-length losses are one copy of 1 and 53 copies of 2 , so that the loss function is given by Δ (n) = 2n − 1 when 1 ≤ n ≤ 54 and Δ (0) = 0 . If we consider (unquantized) size 9 coefficients, the 15 largest local code-length gains are 15 times the same value 3 , so that the gain function is given by A(a ) = 3a ( a ≤ 15 ). Similarly, for unquantized size 10 we obtain the gain function B (b) = 6b ( b ≤ 3 ). The total function A( a ) + B (b) − Δ (3a + 15b) is zero when a = b = 0 and equals − 3a − 24b + 1 when a > 0 or b > 0 . Its maximum lies at a = b = 0 , so that the upper code-length limit for chrominance AC coefficients is just given by the 24 of 27 09/01/20 K. Horie OIMC 07P03556 length of the reference configuration, Λ = len( FR ) = 349 . This result means that the reference FR is in fact a maximum code-length configuration. Upper limits for scale factors other than SF = 1 and for luminance AC coefficients can be calculated in a similar fashion. Table III shows calculation results for a variety of scale factors. Table III: Upper limits of AC code-lengths in bits for different scale factors Scale factor Luminance AC Chrominance AC 1/64 1134 1071 1/16 956 797 1/8 812 670 1/6 715 603 1/4 654 593 1/2 517 468 1/1 429 349 VI. DISCUSSION The results in Table III are much smaller than the crude upper bound l0 = 1642 (5). In order to further illustrate the closeness of our result for the case SF = 1 / 64 , we consider the 8x8 image block of Table IV. Table IV: Pixel values for an example 8x8 block 252 61 199 116 120 203 71 99 61 18 34 231 2 254 111 68 199 34 229 165 192 247 250 53 116 231 165 244 136 9 59 4 120 2 192 136 233 252 27 59 203 254 247 9 252 4 16 174 71 111 250 59 27 16 247 11 99 68 53 4 59 174 11 1 After DCT and quantization, the configuration has 18 positions with size 8 and all other positions having size 7, which would yield 999 bits were these coefficients luminance AC and 936 bits were these chrominance AC. Our results in Table III for SF = 1 / 64 are as close as 14% to these bit-lengths and so must be even closer to the corresponding (unknown) maximum code-lengths, despite the enlargement of the configuration space J to the Ball B . In the derivation of the strategy, we have used a few features specific to the quantization table and the Huffman codes of the JPEG specification. Without these features some of the nice properties of 25 of 27 09/01/20 K. Horie OIMC 07P03556 the calculation will be lost, but the proposed method can be adapted and generalized to cope with any quantization table and/or Huffman codes. In particular, transformations other than the DCT may be handled as well by slightly adjusting the Ball B and the AC Ball condition. Furthermore, the outlined general strategy can be adapted to DCT coefficients in MPEG 2 and 4 video compressions, both of which use quantization and VLC coding methods similar to JPEG Baseline coding. Although the coding methods of the transform coefficients are significantly different in H.264/AVC and VC-1 video compressions and in JPEG XR still image compression, some of the general ideas in this work might be useful in investigating the code-length behaviors of these compression methods. Details are left for future work. REFERENCES [1] Information Technology - Digital Compression Coding of Continuous Tone Still Images - Requirements and Guidelines, ISO 10918-1 JPEG International Standard / ITU-T Recommendation T.81, 1992. [2] W. B. Pennebaker and J. L. Mitchell, JPEG Still Image Data Compression Standard. New York: Van Nostrand Reinhold, 1993. [3] N. Kingsbury. (2004, January 22). Image Coding. Connexions Web. [Online]. Available: http://cnx.org/content/col10206/1.3/ [4] S.R. Smoot, L.A. Rowe, “Study of DCT coefficient distributions” in Proc. SPIE Symp. Electronic Imaging, vol. 2657, San Jose, CA, 1996, pp. 403-411. [5] T. Eude, R. Grisel, H. Cherifi, R. Debrie, “On the distribution of the DCT coefficients,” in Proc. IEEE Int. Conf. ASSP, Adelaide, Australia, April 1994, pp. V.365-V.368. [6] S. D. Miller, P. Smidth, C. H. Coleman, “Data compression using a feedforward quantization estimator,” U.S. patent 5 146 324, Sep. 8, 1992. [7] A. Bruna, S. Smith, F. Vella, and F. Naccari, “JPEG Rate Control Algorithm for Multimedia,” in Proc. IEEE Int. Symp. Consumer Electronics, Reading, UK, Sep. 2004, pp. 114-117. 26 of 27 09/01/20 K. Horie OIMC 07P03556 [8] T. Ohta, T. Ikuma, “Electronic still camera and method operating same,” US patent application publication 2001/0000969A1, May 10, 2001. [9] E. Hamilton, “JPEG File Interchange Format Version 1.02,” C-Cube Microsystems, Sep. 1, 1992, http://www.jpeg.org/public/jfif.pdf. [10] “Digital Still Camera Image File Format Proposal (Exif) Version 1.0 Mar. 24,” JEIDA / Electronic Still Camera Working Group, 1995. [11] G. Strang, “The Discrete Cosine Transform,” SIAM Review, vol. 41, pp. 135-47, 1999. 27 of 27 09/01/20
5cs.CE
1 Asynchronous Distributed Optimization via Randomized Dual Proximal Gradient arXiv:1509.08373v2 [cs.SY] 24 Jun 2016 Ivano Notarnicola and Giuseppe Notarstefano Abstract In this paper we consider distributed optimization problems in which the cost function is separable, i.e., a sum of possibly non-smooth functions all sharing a common variable, and can be split into a strongly convex term and a convex one. The second term is typically used to encode constraints or to regularize the solution. We propose a class of distributed optimization algorithms based on proximal gradient methods applied to the dual problem. We show that, by choosing suitable primal variable copies, the dual problem is itself separable when written in terms of conjugate functions, and the dual variables can be stacked into non-overlapping blocks associated to the computing nodes. We first show that a weighted proximal gradient on the dual function leads to a synchronous distributed algorithm with local dual proximal gradient updates at each node. Then, as main paper contribution, we develop asynchronous versions of the algorithm in which the node updates are triggered by local timers without any global iteration counter. The algorithms are shown to be proper randomized block-coordinate proximal gradient updates on the dual function. I. I NTRODUCTION Several estimation, learning, decision and control problems arising in cyber-physical networks involve the distributed solution of a constrained optimization problem. Typically, computing processors have only a partial knowledge of the problem (e.g. a portion of the cost function or a subset of the constraints) and need to cooperate in order to compute a global solution of the whole problem. A key challenge when designing distributed optimization algorithms in Ivano Notarnicola and Giuseppe Notarstefano are with the Department of Engineering, Università del Salento, Via Monteroni, 73100 Lecce, Italy, [email protected]. This result is part of a project that has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 638992 - OPT4SMART). March 15, 2018 DRAFT 2 peer-to-peer networks is that the communication among the nodes is time-varying and possibly asynchronous. Early references on distributed optimization algorithms involved primal and dual subgradient methods and Alternating Direction Method of Multipliers (ADMM), designed for synchronous communication protocols over fixed graphs. More recently time-varying versions of these algorithmic ideas have been proposed to cope with more realistic peer-to-peer network scenarios. A Newton-Raphson consensus strategy is proposed in [1] to solve unconstrained, convex optimization problems under asynchronous, symmetric gossip communications. In [2] a primal, synchronous algorithm, called EXTRA, is proposed to solve smooth, unconstrained optimization problems. In [3] the authors propose accelerated distributed gradient methods for unconstrained optimization problems over symmetric, time-varying networks connected on average. In order to deal with time-varying and directed graph topologies, in [4] a push-sum algorithm for average consensus is combined with a primal subgradient method in order to solve unconstrained convex optimization problems. Paper [5] extends this algorithm to online distributed optimization over time-varying, directed networks. In [6] a novel class of continuous-time, gradient-based distributed algorithms is proposed both for fixed and time-varying graphs and conditions for exponential convergence are provided. A distributed (primal) proximal-gradient method is proposed in [7] to solve (over time-varying, balanced communication graphs) optimization problems with a separable cost function including local differentiable components and a common nondifferentiable term. In [8] experiments of a dual averaging algorithm are run for separable problems with a common constraint on time-varying and directed networks. For general constrained convex optimization problems, in [9] the authors propose a distributed random projection algorithm, that can be used by multiple agents connected over a (balanced) time-varying network. In [10] the author proposes (primal) randomized block-coordinate descent methods for minimizing multi-agent convex optimization problems with linearly coupled constraints over networks. In [11] an asynchronous ADMM-based distributed method is proposed for a separable, constrained optimization problem. The algorithm is shown to converge at the rate O(1/t) (being t the iteration counter). In [12] the ADMM approach is proposed for a more general framework, thus yielding a continuum of algorithms ranging from a fully centralized to a fully distributed. In [13] a method, called ADMM+, is proposed to solve separable, convex optimization problems with a cost function written as the sum of a smooth and a non-smooth March 15, 2018 DRAFT 3 term. Successive block-coordinate updates are proposed in [14], [15] to solve separable optimization problems in a parallel big-data setting. Another class of algorithms exploits the exchange of active constraints among the network nodes to solve constrained optimization problems [16]. This idea has been combined with dual decomposition and cutting-plane methods to solve robust convex optimization problems via polyhedral approximations [17]. These algorithms work under asynchronous, direct and unreliable communication. The contribution of the paper is twofold. First, for a fixed graph topology, we develop a distributed optimization algorithm (based on a centralized dual proximal gradient idea introduced in [18]) to minimize a separable strongly convex cost function. The proposed distributed algorithm is based on a proper choice of primal constraints (suitably separating the graph-induced and node-local constraints), that gives rise to a dual problem with a separable structure when expressed in terms of local conjugate functions. Thus, a proximal gradient applied to such a dual problem turns out to be a distributed algorithm where each node updates: (i) its primal variable through a local minimization and (ii) its dual variables through a suitable local proximal gradient step. The algorithm inherits the convergence properties of the centralized one and thus exhibits an O(1/t) (being t the iteration counter) rate of convergence in objective value. We point out that the algorithm can be accelerated through a Nesterov’s scheme, [19], thus obtaining an O(1/t2 ) convergence rate. Second, as main contribution, we propose an asynchronous algorithm for a symmetric eventtriggered communication protocol. In this communication set-up, a node is in idle mode until its local timer triggers. When in idle, it continuously collects messages from neighboring nodes that are awake and, if needed, updates a primal variable. When the local timer triggers, it updates local primal and dual variables and broadcasts them to neighboring nodes. Under mild assumptions on the local timers, the whole algorithm results into a uniform random choice of one active node per iteration. Using this property and showing that the dual variables can be stacked into separate blocks, we are able to prove that the distributed algorithm corresponds to a block-coordinate proximal gradient, as the one proposed in [20], performed on the dual problem. Specifically, we are able to show that the dual variables handled by a single node represent a single variableblock, and the local update at each triggered node turns out to be a block-coordinate proximal gradient step (in which each node has its own local step-size). The result is that the algorithm March 15, 2018 DRAFT 4 inherits the convergence properties of the block-coordinate proximal gradient in [20]. An important property of the distributed algorithm is that it can solve fairly general optimization problems including both composite cost functions and local constraints. A key distinctive feature of the algorithm analysis is the combination of duality theory, coordinate-descent methods, and properties of the proximal operator when applied to conjugate functions. To summarize, our algorithms compare to the literature in the following way. Works in [1], [3]–[6] do not handle constrained optimization and use different methodologies. In [7], [9], [10] primal approaches are used. Also, local constraints and regularization terms cannot be handled simultaneously. In [7] a proximal operator is used, but only to handle a common, non-smooth cost function (known by all agents) directly on the primal problem. The algorithm in [10] uses a coordinate-descent idea similar to the one we use in this paper, but it works directly on the primal problem, does not handle local constraints and does not make use of proximal operators. In this paper we propose a flexible dual approach to take into account both local constraints and regularization terms. The problem set-up in [11]–[13] is similar to the one considered in this paper. Differently from our approach, which is a dual method, ADMM-based algorithms are proposed in those references. This difference results in different algorithms as well as different requirements on the cost functions. Moreover, compared to these algorithms we are able to use constant, local step-sizes (which can be locally computed at the beginning of the iterations) so that the algorithm can be run asynchronously and without any coordination step. On this regard, we propose an algorithmic formulation of the asynchronous protocol that explicitly relies on local timers and does not need any global iteration counter in the update laws. The paper is organized as follows. In Section II we set-up the optimization problem, while in Section III we derive an equivalent dual problem amenable for distributed computation. The distributed algorithms are introduced in Section IV and analyzed in Section V. Finally, in Section VI we highlight the flexibility of the proposed algorithms by showing important optimization scenarios that can be addressed, and corroborate the discussion with numerical computations. Notation: Given a closed, nonempty convex set X, the indicator function of X is defined as IX (x) = 0 if x ∈ X and IX (x) = +∞ otherwise. Let ϕ : Rd → R ∪ {+∞}, its conjugate  function ϕ∗ : Rd → R is defined as ϕ∗ (y) := supx y > x−ϕ(x) . Let ϕ : Rd → R∪{+∞} be a closed, proper, convex function and α a positive scalar, the proximal operator proxαϕ : Rd → Rd March 15, 2018 DRAFT 5  is defined by proxαϕ (v) := argminx ϕ(x) + 1 kx 2α − vk2 . We also introduce a generalized version of the proximal operator. Given a positive definite matrix W ∈ Rd×d , we define n o 1 2 proxW,ϕ (v) := argmin ϕ(x) + x − v W −1 . 2 x II. P ROBLEM SET- UP AND NETWORK STRUCTURE We consider the following optimization problem n   X min fi (x) + gi (x) , x (1) i=1 d where fi : R → R ∪ {+∞} are proper, closed and strongly convex extended real-valued functions with strong convexity parameter σi > 0 and gi : Rd → R ∪ {+∞} are proper, closed and convex extended real-valued functions. Note that the split of fi and gi may be non-unique and depend on the problem structure. Intuitively, on fi an easy minimization step can be performed (e.g., by division free operations, [21]), while gi has an easy expression of its proximal operator. Remark II.1 (Min-max problems). Our setup does not require fi to be differentiable, thus one can also consider each strongly convex function fi given by fi (x) := maxj∈{1,...,mi } fij (x), mi ∈ N, where {fij (x) | j ∈ {1, . . . , mi }} is a nonempty collection of strongly convex functions.  Since we will work on the dual of problem (1), we introduce the next standard assumption which guarantees that the dual is feasible and equivalent to the primal (strong duality). Assumption II.2 (Constraint qualification). The intersection of the relative interior of dom P and the relative interior of dom ni=1 gi is non-empty. Pn i=1  We want this optimization problem to be solved in a distributed way by a network of peer processors without a central coordinator. Each processor has a local memory, a local computation capability and can exchange information with neighboring nodes. We assume that the communication can occur among nodes that are neighbors in a given fixed, undirected and connected graph G = ({1, . . . , n}, E), where E ⊆ {1, . . . , n} × {1, . . . , n} is the set of edges. That is, the edge (i, j) models the fact that node i and j can exchange information. We denote by March 15, 2018 DRAFT fi 6 Ni the set of neighbors of node i in the fixed graph G, i.e., Ni := {j ∈ {1, . . . , n} | (i, j) ∈ E}, and by |Ni | its cardinality. In Section IV we will propose distributed algorithms for three communication protocols. Namely, we will consider a synchronous protocol (in which neighboring nodes in the graph communicate according to a common clock), and two asynchronous ones, respectively nodebased and edge-based (in which nodes become active according to local timers). To exploit the sparsity of the graph we introduce copies of x and their coherence (consensus) constraint, so that the optimization problem (1) can be equivalently rewritten as n   X min fi (xi ) + gi (xi ) x1 ,...,xn (2) i=1 subj. to xi = xj ∀ (i, j) ∈ E with xi ∈ Rd for all i ∈ {1, . . . , n}. The connectedness of G guarantees the equivalence. Since we propose distributed dual algorithms, next we derive the dual problem and characterize its properties. III. D UAL PROBLEM DERIVATION We derive the dual version of the problem that will allow us to design our distributed dual proximal gradient algorithms. To obtain the desired separable structure of the dual problem, we set-up an equivalent formulation of problem (2) by adding new variables zi , i ∈ {1, . . . , n}, i.e., n   X min fi (xi ) + gi (zi ) x ,...,x 1 n z1 ,...,zn i=1 subj. to xi = xj xi = zi ∀ (i, j) ∈ E (3) ∀ i ∈ {1, . . . , n}. > > > > > Let x = [x> 1 . . . xn ] and z = [z1 . . . zn ] , the Lagrangian of primal problem (3) is given March 15, 2018 DRAFT 7 by n  X L(x, z, Λ, µ) = fi (xi ) + gi (zi ) i=1 X + > λji j∈Ni (xi − xj ) + µ> i (xi  − zi ) n  X X  j > = fi (xi ) + λi (xi − xj ) + µ> i xi i=1 j∈Ni + gi (zi ) − µ> i zi  , where Λ and µ are respectively the vectors of the Lagrange multipliers λji , (i, j) ∈ E, and µi , i ∈ {1, . . . , n}, and in the last line we have separated the terms in x and z. Since G is undirected, the Lagrangian can be rearranged as n   X X j i − λ ) + µ L(x, z, Λ, µ) = fi (xi ) + x> (λ i j i i i=1 j∈Ni + gi (zi ) − zi> µi  . The dual function is q(Λ, µ) := min L(x, z, Λ, µ) x,z  X n  X j > i = min fi (xi ) + xi (λi − λj ) + µi x i=1 j∈Ni + min z = n X i=1 n  X i=1 gi (zi ) − zi> µi   X  j > i min fi (xi ) + xi (λi − λj ) + µi + xi n X i=1 j∈Ni   min gi (zi ) − zi> µi , zi where we have used the separability of the Lagrangian with respect to each xi and each zi . Then, by using the definition of conjugate function (given in the Notation paragraph), the dual function can be expressed as q(Λ, µ) = n X i=1 March 15, 2018 !  X  j −fi∗ − (λi − λij ) − µi − gi∗ (µi ) . j∈Ni DRAFT 8 The dual problem of (3) consists of maximizing the dual function with respect to dual variables Λ and µ, i.e.    X n  X j ∗ i ∗ (λi − λj ) − µi − gi (µi ) . max −fi − Λ,µ i=1 (4) j∈Ni By Assumption II.2 the dual problem (4) is feasible and strong duality holds, so that (4) can be equivalently solved to get a solution of (3). In order to have a more compact notation for problem (4), we stack the dual variables as y = [y1> . . . yn> ]> , where   Λi yi =   ∈ Rd|Ni |+d µi (5) with Λi ∈ Rd|Ni | a vector whose block-component associated to neighbor j is λji ∈ Rd . Thus, changing sign to the cost function, dual problem (4) can be restated as min Γ(y) = F ∗ (y) + G∗ (y), (6) y where ∗ F (y) := n X i=1 fi∗ n   X X  j ∗ i gi∗ µi . (λi − λj ) − µi , G (y) := − j∈Ni i=1 IV. D ISTRIBUTED DUAL PROXIMAL ALGORITHMS In this section we derive the distributed optimization algorithms based on dual proximal gradient methods. A. Synchronous Distributed Dual Proximal Gradient We begin by deriving a synchronous algorithm on a fixed graph. We assume that all the nodes share a common clock. At each time instant t ∈ N, every node communicates with its neighbors in the graph G = ({1, . . . , n}, E) (defined in Section II) and updates its local variables. First, we provide an informal description of the distributed optimization algorithm. Each node i ∈ {1, . . . , n} stores a set of local dual variables λji , j ∈ Ni , and µi , updated through a local proximal gradient step, and a primal variable x?i , updated through a local minimization. Each node uses a properly chosen, local step-size αi for the proximal gradient step. Then, the updated primal and dual values are exchanged with the neighboring nodes. The local dual variables at node i are initialized as λji0 , j ∈ Ni , and µi0 . A pseudo-code of the local update at each node of the distributed algorithm is given in Algorithm 1. March 15, 2018 DRAFT 9 Algorithm 1 Distributed Dual Proximal Gradient Processor states: x?i , λji for all j ∈ Ni and µi Initialization: λji (0) = λji0 for all j ∈ Ni , µi (0) = µi0   o  n P j i +µ +f (x ) λ − λ x?i (0) = argminxi x> i0 i i j0 i i0 j∈Ni Evolution: FOR : t = 1, 2, . . . DO receive x?j (t − 1) for each j ∈ Ni , update   λji (t) = λji (t − 1) + αi x?i (t − 1) − x?j (t − 1) and update µ̃i = µi (t − 1) + αi x?i (t − 1)   µ̃i µi (t) = µ̃i − αi prox 1 gi αi αi receive λij (t) for each j ∈ Ni and compute   X    j i x?i (t) = argmin x> λ (t) − λ (t) +µ (t) +f (x ) i i i i j i xi j∈Ni Remark IV.1. In order to start the algorithm, a preliminary communication step is needed in which each node i receives from each neighbor j its λij0 (to compute x?i (0)) and the convexity parameter σj of fj (to set αi , as it will be clear from the analysis in Section V-B). This step can be avoided if the nodes agree to set λji0 = 0 and know a bound for αi . Also, it is worth noting that, differently from other algorithms, in general λji (t) 6= −λij (t).  We point out once more that to run the Distributed Dual Proximal Gradient algorithm the nodes need to have a common clock. Also, it is worth noting that, as it will be clear from the analysis in Section V-B, to set the local step-size each node needs to know the number of nodes, n, in the network. In the next sections we present two asynchronous distributed algorithms which overcome these limitations. March 15, 2018 DRAFT 10 B. Asynchronous Distributed Dual Proximal Gradient Next, we propose a node-based asynchronous algorithm. We consider an asynchronous protocol where each node has its own concept of time defined by a local timer, which randomly and independently of the other nodes triggers when to awake itself. Between two triggering events the node is in an idle mode, i.e., it continuously receives messages from neighboring nodes and, if needed, runs some auxiliary computation. When a trigger occurs, it switches into an awake mode in which it updates its local (primal and dual) variables and transmits the updated information to its neighbors. Formally, the triggering process is modeled by means of a local clock τi ∈ R≥0 and a randomly generated waiting time Ti . As long as τi < Ti the node is in idle mode. When τi = Ti the node switches to the awake mode and, after running the local computation, resets τi = 0 and draws a new realization of the random variable Ti . We make the following assumption on the local waiting times Ti . Assumption IV.2 (Exponential i.i.d. local timers). The waiting times between consecutive triggering events are i.i.d. random variables with same exponential distribution.  When a node i wakes up, it updates its local dual variables λji , j ∈ Ni and µi by a local proximal gradient step, and its primal variable x?i through a local minimization. The local stepsize of the proximal gradient step for node i is denoted by αi . In order to highlight the difference between updated and old variables at node i during the “awake” phase, we denote the updated ones as λji + and µi + respectively. When a node i is in idle it receives messages from awake neighbors. If a dual variable λij is received it computes an updated value of x?i and broadcasts it to its neighbors. It is worth noting that, being the algorithm asynchronous, there is no common clock as in the synchronous version. C. Edge-based asynchronous algorithm In this section we present a variant of the Asynchronous Distributed Dual Proximal Gradient in which an edge becomes active uniformly at random, rather than a node. In other words, we assume that timers are associated to edges, rather than to nodes, that is a waiting time Tij is extracted for edge (i, j). March 15, 2018 DRAFT 11 Algorithm 2 Asynchronous Distributed Dual Proximal Gradient Processor states: x?i , λji for all j ∈ Ni and µi Initialization: λji = λji0 for all j ∈ Ni , µi = µi0 and n P    o j ? > i xi = argmin xi j∈Ni λi0 − λj0 +µi0 +fi (xi ) xi set τi = 0 and get a realization Ti Evolution: IDLE: WHILE : τi < Ti DO : receive x?j and/or λij from each j ∈ Ni . IF : λij is received THEN : compute and broadcast x?i  = argmin xi x> i  X  λki k∈Ni − λik   + µi  + fi (xi ) go to AWAKE. AWAKE:  + update and broadcast λji = λji + αi x?i − x?j , ∀ j ∈ Ni update µ̃i = µi + αi x?i µ+ i  µ̃i αi   + µ+ i = µ̃i − αi prox 1 gi αi compute and broadcast x?i  = argmin xi x> i X j∈Ni + λji − λij   + fi (xi ) set τi = 0, get a new realization Ti and go to IDLE. The processor states and their initialization stay the same except for the timers. Notice that here we are assuming that nodes i and j have a common waiting time Tij and, consistently, a local timer τij . Each waiting time Tij satisfies Assumption IV.2. In Algorithm 3 we report (only) the evolution for this modified scenario. In this edge-based algorithm the dual variable µi (associated to the constraint xi = zi ) cannot be updated every March 15, 2018 DRAFT 12 time an edge (i, j), j ∈ Ni , becomes active (otherwise it would be updated more often than the variables λji , j ∈ Ni ). Thus, we identify one special neighbor jµi ∈ Ni and update µi only when the edge (i, jµi ) is active. Algorithm 3 Edge-Based Formulation of Algorithm 2 (evolution) IDLE: WHILE τij < Tij DO: nothing go to AWAKE. AWAKE: send x?i to j and receive x?j from j + update and send to j, λji = λji + αi x?i − x?j IF : j = jµi THEN :  ? update µ+ i = proxαi gi∗ µi + αi xi  compute  X    j+ + i x?i = argmin x> λ − λ + µ + f (x ) i i i j i i xi j∈Ni set τij = 0, deal on a new realization Tij and go to IDLE. V. A NALYSIS OF THE DISTRIBUTED ALGORITHMS To start the analysis we first introduce an extended version of (centralized) proximal gradient methods. A. Weighted Proximal Gradient Methods Consider the following optimization problem min Γ(y) := Φ(y) + Ψ(y), y∈RN (7) where Φ : RN → R and Ψ : RN → R ∪ {+∞} are convex functions. We decompose the decision variable as y = [y1> . . . yn> ]> and, consistently, we decompose the space RN into n subspaces as follows. Let U ∈ RN ×N be a column permutation of the N × N identity matrix and, further, let U = [U1 U2 . . . Un ] be a decomposition of U into P n submatrices, with Ui ∈ RN ×Ni and i Ni = N . Thus, any vector y ∈ RN can be uniquely P written as y = i Ui yi and, viceversa, yi = Ui> y. We let problem (7) satisfy the following assumptions. March 15, 2018 DRAFT 13 Assumption V.1 (Block Lipschitz continuity of ∇Φ). The gradient of Φ is block coordinate-wise Lipschitz continuous with positive constants L1 , . . . , Ln . That is, for all y ∈ RN and si ∈ RNi it holds k∇i Φ(y + Ui si ) − ∇i Φ(y)k ≤ Li ksi k, where ∇i Φ(y) is the i-th block component of ∇Φ(y).  Assumption V.2 (Separability of Ψ). The function Ψ is block-separable, i.e., it can be decomP posed as Ψ(y) = ni=1 ψi (yi ), with each ψi : RNi → R∪{+∞} a proper, closed convex extended real-valued function.  Assumption V.3 (Feasibility). The set of minimizers of problem (7) is non-empty.  1) Deterministic descent: We first show how the standard proximal gradient algorithm can be generalized by using a weighted proximal operator. Following the same line of proof as in [18], we can prove that a generalized proximal gradient iteration, given by   y(t + 1) = proxW,Ψ y(t) − W ∇Φ y(t)    1 = argmin Ψ(y)+ y− y(t) − W ∇Φ y(t) 2 y (8) 2  , W −1 converges in objective value to the optimal solution of (7) with rate O(1/t). In order to extend the proof of [18, Theorem 3.1] we need to use a generalized version of [18, Lemma 2.1]. To this end we can use a result by Nesterov, given in [22], which is recalled in the following lemma for completeness. Lemma V.4 (Generalized Descent Lemma). Let Assumption V.1 hold, then for all s ∈ RN Φ(y + s) ≤ Φ(y) + s> ∇Φ(y) + where W := diag(w1 , . . . , wn ) satisfies wi ≤ 1 nLi 1 s 2 2 , W −1 for all i ∈ {1, . . . , n}.  Tighter conditions than the one given above can be found in [14] and in [23]. March 15, 2018 DRAFT 14 Theorem V.5. Let Assumption V.1 and V.3 hold and let {y(t)} be the sequence generated by iteration (8) applied to problem (7). Then for any t ≥ 1 y0 − y ? Γ(y(t)) − Γ(y ) ≤ 2t ? where W := diag(w1 , . . . , wn ) with wi ≤ 1 , nLi 2 W −1 , y0 is the initial condition and y ? is any minimizer of problem (7). Proof: The theorem is proven by following the same arguments as in [18, Theorem 3.1], but using Lemma V.4 in place of [18, Lemma 2.1]. 2) Randomized block coordinate descent: Next, we present a randomized version of the weighted proximal gradient, proposed in [20, Algorithm 2] as Uniform Coordinate Descent for Composite functions (UCDC) algorithm. Algorithm 4 UCDC Initialization: y(0) = y0 FOR : t = 0, 1, 2, . . . DO choose it ∈ {1, . . . , n} with probability 1 n compute n o  T (it ) y(t) = argmin Vit (y(t), s) Ni s∈R (9a) t where Vit (y, s) := ∇it Φ(y)> s+ Lit ksk2 +ψit (yit +s) 2 (9b) update  y(t + 1) = y(t) + Uit T (it ) y(t) . (10) The convergence result for UCDC is given in [20, Theorem 5], here reported for completeness. Theorem V.6 ( [20, Theorem 5]). Let Assumptions V.1, V.2 and V.3 hold. Let Γ? denote the   optimal cost of problem (7). Then, for any ε ∈ 0, Γ(y0 ) − Γ? , there exists t̄(ε, ρ) > 0 such that if y(t) is the random sequence generated by UCDC (Algorithm 4) applied to problem (7), March 15, 2018 DRAFT 15 then for all t ≥ t̄ it holds that   Pr Γ(y(t)) − Γ? ≤ ε ≥ 1 − ρ, where y0 ∈ RN is the initial condition and ρ ∈ (0, 1) is the target confidence.  B. Analysis of the synchronous algorithm We start this section by recalling some useful properties of conjugate functions that will be useful for the convergence analysis of the proposed distributed algorithms. Lemma V.7 ( [24], [25]). Let ϕ be a closed, strictly convex function and ϕ∗ its conjugate function. Then   ∇ϕ∗ (y) = argmax y > x−ϕ(x) = argmin ϕ(x)−y > x . x x Moreover, if ϕ is strongly convex with convexity parameter σ, then ∇ϕ∗ is Lipschitz continuous with Lipschitz constant given by σ1 .  In the next lemma we establish some important properties of problem (6) that will be useful to analyze the proposed distributed dual proximal algorithms. Lemma V.8. Let Φ(y) := F ∗ (y) and Ψ(y) := G∗ (y) consistently with the notation of problem (7) in Appendix V-A. Problem (6) satisfies Assumption V.1 (block Lipschitz continuity of ∇Φ), with (block) Lipschitz constants given by s X1 1 1 2 Li = + + , σi2 j∈N σi σj i i ∈ {1, . . . , n}, Assumption V.2 (separability of Ψ) and Assumption V.3 (feasibility).  Proof: The proof is split into blocks, one for each assumption. Block Lipschitz continuity of ∇F ∗ : We show that the gradient of F ∗ is block coordinate-wise Lipschitz continuous. The i-th block component of ∇F ∗ is   ∗ ∇Λi F (y) , ∇i F ∗ (y) =  ∇µi F ∗ (y) March 15, 2018 DRAFT 16 where the block-component of ∇Λi F ∗ associated to neighbor j is given by ∇λj F ∗ and is equal i to ∗ ∇λj F (y) = i ∇fi∗  − − X λki k∈Ni ∇fj∗  − − X k∈Nj   − µi λkj λjk λik −   − µj . By Lemma V.7 both ∇fi∗ and ∇fj∗ are Lipschitz continuous with Lipschitz constants respectively, thus also ∇λj F ∗ is Lipschitz continuous with constant Lij = i 1 σi + 1 σi and 1 σj 1 . σj By using the (Euclidean) 2-norm, we have that ∇Λi F ∗ (y) is Lipschitz continuous with constant qP 2 j∈Ni Lij . Similarly, the gradient of F ∗ with respect to µi is   X  k i ∗ ∗ λi − λk − µi ∇µi F (y) = ∇fi − k∈Ni and is Lipschitz continuous with constant 1 . σi Finally, we conclude that ∇i F ∗ (y) is Lipschitz continuous with constant s s X X1 1 1 1 2 2 Li = L = + + + . σi2 j∈N ij σi2 j∈N σi σj i i P n Separability of G∗ : By definition G∗ (y) = i=1 gi∗ (µi ), where µi is a component of the block yi . P P Thus, denoting G∗i (yi ) := gi∗ (µi ), it follows immediately G∗ (y) = ni=1 gi∗ (µi ) = ni=1 G∗i (yi ). Feasibility: From Assumption II.2 and the convexity condition on fi and gi , strong duality holds, i.e., dual problem (6) is feasible and admits at least a minimizer y ? . Next, we recall how the proximal operators of a function and its conjugate are related. Lemma V.9 (Moreau decomposition, [26]). Let ϕ : Rd → R∪{+∞} be a closed, convex function and ϕ∗ its conjugate. Then, ∀x ∈ Rd , x = proxϕ (x) + proxϕ∗ (x).  Lemma V.10 (Extended Moreau decomposition). Let ϕ : Rd → R ∪ {+∞} be a closed, convex function and ϕ∗ its conjugate. Then, for any x ∈ Rd and α > 0, it holds x  x = proxαϕ x + αprox 1 ϕ∗ . α α March 15, 2018 DRAFT 17 Proof: Let h(x) = αϕ(x), then from the Moreau decomposition in Lemma V.9, it holds x = proxh (x) + proxh∗ (x). To prove the result we simply need to compute proxh∗ (x) in terms  of ϕ∗ . First, from the definition of conjugate function we obtain h∗ (x) = αϕ∗ αx . Then, by using the definition of proximal operator and standard algebraic properties from minimization,  it holds true that proxh∗ (x) = αprox 1 ϕ∗ αx , so that the proof follows. α The next lemma shows how the (weighted) proximal operator of G∗ can be split into local proximal operators that can be independently carried out by each single node. > > with Λi ∈ RD and Lemma V.11. Let y = [y1> . . . yn> ]> ∈ Rn(D+d) where yi = [Λ> i µi ] P µi ∈ Rd , i ∈ {1, . . . , n}. Let G∗ (y) = ni=1 gi∗ (µi ), then for a diagonal weight matrix Dα = diag(α1 , . . . , αn ) > 0, the proximal operator proxDα ,G∗ evaluated at y is given by   Λ1    prox ∗ (µ )  1   α1 g 1    . . .. proxDα ,G∗ y =        Λ n   proxαn gn∗ (µn )  D > > and vi ∈ Rd , be a Proof: Let η = [η1> . . . ηn> ]> ∈ Rn(D+d) , with ηi = [u> i vi ] , ui ∈ R > > variable with the same block structure of y = [y1> . . . yn> ]> ∈ Rn(D+d) , with yi = [Λ> i µi ] (as defined in (5)). By using the definition of weighted proximal operator and the separability of both G∗ and the norm function, we have   proxDα ,G∗ y := argmin η∈Rn(D+d) 1 G (η) + η − y 2 ∗ 2  −1 Dα X  n  1 1 ∗ 2 2 = argmin gi (vi ) + kui −Λi k + kvi −µi k . 2α 2α η i i i=1 March 15, 2018 DRAFT 18 The minimization splits on each component ηi of η, giving   2 argmin ku1 − Λ1 k    n u1 o 1   kv1 − µ1 k2   argmin g1∗ (v1 ) +   v1 2α1    ..   proxDα ,G∗ y =  .    2   argmin ku − Λ k n n   un   n o   1 ∗ 2 argmin gn (vn ) + kvn − µn k 2αn vn so that the proof follows from the definition of proximal operator. We are ready to show the convergence of the Distributed Dual Proximal Gradient introduced in Algorithm 1. Theorem V.12. For each i ∈ {1, . . . , n}, let fi be a proper, closed and strongly convex extended real-valued function with strong convexity parameter σi > 0, and let gi be a proper convex extended real-valued function. Suppose that in Algorithm 1 the local step-size αi is chosen such that 0 < αi ≤ 1 , nLi with Li given by s X1 1 2 1 Li = + , + σi2 j∈N σi σj i ∀ i ∈ {1, . . . , n}. (11) > Then the sequence y(t) = [y1 (t) . . . yn (t)> ]> generated by the Distributed Dual Proximal Gradient (Algorithm 1) satisfies ? Γ(y(t)) − Γ(y ) ≤ y0 − y ? 2t 2 −1 Dα , where y ? is any minimizer of (6), y0 = [y1 (0)> . . . yn (0)> ]> is the initial condition and Dα := diag(α1 , . . . , αn ). Proof: To prove the theorem, we proceed in two steps. First, from Lemma V.8, problem (6) satisfies the assumptions of Theorem V.5 and, thus, a (deterministic) weighted proximal gradient solves the problem. Thus, we need to show that Distributed Dual Proximal Gradient (Algorithm 1) is a weighted proximal gradient scheme. The weighted proximal gradient algorithm, (8), applied to problem (6), with W := Dα , is given by   y(t + 1) = proxDα ,G∗ y(t) − Dα ∇F ∗ y(t) . March 15, 2018 (12) DRAFT 19 Now, by Lemma V.8, Li given in (11) is the Lipschitz constant of the i-th block of ∇F ∗ . Thus, using the hypothesis αi ≤ 1 , nLi we can apply Theorem V.5, which ensures convergence with a rate of O(1/t) in objective value. In order to disclose the distributed update rule, we first split (12) into two steps, i.e., ỹ = y(t) − Dα ∇F ∗ y(t)  y(t + 1) = proxDα ,G∗ ỹ  (13a) (13b) and, then, compute explicitly each component of both the equations. Focusing on (13a) and considering that Dα is diagonal, we can write the i-th block component of ỹ as    Λ̃i ỹi =   = yi (t) − αi ∇i F ∗ y(t) , µ̃i where the explicit update of the block-component of Λ̃i associated to neighbor j is ∂F ∗ (y) ∂λji y=y(t)    X j k i ∗ (λi (t) − λk (t)) − µi (t) = λi (t) + αi ∇fi − λ̃ji = λji (t) − αi (14) k∈Ni − ∇fj∗  − X (λkj (t) k∈Nj − λjk (t)) − µj (t)  and the explicit update of µ̃i is ∂F ∗ (y) ∂µi y=y(t)  X  = µi (t) + αi ∇fi∗ − (λki (t) − λik (t)) − µi (t) . µ̃i = µi (t) − αi (15) k∈Ni Now, denoting x?i (t) := ∇fi∗  − X k∈Ni (λki (t) − λik (t))  − µi (t) , from Lemma V.7 it holds x?i (t) = argmin xi March 15, 2018   X    > k i xi λi (t)−λk (t) +µi (t) +fi (xi ) . k∈Ni DRAFT 20 Thus, we can rewrite (14) and (15) in terms of x?i obtaining h i λ̃ji = λji (t) + αi x?i (t) − x?j (t) µ̃i = µi (t) + αi x?i (t). Finally, the last step consists of applying the rule (13b) to ỹ. In order to highlight the distributed update, we rewrite (13b) in a component-wise fashion, i.e.     Λ1 (t + 1) Λ̃1      µ (t + 1)   µ̃   1   1      .  = proxD ,G∗  ...  , .. y(t + 1) =  α         Λn (t + 1) Λ̃n      µn (t + 1) µ̃n and applying Lemma V.11 and Lemma V.10 with ϕi = gi∗ , we obtain     Λ̃1 Λ̃1       prox ∗ (µ̃ )   µ̃1    1   µ̃1 −α1 prox 1 g1 α α1 g 1 1 α     1     .. .. y(t + 1) = , = . .             Λ̃n Λ̃ n      µ̃n −αn prox 1 gn αµ̃nn proxα g∗ (µ̃n ) αn n n so that the proof follows. Remark V.13 (Nesterov’s acceleration). We can include a Nesterov’s extrapolation step in the algorithm, which accelerates the algorithm (further details in [19]), attaining a faster O(1/t2 ) convergence rate in objective value. In order to implement the acceleration, each node needs to store a copy of the dual variables at the previous iteration. Thus, the update law in (13) would be changed in the following ỹ = y(t) − Dα ∇F ∗ y(t)  ŷ(t) = proxDα ,G∗ ỹ  y(t + 1) = ŷ(t) + θt (ŷ(t) − ŷ(t − 1)) . where θt represents the Nesterov overshoot parameter. March 15, 2018  DRAFT 21 C. Analysis of the node-based asynchronous algorithm In order to analyze the algorithm we start recalling some properties of i.i.d. exponential random variables. Let it ∈ {1, . . . , n}, t = 1, 2, . . . be the sequence identifying the generic tth triggered node. Assumption IV.2 implies that it is an i.i.d. uniform process on the alphabet {1, . . . , n}. Each triggering will induce an iteration of the distributed optimization algorithm, so that t will be a universal, discrete time indicating the t-th iteration of the algorithm itself. Thus, from an external, global perspective, the described local asynchronous updates result into an algorithmic evolution in which, at each iteration, only one node wakes up randomly, uniformly and independently from previous iterations. This variable will be used in the statement and in the proof of Theorem V.14. However, we want to stress that this iteration counter does not need to be known by the agents. Theorem V.14. For each i ∈ {1, . . . , n}, let fi be a proper, closed and strongly convex extended real-valued function with strong convexity parameter σi > 0, and let gi be a proper convex extended real-valued function. Suppose that in Algorithm 2 each local step-size αi is chosen such that 0 < αi ≤ 1 , Li with s Li = X1 1 2 1 + + , σi2 j∈N σi σj i ∀ i ∈ {1, . . . , n}. (16) Then the sequence y(t) = [y1 (t)> . . . yn (t)> ]> generated by the Asynchronous Distributed Dual Proximal Gradient (Algorithm 2) converges in objective value with high probability, i.e., for any  ε ∈ 0, Γ(y0 ) , where y0 = [y1 (0)> . . . yn (0)> ]> is the initial condition, and target confidence 0 < ρ < 1, there exists t̄(ε, ρ) > 0 such that for all t ≥ t̄ it holds   Pr Γ(y(t)) − Γ? ≤ ε ≥ 1 − ρ, where Γ? is the optimal cost of problem (6). Proof: To prove the theorem, we proceed in two steps. First, we show that we can apply the Uniform Coordinate Descent for Composite functions (Algorithm 4) to solve problem (6). Second, we show that, when applied to this problem, Algorithm 4 gives the iterates of our Asynchronous Distributed Dual Proximal Gradient. The first part follows immediately by Lemma V.8, which asserts that problem (6) satisfies the assumptions of Theorem V.6, so that Algorithm 4 solves it. March 15, 2018 DRAFT 22 Next, we show that the two algorithms have the same update. First, by Lemma V.8, Li given in (16) is the Lipschitz constant of the i-th block of ∇F ∗ . Thus, in the rest of the proof, following [20], we set αi = 1 Li (the maximum allowable value). Clearly the convergence is preserved if a smaller stepsize is used. Consistently with the notation in Algorithm 4, let it denote the uniform-randomly selected n o index at iteration t. Thus, T (it ) (y(t)) = argminsi ∈RNit Vit (y(t), sit ) defined in (9) can be t written in terms of a proximal gradient update applied to the it -th block component of y. In fact, by definition, for our function Γ = F ∗ + G∗ , we have n >  (it ) y(t) = argmin ∇it F ∗ y(t) s T Ni s∈R t o Lit 2 ∗ ksk + git yit (t) + s . + 2 In order to apply the formal definition of a proximal gradient step, we add a constant term and introduce a change of variable given by s̄ := yit (t) + s, obtaining n   T (it ) y(t) = −yit (t) + argmin Ui>t F ∗ y(t) Ni s̄∈R t o > Li + ∇it F ∗ y(t) (s̄−yit (t)) + t ks̄−yit (t)k2 + gi∗t (s̄) , 2 which yields T (it )  y(t) = −yit (t)+prox   1 ∗ 1 ∇i F y(t) . gi∗ yit (t)− Li Lit t t t Thus, update (10) in fact changes only the component yit of y, which is updated as  yit (t + 1) = yit (t) + T (it ) y(t)   1 = prox 1 gi∗ yit (t) − ∇it F ∗ y(t) , Li Lit t t (17) while all the other ones remain unchanged, i.e., yi (t + 1) = yi (t) for all i ∈ {1, . . . , n} with i 6= it . Following the same steps as in the proof of Theorem V.12, we split the update in (17) into a gradient and a proximal steps. The gradient step is given by    Λ̃it 1 ỹit =   = yit (t) − ∇it F ∗ y(t) L it µ̃it March 15, 2018 DRAFT 23 where Λ̃it and µ̃it are the same as in (14) and (15) respectively. The proximal operator step turns out to be     Λit (t + 1) Λ̃  = prox 1 g∗  it  . yit (t + 1) =  i Li t t µit (t + 1) µ̃it Applying Lemma V.11 on the it -th block with αit = 1/Lit , we can rewrite (17) as     Λ̃it (t) Λ (t + 1)  it =   , µ̃it − L1i proxLit git Lit µ̃it µit (t + 1) (18) t where each component of Λ̃it is given by λ̃jit = λjit (t) i 1 h ? ? x (t) − xj (t) , + Lit it and µ̃it = µit (t) + 1 ? x (t), Lit it with x?i (t) = argmin xi   X    > k i xi λi (t)−λk (t) +µi (t) +fi (xi ) . k∈Ni Here we have used again the property from Lemma V.7   X (λki (t) − λik (t)) − µi (t) = x?i (t). ∇fi∗ − k∈Ni Now, from Assumption IV.2 a sequence of nodes it , t = 1, 2..., becomes active according to a uniform distribution, so that each node triggering can be associated to an iteration of Algorithm 4 given by the update in (18). That is, only a node it is active in the network, which performs an update of its dual variables yit . In order to perform the local update, the selected node it needs to know the most updated information after the last triggering. As regards the neighbors’ dual variables λijt , j ∈ Nit , they have been broadcast by each j ∈ Nit the last time it has become active. Regarding the primal variables x?j , j ∈ Nit , the situation is a little more tricky. Indeed, x?j , j ∈ Nit , may have changed in the past due to either j or one of its neighbors has become active. In both cases j has to broadcast to it its updated dual variable (either because it has become active or because it has received, in idle, an updated dual variable from one of its neighbors). March 15, 2018 DRAFT 24 Remark V.15. Differently from the synchronous algorithm, in the asynchronous version nodes do not need to know the number of nodes, n, in order to set their local step-size. In fact, each node i can set its parameter αi by only knowing the convexity parameters σi and σj , j ∈ Ni . Remark V.16. If a strongly convex, separable penalty term is added to the dual function Γ = F ∗ +G∗ , then it becomes strongly convex, so that a stronger result from [20, Theorem 7] applies, i.e., linear convergence with high probability is guaranteed. Note that strong convexity of the dual function Γ is obtained if the primal function has Lipschitz continuous gradient, [27, Chapter  X, Theorem 4.2.2]. D. Analysis of the edge-based asynchronous algorithm The convergence of Algorithm 3 relies essentially on the same arguments as in Theorem V.14, but with a different block partition. In fact, we have to split the y variable into |E| blocks (with |E| the number of edges of G). Notice that since the dual variables µi are only n, they need to be associated to a subset of edges. Thus, the variable y is split in blocks yij given by yij = [λji λij µi ] if j = jµi and yij = [λji λij ] otherwise. This is why in the algorithm µi is updated only when neighbor jµi becomes active. VI. M OTIVATING OPTIMIZATION SCENARIOS AND NUMERICAL COMPUTATIONS A. Constrained optimization As first concrete setup, we consider a separable constrained convex optimization problem min x n X hi (x) i=1 subj. to x ∈ n \ i=1 (19) Xi ⊆ Rd where each hi is a strongly convex function and each Xi is a closed, convex set. This problem structure is widely investigated in distributed and large-scale optimization as shown in the literature review. Notice that, as pointed out in our discussion in the introduction, we assume strong convexity of hi , but we do not need to assume smoothness. March 15, 2018 DRAFT 25 We can rewrite this problem by transforming the constraints into additional terms in the objective function, by using indicator functions IXi associated to each Xi , n   X min hi (x) + IXi (x) . x∈Rd i=1 Since each Xi is a convex set, then IXi is a convex function. Thus the problem can be mapped in our distributed setup (1) by setting fi (x) = hi (x) and gi (x) = IXi (x). Treating the local constraints in this way, we have to perform a local unconstrained minimization step when computing x?i , while the local feasibility is entrusted to the proximal operator of gi . In fact, the proximal operator of a convex indicator function reduces to the standard Euclidean projection, i.e. o n  1 proxIX v = argmin IX (x) + kx − vk2 = ΠX (v). 2 x Remark VI.1. When considering quadratic costs hi , we can benefit greatly from a numerical point of view. In fact, an unconstrained quadratic program can be solved via efficient methods, which often result in division-free algorithms (possibly after some off-line precomputations), and can be implemented in fixed-point arithmetic, see [21] for further details.  An attractive feature of our setup is that one can conveniently decide how to rewrite the local constraints. In the formulation above, we suggested to include the local constraint into gi . But it is also reasonable to include the constraint into fi , by consider the indicator function in its definition, i.e., define fi (x) :=   hi (x)  +∞ if x ∈ Xi (20) otherwise. and, thus, have gi identically zero (still convex). This strategy results into an algorithm which is basically an asynchronous distributed dual decomposition algorithm. Notice that with this choice recursive feasibility is obtained provided that the local algorithm solves the minimization in an interior point fashion. Between these two extreme scenarios one could also consider other possibilities. Indeed, it could be the case that one can benefit from splitting each local constraint Xi into two distinct contributions, i.e., Xi = Yi ∪Zi . In this way the indicator function of Yi (e.g. the positive orthant) March 15, 2018 DRAFT 26 could be included into fi , allowing for a simpler constrained local minimization step, while the other constraint could be mapped into the second term as gi (x) = IZi (x). The choice in (20) leads to a simple observation: leaving the gi equal to zero seems to be a waste of a degree of freedom that could be differently exploited, e.g., by introducing a regularization term. B. Regularized and constrained optimization As highlighted in the previous paragraph, the flexibility of our algorithmic framework allows us to handle, together with local constraints, also a regularization cost through the gi . Regularize the solution is a useful technique in many applications as sparse design, robust estimation in statistics, support vector machine (SVM) in machine learning, total variation reconstruction in signal processing and geophysics, and compressed sensing. In these problems, the cost fi is a loss function representing how the predictions based on a theoretical model mismatch the real data. Next, we focus on the most widespread choice for the loss function, which is the least square cost, giving rise to the following optimization problem min x n X i=1 kAi x − bi k2 (21) where Ai are data/regressors and bi are labels/observations. A typical challenge arising in regression problems is due to the fact that problem (21) is often ill-posed and standard algorithms easily incur in over-fitting phenomena. A viable technique to prevent over-fitting consists of adding a regularization cost; usual choices are the `2 -norm, also referred as Tikhonov regularization or ridge regression, or the `1 -norm, which leads to the so called LASSO (Least Absolute Shrinkage and Selection Operator) problem min x n X i=1 kAi x − bi k2 + γkxk1 where γ is a positive scalar. In some cases (as, e.g., in distributed estimation [28]) one may be interested in having the solution bounded in a given box or leaving in a reduced subspace. This gives rise to the so called constrained LASSO problem, see, e.g., [23], [29], [30]. March 15, 2018 DRAFT 27 As discussed before, our setup can simultaneously manage a constrained and regularized problem as the constrained lasso. The first way to map the problem in our setup is by defining   kAi x − bi k2 if x ∈ Xi (22) fi (x) :=  +∞ otherwise and setting gi (x) := γ kxk1 . n The proximal operator of the `1 -norm admits an analytic solution which is well known as soft thresholding operator. When applied to a vector v ∈ Rd (with `-th component v` ), it gives a vector in Rd whose `-th component is    v` − γ, v` > γ     proxγk·k1 (v) = 0, |v` | ≤ γ −γ  `    v + γ, v < −γ ` ` γ i.e., it thresholds the components of v which are in modulus greater then γ, see, e.g., [18], [31]. Alternatively, we may include both the constraint Xi and the regularization term into the gi , obtaining an unconstrained local minimization at each node. This choice is particularly appealing when the constraint Xi is a box, i.e., Xi = {v ∈ Rd | lb` ≤ v` ≤ ub` for all ` ∈ {1, . . . , d}}. In this case the proximal operator of gi becomes a saturated version of the soft-thresholding operator, as depicted in Figure 1. ub` −γ γ lb` Fig. 1. Saturated soft-thresholding operator. March 15, 2018 DRAFT 28 C. Numerical tests In this section we provide a numerical example showing the effectiveness of the proposed algorithms. We test the proposed distributed algorithms on a constrained LASSO optimization problem, min lb≤x≤ub n X i=1 kAi x − bi k2 + γkxk1 , 3 where x ∈ R is the decision variable, and Ai ∈ R150×3 and bi ∈ R150 represent respectively the data matrix and the labels associated with examples assigned to node i. The inequality lb ≤ x ≤ ub is meant component-wise. We randomly generate the LASSO data following the idea suggested in [31]. Each element of Ai is ∼ N (0, 1), and bi s are generated by perturbing a “true” solution xtrue (which has around a half nonzero entries) with an additive noise v ∼ N (0, 10−2 I). Then the matrix Ai and the vector bi are normalized with respect to the number of local samples at each node. The box bounds are h i> h i> set to lb = −0.8 −0.8 −0.8 and ub = 0.8 0.8 0.8 , while the regularization parameter is γ = 0.1. To match the problem with our distributed framework, we introduce copies xi of the decision variable x. Consistently, we define the local functions fi as the least-square costs in (22), where each Xi is the box defined by lb and ub. We let each gi be the `1 -norm regularization term with local parameter γ/n. We initialize to zero the dual variables λji , j ∈ Ni , and µi for all i ∈ {1, . . . , n}, and use as step-sizes αi = Li , where Li has the expression in (11), with σi being the smallest eigenvalue of A> i Ai . We consider an undirected connected Erdős-Rényi graph G, with parameter 0.2, connecting n = 50 nodes. We run both the synchronous and the asynchronous algorithms over this underlying graph and we stop them if the difference between the current dual cost and the optimal value drops below the threshold of 10−6 . Figure 2 shows the difference between the dual cost at each iteration t and the optimal value, Γ(y(t)) − Γ? , in a logarithmic scale. In particular, the rates of convergence of the synchronous (left) and asynchronous (right) algorithms are shown. For the asynchronous algorithm, we normalize the iteration counter t with respect the number of agents n. March 15, 2018 DRAFT 29 100 10−1 10−1 10−2 10−2 Γ(y(t)) − Γ? Γ(y(t)) − Γ? 100 10−3 10−3 10−4 10−4 10−5 10−5 10−6 Fig. 2. 0 10−6 100 200 300 t 0 2 4 t/n 6 8 Evolution of the cost error, in logarithmic scale, for the synchronous (left) and node-based asynchronous (right) distributed algorithms. Then we concentrate on the asynchronous algorithm. In Figure 3 we plot the evolution of the (three) components of the primal variables, x?i (t), i ∈ {1, . . . , n}. The horizontal dotted lines represent the optimal solution. It is worth noting that the optimal solution has a first component slightly below the constraint boundary, a second component equal to zero, and a third component on the constraint boundary. This optimal solution can be intuitively explained by the “simultaneous influence” of the box constraints (which restrict the set of admissible values) and the regularization term (which enforces sparsity of the vector x). In the inset the first iterations for a subset of selected nodes are highlighted, in order to better show the transient, piece-wise constant behavior due to the gossip update and the effect of the box constraint on each component. 0.8 0 −0.5 0.75 0.7 0.65 50 100 150 200 50 100 150 200 250 t 1 0 0.05 −0.5 0 −0.05 −1 50 0 20 40 60 80 100 100 150 200 250 t x?i (t), x? (3rd component) 0.5 −1 Fig. 3. 0.5 x?i (t), x? (2nd component) x?i (t), x? (1st component) 1 0.5 0 −0.5 −1 0 −0.2 −0.4 −0.6 −0.8 50 0 20 40 60 80 100 100 150 200 250 t Evolution of the three components of the primal variables x?i (t), i ∈ {1, . . . , n}, for the node-based Asynchronous Distributed Dual Proximal Gradient. March 15, 2018 DRAFT 30 Specifically, for the first component, it can be seen how the temporary solution of some agents hits the boundary in some iterations (e.g., one of them hits the boundary from iteration 50 to iteration 100) and then converges to the (feasible) optimal value. The second components are always inside the box constraint and converge to zero, while the third components start inside the box and then in a finite number of iterations hit the boundary. VII. C ONCLUSIONS In this paper we have proposed a class of distributed optimization algorithms, based on dual proximal gradient methods, to solve constrained optimization problems with non-smooth, separable objective functions. The main idea is to construct a suitable, separable dual problem via a proper choice of primal constraints and solve it through proximal gradient algorithms. Thanks to the separable structure of the dual problem in terms of local conjugate functions, a weighted proximal gradient update results into a distributed algorithm, where each node performs a local minimization on its primal variable, and a local proximal gradient update on its dual variables. As main contribution, two asynchronous, event-triggered distributed algorithms are proposed, in which the local updates at each node are triggered by a local timer, without relying on any global iteration counter. The asynchronous algorithms are shown to be special instances of a randomized, block-coordinate proximal gradient method on the dual problem. The convergence analysis is based on a proper combination of duality theory, coordinate-descent methods, and properties of proximal operators. R EFERENCES [1] F. Zanella, D. Varagnolo, A. Cenedese, G. Pillonetto, and L. Schenato, “Asynchronous Newton-Raphson consensus for distributed convex optimization,” in 3rd IFAC Workshop on Distributed Estimation and Control in Networked Systems, 2012. [2] W. Shi, Q. Ling, G. Wu, and W. Yin, “Extra: An exact first-order algorithm for decentralized consensus optimization,” SIAM Journal on Optimization, vol. 25, no. 2, pp. 944–966, 2015. [3] D. Jakovetic, J. M. Freitas Xavier, and J. M. Moura, “Convergence rates of distributed Nesterov-like gradient methods on random networks,” IEEE Transactions on Signal Processing, vol. 62, no. 4, pp. 868–882, 2014. [4] A. Nedić and A. Olshevsky, “Distributed optimization over time-varying directed graphs,” vol. 60, no. 3, pp. 601–6015, 2015. [5] M. Akbari, B. Gharesifard, and T. Linder, “Distributed subgradient-push online convex optimization on time-varying directed graphs,” in IEEE 52nd Annual Allerton Conference on Communication, Control, and Computing (Allerton), 2014. March 15, 2018 DRAFT 31 [6] S. S. Kia, J. Cortés, and S. Martı́nez, “Distributed convex optimization via continuous-time coordination algorithms with discrete-time communication,” Automatica, vol. 55, pp. 254–264, 2015. [7] A. I. Chen and A. Ozdaglar, “A fast distributed proximal-gradient method,” in IEEE 50th Annual Allerton Conference on Communication, Control, and Computing (Allerton), 2012, pp. 601–608. [8] K. I. Tsianos, S. Lawlor, and M. G. Rabbat, “Consensus-based distributed optimization: Practical issues and applications in large-scale machine learning,” in IEEE 50th Annual Allerton Conference on Communication, Control, and Computing (Allerton), 2012, pp. 1543–1550. [9] S. Lee and A. Nedić, “Distributed random projection algorithm for convex optimization,” IEEE Journal of Selected Topics in Signal Processing, vol. 7, no. 2, pp. 221–229, 2013. [10] I. Necoara, “Random coordinate descent algorithms for multi-agent convex optimization over networks,” IEEE Transactions on Automatic Control, vol. 58, no. 8, pp. 2001–2012, 2013. [11] E. Wei and A. Ozdaglar, “On the O(1/k) convergence of asynchronous distributed alternating direction method of multipliers,” in IEEE Global Conference on Signal and Information Processing (GlobalSIP), 2013, pp. 551–554. [12] F. Iutzeler, P. Bianchi, P. Ciblat, and W. Hachem, “Explicit convergence rate of a distributed alternating direction method of multipliers,” arXiv preprint arXiv:1312.1085, 2013. [13] P. Bianchi, W. Hachem, and F. Iutzeler, “A stochastic coordinate descent primal-dual algorithm and applications,” in IEEE International Workshop on Machine Learning for Signal Processing (MLSP), 2014, pp. 1–6. [14] P. Richtárik and M. Takáč, “Parallel coordinate descent methods for big data optimization,” Mathematical Programming, pp. 1–52, 2015. [15] F. Facchinei, G. Scutari, and S. Sagratella, “Parallel selective algorithms for nonconvex big data optimization,” IEEE Transactions on Signal Processing, vol. 63, no. 7, pp. 1874–1889, 2015. [16] G. Notarstefano and F. Bullo, “Distributed abstract optimization via constraints consensus: Theory and applications,” vol. 56, no. 10, pp. 2247–2261, 2011. [17] M. Bürger, G. Notarstefano, and F. Allgöwer, “A polyhedral approximation framework for convex and robust distributed optimization,” IEEE Transactions on Automatic Control, vol. 59, no. 2, pp. 384–395, 2014. [18] A. Beck and M. Teboulle, “A fast iterative shrinkage-thresholding algorithm for linear inverse problems,” SIAM journal on imaging sciences, vol. 2, no. 1, pp. 183–202, 2009. [19] Y. Nesterov, “Gradient methods for minimizing composite functions,” Mathematical Programming, vol. 140, no. 1, pp. 125–161, 2013. [20] P. Richtárik and M. Takáč, “Iteration complexity of randomized block-coordinate descent methods for minimizing a composite function,” Mathematical Programming, vol. 144, no. 1-2, pp. 1–38, 2014. [21] B. O’Donoghue, G. Stathopoulos, and S. Boyd, “A splitting method for optimal control,” IEEE Transactions on Control Systems Technology, vol. 21, no. 6, pp. 2432–2442, 2013. [22] Y. Nesterov, “Efficiency of coordinate descent methods on huge-scale optimization problems,” SIAM Journal on Optimization, vol. 22, no. 2, pp. 341–362, 2012. [23] I. Necoara and D. Clipici, “Parallel random coordinate descent method for composite minimization: Convergence analysis and error bounds,” SIAM Journal on Optimization, vol. 26, no. 1, pp. 197–226, 2016. [24] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge university press, 2004. [25] A. Beck and M. Teboulle, “A fast dual proximal gradient algorithm for convex minimization and applications,” Operations Research Letters, vol. 42, no. 1, pp. 1–6, 2014. March 15, 2018 DRAFT 32 [26] H. H. Bauschke and P. L. Combettes, Convex analysis and monotone operator theory in Hilbert spaces. Springer Science & Business Media, 2011. [27] J.-B. Hiriart-Urruty and C. Lemaréchal, “Convex analysis and minimization algorithms ii: Advanced theory and bundle methods,” Grundlehren der mathematischen Wissenschaften, vol. 306, 1993. [28] S. Kar and J. M. Moura, “Convergence rate analysis of distributed gossip (linear parameter) estimation: Fundamental limits and tradeoffs,” IEEE Journal of Selected Topics in Signal Processing, vol. 5, no. 4, pp. 674–690, 2011. [29] G. M. James, C. Paulson, and P. Rusmevichientong, “The constrained LASSO,” Citeseer, Tech. Rep., 2012. [30] H. Xu, D. J. Eis, and P. J. Ramadge, “The generalized lasso is reducible to a subspace constrained lasso,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2013, pp. 3268–3272. [31] N. Parikh and S. Boyd, “Proximal algorithms,” Foundations and Trends in Optimization, vol. 1, no. 3, pp. 123–231, 2013. Giuseppe Notarstefano has been an Assistant Professor (Ricercatore) at the Università del Salento (Lecce, Italy) since February 2007. He received the Laurea degree “summa cum laude” in Electronics Engineering from the Università di Pisa in 2003 and the Ph.D. degree in Automation and Operation Research from the Università di Padova in April 2007. He has been visiting scholar at the University of Stuttgart, University of California Santa Barbara and University of Colorado Boulder. His research interests include distributed optimization, motion coordination in multi-agent networks, applied nonlinear optimal control, and modeling, trajectory optimization and aggressive maneuvering of aerial and car vehicles. He serves as an Associate Editor in the Conference Editorial Board of the IEEE Control Systems Society and for the European Control Conference, IFAC World Congress and IEEE Multi-Conference on Systems and Control. He coordinated the VI-RTUS team winning the International Student Competition Virtual Formula 2012. He is recipient of an ERC Starting Grant 2014. Ivano Notarnicola has been a Ph.D. student in Engineering of Complex Systems at the Università del Salento (Lecce, Italy) since November 2014. He received the Laurea degree “summa cum laude” in Computer Engineering from the Università del Salento in 2014. He was a visiting student at the Institute of System Theory (Stuttgart, Germany) from March to June 2014. His research interests include distributed optimization and randomized algorithms. March 15, 2018 DRAFT
3cs.SY
1 Secure Video Streaming in Heterogeneous Small Cell Networks with Untrusted Cache Helpers arXiv:1707.08050v3 [cs.IT] 29 Jan 2018 Lin Xiang, Student Member, IEEE, Derrick Wing Kwan Ng, Senior Member, IEEE, Robert Schober, Fellow, IEEE, and Vincent W.S. Wong, Fellow, IEEE Abstract—This paper studies secure video streaming in cacheenabled small cell networks, where some of the cache-enabled small cell base stations (BSs) helping in video delivery are untrusted. Unfavorably, caching improves the eavesdropping capability of these untrusted helpers as they may intercept both the cached and the delivered video files. To address this issue, we propose joint caching and scalable video coding (SVC) of video files to enable secure cooperative multiple-input multipleoutput (MIMO) transmission and, at the same time, exploit the cache memory of both the trusted and untrusted BSs for improving the system performance. Considering imperfect channel state information (CSI) at the transmitters, we formulate a two-timescale non-convex mixed-integer robust optimization problem to minimize the total transmit power required for guaranteeing the quality of service (QoS) and secrecy during video streaming. We develop an iterative algorithm based on a modified generalized Benders decomposition (GBD) to solve the problem optimally, where the caching and the cooperative transmission policies are determined via offline (long-timescale) and online (short-timescale) optimization, respectively. Furthermore, inspired by the optimal algorithm, a low-complexity suboptimal algorithm based on a greedy heuristic is proposed. Simulation results show that the proposed schemes achieve significant gains in power efficiency and secrecy performance compared to several baseline schemes. Index Terms—Physical layer security, untrusted nodes, wireless caching, MIMO, non-convex optimization, resource allocation. I. I NTRODUCTION MALL cells are among the most promising solutions for meeting the enormous capacity requirements introduced by video streaming applications in the fifth-generation (5G) wireless networks [2]. By densely deploying low-power base stations (BSs), both the spectral and energy efficiencies of wireless communication systems can be improved significantly. However, to achieve these performance gains, highcapacity secure backhaul links are required to transport the video files from the Internet to the small cell BSs. While S The work of D. W. K. Ng was supported under Australian Research Council’s Discovery Early Career Researcher Award funding scheme (DE170100137). The work of R. Schober was supported by the Alexander von Humboldt Professorship Program. The work of V.W.S. Wong was supported by the Natural Sciences and Engineering Research Council of Canada. Part of this work has been accepted for presentation at the IEEE Global Commun. Conf. (Globecom), Singapore, Dec. 2017 [1]. L. Xiang and R. Schober are with the Institute for Digital Communications, Friedrich-Alexander University of Erlangen-Nuremberg, Erlangen 91058, Germany (Email: {lin.xiang, robert.schober}@fau.de). D. W. K. Ng is with the School of Electrical Engineering and Telecommunications, University of New South Wales, Sydney, NSW 2052, Australia (Email: [email protected]). V.W.S. Wong is with the Department of Electrical and Computer Engineering, University of British Columbia, Vancouver, BC V6T 1Z4, Canada (Email: [email protected]). wireless backhauling is usually preferred for small cells due to its low cost and high flexibility in deployment [3], the capacity provided by wireless backhauling is often insufficient, which deteriorates the overall system performance [4] and limits the maximum number of concurrent streaming users/connections. Moreover, since wireless transmission is susceptible to eavesdropping, the security of wireless backhauling is a fundamental concern for 5G wireless networks. Recently, wireless caching has been proposed as a viable solution to enhance the capacity of small cell backhauling [5]–[15]. Built upon the content-centric networking paradigm, in wireless caching, the most popular contents are pre-stored at the access points or BSs in close proximity of the user equipments (UEs). Consequently, the backhaul traffic is offloaded by reusing the cached content [5], [6]. Caching as an alternative to small cell backhauling was first investigated in [7], where caching was shown to also substantially reduce the average downloading delay. Besides, caching improves the energy efficiency of wireless backhauling systems as was shown in [8]. In [9], caching was optimized to facilitate powerefficient cooperative multiple-input multiple-output (MIMO) transmission in small cell networks. In [10], joint caching and buffering for small cell networks was proposed to overcome the backhaul capacity bottleneck and the half-duplex transmission constraint simultaneously to enable fast downloading of video files. In [11], [12], coded caching was introduced, which reduces the backhaul load by exploiting coded multicast transmission for simultaneous delivery of different files. Coded caching was extended to various network scenarios, see [13]– [15] and references therein. On the other hand, although communication secrecy is of high importance in wireless networks, providing security for networks employing wireless caching has been a major challenge. This is because current video streaming applications, e.g. YouTube and Netflix, mainly rely on end-toend encryption schemes such as the hypertext transfer protocol secure (HTTPS) [23] to ensure communication security. However, with such schemes, the benefits of content-centric networking vanish as encrypted contents are uniquely defined for each user request and cannot be reused to serve other user requests [5]. For this reason, caching was mainly considered for content without security restrictions in the literature [5]– [10]. To overcome this limitation, physical layer security (PLS) schemes for wireless caching were proposed in [16]–[18]. As PLS techniques rely on wiretap channel coding instead of source encryption, content can still be reused at the wireless edge for secure wireless transmission, and hence, caching and PLS are compatible. In [16], cache-enabled cooperative MIMO 2 transmission was shown to be an effective physical layer mechanism for increasing the secrecy rate for video delivery in homogeneous cellular networks. However, a secure backhaul for cache placement was required in [16], which cannot always be guaranteed with wireless backhauling in practice. Considering an insecure backhaul, a secure cache placement strategy for heterogeneous cellular networks (HetNets) was developed in [17], whereby eavesdroppers tapping the insecure backhaul can be prevented from obtaining a sufficient number of coded packets for successful recovery of the video file. Assuming caching at the end users, the authors of [18] proposed a secure coded multicast scheme for relay networks to prevent end users and external eavesdroppers from intercepting the nonrequested and the delivered files, respectively. However, [16]–[18] optimistically assumed that the (cache) helpers can be trusted for cooperation and that the cache cannot be exploited for eavesdropping purposes. These assumptions may be unrealistic for HetNets. In particular, due to the distributed network architecture, cache-enabled small cell BSs can be untrusted helpers, i.e., they may be potential eavesdroppers1, and hence, may not cooperate altruistically [20]– [22]. Examples of untrusted helpers include home-owned and open-access small cell BSs which can be easily manipulated by third parties to eavesdrop premium video streaming services, for which they have not paid, and/or users’ private video files. In contrast to trusted small cell BSs deployed and owned by the service provider, at these untrusted small cell BSs, user data is left unprotected and prone to eavesdropping because the small cell BS itself is responsible for encrypting and decrypting the user data before forwarding it to the macro BS and the intended users, respectively [20]–[22]. Moreover, different from the case of cache-disabled eavesdroppers2 considered in [16], [17], the cache memory equipped at the untrusted helpers unfavorably enhances their eavesdropping capability as they can intercept both the cached and the delivered video data, and utilize the cached video data as side information to improve reception. Two fundamental questions need to be addressed when cache helpers are untrusted: (a) Can cooperation with untrusted helpers still yield secrecy benefits? That is, can the cache deployed at the untrusted helpers be utilized to improve the system performance? If so, (b) how to cache and cooperate intelligently to reap the possible performance gains? To our knowledge, state-of-the-art small cell networks perform only passive authentication of BSs and completely exclude untrusted BSs from participating in cooperative transmission [21], [22]. However, in this case, untrusted BS cannot be used to improve the system performance3. In [25], untrusted helpers without caching have been investigated for relay networks. It was shown that cooperation with compress-and-forward relays yields a positive secrecy rate even if the relays are untrusted. However, the problem studied in this paper is more challenging as the untrusted helpers can cache content to enhance their eavesdropping capability. Thus, the solutions proposed in [25] are not applicable and a new study is needed. In this paper, to facilitate secure cooperative transmission with untrusted cache helpers, we propose an advanced caching scheme that combines scalable video coding (SVC) and cooperative MIMO transmission. Specifically, each video file is encoded by SVC into base-layer subfiles, containing basic-quality and independently decodable video information, and enhancement-layer subfiles, containing high-quality video information which is decodable only after the base layer has been successfully decoded. By caching the enhancement-layer subfiles across all BSs and the baselayer subfiles only across trusted BSs, secure cooperation of all BSs is enabled by exploiting the encoding and decoding structure of SVC. Thereby, the large virtual transmit antenna array formed by all BSs that have cached the same subfile introduces additional degrees of freedom (DoFs) which may be utilized for secure and power-efficient video streaming. To reap the cache-enabled secrecy benefits, a centralized framework for caching and delivery optimization is adopted. Hence, the proposed architecture follows the cloud radio access network (CRAN) philosophy [4], [26] which has been advocated for next-generation HetNets for cooperative MIMO transmission [27], [28] and advanced resource allocation [29]. In the conference version of this paper [1], we investigated cache-enabled secure transmission by assuming perfect knowledge of all channels. However, in practice, the channel state information (CSI) gathered at the central controller, e.g. the macro BS, is imperfect due to quantization noise and feedback delay, which deteriorates the system performance and has to be taken into account for system design. To mitigate the information leakage from the trusted BSs to the untrusted BSs, artificial noise (AN) based jamming is applied in this paper. In the literature [29]–[31], AN has been employed to effectively reduce the receive signal-to-interference-plus-noise ratio (SINR) at the eavesdropper without interfering the desired users. In this paper, we consider cooperative AN transmission by the trusted BSs for power-efficient jamming to combat the eavesdropping of the untrusted BSs. By considering untrusted BSs and imperfect CSI, we jointly optimize caching and cooperative data and AN transmission for a secure and powerefficient system design. In particular, a two-timescale robust optimization problem is formulated for minimization of the transmit power required for secure video streaming under imperfect CSI. The main contributions of this paper can be summarized as follows: • 1 In this paper, we only consider passive eavesdroppers which remain silent during eavesdropping. Studying the case of active eavesdroppers such as jamming and spoofing attackers [19] is an interesting topic for future work. 2 The case considered in this paper is more general than that in [16], [17]. In fact, the eavesdroppers in [16], [17] can be viewed as untrusted helpers with zero cache capacity. 3 We note that, as untrusted BSs present the man-in-the-middle threat to wireless networks, HTTPS also cannot facilitate secure cooperative transmission [24]. • We study a new secrecy threat in small cell networks originating from untrusted cache helpers, i.e., cacheenabled eavesdropping small cell BSs. To facilitate secure cooperative MIMO transmission of trusted and untrusted small cell BSs, we propose a secure caching scheme based on SVC. We optimize the caching and the cooperative delivery policies for minimization of the transmit power while 3 guaranteeing quality-of-service (QoS) and communication secrecy under imperfect CSI. We show that the optimal delivery decisions can be obtained by semidefinite programming (SDP) relaxation with probability one under mild conditions. For the optimal caching decisions, an optimal iterative algorithm is developed based on a modified generalized Benders decomposition (GBD). To reduce the computational complexity, a polynomial-time suboptimal greedy scheme is also proposed. • Our simulation results show that the proposed robust schemes can efficiently exploit the cache capacities of both trusted and untrusted small cell BSs to enable powerefficient and secure video streaming in heterogeneous small cell networks. The remainder of this paper is organized as follows. In Section II, we present the system model for cooperative secure video streaming in the presence of untrusted cache helpers. The formulation and solution of the proposed optimization problem are provided in Sections III and IV, respectively. In Section V, we evaluate the performance of the proposed schemes and compare it to that of several baseline schemes. Finally, Section VI concludes the paper. Notations: R and C denote the sets of real and complex numbers, respectively; ℜ{z} denotes the real part of z ∈ C; IL is an L × L identity matrix; 1M×N and 0M×N are M × N all-one and all-zero matrices, respectively; (·)T and (·)H are the transpose and complex conjugate transpose operators, respectively; k·kℓ denotes the ℓ-norm of a vector; k·kF , tr(·), rank(·), det(·), and λmax (·) denote the Frobenius norm, the trace, the rank, the determinant, and the maximal eigenvalue of a square matrix, respectively; E(·) is the expectation operator; the circularly symmetric complex Gaussian distribution is denoted by CN (µ, C) with mean vector µ and covariance matrix C; ∼ stands for “distributed as”; diag(x) is a diagonal matrix with the diagonal elements given by vector x; |X | represents the cardinality of set X ; A  0 and A ≻ 0 indicate that matrix A is positive semidefinite and positive definite, respectively; finally, [x]+ stands for max{0, x}. II. S YSTEM M ODEL A. Network Topology We consider downlink wireless video streaming in a heterogeneous small cell network, where M small cell BSs, max each equipped with a cache memory of size Cm bits, are distributed in the coverage area of a macro BS, see Fig. 1(a). For convenience, a list with key notations used in this paper is provided in Table I. Let m ∈ M , {0, 1, . . . , M } be the BS index, where m = 0 refers to the macro BS. The macro BS is connected to the video server on the Internet via a dedicated secure high-capacity backhaul link such as optical fiber. For simplicity of notation, the backhaul to the macro BS is modeled as a cache with an equivalent capacity of C0max bits. In contrast, the small cell BSs are connected to the macro BS via wireless backhaul links for convenience of deployment. Assume that BSm , m ∈ M, is equipped with Nm antennas. P The total number of transmit antennas is denoted by N , m∈M Nm . The video server owns a library of F video files, indexed by F , {1, . . . , F }, to be streamed to K single-antenna UEs, indexed by K , {1, . . . , K}. The size of file f is Vf bits. Employing SVC coding, as utilized e.g. for wireless video delivery in the H.264/Moving Picture Expert Group (MPEG)4 standard [32]–[34], each video file f ∈ F is encoded into one base-layer subfile, (f, 0), and L − 1 enhancementlayer subfiles, (f, l), l ∈ {1, . . . , L − 1}, where the information embedded in enhancement layer l is used to refine the information contained in the previous layers 0, . . . , l − 1. Let L , {0, . . . , L − 1} be the index set of all layers. The size of subfile (f, l) is Vf,l bits. The base layer can be decoded independent of the enhancement layers. In contrast, enhancement layer l ∈ L\ {0} can be decoded only after layers 0, . . . , l − 1 have already been decoded. Therefore, the layers have to be decoded in a sequential manner [32]. Due to this specific encoding and decoding structure, only the base layer has to be protected in order to ensure communication secrecy. An eavesdropper, who cannot decode the base layer, will also not be able to decode any of the enhancement layers. The small cell BSs serve as helpers of the macro BS in delivering the video files. However, a subset of the small cell BSs are untrusted. These BSs may leak the cached video data and eavesdrop the transmitted video data while utilizing the cached data as side information. Let MT , {0, 1, . . . , J} and MU , {J + 1, . . . , M } denote the sets of Ptrusted and untrusted BSs having a total number of NT , m∈MT Nm P and NU , m∈MU Nm antennas, respectively, where J ≤ M and NT + NU = N . In this paper, we assume that the set of untrusted BSs, MU , is known. In practice, untrusted BSs may largely be home-owned and open-access small cell BSs which have insufficient security protection and can easily be compromised by third parties. Due to the eavesdropping and intensive processing, untrusted BSs may consume a large power and/or experience a long end-to-end latency even if the uplink and downlink throughputs are small. Hence, the power/delay versus throughput pattern of untrusted BSs is statistically different from that of trusted operator-owned BSs such that they constitute outliers. Therefore, by exploiting the power, delay, and throughput records of all BSs collected by the service providers, the set of untrusted BSs can be estimated by applying state-of-the-art outlier detection methods, e.g., supervised and unsupervised learning techniques [35]–[37]. The considered system is time slotted and the duration of a time slot is smaller than the channel coherence time. We consider a two-timescale control for caching and delivery. As shown in Fig. 1(b), the video files in the cache are updated every T0 time slots, referred as one period, based on the historical profiles of user preferences and CSI. In contrast, the video delivery decisions are determined in each time slot based on the actual requests of the users and instantaneous CSI. We have T0 ≫ 1, as the users’ preferences vary on a much slower scale (e.g., from day to day) than the user requests. For notational simplicity, we consider the system only during one typical period T0 , {1, . . . , T0 } and the corresponding time slots are indexed by t ∈ T0 . 4 TABLE I L IST OF K EY N OTATIONS . M, MT , MU ... Sets of M BSs, J trusted BSs, and M − J untrusted BSs Total number of antennas at all BSs, trusted BSs, and untrusted BSs Subset of cooperating BSs for delivery of subfile (f, l) Setstime of K users, F video files, L layer subfiles per file, and T0 time slots Request of user k for file f and set of user requests ... Requesting UE and requested file corresponding to ρ Binary cooperative delivery decisions for subfile (f, l) at BS m Source symbol of subfile (f, l) for serving request ρ at time t Beamforming vectors of BS m and BS set M for sending symbol sρ,l,t AN and its covariance matrix at time t Cache size at BS m Achievable rate and achievable secrecy rate at user κ(ρ) for decoding sρ,l,t Capacity of untrusted BS j for eavesdropping symbol sρ,l,t ... offline caching control N , NT , NU user requests Coop CSI ... Video information Mf,l K, F, L, T0 ρ , (k, f ), S ... κ(ρ), f (ρ) online delivery control qf,l,m ∈ {0, 1} sρ,l,t wρ,l,m,t , wρ,l,t vt , Vt max Cm sec Rρ,l,t , Rρ,l,t Rj,ρ,l,t Ar tific ia ln offline caching ois e control offline caching control online delivery online delivery control Macro BS control Artifi ... ... ... ... c rm fo in n io at ... CSI time cache user requests Untrusted small BS UE 1 ... UE K 1 offline caching control (a) offline caching control online delivery control ... ... online delivery control ... ... 2 T0 T0+1 CSI user requests ... 2T0 2T0+1 time CSI user requests and the capacity constraint X max qf,l,m Vf,l ≤ Cm , m ∈ M. C2: (2) (f,l)∈F ×L at io n user requests in fo rm o de Trusted small BS CSI de o Vi cache user requests ial no ise Vi ... C1: qf,l,m ∈ {0, 1} , ∀(f, l) ∈ F × L, ∀m ∈ M, and qf,0,m = 0, ∀f ∈ F , m ∈ MU , (1) user requests (b) Fig. 1. (a) System model for secure video delivery in a heterogeneous network, where a trusted and an untrusted small cell BS are distributed in the coverage area of a macro BS; (b) caching and delivery control are performed in two timescales. During data delivery, the set of BSs cooperating for the delivery of subfile (f, l) is denoted by MCoop , f,l {m ∈ M | qf,l,m = 1}. Assume that a UE requests one file but possibly multiple layers of the file at a time. We denote a request from user κ for file f by ρ , (κ, f ) and the set of requests by S ⊆ K × F . For convenience, the requesting UE and the requested file corresponding to ρ are denoted by κ(ρ) and f (ρ), respectively. Moreover, user κ(ρ) may request Lρ layers, indexed by Lρ , {0, 1, . . . , Lρ − 1}. We assume a frequency flat fading channel for video data transmission. As the worst case, we assume that the untrusted BSs are full-duplex, i.e., they can simultaneously eavesdrop the video information intended for the UEs and participate in the cooperative delivery of the cached files. In time slot t ∈ T0 , the self-interference [38], [39] at BS j ∈ MU caused by simultaneous reception and transmission at the same frequency is denoted by cj,t . Let xt ∈ CN ×1 denote the joint transmit signal of BS set M. The received signals at user κ(ρ) and the untrusted BSs, denoted by yρ,t ∈ C and yU ,j,t ∈ CNj ×1 , j ∈ MU , respectively, are given by H yρ,t = hH ρ,t xt + zρ,t and yU ,j,t = Gj,t xt + cj,t + zj,t , (3) B. Secure Video Caching and Delivery As the cache helpers in set MU are untrusted, only the enhancement layers are cached at BSs m ∈ MU . Hence, the cached subfiles cannot be used by the untrusted BSs to reconstruct the original video files as long as they do not have access to the base-layer subfiles. Meanwhile, BSs that have the same base-layer or enhancement-layer subfile cached, can employ cooperative transmission for power-efficient and secure delivery of the subfile to the UEs. On the other hand, video files that are uncached at the small cell BSs can be delivered only by the macro BS. Let qf,l,m = 1 indicate that subfile (f, l) is cached at BS m, and qf,l,m = 0 otherwise. The cache placement has to satisfy the condition H H where hρ,t = [hH ∈ CN ×1 and ρ,0,t , . . . , hρ,M,t ] H H H H H Gj,t = [Gj,0,t , . . . , Gj,j−1,t , 0Nj ×Nj ,Gj,j+1,t , . . . , GH j,M,t ] N ×Nj are the channel vectors/matrices from BS set M ∈ C to user κ(ρ) and BS j, respectively. hρ,m,t ∈ CNm ×1 and Gj,m,t ∈ CNm ×Nj model the channels between BS m ∈ M and the respective receivers. The term 0H Nj ×Nj in the definition of Gj,t accounts for the fact that the self-interference at BS j is included in cj,t . Furthermore, zρ,t ∼ CN (0, σ 2 ) and zj,t ∼ CN (0, σj2 INj ) are the zero-mean complex Gaussian noises at the users and the BSs with variance σ 2 and covariance matrix σj2 INj , respectively. The source symbols of subfile (f, l) for serving request ρ in time slot t, denoted by sρ,l,t ∈ C, l ∈ Lρ , are 5 complex Gaussian random variables with sρ,l,t ∼ CN (0, 1). H H Let wρ,l,t , [wρ,l,0,t , . . .,wρ,l,M,t ]H ∈ CN ×1 denote the joint beamforming vector for transmit symbol sρ,l,t , where wρ,l,m,t ∈ CNm ×1 is the individual beamforming vector used by BS m ∈ M in time slot t. Then, the joint transmit signal of BS set M in time slot t ∈ T0 is given by X X wρ,l,t sρ,l,t + vt , (4) xt = ρ∈S l∈Lρ where superposition coding is used to superimpose the Lρ layers intended for user κ(ρ) [40]. Herein, complex Gaussian distributed AN, vt ∈ CN ×1 , is sent cooperatively by the trusted BSs in set MT to proactively interfere the reception of the untrusted BSs in set MU [30]. We assume vt ∼ CN (0, Vt ), where Vt ∈ CN ×N is the covariance matrix of the AN, i.e., Vt , E[vt vtH ]  0. As vt is cooperatively injected only by the trusted BS set MT , we require ΛU Vt = 0, where ΛU is an N × N diagonal matrix given by ΛU = diag 0TNT ×1 , 1TNU ×1 , to ensure that the components of Vt which correspond to untrusted BSs are equal to zero. Moreover, for BS m ∈ M, participating in cooperative transmission of sρ,l,t is possible only if the requested subfile is cached at the BS, i.e., we require  max H ≤ qf (ρ),l,m Pm , C3: tr Λm wρ,l,t wρ,l,t m ∈ M, ρ ∈ S, l ∈ Lρ , t ∈ T0 , (5) max is the maximum transmit power at BS where Pm m, and Λm is an N × N diagonal matrix given  by T TP TP such Λm = diag 0( m−1 N )×1 , 1Nm ×1 , 0( M j j=m+1 N j=0 j )×1   H H that tr Λm wρ,l,t wρ,l,t ≡ tr wρ,l,m,t wρ,l,m,t holds. C3 H = 0, i.e., wρ,l,m,t = 0, when enforces Λm wρ,l,t wρ,l,t Coop qf,l,m = 0, i.e., m ∈ / Mf,l . Otherwise, when qf,l,m = 1, i.e., m ∈ MCoop , C3 ensures that the maximum transmit power, f,l max , of BS m ∈ M is not exceeded. A constraint of the form Pm of C3 is also referred to as a big-M constraint [41]. Based on C1 and C3, we have wρ,0,m,t ≡ 0, ∀m ∈ MU , i.e., the baselayer subfiles cannot be transmitted by untrusted BSs. C. Achievable Secrecy Rate Each user employs successive interference cancellation (SIC) at the receiver [40]. The base-layer subfile is decoded first, as it is required for the decoding of the other layers. In decoding the subfile of layer l ∈ Lρ \ {0}, the previously decoded lower layers 0, . . . , l − 1 are first removed from the received signal for interference cancellation. This process continues until layer Lρ − 1 is decoded [32]. Define the ρ′ ,l′ interference cancellation coefficient aρ,l ∈ {0, 1}, where ′ The instantaneous achievable rate (bits/s/Hz) for layer l ∈ Lρ at user κ(ρ) is given by ! 2 1 H σ2 hρ,t wρ,l,t , (7) Rρ,l,t = log2 1 + 1 + σ12 Iρ,l,t + σ12 hH ρ,t Vt hρ,t X ′ ′ 2 aρρ,l,l hH Iρ,l,t = (8) ρ,t wρ′ ,l′ ,t , (ρ′ ,l′ )6=(ρ,l) where Iρ,l,t is the residual interference term for decoding layer l of user κ(ρ) and (ρ, l) 6= (ρ′ , l′ ) indicates ρ 6= ρ′ and/or l 6= l′ . On the other hand, the untrusted BSs may eavesdrop the video information intended for the users. For guaranteeing communication secrecy, the proposed secure delivery scheme is designed to avoid information leakage even under worst-case conditions. Specifically, we assume that BS j ∈ MU can fully cancel the self-interference power cj,t during eavesdropping4, and hence, achieves the full-duplex capacity upper bound for layer l of the signal intended for user κ(ρ) given by   H H G w w G , Rj,ρ,l,t = log2 det INj + σ12 Z−1 ρ,l,t j,t ρ,l,t j,ρ,l,t j,t j (9) 1 GH j,t Vt Gj,t σj2 1 Ψj,ρ,l,t σj2 + ≻ 0, (10) Zj,ρ,l,t = INj + X ρ′ ,l′  H aρ,l 1 − qf (ρ′ ),l′ ,j Gj,t wρ′ ,l′ ,t wρH′ ,l′ ,t Gj,t . Ψj,ρ,l,t = (ρ′ ,l′ )6=(ρ,l) (11) ′ Note that if subfile (f, l ) is cached at BS j ∈ MU , we have 1 − qf,l′ ,j = 0 in (11). That is, in addition to SIC, BS j ∈ MU can also utilize the cached video data as side information to suppress the interference caused by subfile (f, l′ ). The secrecy rate achievable at user κ(ρ) for decoding layer l ∈ Lρ in time slot t ∈ T0 is given by i+ h sec (12) = Rρ,l,t − max Rj,ρ,l,t . Rρ,l,t j∈MU Remark 1. Note that a passive eavesdropper, as considered for non-caching networks in [29]–[31] and caching networks in [16], [17], can be cast as an untrusted BS having no cache memory or no data cached. Considering C1–C3, such an eavesdropper will not participate in the cooperative transmission of the video files. Thus, the untrusted cache helpers considered in this paper correspond to a more general eavesdropping model than that investigated in the literature [16], [17], [29]–[31]. III. P ROBLEM F ORMULATION In this section, we first present the adopted imperfect CSI model for video delivery. Then, a two-timescale robust optimization problem is formulated for minimization of the total BS transmit power required for video streaming under ′ ρ ,l aρ,l = 1 indicates that the transmission of subfile (f (ρ′ ), l′ ) ′ ′ interferes that of subfile (f (ρ), l), and aρρ,l,l = 0 otherwise. By adopting SIC decoding for the SVC video files, we have ( 0, if ρ = ρ′ , l ≥ l′ ; ρ′ ,l′ aρ,l = (6) 1, otherwise. 4 In practice, if self-interference is not perfectly canceled, the residual self-interference impairs the eavesdropping at the untrusted BSs, and hence, improves communication secrecy. However, estimating the residual selfinterference at the central controller (e.g., the macro BS), which is responsible for resource allocation, may not be possible. Hence, we make the worst-case assumption of zero self-interference in this paper and the obtained results provide a lower bound on the performance for the case of imperfect selfinterference cancellation. 6 QoS and secrecy constraints. Note that low transmit power is desirable to minimize the interference caused in other cells and to reduce the network operation cost. For a given cache status, the cooperative transmission decisions for each time slot are optimized online based on instantaneous CSI estimates. However, due to time causality and computational complexity constraints, the cache placement for each period is optimized offline based on historical user requests and CSI [10], [16]. collected during the time period6. For the considered typical period T0 , the caching optimization problem is formulated as: X P0: minimize UTP (q, wρ,t , Vt ) (15) q,wρ,t ,Vt t∈T0 subject to C1, C2, C3, C4: Vt  0, ΛU Vt = 0,    X H max C5: tr Λm  wρ,l,t wρ,l,t +Vt  ≤ Pm , ρ,l C6: A. Channel State Information min ∆hρ,t ∈Ωρ,t C7: max req Rρ,l,t ≥ Rρ,l , ρ ∈ S, l ∈ Lρ , max j∈MU ∆Gj,t ∈Ωj,t At the beginning of each time slot, the CSI hρ,t and Gj,t has to be acquired5 at the centralized controller, i.e., the macro BS, for computing the resource allocation. The estimates of hρ,t b ρ,t ∈ CN ×1 and Gj,t gathered at the macro BS, denoted by h b j,t ∈ CN ×Nj , respectively, will in general be imperfect. and G b ρ,t + ∆hρ,t That is, the actual channels are given by hρ,t = h b and Gj,t = Gj,t + ∆Gj,t , where ∆hρ,t and ∆Gj,t represent the respective channel estimation errors caused by quantization errors, imperfect feedback channels, as well as outdated and noisy estimates. In fact, the estimation errors ∆hρ,t and ∆Gj,t may be enhanced by the actions of the untrusted BSs which may not fully cooperate with the macro BS during channel estimation and feedback. The specific values of ∆hρ,t and ∆Gj,t are not known at the macro BS. To model the imperfect CSI, we assume that the possible values of ∆hρ,t and ∆Gj,t lie in ellipsoidal uncertainty regions [42] given by n o 2 Ωρ,t , ∆hρ,t ∈ CN ×1 | ∆hH Ξ ∆h ≤ ε ρ ρ,t ρ,t ρ , Ωj,t ρ ∈ S, t ∈ T0 , (13) o n  2 , ∆Gj,t ∈ CN ×Nj | tr ∆GH j,t Ξj ∆Gj,t ≤ εj , j ∈ MU , t ∈ T0 . (14) Here, ερ > 0 and εj > 0 represent the radii of uncertainty regions Ωρ,t and Ωj,t , respectively; Ξρ ∈ CN ×N and Ξj ∈ CN ×N denote the orientations of the uncertainty regions, respectively, where Ξρ ≻ 0 and Ξj ≻ 0. In practice, the values of ερ , εj , Ξρ , and Ξj depend on the channel coherence time and the adopted channel estimation methods. tol Rj,ρ,0,t ≤ Rρ,0 , ρ ∈ S, where UTP (q, wρ,t , Vt ) ,  P P P H denotes tr ρ∈S l∈Lρ wρ,l,t wρ,l,t + Vt m∈M Λm the total BS transmit power in time slot t ∈ T0 . Constraint C5 max limits the maximal transmit power of BS m ∈ M to Pm . req C6 guarantees the minimum video delivery rate, Rρ,l , in each time slot t ∈ T0 to provide QoS in delivering layer l ∈ Lρ for serving request ρ ∈ S. C7 constrains the maximum tol data rate leaked to the untrusted BSs in set MU to Rρ,0 in each time slot t to ensure communication secrecy. Since the untrusted BSs are unable to decode the enhancement layers without base-layer information, secrecy can be ensured by imposing C7 only on the delivery of the base-layer subfiles. Due to the imperfect CSI, the minimum/maximum data rate in C6/C7 is guaranteed/constrained for all possible estimation error vectors/matrices in the respective uncertainty sets in order to facilitate robustness with respect to (w.r.t.) communication secrecy. This robust optimization approach has been commonly adopted for studying PLS in the literature, see [29]–[31] and references therein. Constraints C6 and C7 jointly guarantee a minimum achievable secrecy req sec tol + , t ∈ T0 , for delivering the rate of Rρ,0,t = Rρ,0 − Rρ,0 base-layer subfiles for request ρ, provided that problem P0 is feasible. Problem P0 is a non-convex mixed-integer nonlinear program (MINLP)7 due to the binary caching decision vector q and the non-convex constraints C6 and C7. This type of problem is NP-hard [41]. Yet, since P0 is solved offline for a large timescale, we adopt a global optimization method to solve P0 optimally in Section IV-A. The obtained solution defines a performance benchmark for low-complexity suboptimal schemes, cf. Section IV-B. B. Caching Optimization C. Delivery Optimization Let q , [q1,0,0 , . . . , qf,l,m , . . . , qF,L−1,M ] and wρ,t , H H ]H be the caching and the transmitter [wρ,0,t , . . . , wρ,L ρ −1,t beamforming optimization vectors, respectively. Considering the two-timescale control in Fig. 1(b), the caching decision q is made (at the end of) every T0 time slots based on the historical profiles of user requests and CSI that have been Assume that the instantaneous CSI estimates are given. Moreover, the cache status q for T0 has been determined at the end of the previous time period. Then, the cooperative transmission policy {wρ,t , Vt } for time t ∈ T0 is optimized online by solving the following problem 5 For example, by exploiting channel reciprocity in time division duplex systems, hρ,t and Gj,t can be estimated in the uplink at the small cell and macro BSs based on pilots emitted by the UEs and the untrusted BSs, respectively. Then, the estimated CSI obtained at the small cell BSs is fed back to the macro BS via the X2 interface [53]. 6 Prediction of the users’ future requests based on historical user profiles can further improve the cache placement at the cost of an increased computational complexity. 7 For a non-convex MINLP, even if the binary constraints are relaxed into convex ones, the problem remains non-convex [41]. 7 Q0: minimize wρ,t ,Vt subject to UTP (q, wρ,t , Vt ) (16) C3, C4, C5, C6, C7. Problem Q0 is non-convex due to constraints C6 and C7. However, we will show that Q0 can be optimally solved by employing SDP relaxation, cf. Section IV-C. IV. P ROBLEM S OLUTION In this section, the caching problem P0 is tackled first. We show that P0 can be transformed into a convex MINLP by SDP relaxation and further solved optimally by an iterative algorithm. Inspired by the optimal algorithm, a low-complexity suboptimal caching scheme is developed to balance between optimality and computational complexity. Moreover, we show that the delivery problem Q0 can be optimally and efficiently solved. A. Optimal Caching Scheme 1) Problem Transformation: To reformulate problem P0 as a convex MINLP, C6 and C7 have to be transformed H into convex constraints. Let Wρ,l,t = wρ,l,t wρ,l,t  0 be the beamforming matrix subject to rank (Wρ,l,t ) ≤ 1. By substituting Wρ,l,t and employing elementary arithmetic operations, C6 is equivalently reformulated as an affine inequality constraint that is jointly convex w.r.t. {Wρ,l,t , Vt }, C6: hH ρ,t Tρ,l,t hρ,t where Tρ,l,t , req 1 req ηρ,l 2 ≥ σ , ∀∆hρ,t ∈ Ωρ,t , (17) ′ ′ P ρ ,l Wρ′ ,l′ ,t − Vt Wρ,l,t − (ρ′ ,l′ )6=(ρ,l) aρ,l req and ηρ,l , 2Rρ,l − 1. However, as Ωρ,t is a continuous set, C6 is semi-infinite, i.e., it represents infinitely many inequalities for Tρ,l,t , and hence, is still intractable for optimization. To overcome this issue, C6 is transformed into a finite number of b ρ,t + convex constraints. To this end, we substitute hρ,t = h ∆hρ,t in C6 and apply the S-procedure from [43, Appendix B], which leads to o n bH C6: ∆hH ρ,t Tρ,l,t ∆hρ,t + 2ℜ hρ,t Tρ,l,t ∆hρ,t b ρ,t − σ 2 ≥ 0, b H Tρ,l,t h +h ρ,t ∀∆hρ,t ∈ Ωρ,t , f UH Tρ,l,t Uρ,t  ⇐⇒ C6: "ρ,t # −δρ,l,t Ξρ 0 , δρ,l,t ≥ 0, (18) 0H σ 2 + δρ,l,t ε2ρ b ρ,t ] ∈ C(N +1)×(N +1) and δρ,l,t is an where Uρ,t , [IN , h auxiliary optimization variable. Next, let Wρ,l,j,t = (1 − qf (ρ),l,j )Wρ,l,t  0 be an auxiliary optimization matrix. We have   X ρ′ ,l′ ′ ,l′ ,j,t Gj,t , (19) a W Ψj,ρ,l,t = GH ρ j,t ρ,l (ρ′ ,l′ )6=(ρ,l) if and only if rank(Wρ,l,j,t ) ≤ 1 and the following constraints hold, C8: tr(Wρ,l,t − Wρ,l,j,t )  qf (ρ),l,j Pmax , C9: tr(Wρ,l,j,t )  (1 − qf (ρ),l,j )Pmax , C10: Wρ,l,t  Wρ,l,j,t , Wρ,l,j,t  0, (20) P max where Pmax , m∈M Pm . Here, C8 and C9 guarantee that Wρ,l,j,t = 0 if qf (ρ),l,j = 1, and Wρ,l,j,t = Wρ,l,t otherwise. By substituting Wρ,l,j,t and Ψj,ρ,l,t , C7 can be reformulated into an LMI as follows H H tot C7 ⇐⇒ σ12 wρ,0,t Gj,t Z−1 j,ρ,0,t Gj,t wρ,0,t ≤ ηρ,0 j  2 tot H ⇐⇒ tr Z−1 j,ρ,0,t Gj,t Wρ,0,t Gj,t ≤ σj ηρ,0 ,   1 (a) −1/2 − /2 2 tot ⇐⇒ λmax Zj,ρ,0,t GH j,t Wρ,0,t Gj,t Zj,ρ,0 ≤ σj ηρ,0 , 2 ⇐⇒ C7: GH j,t Tρ,0,j,t Gj,t  σj INj , ∀j ∈ MU , ∀∆Gj,t ∈ Ωj,t , tot where ηρ,0 P , 2 ′ ρ ,l (ρ′ ,l′ )6=(ρ,0) aρ,0 ′ Rtot ρ,0 − 1, Tρ,0,j,t , (21) 1 tot Wρ,0,t ηρ,0 − Wρ′ ,l′ ,j,t − Vt , and (a) holds due to rank(Wρ,l,t ) ≤ 1. In fact, as Ωj,t is a continuous set, C7 in (21) represents infinitely many LMIs that are jointly  convex w.r.t. Wρ,0,t , Vt , Wj,t . For tractability, C7 has to be transformed into a finite number of convex constraints. This can be accomplished by exploiting the robust quadratic matrix inequality in [45, Theorem 3.3]. Thereby, we obtain f UH Tρ,0,j,t Uj,t  C7 ⇐⇒ C7: j,t # " 2 0 (σj − δρ,0,j,t )INj , δρ,0,j,t ≥ 0, δρ,0,j,t Ξj 0 ε2 (22) j b j,t IN ] ∈ CN ×(N +Nj ) and δρ,0,j,t is an where Uj,t , [G auxiliary optimization variable. Finally, by defining the delivery variable Dt , [Wρ,l,t , Wρ,l,j,t , Vt ] and applying the above transformations, the original problem P0 is equivalently reformulated as X minimize UTP (q, Dt ) (23) q,Dt subject to t∈T0 f C7, f C8, C9, C10, C1, C2, C4, C6, max C3: tr(Λm Wρ,l,t ) ≤ qf (ρ),l,m Pm ,  X  max ≤ Pm , C5: tr Λm Wρ,l,t + Vt ρ,l C11: rank(Wρ,l,t ) ≤ 1, ρ ∈ S, l ∈ Lρ . Here, constraint rank(Wρ,l,t ) ≤ 1 is dropped due to C10 and C11. Let P1 denote the SDP relaxation of problem (23), obtained by dropping C11 in (23). Then, problem P1 is a convex MINLP, i.e., by relaxing the binary constraints of P1 into convex ones, we arrive at a convex problem. The GBD algorithm is a simple iterative method to handle convex MINLPs [41, Section 6.3]. In each GBD iteration, upper and lower bounds on the optimal value are generated by solving a primal subproblem and a master problem, respectively. To ensure convergence, optimality and feasibility cuts are successively added to tighten the bounds and eliminate the infeasible solutions possibly obtained during the iterations, respectively. The GBD algorithm is attractive for solving P1 as it can be efficiently implemented exploiting the structure of P1. In particular, the resulting primal subproblem is a convex problem where strong duality holds while the master problem is a mixed-integer linear program (MILP), and both problems are easy to handle using off-the-shelf numerical solvers such 8 as CVX [46] and MOSEK [47]. However, the GBD algorithm typically suffers from slow convergence. This is because, when an infeasible solution is obtained in an iteration of the GBD, the resulting feasibility cut is usually ineffective in improving the solution. If the problem is infeasible, the GBD algorithm terminates only after having performed an exhaustive search over all possible candidate solutions. To remedy this issue, an improved GBD algorithm8 is proposed below. 2) Problem Decomposition: The proposed modified GBD algorithm applies a two-layer decomposition of problem P1 and solves a binary caching optimization problem for q in the outer layer and a continuous delivery optimization problem for Dt in the inner layer. However, q and Dt are coupled via constraints C3, C8, and C9. To facilitate the decomposition, we perturb the right-hand sides of C3, C8, and C9 by introducing C8 C9 slack variables sC3 ρ,l,m,t ≥ 0, sρ,l,j,t ≥ 0, and sρ,l,j,t ≥ 0, C3 C8 C9 respectively. Let st , [sρ,l,m,t , sρ,l,j,t , sρ,l,j,t ] be the perturbation vector and st  0. Moreover, in the objective function, we add an ℓ1  -norm (exact) penalty cost function for  st , Lq (Dt , st ; λt ) = f1 (q; λt ) + f2 (Dt , st ; λt ) , (27)  P C8 where f1 (q; λt ) = ρ,l,j λC9 ρ,l,j,t − λρ,l,j,t qf (ρ),l,j Pmax − P C3 max ρ,l,m,t λρ,l,m,t qf (ρ),l,m Pm , and with penalty factor µ ≫ 1. Consequently, problem P1 decomposes into T0 SDP subproblems in the inner layer, i.e., one subproblem for each time slot t ∈ T0 , and an MILP in the outer layer, which are shown in (25) and (26) at the top of the next page, respectively. Problem (26) is referred as the master problem. Thereby, problems (25) and (26) are equivalent to P1 when µ ≫ 1, as stated in Proposition 1. where ξt (q; λt ) , minDt ∈Dt , st 0 Lq (Dt , st ; λt ). Although problem (29) still contains an infinite number of constraints (w.r.t. λt ) and undetermined functions ξt (·; ·), it is readily solvable by an iterative relaxation method as will be explained in the following. 3) Optimal Iterative Solution: The proposed iterative algorithm is given in Algorithm 1. Let k be the iteration index. We start from one constraint at k = 1, which defines a cutting plane (also referred as an optimality cut [41]). Then, the number of constraints/cuts are increased sequentially as the iteration proceeds. Specifically, for given dual variables λjt , j = 1, . . . k − 1, the following master problem is solved in iteration k, minimize α (30) q∈Q,α X  subject to α ≥ ξt q; λjt , j = 1, . . . , k − 1. fPen (st ) , µ  X ρ,l,m sC3 ρ,l,m,t + X sC8 ρ,l,j,t ρ,l,j + sC9 ρ,l,j,t   , (24) Proposition 1. For µ ≫ 1, problems P1 and (25), (26) are equivalent such that: i) if P1 is feasible, then SDP subproblem (25) is always feasible for q ∈ Q; moreover, the optimal solution of q for P1 solves the master problem (26); ii) if problem (25) is infeasible, i.e., ν (q) = +∞, or its optimal solution satisfies st′ 6= 0 for some t′ ∈ T0 , then problem P1 is infeasible. Proof: Please refer to Appendix A. Remark 2. By perturbation, the feasible set of ν (q) in (25), (26) is extended to Q ⊆ {0, 1}F ×L×M . Consequently, based on Proposition 1, infeasible solutions can be avoided if problem P1 is feasible, cf. i), or identified easily if problem P1 is infeasible, cf. ii). These properties facilitate an efficient implementation of the GBD algorithm in the sequel to optimally solve (25) and (26). SDP subproblem (25) is convex and can be solved by interior point methods [43], [48], i.e., numerical solvers such as CVX [46] are applicable. Meanwhile, exploiting the convexity (and strong duality) of (25), we can further simplify the formuC8 lation of the master problem (26). Let λC3 ρ,l,m,t ≥ 0, λρ,l,j,t ≥ 0, and λC9 ρ,l,j,t ≥ 0 be the Lagrange multipliers for C3, C8, and C8 C9 C9, respectively, and define λt , [λC3 ρ,l,m,t , λρ,l,j,t , λρ,l,j,t ]  0. The Lagrangian of (25) can be written as 8 A similar approach as in the proposed improved GBD algorithm has been successfully applied to accelerate the outer approximation algorithm in [41, Section 6.6]. f2 (Dt , st ; λt ) = fPen (st ) + X hX X  i λC3 tr λC8 ρ,l,m,t Λm + 1 + ρ,l,j,t IN Wρ,l,t ρ,l m + λC9 ρ,l,j,t X ρ,l,j − X ρ,l,m j − λC8 ρ,l,j,t C3 λC3 ρ,l,m,t sρ,l,m,t   X C9 tr Wρ,l,j,t − λρ,l,j,t Pmax + X ρ,l,j C8 λC8 ρ,l,j,t sρ,l,j,t ρ,l,j  C9 − λC9 ρ,l,j,t sρ,l,j,t . Lq (Dt , st ; λt ) is separable w.r.t. {q} and {Dt , st }. Since, for given q ∈ Q, problem (25) is convex and fulfills Slater’s condition, the following result holds due to strong duality: νt (q) = max min Lq (Dt , st ; λt ) , ∀q ∈ Q. (28) λt 0 Dt ∈Dt , st 0 Consequently, the master problem is reformulated as minimize α q∈Q,α X subject to α ≥ ξt (q; λt ) , ∀λt  0, (29) t∈T0 t∈T0 Problem (30) is a relaxation of problem (29). Due to the enlarged feasible set, the optimal value of problem (30) gives a lower bound on that of problem (29). The relaxation solution, denoted by (qk , αk ), is optimal for problem (29) if it is feasible for problem (29). Otherwise, we add another optimality cut to the feasible set of (30) in the next iteration to tighten the relaxation. As this process continues, we obtain a non-decreasing sequence of lower bounds until the relaxed solution becomes feasible, i.e., solves problem (29) optimally, or until the problem is known to be infeasible. Two remarks regarding Algorithm 1 are in order. First, as can be observed in Algorithm 1 (lines 5–10), the feasibility or optimality of qk is verified by solving SDP subproblem (25). This is because, if qk is optimal, we know that solving problem (25) for q = qk in line 5 will return the optimal value of αk , i.e., ν(qk ) = αk , owing to the strong duality of problem (25). Otherwise, ν(qk ) gives an upper bound on the optimal value, and thus, ν(qk ) ≥ αk . By keeping  the lowest upper bound obtained so far, i.e., U B ← min U B, ν(qk ) 9 νt (q) , minimize UTP (q, Dt ) + fPen (st ) Dt ,st 0  subject to Dt ∈ Dt , Dt | C4, C5, C6, C7, C10 , minimize q,α subject to α α ≥ ν (q) , X t∈T0 νt (q) , 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: for given q0 and determine D1t , s1t , λ1t ; set tolerance ε ≥ 0, U B ← ν(q0 ), LB ←− ∞, k ← 1; while (U B > LB + ε) do Solve the relaxed master problem (30) for given Dkt , skt , λkt and determine the solution (qk , αk ); Update lower bound and solution: LB ← αk , q∗ ← qk ; Solve SDP subproblem (25) for given qk and determine the primal and the dual solutions Dk+1 , sk+1 , λk+1 ; t t t k if (ν(q ) = +∞, i.e., (25) is infeasible, OR ν(qk ) ≤ αk + ε) then and exit the while loop; , s∗t ← sk+1 Set D∗t ← Dk+1 t t else if (ν(qk ) < U B) then Update upper bound and solution: U B ← ν(qk ), D∗t ← Dk+1 , s∗t ← sk+1 ; t t end if Update iteration index: k ← k + 1; end while if (s∗t = 0) then Return the optimal solutions q∗ and D∗t ; else Return the infeasible problem P0/P1. end if (cf. line 9), the optimality condition is satisfied when the gap between U B and the lower bound vanishes. Second, for computational convenience, the values of λkt in iteration k can be intelligently chosen as the optimal dual solutions of problem (28) or (25) for q = qk in line 5. In this case, the constraint function ξt (·; ·) can be easily computed as explained in the following proposition. Proposition 2. Let (Dkt , skt ) and λkt be the optimal primal and dual solutions of (25) for ν(qk ) at iteration k, respectively. Then, we have i) (Dkt , skt ) also solves the minimization problem in the optimality cut of iteration k + 1 (cf. (30)), i.e., (Dkt , skt ) ∈ arg minDt ∈Dt , st ≥0 Lq (Dt , st ; λkt ). ii) By choosing λt = λkt , function ξt q; λkt reduces to an affine function given by   X   C8,k  k λC9,k ξt q; λkt = ρ,l,j,t − λρ,l,j,t Pmax qf (ρ),l,j − qf (ρ),l,j ρ,l,j − X ρ,l,m   C3,k max Pm qf (ρ),l,m − qfk(ρ),l,m λρ,l,m,t   + UTP qk , Dkt + fPen skt . Proof: Please refer to Appendix B. C9: tr(Wρ,l,j,t ) − (1 − qf (ρ),l,j )Pmax ≤ sC9 ρ,l,j , (26) q ∈ Q , {q | C1, C2} . 1: Initialization: Given q0 ← 0. Solve the SDP subproblem (25) 4: 5: max C3: tr(Λm Wρ,l,t )−qf (ρ),l,m Pm ≤ sC3 ρ,l,m , C8: tr(Wρ,l,t − Wρ,l,j,t ) − qf (ρ),l,j Pmax ≤ sC8 ρ,l,j , Algorithm 1 Optimal iterative algorithm for solving P1 and P0 2: 3: (25) (31) Based on Proposition 2, the relaxed master problem (30) is an MILP and can be solved optimally, e.g., using the numerical solver MOSEK [47]. Similar to the conventional GBD method, Algorithm 1 converges in a finite number of iterations as shown in Proposition 3. The obtained solution is globally optimal for problem P1. In general, the solution of P1 gives a lower bound for problem P0. However, by inspecting the rank of the SDP solution of problem P1, we can further show that the SDP relaxation is tight. Proposition 3. Algorithm 1 converges in a finite number b ρ,t , iterations. Moreover, assuming that the channel vectors h ρ ∈ S, can be modeled as statistically independent random vectors, problems P1 and P0 are equivalent in the sense that whenever P0 is feasible, the solution of P1 is also (globally) optimal for P0 with probability one, and the optimal beamformer is given by the principal eigenvector of Wρ,l,t . Proof: Please refer to Appendix C. Due to perturbation, only optimality cuts, cf. (30), need to be generated by Algorithm 1 in each iteration, cf. Remark 2. This is different from the classical GBD algorithm [41, Section 6.3] where feasibility cuts are also required to exclude infeasible solutions during intermediate iterations. Since the optimality cuts can successively improve the lower bounds, Algorithm 1 is expected to converge faster than the classic GBD algorithm if P1 is feasible. On the other hand, even if P1 is infeasible, the perturbed problem is generally feasible. Then, optimality cuts can be still generated to iteratively improve the solutions and reduce the required number of iterations with a high probability. 4) Computational Complexity: Assume that the interiorpoint method [43], [48] is applied to solve the SDP subproblems in each iteration of the GBD algorithm. The computational complexity of solving each SDP subproblem w.r.t. the number of UEs, K, the number of BSs, M , the number of BS antennas, N , and the number of SVC layers, L, can be approximated as [49, Theorem 3.12]    4 3 Θsdp = O (M KL) N 3 + N 2 + 2 + (M KL) {z } | Complexity per iteration  M KLN log ǫ−1 {z } | Number of iterations   4.5 = O (M KL) N 3.5 log ǫ−1 , × √ (32) where ǫ > 0 is the solution accuracy specified by the numerical solver and O(·) is the big-O notation. Although the SDP 10 Algorithm 2 Suboptimal iterative algorithm for solving P1 and P0 1: Initialization: Given q1m ← 0, ∀m ∈ M; k ← 1; 2: while I k 6= ∅ do 3: for each i ∈ I k do 4: Solve SDP subproblem (25) for each given {qm } satisfying 5: 6: 7: 8: qi ∈ Qki ∩ Q; Determine qk+1 in (34); i end for k ← k + 1. end while subproblems can be solved in polynomial time in line 5, cf. (32), the overall computational complexity of Algorithm 1 grows non-polynomially with the size of problem P0. This is because the MILP solver in line 3 may incur an exponentialtime computational complexity, O(2F LM ), in the worst case [41], even though the likelihood that the worst case occurs is low due to the employed perturbation, cf. Remark 2. Thus, only offline cache optimization may be feasible in practical implementations. B. Suboptimal Caching Scheme For systems with limited computing resources, Algorithm 1 may not be applicable due to its worst-case exponential-time computational complexity. Instead, polynomial-time suboptimal schemes facilitating a better trade-off between system performance and computational complexity may be preferable. Based on Proposition 3, P0 can be solved via its equivalent convex MINLP, P1. As is also evident from (25), for given q, P1 reduces to an SDP and can be solved optimally in polynomial time, cf. (32). Therefore, by additionally adjusting q via a greedy iterative search, we obtain the low-complexity suboptimal scheme in Algorithm 2. Let FS and qm be the set of files requested by S (the set of requests) and the caching vector at BS m, respectively, where FS ⊆ F . We n define o 2 k Qm , qm ∈ {0, 1}|FS |×L | qm − qkm 2 ≤ 1 (33) as the set of binary vectors within a distance of one from  qkm . Besides, I k , m ∈ M | Qkm ∩ Q > 1 defines the set of BS indices where Qkm and Q have non-unique intersection points. During iteration k, the vector in set Qki ∩Q that minimizes the objective value of primal problem (25) is chosen as the new caching vector at BS i ∈ I k , i.e., qk+1 = arg min i ν (q1 , . . . , qM ) . By applying the same transformation techniques as for problem P0 in Section IV-A and relaxing the rank constraint, cf. C11, problem Q0 can be reformulated as an SDP, which is equivalent to problem (25). Since, based on Proposition 3, the solution of problem (25) fulfills rank constraint C11 with probability one, delivery problem Q0 can be solved optimally via SDP subproblem (25), as stated in Corollary 1. Corollary 1. SDP subproblem (25) (for the respective instantaneous CSI) and the delivery optimization problem Q0 are equivalent in the sense that the solution of (25) is also optimal for Q0 whenever Q0 is feasible. Proof: The proof is similar to that of Propositions 1 and 3 and is omitted for brevity. Therefore, delivery optimization problem Q0 can be solved in polynomial time with computational complexity Θsdp , cf. (32), which is desirable for online implementation [42]. Moreover, delivery optimization incurs a signaling overhead of O(M KLN ) for collecting the CSI at the macro BS and distributing the optimization results to the small cell BSs. Remark 4. Although the total number of BSs in dense small cell networks may be large [44], the coverage areas of most small cell BSs will not overlap. Therefore, the proposed caching and delivery algorithms may be applied to several small groups of small cell BSs with overlapping coverage areas rather than jointly to all small cell BSs. This considerably reduces the computational complexity and signaling overhead. Remark 5. The proposed caching and delivery optimization framework can be extended to integrate (centralized and decentralized) coded caching at the end users [11]–[15]. For example, multicast codewords can be cooperatively transmitted by a subset of the BSs if these BSs have cached the subfiles required for coded multicast transmission. Such a design reaps the performance gains of both coded caching and cache-enabled cooperative MIMO transmission. However, the resulting cache optimization problem would involve a large number of binary caching variables, which have to be defined per subfile, and SDP relaxation of the delivery optimization problem may not yield the optimal solution anymore [51]. Hence, extending the proposed framework to coded caching is an interesting topic for future research. (34) qi ∈Qk i ∩Q qk+1 , i time. The obtained solution is ensured to be feasible for P1. Moreover, it is often close-to-optimal due to the iterative minimization in (34) [50], as will be shown in Section V. C. Optimal Delivery Solution That is, the cache vectors within a distance of one from qki , are iteratively updated to successively reduce the objective value. The iteration continues until Qki ∩ Q becomes unique, i.e., no further reduction in the objective function is possible, which yields the solution. Hence, the number of problem 2 instances of (25) to be solved is bounded by M L2 |FS | . Consequently, the overall computational complexity of Algorithm   2 2 is approximated as O M L2 |FS | Θsdp , which grows only polynomially with the problem size. Remark 3. By adopting the greedy heuristic, searching over the non-convex set Q of P1 can be done in polynomial V. S IMULATION R ESULTS In this section, we evaluate the performance of the proposed optimal and suboptimal schemes. Consider a cell of radius R1 = 1 km, where the macro BS is located at the center of the cell and three small cell BSs are uniformly distributed within the cell. The number of untrusted small cell BSs is set to Mu , |MU | = 1, unless stated otherwise. To gain insight, in Figs. 2–5, we consider a small network with only three small cell BSs. A larger network is considered in Fig. 6. The macro BS is equipped with N0 = 6 antennas while each small cell BS has Nm = 2 antennas. We assume that F = 10 video files, each of duration 45 minutes and size 500 MB 11 TABLE II S IMULATION PARAMETERS . Settings 5 MHz 10 ms 45 mins P0max = 46 dBm max Pm = 39 dBm −172.6 dBm/Hz C0max = 1000 MB 10 (Bytes), are delivered to K = 6 single-antenna UEs. Each user requests the files independent of the other users. Let θ = [θ1 , . . . , θF ] be the probability distribution of the requests for different files. We set θ according to a Zipf distribution with γ = 1.1. In particular, assuming that file f ∈ F is the σf th most popular file for the UEs, thePprobability of file f ∈ F being requested is given by θf = σ1γ / f ∈F σ1γ [52]. We adopt f f an SVC codec with L = 2. That is, each video file is encoded into a base-layer subfile (l = 0) and an enhancement-layer subfile (l = 1), each of size Vf,l = 250 MB. The minimum streaming rate and the secrecy rate threshold for the base-layer req req tol subfiles are Rρ,0 = 825 kbps and Rρ,0 = 0.1Rρ,0 = 82.5 kbps, respectively. Therefore, if problem Q0 is feasible, a secrecy sec streaming rate of Rρ,0,t = 742.5 kbps can be guaranteed for secure and uninterrupted video streaming for each user as sec Rρ,0,t ≥ 250×8×106/(45×60) = 741 kbps. The streaming rate of req sec the enhancement-layer subfiles is Rρ,1 = 2Rρ,0,t = 1.5 Mbps. The users are randomly distributed in the system. Based on the locations of the BSs and users, the path loss is calculated using the 3GPP model for the “urban macro non-line-of-sight” scenario [53]. The small-scale fading coefficients are independent and identically distributed (i.i.d.) Rayleigh random variables. We employ Euclidean spheres for modeling the uncertainty regions Ωρ,t and Ωj,t by setting Ξρ = Ξj = IN . Meanwhile, we define the maximum normalized channel 2esε timation error variances of hρ,t and Gj,t as σρ2 = kh ρ k2 ε2 • • Baseline 1 (Random caching): The video (sub)files are randomly cached until the cache capacity is reached. Baseline 2 (Preference based caching): The most popular (sub)files are cached. In trusted BSs, since the base-layer subfiles are more important, they are cached with higher priority than the enhancement-layer subfiles of the same video file. For Baselines 1 and 2, the optimal delivery decisions are obtained by solving problem Q0. Baseline 3 (No cooperation with untrusted BSs): No video files are cached at the untrusted BSs, which act as pure eavesdroppers. Hence, the untrusted BSs are not allowed to cooperate for delivery of the video files. This approach is adopted in state-of-the-art cellular networks. -1 -2 300 600 900 1200 1500 Cache capacity per small cell BS [MB] 1800 2100 (a) (b) Fig. 2. (a) Total BS transmit power and (b) secrecy outage probability versus cache capacity for different caching and delivery schemes. ρ,t 2 and σj2 = kG j k2 , respectively. Unless otherwise specified, j,t F we assume σρ2 = 0.01, ρ ∈ S and σj2 = 0.05, j ∈ MU . All other relevant system parameters are given in Table II. For comparison, we consider two heuristic caching schemes and a non-cooperative and a non-robust delivery scheme as baselines: • 10 Secrecy outage probability Parameters System bandwidth Duration of time slot Duration of delivery period Macro BS transmit power Small BS transmit power Noise power density Cache capacity at macro BS Optimal Suboptimal Baseline 1 Baseline 2 Baseline 3 Total BS transmit power [dBm] Secrecy outage probability 10 • The optimal caching and delivery decisions are obtained max from problems P0 and Q0, respectively, with Cm = 0, ∀m ∈ MU . Baseline 4 (Non-robust transmission): Different from the proposed schemes, the macro BS treats the channel b ρ,t and G b j,t as accurate. The optimal caching estimates h and delivery decisions are obtained by solving problems b ρ,t and P0 and Q0, respectively, after setting hρ,t = h b Gj,t = Gj,t . Figs. 2(a) and 2(b) illustrate the performance of the considered caching and delivery schemes as functions of the cache capacity. Herein, the system performance is evaluated during the online delivery of the video files, cf. problemP Q0. The sesec crecy outage probability, defined as pout , Pr ρ Rρ,0,t <   P  req tol + , characterizes the likelihood that probρ Rρ,0 − Rρ,0 lem Q0 is infeasible because either the QoS constraint C6 or the secrecy constraint C7 cannot be satisfied. As can be observed from Figs. 2(a) and 2(b), for all considered schemes, a larger cache capacity leads to both a lower total BS transmit power and a smaller secrecy outage probability as larger virtual transmit antenna arrays can be formed among the trusted and Secrecy outage probability 10 12 -1 Optimal Total BS transmit power [dBm] Average number of cooperating BSs 10 10 max max max = 0 MB Suboptimal = 600 MB Baseline 4 = 1200 MB -2 0.01 0.02 0.03 0.04 0.05 0.06 0.07 Normalized channel estimation error 0.08 0.09 Optimal Suboptimal Baseline Fig. 3. Average number of cooperating BSs versus cache capacity for (a) untrusted BSs for cooperative beamforming transmission of the base-layer and enhancement-layer subfiles, respectively. There is a non-negligible performance gap between the optimal scheme and Baseline 3, particularly in the high cache capacity regime. This is because the proposed caching scheme can exploit the cache resources of the untrusted helpers for delivering enhancement-layer subfiles while Baseline 3 cannot. The performance gap between the proposed optimal scheme and Baselines 1 and 2 is small for small (large) cache capacities because of insufficient (saturated) BS cooperation. For medium cache capacities, however, the proposed optimal scheme achieves considerable performance gains due to its ability to exploit information regarding the user requests and CSI for cache placement. We note that the proposed suboptimal scheme attains good performance in all regimes despite its low computational cost. To provide more insight into how the BSs cooperate, Fig. 3 shows the average numbers of cooperating (small cell and macro) BSs for transmission of the base-layer and enhancement-layer subfiles, denoted as N BL and N EL , respectively, if Algorithm 1 is employed for cache optimization. Recall that the proposed caching scheme does not cache base-layer subfiles at untrusted BSs, cf. constraint C1. Consequently, for a given cache capacity, N BL ≤ N EL holds. Interestingly, the behavior of N BL and N EL is not monotonic as the cache capacity increases. In the small and medium cache capacity regime, N BL and N EL monotonically increase with the cache capacity since the number of subfiles that can be cached at the small cell BSs and the number of BSs that can participate in cooperative transmission increase. For example, N BL and N EL are 0.5 and 0.7 for 300 MB cache capacity per small cell BS, respectively, and increase to 1.7 and 2.7 for 1200 MB cache capacity per small cell BS, respectively. In the large cache capacity regime, the performance gains saturate as the available DoFs for the transmission of the base-layer and enhancement-layer subfiles saturate, cf. Fig. 2. However, N BL and N EL decrease by 1 before reaching the stationary BS cooperation topology. This is because, when the cache-enabled DoFs are sufficient, the optimal caching scheme Optimal Secrecy outage probability transmission of the base-layer and enhancement-layer subfiles when Algorithm 1 is employed for cache optimization. Suboptimal Baseline 4 (b) Fig. 4. (a) Total BS transmit power and (b) secrecy outage probability versus normalized channel estimation error variance for the proposed optimal scheme (solid line), the suboptimal scheme (dotted line), and Baseline 4 (dashed line). selects preferred trusted cooperating (small cell and/or macro) BSs, e.g., based on their channel conditions and cache status, instead of exploiting all BSs available for cooperation. Next, we evaluate the robustness of the proposed schemes w.r.t. channel estimation errors. Figs. 4(a) and 4(b) show the performance of the proposed schemes and Baseline 4 as functions of the normalized channel estimation error variances, σj2 , where Cmax denotes the cache capacity per small cell BS. We observe that, compared to Baseline 4, the proposed schemes achieve a lower secrecy outage probability at the cost of a slightly higher transmit power consumption. Specifically, to achieve robustness in meeting QoS constraint C6 under imperfect CSI, the proposed schemes employ wide beams for transmitting the base-layer subfiles, which may lead to information leakage to the untrusted BSs. Hence, to ensure communication secrecy in constraint C7, the proposed schemes also have to transmit a non-negligible amount of AN to degrade the reception of the untrusted BSs. On the other hand, the wide beams and the interference caused by the AN to the 10 13 =1 =2 -1 Total BS transmit power [dBm] 10 10 -2 300 600 900 1200 1500 Cache capacity per small cell BS [MB] 1800 2100 Optimal Suboptimal (a) Secrecy outage probability legitimate users, cf. (7), have to be compensated by increasing the transmit power of the beamforming vectors. Therefore, the total transmit power increases as the CSI uncertainty increases. In contrast, by treating the imperfect CSI as perfect in C6, Baseline 4 employs narrow transmit beams to save transmit power but this leads to the highest secrecy outage probability in Fig. 4(b). The impact of the number of trusted and untrusted BSs on cache-enabled secrecy is studied in Figs. 5(a) and 5(b). Fig. 5(a) reveals that the required transmit power increases with the number of untrusted BSs Mu , if the total number of BSs is kept constant. This is because, as more helpers become untrusted, fewer (trusted) BSs are available for cooperative transmission of the base-layer subfiles and, at the same time, the trusted BSs have to transmit a larger amount of AN to combat the increasing number of potential eavesdroppers. On the other hand, as the base layers are not cached at the untrusted BSs, for a larger Mu , more cache capacity can be utilized to transmit the enhancement-layer subfiles. Hence, the transmit power of the optimal/suboptimal scheme is only enlarged moderately when Mu increases from 1 to 2; in the high cache capacity regime, the increase in transmit power is even negligible. Due to the cooperative transmission of the base-layer and enhancement-layer subfiles, in Fig. 5(b), the secrecy outage probability monotonically decreases with the cache capacity for Mu ≤ 2. However, when Mu increases from 2 to 3, both the transmit power and the secrecy outage probability are increased significantly; particularly, their values saturate at high levels for cache capacities exceeding 600 MB. This is because, for Mu = 3, the total number of antennas equipped at the untrusted BSs equals the total number of antennas equipped at the trusted BSs; hence, the available DoFs for secure transmission of the base-layer subfiles are limited, irrespective of the cache capacities, as the cache of the untrusted small cell BSs can only facilitate the cooperative transmission of enhancement-layer subfiles. Hence, for secure delivery of the base-layer subfiles, the system has to allocate large amounts of power to AN transmission to degrade the reception quality of the untrusted BSs and, at the same time, to the users’ signals to mitigate the impact of the interference caused by AN to the legitimate users. Finally, Figs. 6(a) and 6(b) show the performance of the proposed schemes and Baseline 3 for larger networks, where the number of users, K, the number of small cells, M − 1, and the number of untrusted small cell BSs, Mu , satisfy M = K + 1 and Mu = K/10. We consider a system bandwidth of 7 MHz to ensure that, despite the large number of users, the QoS requirements of each user can be fulfilled with high probability. For Baseline 3, the caching decisions are determined by employing Algorithms 1 and 2, and the resulting schemes are referred to as “Baseline 3 optimal” and “Baseline 3 suboptimal”, respectively. For K = 50, only the performance of the proposed suboptimal and baseline suboptimal schemes is shown because of the high computational complexity of the optimal schemes. As K increases, more untrusted BSs, Mu = K/10, are present in the network and each untrusted BS can eavesdrop a larger number of users. Hence, the available DoFs per BS for the delivery of the base- Secrecy outage probability =3 Optimal Suboptimal (b) Fig. 5. (a) Total BS transmit power and (b) secrecy outage probability of the proposed optimal (solid line) and suboptimal (dotted line) schemes versus cache capacity for different numbers of untrusted BSs. layer subfiles is reduced and the likelihood of data leakage is increased. Therefore, to ensure secure video streaming, for larger K, a larger average transmit power per BS is needed for AN transmission. On the other hand, exploiting both trusted and untrusted BSs for cooperative transmission, the proposed schemes significantly outperform Baseline 3, particularly for networks with large numbers of users and large cache capacities. VI. C ONCLUSION In this paper, secure video streaming was investigated for small cell networks with untrusted small cell BSs which can intercept both cached and delivered video data. SVC coding and caching were jointly exploited to facilitate secure cooperative MIMO transmission and to not only mitigate the negative impact of the untrusted BSs but to exploit them for secrecy enhancement. A two-timescale non-convex robust optimization problem was formulated to optimize caching and delivery for minimization of the total BS transmit power required for secure video streaming with imperfect CSI knowledge. In the Average transmit power per BS [dBm] 14 Secrecy outage probability (a) where Dt is given in (25). Problems P2 and {(25), (26)} are equivalent as minx∈X , y∈Y f (x, y) = miny∈Y [minx∈X f (x, y)] [43]. Because of the perturbation, the feasible set of problem P2 is a superset of the feasible set of problem P1. Thus, if P1 is feasible, so is P2 (and (25)). Moreover, the inequality constraint functions on the left hand sides of the big-M constraints C3, C8, and C9 are bounded from above, e.g., max max tr(Λm Wρ,l,t ) − qf (ρ),l,m Pm ≤ Pm for C3. Considering st  0 on the right hand sides of C3, C8, and C9, the feasibility statement in part i) of Proposition 1 thus always holds for any q ∈ Q. Next, we show the optimality statement in part i) by contradiction. Assume that (q∗ , D∗t ) solves P1. Then (q∗ , D∗t , st ) with st  0 is feasible for P2. Denote the objective function of P2 by f (q, Dt , st ). We have f (q, Dt , 0) ≥ + f (q∗ , D∗t , 0) , ∀(q, Dt ). Besides, let (q+ , D+ t , st ) be the + optimal solution of P2. If (q+ , Dt ) 6= (q∗ , D∗t ), then s+ t 6=  0 > necessarily holds. However, since µ ≫ 1,we have fPen s+ t + ∗ ∗ f (q∗ , D∗t , 0) and thus f q+ , D+ , s > f (q , D , 0), t t t  + which contradicts the optimality of q+ , D+ t , st . Therefore, part i) is proved. Finally, we prove part ii). Obviously, i) implies that problem P1 is infeasible when ν (q) = +∞. Assume that the optimal + + solution of problem P2, q+ , D+ t , st , satisfies st 6= 0. Then, a feasible solution of the form (q,  Dt , 0) does not exist for P2, + since otherwise, f q+ , D+ , s ≤ f (q, Dt , 0) has to hold. t t Therefore, P1 is also infeasible, which completes the proof. B. Proof of Proposition 2 (b) Fig. 6. (a) Average transmit power per BS and (b) secrecy outage probability versus cache capacity for the proposed optimal and suboptimal schemes and Baseline 3, where M = K + 1 and Mu = K/10. large timescale, the caching optimization problem was solved offline by a modified GBD algorithm. To reduce the computational complexity, a suboptimal caching algorithm was also studied. In the short timescale, the delivery optimization problem for a given cache status was solved online by SDP. Simulation results revealed that, compared to several baseline schemes, the proposed optimal and suboptimal schemes can significantly enhance both the secrecy and the power efficiency of video streaming in small cell networks as long as the total number of antennas at the trusted BSs exceeds that at the untrusted BSs. A PPENDIX A. Proof of Proposition 1 We begin the proof by defining an auxiliary optimization problem: X [UTP (q, Dt ) + fPen (st )] (35) P2: minimize Dt , st , q subject to As strong duality holds for (25), we know that (Dkt , skt ) minimizes the Lagrangian, i.e., (Dkt , skt ) ∈ arg minDt ∈Dt , st 0 Lqk (Dt , st ; λkt ). Since Lqk (Dt , st ) = f1 (qk ; λkt ) + f2 (Dt , st ; λkt ), where f1 (qk ; λkt ) is a constant, we also have (Dkt , skt ) ∈ arg minDt ∈Dt , st 0 f2 (Dt , st ; λkt ). Finally, i) has to hold as Lq (Dt , st ; λt ) is separable w.r.t. q and {Dt , st } in (30), cf. (27). Meanwhile, according to the Karush-Kuhn-Tucker (KKT) conditions for (25), we have C3,k max k − sC3,k [tr(Λm Wρ,l,t ) − qfk(ρ),l,m Pm λρ,l,m,t ρ,l,m,t ] = 0, k C8,k k λρ,l,j,t [tr(Wρ,l,t − Wρ,l,j,t ) − qfk(ρ),l,j Pmax − sC8,k ρ,l,j,t ] = 0, k C9,k k λC9,k ρ,l,j,t [tr(Wρ,l,j,t ) − (1 − qf (ρ),l,j )Pmax − sρ,l,j,t ] = 0, Therefore, min Dt ∈Dt , st 0 f2 (Dt , st ; λkt ) = f2 (Dkt , skt ; λkt )   = UTP qk , Dkt + fPen skt − f1 (qk ; λkt ). By substituting (36) into This completes the proof. Lq (Dt , st ; λkt ), (36) (31) is established. C. Proof of Proposition 3 b ∈ Q be a solution of (30). By Algorithm 1, the Let q b t ) is then generated for q = q t∈T0 b. optimality cut α ≥ ξ(q; λ b C3, C8, C9, C12, Dt ∈ Dt , st  0, q ∈ Q, We have ξ(b q; λt ) = ν(b q) due to the strong duality of the 15 b. If (q, α) were to respective SDP subproblem (25) for q = q b, then α ≥ solve (30) again in another iteration with q = q b t ) and α ≥ ν(b q) = ν(q) would hold, which lead to the ξ(q; λ termination of Algorithm 1 since LB ≥ U B, cf. line 2. This b does not repeat itself in intermediate iterations. implies that q Since set Q ⊆ {0, 1}F ×L×M is finite, Algorithm 1 has to converge in a finite number of iterations. To prove the equivalence between P1 and P0, we show here that the solution of the relaxed problem P1 satisfies rank(Wρ,l,t ) = 1 with probability one. Let Υρ,l,t  0, Φρ,l,t  0, and Θρ,l,t  0 be the Lagrange multipliers f C10, and C12: Wρ,l,t  0, associated with constraints C6, respectively, where C12 is implied by C10. The Lagrangian of problem P1 is given by, minimum value of (38) is obtained for β → ∞, since hρ,t is α H b b H e ρ,l,t w e ρ,l,t statistically independent and −β ηρ,l,t hρ,t → req hρ,t w ρ,l −∞ with probability one. That is, the dual problem (38) is unbounded from below, and consequently, the primal problem is infeasible, which is also a contradiction. Therefore, Bρ,l,t ≻ 0 is proved. Finally, based on (40), (41), and Bρ,l,t ≻ 0, we have, (a) rank(Wρ,l,t ) = rank(Bρ,l,t Wρ,l,t )  α b bH = rank ηρ,l,t req Wρ,l,t hρ,t hρ,t ρ,l n (c) o  α b bH ≤ min rank ηρ,l,t req Wρ,l,t , rank hρ,t hρ,t (b) ρ,l ≤ 1, (43) where (a) is due to Bρ,l,t ≻ 0, (b) is a result of (40) and (41), L(Wρ,l,t ; Υρ,l,t , Φρ,l,t , Θρ,l,t ) = i  X h and (c) follows from the basic rank inequality rank(AB) ≤ 1 Uρ,t Υρ,l,t UH tr Bρ,l,t − Θρ,l,t − ηreq ρ,t Wρ,l,t +∆2 , min {rank(A), rank(B)}. On the other hand, since W ρ,l,t 6= ρ,l ρ,l 0, the condition rank(Wρ,l,t ) = 1 holds with probability one. (37) This completes the proof. where Bρ,l,t , I + ∆1 − Φρ,l,t ; and ∆1  0 and ∆2 ∈ R denote the collection of terms that are relevant and irrelevant R EFERENCES to Wρ,l,t , respectively. Hence, the dual problem of P1 is given [1] L. Xiang, D. W. K. Ng, R. Schober, and V. W. S. Wong, “Secure by video streaming in heterogeneous small cell networks with untrusted min L(Wρ,l,t ; Υρ,l,t , Φρ,l,t , Θρ,l,t ). max Υρ,l,t 0,Φρ,l,t 0,Θρ,l,t 0 Wρ,l,t (38) We now define " Υρ,l,t Υρ,l,t , γH ρ,l,t γ ρ,l,t αρ,l,t # [3] ∈C (N +1)×(N +1) , (39) [4] where Υρ,l,t  0, αρ,l,t ≥ 0 and γ ρ,l,t is chosen to ensure Υρ,l,t  0. If P0 is feasible, so is P1, and then Wρ,l,t 6= 0. Moreover, as strong duality holds for SDP subproblem (25), the optimal beamformers and the optimal dual solutions satisfy the KKT optimality conditions. In particular, by substituting b ρ,t ] and (39) into (37), we have Uρ,t = [IN , h α b b H = Θρ,l,t , (40) Bρ,l,t − ρ,l,t req hρ,t h ηρ,l [2] [5] [6] [7] ρ,t Θρ,l,t Wρ,l,t = 0, (41) where Bρ,l,t =   1 H b H b Υρ,l,t + γ ρ,l,t hρ,t + γ ρ,l,t hρ,t . η req − Bρ,l,t [8] [9] ρ,l Next, we show by contradiction that Bρ,l,t ≻ 0 holds with probability one. Assume that Bρ,l,t has at least one nonpositive eigenvalue τ ≤ 0 and the corresponding eigenvece ρ,l,t , i.e., (Bρ,l,t − τ I)w e ρ,l,t = 0. Let Wρ,l,t = tor is w H e ρ,l,t e ρ,l,t w  0, where β > 0. By substituting Wρ,l,t into βw (38), we further have X H e ρ,l,t e ρ,l,t L(Wρ,l,t ; Υρ,l,t , Φρ,l,t , Θρ,l,t ) = β τ w w −β Xα ρ,l,t ρ,l req ηρ,l | ρ,l {z ≤0 } H b bH w e ρ,l,t hρ,t +∆2 . (42) h ρ,t e ρ,l,t w e ρ,l,t = 0 necessarily holds Hence, if αρ,l,t = 0, then w to ensure L > −∞; yet this contradicts the condition that Wρ,l,t 6= 0. On the other hand, if αρ,l,t > 0, then the [10] [11] [12] [13] [14] [15] cache helpers,” in Proc. IEEE Global Commun. Conf. (GLOBECOM), Singapore, Dec. 2017. V. W. S. Wong, R. Schober, D. W. K. Ng, and L.-C. Wang, Key Technologies for 5G Wireless Systems. Cambridge University Press, 2017. X. Ge, H. Cheng, M. Guizani, and T. Han, “5G wireless backhaul networks: Challenges and research advances,” IEEE Netw., vol. 28, no. 6, pp. 6–11, Nov./Dec. 2014. P. Rost et al., “Cloud technologies for flexible 5G radio access networks,” IEEE Commun. Mag., vol. 52, no. 5, pp. 68–76, May 2014. G. Paschos, E. Baştuğ, I. Land, G. Caire, and M. Debbah, “Wireless caching: Technical misconceptions and business barriers,” IEEE Commun. Mag., pp. 16–22, Aug. 2016. M. Chen, M. Mozaffari, W. Saad, C. Yin, M. Debbah, and C. S. Hong, “Caching in the sky: Proactive deployment of cache-enabled unmanned aerial vehicles for optimized quality-of-experience,” IEEE J. Sel. Areas Commun., vol. 35, no. 5, pp. 1046–1061, May 2017. K. Shanmugam, N. Golrezaei, A. Dimakis, A. Molisch, and G. Caire, “FemtoCaching: Wireless content delivery through distributed caching helpers,” IEEE Trans. Inf. Theory, vol. 59, no. 12, pp. 8402–8413, Dec. 2013. D. Liu and C. Yang, “Energy efficiency of downlink networks with caching at base stations,” IEEE J. Sel. Areas Commun., vol. 34, no. 4, pp. 907–922, Apr. 2016. A. Liu and V. Lau, “Cache-enabled opportunistic cooperative MIMO for video streaming in wireless systems,” IEEE Trans. Signal Process., vol. 62, no. 2, pp. 390–402, Jan. 2014. L. Xiang, D. W. K. Ng, T. Islam, R. Schober, V. W. S. Wong, and J. Wang, “Cross-layer optimization of fast video delivery in cache- and buffer-enabled relaying networks,” IEEE Trans. Veh. Technol., vol. 66, no. 12, pp. 11366–11382, Dec. 2017.. M. A. Maddah-Ali and U. Niesen, “Fundamental limits of caching,” IEEE Trans. Inf. Theory, vol. 60, no. 5, pp. 2856–2867, May 2014. ——, “Decentralized coded caching attains order-optimal memory-rate tradeoff,” IEEE/ACM Trans. Netw., vol. 23, no. 4, pp. 1029–1040, Aug. 2015. J. Zhang, X. Lin, and X. Wang, “Coded caching under arbitrary popularity distributions,” in Proc. Inf. Theory and App. Workshop (ITA), 2015, San Diego, CA, Feb. 2015. N. Naderializadeh, M. A. Maddah-Ali, and A. S. Avestimehr, “On the optimality of separation between caching and delivery in general cache networks,” in Proc. Int. Symp. Inf. Theory (ISIT), Aachen, Germany, Jun. 2017. A. A. Zewail and A. Yener, “Coded caching for combination networks with cache-aided relays,” in Proc. Int. Symp. Inf. Theory (ISIT), Aachen, Germany, Jun. 2017. 16 [16] L. Xiang, D. W. K. Ng, R. Schober, and V. W. S. Wong, “Cache-enabled physical layer security for video streaming in backhaul-limited cellular networks,” to be published in IEEE Trans. Wireless Commun.. [17] F. Gabry, V. Bioglio, and I. Land, “On edging caching with secrecy constraints,” in Proc. IEEE Int’l Conf. Commun. (ICC), Kuala Lumpur, Malaysia, May 2016. [18] A. A. Zewail and A. Yener, “Coded caching for resolvable networks with security requirements,” in Proc. Commun. and Netw. Security (CNS), Philadelphia, PA, Oct. 2016. [19] L. Xiao, C. Xie, T. Chen, H. Dai, and H. V. Poor, “A Mobile Offloading Game Against Smart Attacks”, IEEE Access, vol. 4, pp. 2281–2291, 2016. [20] J. Wright and J. Cache, Hacking Exposed Wireless: Wireless Security Secrets & Solutions, 3rd ed. McGraw-Hill Education Group, 2015. [21] G. Mantas, N. Komninos, J. Rodriuez, E. Logota, and H. Marques, “Security for 5G communications,” in Fundamentals of 5G Mobile Networks, J. Rodriguez, Ed. John Wiley & Sons, Ltd., 2015. [22] 3GPP TR 33.820 v8.3.0, “Technical specification group service and system aspects: Security of H(e)NB (Release 8),” Dec. 2009. [23] E. Rescorla, “HTTP over TLS,” IETF RFC 2818, May 2000. [24] F. Callegati, W. Cerroni, and M. Ramilli, “Man-in-the-middle attack to the HTTPS protocol,” IEEE Security Privacy, vol. 7, no. 1, pp. 78–81, Jan./Feb. 2009. [25] X. He and A. Yener, “Cooperation with an untrusted relay: A secrecy perspective,” IEEE Trans. Inf. Theory, vol. 56, no. 8, pp. 3807–3827, Aug. 2010. [26] A. Checko, H. L. Christiansen, Y. Yan, L. Scolari, G. Kardaras, M. S. Berger, and L. Dittmann, “Cloud RAN for mobile networks: A technology overview,” IEEE Commun. Surveys & Tuts., vol. 17, no. 1, pp. 405–426, First Quarter 2015. [27] J. Zhao, T. Q. Quek, and Z. Lei, “Coordinated multipoint transmission with limited backhaul data transfer,” IEEE Trans. Wireless Commun., vol. 12, no. 6, pp. 2762–2775, Jun. 2013. [28] F. Zhuang and V. K. Lau, “Backhaul limited asymmetric cooperation for MIMO cellular networks via semidefinite relaxation,” IEEE Trans. Signal Process., vol. 62, no. 3, pp. 684–693, Feb. 2014. [29] D. W. K. Ng and R. Schober, “Secure and green SWIPT in distributed antenna networks with limited backhaul capacity,” IEEE Trans. Wireless Commun., vol. 14, no. 9, pp. 5082–5097, Sept 2015. [30] S. Goel and R. Negi, “Guaranteeing secrecy using artificial noise,” IEEE Trans. Wireless Commun., vol. 7, no. 6, pp. 2180–2189, Jun. 2008. [31] Q. Li and W.-K. Ma, “Optimal and robust transmit designs for MISO channel secrecy by semidefinite programming,” IEEE Trans. Signal Process., vol. 59, no. 8, pp. 3799–3812, Aug. 2011. [32] J.-R. Ohm, “Advances in scalable video coding,” Proc. IEEE, vol. 93, no. 1, pp. 42–56, Jan. 2005. [33] H. Schwarz, D. Marpe, and T. Wiegand, “Overview of the scalable video coding extension of the H. 264/AVC standard,” IEEE Trans. Circuits Syst. Video Technol., vol. 17, no. 9, pp. 1103–1120, Sep. 2007. [34] S. de la Fuente et al., “iDASH: improved dynamic adaptive streaming over HTTP using scalable video coding,” in Proc. ACM MMSys, California, CA, Feb. 2011. [35] Y. Zhang, N. Meratnia, and P. Havinga, “Outlier detection techniques for wireless sensor networks: A survey,” IEEE Commun. Surveys Tuts., vol. 12, no. 2, pp. 159–170, 2010. [36] P. Gogoi, D. Bhattacharyya, B. Borah, and J. K. Kalita, “A survey of outlier detection methods in network anomaly identification,” The Comput. J., vol. 54, no. 4, pp. 570–588, Apr. 2011. [37] V. Hodge and J. Austin, “A survey of outlier detection methodologies,” Artif. Intell. Review, vol. 22, no. 2, pp. 85–126, Oct. 2004. [38] D. Bharadia, E. McMilin, and S. Katti, “Full duplex radios,” ACM SIGCOMM Comput. Commun. Review, vol. 43, no. 4, pp. 375–386, 2013. [39] B. P. Day, A. R. Margetts, D. W. Bliss, and P. Schniter, “Full-duplex bidirectional MIMO: Achievable rates under limited dynamic range,” IEEE Trans. Signal Process., vol. 60, no. 7, pp. 3702–3713, Jul. 2012. [40] D. Tse and P. Viswanath, Fundamentals of Wireless Communication. Cambridge University Press, 2005. [41] C. A. Floudas, Nonlinear and Mixed Integer Optimization: Fundamentals and Applications. Oxford University Press, 1995. [42] D. P. Palomar and Y. C. Eldar, Convex Optimization in Signal Processing and Communications. Cambridge University Press, 2010. [43] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge University Press, 2004. [44] X. Ge, S. Tu, G. Mao, C.-X. Wang and T. Han, “5G Ultra-Dense Cellular Networks,” IEEE Wireless Commun., vol. 23, no. 1, pp. 72– 79, Feb. 2016. [45] Z.-Q. Luo, J. F. Sturm, and S. Zhang, “Multivariate nonnegative quadratic mappings,” SIAM J. Optimization, vol. 14, no. 4, pp. 1140– 1162, Jul. 2004. [46] M. Grant and S. Boyd, “CVX: Matlab software for disciplined convex programming, version 2.1,” [Online] Available: http://cvxr.com/cvx, Dec. 2016. [47] Mosek ApS, “The MOSEK optimization software, version 8.0.0.45,” [Online] Available: http://www.mosek.com, Nov. 2016. [48] Y. Ye, Interior Point Algorithms: Theory and Analysis. John Wiley & Sons, 1997. [49] I. Pólik and T. Terlaky, “Interior point methods for nonlinear optimization,” in Nonlinear Optimization. Springer Berlin Heidelberg, 2010, pp. 215–276. [50] B. Korte and J. Vygen, Combinatorial Optimization, 5th ed. Springer, 2012. [51] N. D. Sidiropoulos, T. .N. Davidson, and Z. Luo, “Transmit beamforming for physical-layer multicasting,” IEEE Trans. Signal Process., vol. 54, no. 6, pp. 2239–2251, Jun. 2006. [52] L. Breslau, P. Cao, L. Fan, G. Phillips, and S. Shenker, “Web caching and Zipf-like distributions: Evidence and implications,” in Proc. IEEE INFOCOM, New York, NY, Mar. 1999. [53] 3GPP TR 36.814 v9.0.0, “Further advancements for E-UTRA physical layer aspects (Release 9),” Mar. 2010.
7cs.IT
arXiv:1706.02420v1 [math.PR] 8 Jun 2017 Berry-Esséen bounds for parameter estimation of general Gaussian processes Soukaina Douissi1 , Khalifa Es-Sebaiy2 and Frederi G. Viens Cadi Ayyad University and Michigan State University 3 Abstract: We study rates of convergence in central limit theorems for the partial sum of squares of general Gaussian sequences, using tools from analysis on Wiener space. No assumption of stationarity, asymptotically or otherwise, is made. The main theoretical tool is the so-called Optimal Fourth Moment Theorem [19], which provides a sharp quantitative estimate of the total variation distance on Wiener chaos to the normal law. The only assumptions made on the sequence are the existence of an asymptotic variance, that a least-squares-type estimator for this variance parameter has a bias and a variance which can be controlled, and that the sequence’s auto-correlation function, which may exhibit long memory, has a no-worse memory than that of fractional Brownian motion with Hurst parameter H < 3/4. Our main result is explicit, exhibiting the trade-off between bias, variance, and memory. We apply our result to study drift parameter estimation problems for subfractional Ornstein-Uhlenbeck and bifractional Ornstein-Uhlenbeck processes with fixed-time-step observations. These are processes which fail to be stationary or self-similar, but for which detailed calculations result in explicit formulas for the estimators’ asymptotic normality. Key words: Central limit theorem; Berry-Esseen; stationary Gaussian process; NourdinPeccati analysis; parameter estimation; fractional Brownian motion; long memory. 2010 Mathematics Subject Classification: 60F05; 60G15; 60H05; 60H07, 62F12. 1 Introduction This paper presents the estimation of the asymptotic variance, when it exists, of a general Gaussian sequence, and applies it to non-stationary stochastic processes based on fractional noises, which can have long memory. We choose to work with a simple method based on empirical second moments, which is akin to the discretization of a least-squares method for 1 Cadi Ayyad University, Marrakesh, Morocco. Email: [email protected] National School of Applied Sciences - Marrakesh, Cadi Ayyad University, Marrakesh, Morocco. Email: [email protected] 3 Department of Statistics and Probability, Michigan State University 619 Red Cedar Rd East Lansing, MI 48824, USA. E-mail: [email protected] 2 1 the corresponding processes using discrete observations with no in-fill assumption (fixed time step). Our work responds to the recent preprint [12], which was the first instance in which this type of estimation was performed without in-fill assumptions on the data, but which insisted on keeping the data stationary or asymptotically so. That paper [12], available on arXiv, contains an extensive description of the literature on parameter estimation for Gaussian stochastic processes. We do not repeat that analysis here. Instead, we list a number of recent articles which deal with various versions of least-squares-type parameter estimation for fractional-noise-driven Ornstein-Uhlenbeck processes (fOU-type), since that class of processes is similar to the various examples we consider in this article: [14], [1], [2], [11], [15], [5], [18], [3], [9] and [10]. respectively. Those papers are described in the context of our motivations in [12]. We now highlight the distinction between our paper and [12]. In that work, the assumption of stationarity of the data was weakened by asking for asymptotic stationarity, where the deviation from a stationary model converged to 0 exponentially fast. This strong assumption only allowed to work with specific examples which are close to stationary. In that article, Ornstein-Uhlenbeck models were covered, as long as the driving noise was stationary, as it would be for standard fractional Brownian motion. That paper was the first instance in which parameters for fOU-type processes could be estimated with asymptotic normality using only discrete data with fixed time step. However, this range of application in [12] extended to non-stationary processes which benefitted from an exponential ergodicity property. One leaves this range of applicabilith when driving OU equations with noises which are not stationary to begin with. This is our primary motivation in this article: to develop a framework in which no assumption of stationarity is made at the level of driving noise processes. Instead, we investigate a set of arguably minimal assumptions on how to estimate an asymptotic variance, if it exists, for any discretely observed Gaussian process. In this article, we make no a-priori assumption about the speed of convergence of the data to a stationary model, only asking that convergence occur. This implies that the main result of the paper, our estimator’s asymptotic normality, applies to some of the more exotic non-stationary covariance structures, for which the estimator properties cannot be handled by any classical or current method. Our main result isolates three terms in order to construct estimates which are fully quantitative. We show that the Wasserstein distance between the CLT scaling of our discrete-observation leastsquares estimator of the asymptotic variance and a normal law is bounded above by the sum of a bias term, a variance term, and a term which comes from an application of the optimal fourth moment theorem of Nourdin and Peccati [19], and depends only on an upper bound on the covariance structure of the discrete observations. This allows us to formulate clear assumptions on what is needed to turn this estimate into a quantiative asymptotic normality theorem for the estimator, in Wasserstein distance. The advantage of working in this fashion is that one has an immediate way of identifying possible mean-variance-correlation trade-offs, since the corresponding three terms are explict and are simply added together as 2 a measure of how fast the estimator converges to a normal. There are additional features to our method which are described in Section 3, such as our ability to separate our minimal convergence assumptions between what is needed for strong consistency and what is needed in addition to obtain asymptotic normality: see Hypotheses (H1) thru (H4) in Section 3, the bullet points that follow, and the statements of Theorems 4 and 6 in the same section. When applying our general method from Section 3 to specific examples in Section 4, we consider the solutions of the OU equation driven by sub-fractional Brownian motion and by bi-fractional Brownian motion, both of which have somewhat non-trivial covariance structures with non-stationary increments, which are then inherited by the corresponding fOU-type processes whose drift parameter θ is of interest. These fOU-type processes have been well studied in the literature over the last decade, including efforts to understand their statistical inference. We refer only to [9] for information about these processes and additional references. We compute the asymptotic variances of these fOU-type processes, which are explicit functions f (θ), depending only on θ (and on the Hurst-type parameters which are assumed to be known), and are in fact power functions of θ. The normal asymptotics for the estimators of each of these two f (θ)’s can be converted to similar results for θ itself; see [12] for instance for an analysis, which we do not explore further here. Part of being able to apply the strategy of Section 3 successfully in the examples of Section 4, relies on three things, the first two which cannot be avoided, the third which enables an efficient estimation of convergence speeds: 1. being able to identify f (θ), preferably in an explicit way, thus giving access to θ, but at least to show that the data stream’s variance converges; (a) we do this for both the sub-fractional and bi-fractional OU processes; (b) the expressions for f (θ) are simple (see the formulas for fH (θ) and fH,K (θ) for each process respectively, in (15) and (20)) but proving convergence and evaluating the speed therein is slightly non-trivial; 2. showing that the CLT-normalized estimator of f (θ) has an asymptotic variance σ 2 ; this does not need to be found explicitly, as long as one can prove that the variances converge; (a) we succeed in this task for both the sub-fractional and bi-fractional OU processes, and are moreover able to compute σ 2 semi-explicitly, using a series representation based on the covariance function of discrete-time fractional Gaussian noise (fGn); see formulas (19) and (21) respectively; (b) these expressions comes as something of a surprise, since a priori there was no reason to expect that the estimator’s asymptotic variance could be expressed using the fGn’s covariance rather than the non-stationary covariance structure of the increments of the sub- or bi-fractional processes; 3 3. computing speeds of convergence of both the estimator’s bias and its asymptotic variance, as explicitly as possible; (a) this is done for both the sub-fractional and bi-fractional OU processes, in a rather explicit fashion, and requires a certain amount of elbow grease, where most calculations are relegated to lemmas in the Appendix; (b) we do not claim to have performed these estimations as efficiently as can possibly be done; however, we think that the power orders of these convergences are probably optimal because: (i) we have based the estimations on the optimal fourth moment theorem of Nourdin and Peccati [19], (ii) we have confirmed a phenomenon observed in a simpler (stationary) context in [17] (third moment theorem) by which the speed of convergence is given by that of an absolute third moment rather than by a fourth cumulant, and (iii) our final result for the suband bi-fractional processes identify a well-known threshold of a Hurst parameter not exceeding 3/4 as the upper endpoint of the range where quadratic variations are asymptotically normal. To finish this introduction, we note the general structure of this paper. The next section presents preliminaries regarding the analysis of Wiener space, and other convenient facts, which are needed in some technical parts of the paper. The reader interested in our results more than our proofs can safely skip this Section 2 upon a first reading. Section 3 presents the general framework of how to estimate the asymptotic variance of a Gaussian process observed in discrete time. Section 4 shows how this method can be implemented in two cases of discretely observed OU processes driven by two well-known non-stationary Gaussian processes with medium-to-long memory. The Appendix collects calculations useful in Section 4. 2 Preliminaries, and elements of analysis on Wiener space This section provides essential facts from basic probability and the analysis on Wiener space and the Malliavin calculus. These facts and their corresponding notation underlie all the results of this paper, but most of our results and arguments can be understood independently of this section. The interested reader can find more details in [21, Chapter 1] and [20, Chapter 2]. 2.1 A convenient lemma The following result is a direct consequence of the Borel-Cantelli Lemma (the proof is elementary; see e.g. [16]). It is convenient for establishing almost-sure convergences from Lp convergences. 4 Lemma 1 Let γ > 0. Let (Zn )n∈N be a sequence of random variables. If for every p > 1 there exists a constant cp > 0 such that for all n ∈ N, kZn kLp (Ω) 6 cp · n−γ , then for all ε > 0 there exists a random variable ηε which is almost such that |Zn | 6 ηε · n−γ+ε almost surely for all n ∈ N. Moreover, E|ηε |p < ∞ for all p > 1. A typical application is when a sequence in a finite sum of Wiener chaoses converges in the mean square. By the equivalence of all Lp norms in Wiener chaos, which is a consequence of the hypercontractivity of the Ornstein-Uhlenbeck semigroup on Wiener space (see [20, Section 2.8], and Section 2.3 below), the lemma’s assumption is then automatically satisfied if the speed of convergence in the mean square is of the same power form. The random variable ηε can typically be chosen more explicitly than what the lemma guarantees. For instance, it is well known that if Z is Gaussian, then ηε can be chosen as a gauge function of Z, and is therefore sub-Gaussian (tails no larger than Gaussian; see [13]). A similar property holds for processes in a finite sum of Wiener chaoses, as can be inferred from [23]. We will not pursue these refinement, since the above lemma will be sufficient for our purposes herein. 2.2 The Young integral Fix T > 0. Let f, g : [0, T ] −→ R be Hölder continuous functions of orders α ∈ (0, 1) and β ∈ (0, 1) respectively with α+β R . > 1. Young R . [24] proved that the Riemann-Stieltjes integrals (so-called Young integrals) 0 fu dgu and 0 gu dfu exist as limits of the usual discretization. Morever, for all t ∈ [0, T ], integration by parts (product rule) holds: Z t Z t ft gt = f0 g0 + gu dfu + fu dgu . (1) 0 0 This integral is convenient to define stochastic integrals in a pathwise sense with respect to processes f which are smoother than Brownian motion, in the sense that they are almost surely α-Hölder continuous for some α > 1/2. A typical example of such a process is the fractional Brownian motion with Hurst parameter H > 1/2. In that case, any α can be chosen in (1/2, H), and g can be any process with the same regularity as the fBm, and thus with β = α; this enables a stochastic calculus immediately, in which no “Itô”-type correction terms occur, owing to (1). For instance, for any Lipshitz non-random function g, this integration-by-parts formula holds for any a.s. Hölder-continuous Gaussian process f , since then α > 0 and β = 1. In particular, the first integral in (1) coincides with the Wiener integral, the second is an ordinary Riemann-Stieltjes integral, and no Itô correction occurs since one of the processes is of finite variation. 5 2.3 Elements of Analysis on Wiener space With (Ω, F , P) denoting the Wiener space of a standardRWiener process W , for a deterministic function h ∈ L2 (R+ ) =: H, the Wiener integral R+ h (s) dW (s) is also denoted by R W (h). The inner product R+ f (s) g (s) ds will be denoted by hf, giH . • The Wiener chaos expansion. For every q > 1, Hq denotes the qth Wiener chaos of W , defined as the closed linear subspace of L2 (Ω) generated by the random variables {Hq (W (h)), h ∈ H, khkH = 1} where Hq is the qth Hermite polynomial. All Wiener chaos random variable are orthogonal in L2 (Ω). The so-called Wiener chaos expansion is the fact that any X ∈ L2 (Ω) can be written as X = EX + ∞ X Xq (2) q=0 for some Xq ∈ Hq for every q > 1. This is summarized in the direct-orthogonal-sum notation L2 (Ω) = ⊕∞ q=0 Hq . Here H0 denotes the constants. • Relation with Hermite polynomials. Multiple Wiener integrals. The mapping Iq (h⊗q ) : = q!Hq (W (h)) is a linear isometry between the symmetric tensor product H⊙q (equipped with the modified norm k.kH⊙q = √1q! k.kH⊗q ) and Hq . To relate this to standard stochastic calculus, one first notes that Iq (h⊗q ) can be interpreted as the multiple Wiener integral of h⊗q w.r.t. W . By this we mean that the Riemann-Stieltjes approximation of such an integral converges in L2 (Ω) to Iq (h⊗q ). This is an elementary fact from analysis on Wiener space, which can also be proved using standard stochastic calculus for square-integrable martingales, because the multiple integral interpretation of Iq (h⊗q ) as a Riemann-Stieltjes integral over (R+ )q can be further shown to coincinde with q! times the iterated Itô integral over the first simplex in (R+ )q . More generally, for X and its Wiener chaos expansion (2) above, each term Xq can be interpreted as a multiple Wiener intergral Iq (fq ) for some fq ∈ H⊙q . • The product formula - Isometry property. For every f, g ∈ H⊙q the following extended isometry property holds E (Iq (f )Iq (g)) = q!hf, giH⊗q . Similarly as for Iq (h⊗q ), this formula is established using basic analysis on Wiener space, but it can also be proved using standard stochastic calculus, owing to the coincidence of Iq (f ) and Iq (g) with iterated Itô integrals. To do so, one uses Itô’s version of integration by parts, in which iterated calculations show coincidence of the expectation of the bounded variation term with the right-hand side above. What is typically referred to 6 as the Product Formula on Wiener space is the version of the above formula before taking expectations (see [20, Section 2.7.3]). In our work, beyond the zero-order term in that formula, which coincides with the expectation above, we will only need to know the full formula for q = 1, which is I1 (f )I1 (g) = 2−1 I2 (f ⊗ g + g ⊗ f ) + hf, giH. • Hypercontractivity in Wiener chaos. For h ∈ H⊗q , the multiple Wiener integrals Iq (h), which exhaust the set Hq , satisfy a hypercontractivity property (equivalence in Hq of all Lp norms for all p > 2), which implies that for any F ∈ ⊕ql=1 Hl (i.e. in a fixed sum of Wiener chaoses), we have  1/p  1/2 E |F |p 6 cp,q E |F |2 for any p > 2. (3) It should be noted that the constants cp,q above are known with some precision when F is a single chaos term: indeed, by Corollary 2.8.14 in [20], cp,q = (p − 1)q/2 . • Malliavin derivative. The Malliavin derivative operator D on Wiener space is not needed explicitly in the paper. However, because of the fundamental role D plays in evaluating distances between random variables, it is helpful to introduce it, to justify the estimates (5) and (6) below. For any function Φ ∈ C 1 (R) with bounded derivative, and any h ∈ H, the Malliavin derivative of the random variable X := Φ (W (h)) is defined to be consistent with the following chain rule: DX : X 7→ Dr X := Φ′ (W (h)) h (r) ∈ L2 (Ω × R+ ) . A similar chain rule holds for multivariate Φ. One then extends D to the so-called Gross-Sobolev subset D1,2 & L2 (Ω) by closing D inside L2 (Ω) under the norm defined by  Z  2 2 2 |Dr X| dr . kXk1,2 = E X + E R+ All Wiener chaos random variable are in the domain D1,2 of D. In fact can P this domain 1,2 2 be expressed explicitly for any X as in (2): X ∈ D if and only if q qkfq kH⊗q < ∞. • Generator L of the Ornstein-Uhlenbeck semigroup. The linear operator L is defined as being diagonal under the Wiener chaos expansion of L2 (Ω): Hq is the eigenspace of L with eigenvalue −q, i.e. for any X ∈ Hq , LX = −qX. We have Ker(L) = H0 , the constants. The operator −L−1 is the negative pseudo-inverse of L, so that for any X ∈ Hq , −L−1 X = q −1 X. Since the variables we will be dealing with in this article are finite sums of elements of Hq , the operator −L−1 is easy to manipulate thereon. The use of L is crucial when evaluating the total variation distance between the laws of random variables in Wiener chaos, as we will see shortly. 7 • Distances between random variables. Recall that, if X, Y are two real-valued random variables, then the total variation distance between the law of X and the law of Y is given by dT V (X, Y ) = sup |P [X ∈ A] − P [Y ∈ A]| A∈B(R) where the supremum is over all Borel sets. If X, Y are two real-valued integrable random variables, then the Wasserstein distance between the law of X and the law of Y is given by dW (X, Y ) = sup |Ef (X) − Ef (Y )| f ∈Lip(1) where Lip(1) indicates the collection of all Lipschitz functions with Lipschitz constant 6 1. Let N denote the standard normal law. • Malliavin operators and distances between laws on Wiener space. There are two key estimates linking total variation distance and the Malliavin calculus, which were both obtained by Nourdin and Peccati. The first one is an observation relating an integration-by-parts formula on Wiener space with a classical result of Ch. Stein. The second is a quantitatively sharp version of the famous 4th moment theorem of Nualart and Peccati. – The observation of Nourdin and Peccati. Let X ∈ D1,2 with E [X] = 0 and V ar [X] = 1. Then (see [19, Proposition 2.4], or [20, Theorem 5.1.3]), for f ∈ Cb1 (R),   E [Xf (X)] = E f ′ (X) DX, −DL−1 X H and by combining this with properties of solutions of Stein’s equations, one gets dT V (X, N) 6 2E 1 − DX, −DL−1 X H . (4) One notes in particular that when X ∈ Hq , since −L−1 X = q −1 X, we obtain conveniently (5) dT V (X, N) 6 2E 1 − q −1 kDXk2H . It is this last observation which leads to a quantitative version of the fourth moment theorem of Nualart and Peccati, which entails using Jensen’s inequality to note that the right-hand side of (4) is bounded above by the variance of hDX, −DL−1 XiH , and then relating that variance in the case of Wiener chaos with the 4th cumulant of X. However, this strategy was superceded by the following, which has roots in [4]. – The optimal fourth moment theorem. For each integer n, let Xn ∈ Hq . Assume V ar [Xn ] = 1 and (Xn )n converges in distribution to a normal law. It 8 is known (original proof in [22], known as the fourth moment theorem) that this convergence is equivalent to limn E [Xn4 ] = 3. The following optimal estimate for dT V (X, N), known as the optimal fourth moment theorem, was proved in [19]: with the sequence X as above, assuming convergence, there exist two constant c, C > 0 depending only on X but not on n, such that           . 6 dT V (Xn , N) 6 C max E Xn4 − 3, E Xn3 c max E Xn4 − 3, E Xn3 (6) Given its importance, the centered fourth moment, also known as the fourth cumulant, of a standardized random variable, warrants the following special notation:   κ4 (X) := E X 4 − 3. Let us also recall that if E[Fn ] = 0, we denote κ3 (Fn ) = E[Fn3 ] its third cumulant. Throughout the paper we use the notation N ∼ N (0, 1). We also use the notation C for any positive real constant, independently of its value which may change from line to line when this does not lead to ambiguity. 3 General Context Let G = {Gt , t > 0} be an underlying Gaussian process, and let X = {Xn }n>0 be a meanzero Gaussian sequence measurable with respect to G. This means that X is a sequence of random variables which can be represented as Xt = I1 (ht ) for every t > 0, i.e. Xt is a Wiener integral with respect to G, where ht ∈ H with H the Hilbert space associated to G. In particular, we do not assume that X is a stationary process. Define n 1X An (X) := E[Xi2 ] n i=1 n 1 X 2 Vn (X) := √ (X − E[Xi2 ]). n i=1 i and As we explained in the introduction, the goal is to estimate the asymptotic variance f of the sequence X, if it exists. Assuming it exists is the first of the following four assumptions, i.e. (H1). The other three give us various levels of additional regularity for the law of X to measure the speed of convergence of a quadratic-variations-based estimator for f : (H1) E[Xn2 ] −→ f as n −→ ∞. (H2) vn := E[Vn (X)2 ] −→ σ 2 > 0, as n −→ ∞. (H3) |E[Xt Xs ]| 6 Cρ (|t − s|) where ρ : R → R is a symmetric positive function such that ρ(0) = 1, and C is a positive constant. 9 (H4) √ n|An (X) − f | −→ 0, as n −→ ∞. We consider the following estimator of f n 1X 2 X . fˆn (X) = n i=1 i (7) We refer to [11] and [12] for the construction of this estimator (7) for some classes of processes and sequences, i.e. how it can be interpreted as a discretization of a least squares estimator of f . The next two subsections study strong consistency and speed of convergence in the asymptotic normality of fˆn (X). We make some comments on the roles and motivations behind the four assumptions above. • As we said, Assumption (H1) states that the asymptotic variance f of the sequence X exists. • Assumption (H2) helps establish strong consistency of fˆn (X) via the Borel-Cantelli lemma and hypercontractivity of Wiener chaos. It states that a proper standartization of fˆn (X) has an asymptotic variance; it is also used to help establish a quantitative upper bound on the total variation distance between a standardized version of fˆn (X) and the standard normal law. • Assumption (H3) formalizes the idea that while the sequence X is not stationary, it may have a covariance structure which is bounded above by one which may be comparable to a stationary one. This assumption itself is largely used as a matter of convenience, since formally by Schwartz’s inequality it can always be assumed to hold for the trivial case ρ ≡ 1. However, by making further quantitative estimates on the rate of decay of ρ to 0, we find convenient explicit upper bound expressions for the total variation distance of between a standardization of fˆn (X) and the normal law. The corresponding results are in Theorem 4. • In particular, Assumption (H4) quantifies the speed of convergence of the mean of fˆn (X) towards f ; combined with Assumption (H2) which determines the speed of convergence of the variance of fˆn (X), and using estimates on Wiener space, one can arrive  √ ˆ at a fully quantified upper bound on the Wasserstein distance between n fn (X) − f and the law N (0, σ 2 ). The corresponding results are in Theorem 6. • The last point above is significant because Theorem 6 does not rely on using vn , which is not observed, to standardize fˆn (X). This theorem also decomposes the distance to the limiting normal law into the sum of three explicit terms: one to account for the bias of fˆn (X) (from Assumption (H4)), one to account for the speed of convergence to the asymptotic variance (from Assumption (H2)), and one from the analysis on Wiener space which uses the speed of decay of the correlations of X (from Assumption (H3)). 10 3.1 Strong consistency The following theorem provides sufficient assumptions to obtain the estimator fˆn (X)’s strong consistency, i.e. almost sure convergence to f . Theorem 2 Assume that (H1) and (H2) hold. Then fˆn (X) −→ f (8) almost surely as n −→ ∞. Proof. It is clear that Vn (X) + An (X). fˆn (X) = √ n The hypothesis (H1) and the convergence of Cesàro means imply that, as n −→ ∞ An (X) −→ f. In order to prove (8) it remains to check that, as n −→ ∞, According to (H2), there exists C > 0 such that for all n > 1 E " Vn (X) √ n 2 #!1/2 Vn (X) √ n −→ 0 almost surely. C 6√ . n Now, using the hypercontractivity property (3) and Lemma 1 the result is obtained. 3.2 Asymptotic normality In this section we study the asymptotic normality of fˆn (X). By the product formula, we can write n n 1 X 2 1 X Vn (X) = √ (Xk − E[Xk2 ]) = √ I2 (fk⊗2 ). n k=1 n k=1 Set vn (X) := E[Vn (X)2 ] = n 2 X (E[Xj Xk ])2 , n j,k=1 where the second equality is from elementary covariance calculation, and Vn (X) Fn (X) := p = I2 (gn ) vn (X) 11 where gn := (nvn (X)) −1/2 n X fk⊗2 . k=1 Let us estimate the third cumulant of Fn (X). We have for every n > 1, by the observation of Nourdin and Peccati (or using calculus on Wiener chaos), we have κ3 (Fn (X)) = 2E[Fn (X)Γ1 (Fn (X))], e 1 gn ) + 2kgn k2H⊗2 . Assume that (H3) holds. We then have where Γ1 (Fn (X)) = 2I2 (gn ⊗ e 1 gn i2H⊗2 κ3 (Fn (X)) = 8hgn , gn ⊗ = 8hgn , gn ⊗1 gn i2H⊗2 n X 8 hfi , fk iH hfi , fj iH hfk , fj iH = (nvn (X))3/2 i,j,k=1 n X 8 = E[Xi Xk ]E[Xi Xj ]E[Xj Xk ] (nvn (X))3/2 i,j,k=1 Therefore, n X 1 |ρ(i − k)ρ(i − j)ρ(j − k)| |κ3 (Fn (X))| E (nvn (X))3/2 i,j,k=1  2 X 1 √  |ρ(k)|3/2  . E vn (X)3/2 n (9) |k|<n The last equality follows from the same argument as for the proof of [4, Proposition 6.3]. The symbol E means we omit multiplicative universal constants. Now for the fourth cumulant, similarly, we have 1 κ4 (Fn (X)) = vn (X)2 n2 n X E[Xk1 Xk2 ]E[Xk2 Xk3 ]E[Xk3 Xk4 ]E[Xk4 Xk1 ] k1 ,k2 ,k3 ,k4 =1 n X 1 |ρ(k − l)ρ(i − j)ρ(k − i)ρ(l − j)| vn (X)2 n2 k ,k ,k ,k =1 1 2 3 4 3  X 4 1  |ρ(k)| 3  E 2 vn (X) n E (10) |k|<n where again the last inequality comes from a similar argument as for the proof of [4, Proposition 6.4]. 12 Remark 3 In [4, Proposition 6.3, Proposition 6.4] the sequence {Xn }n>0 is a centered stationary Gaussian sequence. In our case, it is not necessarily stationary. Our Hypothesis (H3) is sufficient to avoid the assumption of stationarity. Also note that there is no need to take the absolute value of κ4 because the fourth cumulant of a variable in a fixed chaos (Fn is in the 2nd chaos) is known to have a positive 4th cumulant. Theorem 4 Assume that the hypothesis (H3) holds. Then for all n > 1,   3   2   X X 4 1 1 3/2    3 √ |ρ(k)| dT V (Fn (X), N) 6 C max . |ρ(k)| ,  vn (X)3/2 n  vn (X)2 n |k|<n |k|<n −β In particular, if for some β > 1/2, we have ρ (t) = O |t| C > 0 such that, for all n > 1,  1   3   n 2 −3β    C dT V (Fn (X), N) 6 1 3/2 n− 2 log(n)2 vn2 ∧ vn        −1 n 2 (11) for large |t|, then there exists  if β = if β ∈ 1 2 if β = 2 3 1 2 , 2 3  (12) if β > 23 . Proof. The estimate (11) is a direct consequence of the optimal estimate in (6), and the estimates (9) and (10) of the absolute third and fourth cumulants we just computed. For (12), since  √ n if β = 12    3   n1− 2 β if β ∈ 21 , 23    X |ρ(k)|3/2 6 C  log n if β = 32  |k|<n      1 if β > 23 and X |k|<n |ρ(k)|4/3 6 C the desired result is obtained.  n   −4   n 3 β+1     log n       1 13 if β = if β ∈ 1 2 if β = 3 4 if β > 3 4 1 3 , 2 4  Remark 5 The phenomenon observed in the proof of the previous theorem, by which the estimate of the third cumulant dominates that of the fourth cumulant, was observed originally in [4, Proposition 6.3, Proposition 6.4] for stationary sequences, and shown in [17] to be completely general in the stationary case (i.e. not related to the power scale). Here we see that stationarity is not required. This begs the question of whether the phenomenon generalizes to other sequences in the second chaos, for instance without using hypothesis (H3). We do not investigate this question here, since it falls well outside the scope of our motivating topic of parameter estimation. Theorem 6 Assume that the hypothesis (H3) holds. If for some 21 6 β, we have ρ (t) =  O |t|−β for large |t|, then for all n > 1,  1 if β = 21        3   r   r n 2 −3β if β ∈ 21 , 32   n ˆ n C dW fn (X) − f , N 6 |An (X) − f | + 3/2 1 vn vn vn2 ∧ vn    n− 2 log(n)2 if β = 32       1 n− 2 if β > 23 . In addition if (H2) holds, √     √ n ˆ dW fn (X) − f , N 6 C n|An (X) − f | + |vn − σ 2 | σ   3 −3β 1 2 2 , n if β ∈  2 3     1 +C n− 2 log(n)2 if β = 32      −1 n 2 if β > 23 . In particular, if the assumptions (H2)-(H4) hold and for some β > 1/2, we have ρ (t) =  −β O |t| for large |t|, then, as n → ∞ √ law n(fˆn (X) − f ) −→ N (0, σ 2). Proof. Theorem 6 is a direct consequence of Theorem 4, standard properties of the Wasserstein distance, and the fact that r  r  n ˆ n |An (X) − µ(θ)|. fn (X) − f = Fn + vn vn 14 Remark 7 The assumption that |E[Xt Xs ]| 6 C |t|−β for some β > 1/2 corresponds to the notion of moderating how long the memory of the data might be. For instance, when X represents the discrete-time increments of a process based on fBm with Hurst parameter H ∈ (0, 1), we expect that β = 2 − 2H. The restriction β > 1/2 means H < 3/4, a wellknown threshold for the limit of validity of central limit theorems for quadratic variations of long-memory processes. See for instance the excellent treatment of the classical BreuerMajor theorem in [20, Chapter 7]. The above theorem √ shows that the speed of convergence in the CLT reaches the so-called Berry-Esséen rate of 1/ n as√soon as β > 2/3, as long as the terms coming from the bias and variance estimates, namely n|An (X)−f | and |vn −σ 2 |, are √ no greater than that same order 1/ n. The threshold β > 2/3 coincides with H < 2/3 when one translates into the Hurst-parameter memory scale; this had already been identified for the canonical stationary case of fGn in [4], the paper which was the precursor to the optimal fourth moment theorem in [19]. 4 Application to Gaussian Ornstein Uhlenbeck processes In this section we will apply the above results to Ornstein-Uhlenbeck processes driven by a Gaussian process which does not necessarily have stationary increments. More precisely we will study the cases which correspond to sub-fractional Brownian motion and to bifractional Brownian motion. We consider the Gaussian Ornstein-Uhlenbeck process X = {Xt }t>0 defined by the following linear stochastic differential equation dXt = −θXt dt + dGt , X0 = 0; (13) where G is an arbitrary mean-zero Gaussian process which has Hölder continuous paths of strictly positive order, and θ > 0 is an unknown parameter. Our goal is to estimate θ under the discrete observations {X1 , . . . , Xn } as n → ∞. The equation (13) has the following explicit solution Z t Xt = e−θt eθs dGs , t>0 0 where the integral can be understood in the Young sense, since eθs is a Lipshitz function. The Young sense coincide in this case with Wiener integral sense here. As mentioned in the introduction, we will consider two different cases for G: the sub-fractional Brownian motion, and the bi-fractional Brownian motion. 15 4.1 Sub-fractional Brownian motion Consider a sub-fractional Brownian motion (sfBm) S H := {StH , t > 0} with parameter H ∈ (0, 1) : this is the mean-zero Gaussian process with covariance function   1 (t + s)2H + |t − s|2H . RS H (s, t) := E StH SsH = t2H + s2H − 2 1 Note that, when H = 12 , S 2 is a standard Brownian motion. By Kolmogorov’s continuity criterion and the fact that 2 E StH − SsH 6 (2 − 22H−1 )|s − t|2H ; s, t > 0, we deduce that S H has Hölder continuous paths of order H − ε, for every ε ∈ (0, H). In this section we replace G given in (13) by S H . More precisely, we will estimate the drift parameter θ in the following equation dXt = −θXt dt + dStH , (14) X0 = 0. Proposition 8 Suppose that H ∈ (0, 1). Set fH (θ) := HΓ(2H) . θ2H (15) Then for every t > 0, |E[Xt2 ] − fH (θ)| 6 Ct2H−2 . Hence √ C n|An (X) − fH (θ)| 6 √ n n X i=1 i2H−2  −1 n 2      1 6C n− 2 log(n)      2H−3/2 n if 0 < H < 1/2 if H = 1/2 if H > 1/2. In particular, if 0 < H < 3/4, the hypothesis (H4) holds. Rt Proof. Since Xt = e−θt 0 eθs dSsH , t > 0, we can write (see [9]) E[Xt2 ] = ∆H (t) + θI2H (t) − θ2 J2H (t), 2 where −2θt ∆H (t) = 2He Z 0 t s 2H−1 θs −2θt e ds − 2H(2H − 1)e 16 Z t θs dse 0 Z s dreθr (s + r)2H−2 , 0 and −2θt J2H (t) := e Z tZ 0 t θs θr 0 e e |s − r| 2H −θt drds; I2H (t) := e Z t 0 eθs (t − s)2H ds. Hence |E[Xt2 ] − fH (θ)| 6 |∆H (t)| + θ I2H (t) − 2HΓ(2H) θ2 2HΓ(2H) J2H (t) − + . 2H+1 θ 2 θ2H+2 We will check that each term of the right side of the last inequality is less than Ct2H−2 . We can write ∆H (t) = 2HaH (t) − 2H(2H − 1)bH (t) where −2θt aH (t) := e Z t s 2H−1 θs e ds; −2θt bH (t) := e 0 It is clear that |aH (t)| 6 Z t θs dse 0 Z s dreθr (s + r)2H−2 . 0 θt t2H −θt e 6 C e− 2 . 2H On the other hand, Z t Z s Z t Z 2s −2θt θs θr 2H−2 −2θt θs |bH (t)| = e dse dre (s + r) =e e ds eθ(u−s) u2H−2 du 0 0 s 0Z t  Z 2t Z 2t   −2θt u e θu 2H−2 −2θt θu 2H−1 θu 2H−2 e u t∧u− = e du = e u du + (2t − u) e u du 2 2 0 0 t Z t2H −θt 2H−2 2t e +t (2t − u) eθu du 6 Ct2H−2 . 6 4H t Hence, for every t > 0, |∆H (t)| 6 Ct2H−2 . We also have for t > 0 Z t  Z t 1 Γ(2H + 1) Γ(2H + 1) 2H −θu −2θt 2H θu = u e du − e u e du − J2H (t) − θ2H+2 θ θ2H+2 0 0 Z ∞ Z t 1 6 u2H e−θu du − e−2θt u2H eθu du θ t 0 Z ∞  Z t −θt/2 e 2H −θu/2 2H −θu/2 u e du + u e du 6 θ t 0 6 C e−θt/2 . 17 For the last term, let t > 0 Γ(2H + 1) I2H (t) − θ2H+1 Z t Γ(2H + 1) θ2H+1 0 Z +∞ Z +∞ 2H −θu −θt/2 = e−θu/2 u2H du 6 C e−θt/2 u e du 6 e = u2H e−θu du − t t which completes the proof. Proposition 9 For every H ∈ (0, 1), the hypothesis (H3) holds. More precisely, we have for large |t − s| |E[Xt Xs ]| 6 C|t − s|2H−2 . 1 −θ(t−s) Proof. If H = 21 , it is easy to see that |E[Xt Xs ]| 6 2θ e 6 C|t − s|2H−2 for |t − s| > 0. Now, suppose that H ∈ (0, 12 ) ∪ ( 12 , 1). Thanks to Lemma 19 we get Z s Z t θv −θ(t−s) 2 −θt −θs du eθu (v − u)2H−2 dv e E[Xt Xs ] = e E[Xs ] + H(2H − 1) e e 0 Z t Z s s −H(2H − 1) e−θt e−θs dv eθv du eθu (u + v)2H−2 . s 0 Hence, −θ(t−s) |E[Xs Xt ]| 6 e E[Xs2 ] −θt −θs +2e e Define for every α ∈ (0, 1) Ztα := e −θt Z H|2H − 1| Z tZ eθu dBuα , t∈R s s 0 eθu eθv (v − u)2H−2 dudv. t −∞ (16) which is a stationary Gaussian process, where B α is a fractional Brownian motion with Hurst parameter α ∈ (0, 1). Then, we can write for s < t, Z tZ s −θt −θs H(2H − 1) e e eθu eθv (v − u)2H−2 dudv s 0 Z t−s Z 0 −θ(t−s) θy = H(2H − 1) e dy e dx eθx (y − x)2H−2 0 −s Z t−s  Z 0   H   −θ(t−s) θy H θx H H = e E e dBy e dBx = E Zt−s − e−θ(t−s) Z0H Z0H − e−θs Z−s 0 = H E[Zt−s Z0H ] − e−θs E[Z0H ZtH ]  2H−2 −θ(t−s) 6 C (t − s) +e . −s 2 − e−θ(t−s) E[ Z0H ] + e−θt E[Z0H ZsH ] 18 The last inequality comes from the fact that for large r > 0 E[ZrH Z0H ] 6 C|r|2H−2 (see [6], or [12]). Define the following rates of convergence,  −1 n if 0 < α < 1/2    ϕα (n) = (17)    2α−2 n if 1/2 6 α < 1, and ψα (n) =  −1 n       n4α−3 if 0 < α < 1/2 (18) if 1/2 6 α < 3/4. Proposition 10 Let 0 < H < 3/4. Define 2 σH := ρH (0)2 + 2 X ρH (i)2 (19) i∈Z\{0} where ρH (k) := E[ZkH Z0H ], k ∈ N and Z is the process given in (16). Note that ρH (0) = fH (θ). Then 2 |E[Vn (X)2 ] − σH | 6 CψH (n). In particular, the hypothesis (H2) holds. If H = 3/4, we have 9 E[Vn (X)2 ] 6 C log(n)−1 . − log(n) 16θ4 Proof. See Appendix. Propositions 8 and 10 lead to the assumptions (H1) and (H2). Then, applying Theorem 2 we obtain the strong consistency of the estimator fˆn (X) of the form (7). Theorem 11 Let 0 < H < 1. Then we have HΓ(2H) fˆn (X) −→ fH (θ) = θ2H almost surely as n −→ ∞. Now, we will study the asymptotic normality of the estimator fˆn (X) when 0 < H 6 34 . Using (12) with β = 2 − 2H we obtain the following result. 19 Proposition 12 If 0 < H 6 34 , then dT V (Fn (X), N) 6 C  1  n− 2       1    n− 2 log(n)2  9   n6H− 2        log(n)−3/2 if H ∈ 0, 23 if H = if H ∈ 2 3 2 3 , 3 4 if H = 34 .   Combining this with Propositions 8 and 10 we deduce the result. Theorem 13 If 0 < H < 3/4, then √  p n ˆ dW (fn (X) − fH (θ)), N 6 C ψH (n) σH and if H = 3/4, we have dW √ n(fˆn (X) − fH (θ)) p ,N σH log(n) ! 6 C log(n)−1/2 . In particular, if 0 < H < 3/4, then, as n → ∞ √ law 2 n(fˆn (X) − fH (θ)) −→ N (0, σH ) and if H = 3/4, then, as n → ∞ √ ˆ n(fn (X) − fH (θ)) law 2 p −→ N (0, σH ). log(n) 4.2 Bifractional Brownian motion In this section we suppose that G given in (13) is a bifractional Brownian motion (bifBm) B H,K with parameters H ∈ (0, 1) and K ∈ (0, 1]. The B H,K := {BtH,K , t > 0} is the mean-zero Gaussian process with covariance function   1  2H H,K H,K 2H K 2HK E(Bs Bt ) = K t + s . − |t − s| 2 The case K = 1 corresponds to the fBm with Hurst parameter H. The process B H,K verifies,   2 H,K H,K 6 21−K |t − s|2HK , E Bt − Bs so B H,K has (HK −ε)−Hölder continuous paths for any ε ∈ (0, HK) thanks to Kolmogorov’s continuity criterion. 20 Proposition 14 Assume that H ∈ (0, 1) and K ∈ (0, 1]. Then we have for large t > 0 |E[Xt2 ] − fH,K (θ)| 6 Ct2HK−2 where fH,K (θ) := 21−K HKΓ(2HK)/θ2HK . Then √ n|An (X) − µ(θ)| 6 C  1  n− 2  (20) if 0 < HK < 1/2 n2HK−3/2 if HK > 1/2. In particular, if HK < 3/4, the hypothesis (H4) holds. Rt Proof. From [9] and the fact that Xt = e−θt 0 eθs dBsH,K , t > 0, we can write   E Xt2 = ∆H,K (t) + 21−K θI2HK (t) − 2−K θ2 J2HK (t), where Z t eθs s2HK−1 ds 0 Z tZ s 3−K 2 −2θt +2 H K(K − 1) e (sr)2H−1 (s2H + r 2H )K−2 eθr eθs drds ∆H,K (t) = 2 2−K −2θt HK e 0 0 and −2θt J2HK (t) = e Z tZ 0 s θs θr 0 e e (s − r) 2HK −θt drds; I2HK (t) = e Z t 0 eθs (t − s)2HK ds. Hence |E[Xt2 ]−µ(θ)| 6 |∆H (t)|+ θ 2K−1 I2HK (t) − 2HKΓ(2HK) θ2 2HKΓ(2HK) + J2HK (t) − . 2HK+1 K θ 2 θ2HK+2 We will check that each term of the right-hand side is less than Ct2HK−2 . We can write ∆H,K (t) = 22−K HKaH,K (t) + 23−K H 2 K(K − 1)bH,K (t). where aH,K (t) := e −2θt Z t θs e s 2HK−1 ds; −2θt bH,K (t) := e 0 Z tZ 0 It is easy to prove that aH,K (t) 6 C e−θt/2 . 21 s (sr)2H−1 (s2H +r 2H )K−2 eθr eθs drds. 0 On the other hand, using x2 + y 2 > 2|xy|, x, y ∈ R, we get Z tZ s −2θt bH,K (t) = e (sr)2H−1 (s2H + r 2H )K−2 eθr eθs drds 0 0 Z tZ s −2θt K−2 6 e 2 (sr)2H−1 sHK−1r HK−1 eθr eθs drds 0 := 0 b1 (t) + b2 (t) where −2θt b1 (t) = e Z t 2 0 Z s (sr) HK−1 θr θs e e drds; b2 (t) = e −2θt Z tZ t 2 0 s (sr)HK−1 eθr eθs drds. 0 It is easy to see that θt |b1 (t)| 6 Ct2HK e−θt 6 C e− 2 , and |b2 (t)| 6 6 6 6  HK−1 −θt Z t e t r HK−1 eθr dr 2 θ 0 !  HK−1 −θt Z t Z t 2 t e r HK−1 eθr dr r HK−1 eθr dr + t 2 θ 0 2  2HK−1 −θt/2  2HK−2 t e 1 t + 2 2 θ θ 2 2HK−2 Ct . We deduce that ∆H,K (t) 6 Ct2HK−2 . Moreover, by a similar argument as in the proof of Proposition 8 we have 2HKΓ(2HK) | 6 e−θt/2 |I2HK (t) − θ2HK+1 and 2HKΓ(2HK) |6 |J2HK (t) − θ2HK+2 which completes the proof.  2HK+1 2 Γ(2HK + 1), θ  2HK+2 2 Γ(2HK + 1) e−θt/2 θ 22 Proposition 15 For all fixed (H, K) ∈ (0, 1) × (0, 1], with HK 6= 12 , the hypothesis (H3) holds. More precisely, we have for large |t − s|,   |t − s|2HK−2H−1 if 0 < HK < 1/2 |E[Xt Xs ]| 6 C  |t − s|2HK−2 if 1/2 < HK < 1. Proof. Let s < t. Using Lemma 19 we get E[Xt Xs ] = e −θ(t−s) E[Xs2 ] −2 2−K −θt −θs K(1 − K) e +21−K HK(2HK − 1) e−θt e−θs e Z s t eθv Z t Z ss 0 θv e Z s eθu (uv)2H−1(u2H + v 2H )K−2dudv 0 eθu (u − v)2HK−2 dudv. As in the proof of Proposition 9 we have Z t Z s −θt −θs θv δHK := e e e eθu (u − v)2HK−2 dudv 6 C|t − s|2HK−2. s Set λH,K := e −θt −θs e 0 Z If H > 21 , we have for 0 6 u 6 v, t θv e s Z s eθu (uv)2H−1 (u2H + v 2H )K−2 dudv. 0 (uv)2H−1 (u2H + v 2H )K−2 6 v 2HK−2 6 (v − u)2HK−2 . Thus, if we assume that HK 6= 21 , λH,K 6R CδHK 6 C|t − s|2HK−2. s If H < 21 , it is clear that s → e−θs 0 eθu u2H−1 du is bounded. This implies −θt λH,K 6 C e Z t s eθv v 2H−1 v 2HK−4H dv 6 Ct2HK−2H−1 6 C(t − s)2HK−2H−1 which finishes the proof. Using the same arguments as in Proposition 10 we obtain the following result. Proposition 16 For all fixed 0 < HK < 3/4, 2 |E[Vn (X)2 ] − σH,K | 6 CψHK (n) where 2 := 4 σH,K X i∈N∗ (ρH,K (i) − (1 − 21−K ) e−θi ρH,K (0))2 + 23−2K ρH,K (0)2 23 (21) with ρH,K (k) := E[ZkHK Z0HK ], k ∈ N where Z HK is the process given in (16), and ρH,K (0) = HKΓ(2HK) . In particular, the hypothesis (H2) holds. θ 2HK If HK = 3/4, we have E[Vn (X)2 ] 9 6 C log(n)−1 . − log(n) 16θ4 Similarly as in Section 4.1 we obtain the following asymptotic behavior results. Theorem 17 Let H, K ∈ (0, 1). Then we have fˆn (X) −→ fH,K (θ) = 21−K HKΓ(2HK)/θ2HK (22) almost surely as n −→ ∞. Theorem 18 Let HK ∈ (0, 3/4) \ {1/2} and N ∼ N (0, 1), then  √ p n(fˆn (X) − fH,K (θ))/σHK , N 6 C ψHK (n) dW and if HK = 3/4, we have dW √ n(θ̂n (X) − fH,K (θ)) p ,N σHK log(n) ! 6 C log(n)−1/2 . In particular, if HK ∈ (0, 3/4) \ {1/2}, we have as n → ∞ √ law 2 n(θ̂n (X) − fH,K (θ)) −→ N (0, σH,K ) and if HK = 3/4, we have as n → ∞ √ 5 Appendix n(θ̂n (X) − fH,K (θ)) law 2 p −→ N (0, σH,K ). log(n) Lemma 19 Let G be a Gaussian process which has Hölder continuous paths of stictly positive ordre and X is the solution of the equation (13). Define RG (s, t) = E[Gs Gt ], and assume 2R G G that ∂R (s, r) and ∂∂s∂r (s, r) exist on (0, ∞) × (0, ∞). Then for every 0 < s < t, we have ∂s −θ(t−s) E[Xs Xt ] = e E[Xs2 ] +e −θt −θs e 24 Z t e s θv Z 0 s eθu ∂ 2 RG (u, v)dudv. ∂u∂v (23) Proof. Fix s < t. We have    Z s Z t −θs θu −θt θv e dGu e e dGv E[Xs Xt ] = E e 0 0 Z s  Z t −θ(t−s) 2 −θt −θs θu θv = e E[Xs ] + e e E e dGu e dGv . 0 s Moreover, by (1), Z s  Z t −θt −θs θu θv e e E e dGu e dGv 0 s    Z s Z t −θt −θs θs θu θt θs H θv = e e E e Gs − θ e Gu du e Gt − e Gs − θ e Gv dv 0 s  Z t Z s θ(t+s) 2θs θs −θt −θs θv θt e RG (s, t) − e RG (s, s) − θ e = e e e RG (s, v)dv − θ e eθu RG (u, t)du s 0  Z s Z tZ s +θ eθs eθu RG (u, s)du + θ2 eθu eθv RG (u, v)dudv . 0 s 0 Applying again (1) several times we get the last term of (23). Thus the desired result is obtained. 2 Lemma 20 Let X be the solution of (14), and let σH be the constant defined in Proposition 10. If 0 < H < 3/4, we have 2 |E[Vn (X)2 ] − σH | 6 CψH (n) If H = 3/4, we have 9 E[Vn (X)2 ] 6 C log(n)−1 . − 4 log(n) 16θ P Proof. We have Vn (X) = I2 (fn,2 ), with fn,2 = √1n ni=1 fi⊗2 . Then E[Vn (X)2 ] = = n n 2 X 2 X (hfk , fl iH )2 = (E[Xk Xl ])2 n k,l=1 n k,l=1 n n 2X 2 X (E[Xk Xj ])2 . (E[Xk2 ])2 + n k=1 n k,l=1 k6=l We will need the following lemmas. 25 Lemma 21 Let X be the solution of (14). Then, n 1X (E[Xk2 ])2 − fH (θ)2 6 CϕH (n). n k=1 Proof. By Proposition 8 we deduce that n 1X (E[Xk2 ])2 − fH (θ)2 n k=1 n 1X 6 |E[Xk2 ] − fH (θ)|(E[Xk2 ] + fH (θ)) n k=1 n n C X 2H−2 CX 2 |E[Xk ] − fH (θ)| 6 k 6 n k=1 n k=1 6 CϕH (n). Lemma 22 Let X be the solution of (14). If 0 < H < 3/4, we have n X 1 X ρH (i)2 6 CψH (n). (E[Xk Xj ])2 − 2 J(n) := n j,k=1 i∈N∗ k6=j Proof. Suppose that H 6= 1/2. Using (23), we have n 1 X (E[Xk Xj ])2 = E1 (n) + E2 (n) + E3 (n) + E4 (n) + E5 (n) + E6 (n) n k,j=1 where E1 (n) := n n−1 2X X a(k, j)2 , n j=1 E2 (n) := k=j+1 k=j+1 n−1 n 4X X E3 (n) := a(k, j)D(k, j), n j=1 k=j+1 E5 (n) := with for every s 6 t n n−1 4X X a(k, j)A(k, j) n j=1 n−1 n 4X X A(k, j)D(k, j), n j=1 k=j+1 n−1 n 2X X E4 (n) := A(k, j)2 n j=1 k=j+1 E6 (n) := a(s, t) := e−θ(t−s) E[Xs2 ], 26 n−1 n 2X X D(k, j)2 n j=1 k=j+1 −θt −θs A(t, s) := H(2H − 1) e e Z t e s D(t, s) := −H(2H − 1) e−θt e−θs Then, we can write Z t θv Z eθv s s 0 Z eθu (u − v)2H−2 dudv, s eθu (u + v)2H−2 dudv. 0 ∞ X fH (θ)e−2θ 2fH (θ)2 e−2θ −θi E (n) − 4f (θ) e ρ(i) − | + J(n) 6 |E1 (n) − 2 H 1 − e−2θ 1 − e−2θ i=1 + E4 (n) − 2 ∞ X i=1 ρ(i) − fH (θ) e−θi 2 ! + |E3 (n)| + |E5 (n)| + |E6 (n)|. On the other hand, 2fH (θ)2 e−2θ E1 (n) − 1 − e−2θ n−1 n  2 X X −2θ(k−j) e [E(Xj2 )]2 − fH (θ)2 6 n j=1 k=j+1 + n−1 n 2 X X −2θ(k−j) 2fH (θ)2 e−2θ e fH (θ)2 − n j=1 k=j+1 1 − e−2θ ∞ n n−1 n−1 X 2C X X −2θ(k−j) 2H−2 1 X −2θi −2θi 2 6 e + e j + 2fH (θ) ie n j=1 n i=1 i=n k=j+1 # " n−1 n−j 1 X 2H−2 X −2θi −2θn −1 j e +e +n 6 C n j=1 i=1 ! (24) 6 CϕH (n). Furthermore, E2 (n) = = n−1 n 4 X X −θ(k−j) e E(Xj2 )[E[Zk−j Z0 ] − e−θ(k−j) E[Z02 ] − e−θj E[Z0 Zk ] + e−θk E[Z0 Zj ]] n j=1 k=j+1 n−1 n n−1 n 4ρ(0) X X −2θ(k−j) 4 X X −θ(k−j) e E(Xj2 )ρ(k − j) − e E(Xj2 ) n j=1 k=j+1 n j=1 k=j+1 n−1 n n−1 n X X 4 X −θj 4X 2 −θ(k−j) 2 − e E(Xj ) e ρ(k) + ρ(j)E(Xj ) e−θ(k−j) e−θk n j=1 n j=1 k=j+1 k=j+1 := E21 (n) − E22 (n) − E23 (n) + E24 (n). 27 For E21 (n), we have E21 (n) − 4fH (θ) X −θi e ρ(i) i∈N∗ n−1 n 4 X X −θ(k−j) e |E(Xj2 ) − fH (θ)||ρ(k − j)| 6 n j=1 k=j+1 n−1 n ∞ X 1 X X −θ(k−j) e ρ(k − j) − e−θi ρ(i) n j=1 k=j+1 i=1  2H−2 −1 6 C n +n +4|fH (θ)| because ∞ n n−1 X 1 X X −θ(k−j) e−θi ρ(i) e ρ(k − j) − n j=1 i=1 k=j+1 = ∞ X i=n ∞ X n−1 e−θi ρ(i) − 1 X −θi i e ρ(i) n i=1 n−1 1X 6 e |ρ(i)| + i|ρ(i)| e−θi n i=n  i=1 2H−2 −1 6 C n +n −θi and by using a similar argument as in above, we also have n−1 n  4 X X −θ(k−j) e |E(Xj2 ) − fH (θ)||ρ(k − j)| 6 C n2H−2 + n−1 . n j=1 k=j+1 By straightforward calculus, we also have |E22 (n) − 4fH (θ) ∞ X i=1 e−2θi | 6 CϕH (n) For E23 (n), we have |E23 (n)| n−1 n n−1 n X 4 X −θj 4|fH (θ)| X −θj X −θ(k−j) −θ(k−j) 2 6 e |ρ(k)| + e |ρ(k)| e |E(Xj ) − fH (θ)| e n j=1 n j=1 k=j+1 k=j+1 6 Cn−1 . For E24 (n), since n−1 n n−1 C X 4H−4 C 4 X X −θ(k−j) e |ρ(j)||E(Xj2 ) − fH (θ)| e−θk 6 j 6 , n j=1 k=j+1 n j=1 n 28 and we get n−1 n n−1 X C X 2H−2 4X |ρ(j)| e−θ(k−j) e−θk 6 j 6 CϕH (n) n j=1 n j=1 k=j+1 |E24 (n)| 6 CϕH (n). Thus, we conclude E2 (n) − 4fH (θ) ∞ X i=1 fH (θ)e−2θ e−θi ρ(i) − 1 − e−2θ ! 6 CϕH (n). (25) For E3 (n), since for every ∀k > j, |D(k, j)| 6 C(jk)H−1 , we can write n−1 n C X X −θ(k−j) |E3 (n)| 6 e |E(Xj2 )|(kj)H−1 n j=1 k=j+1 " n−1 n # n−1 X n X C X X −θ(k−j) 6 e |E(Xj2) − fH (θ)|(kj)H−1 + fH (θ) e−θ(k−j) (kj)H−1 n j=1 k=j+1 j=1 k=j+1 # " n n−1 C X 4H−4 X 2H−2 j + j 6 n j=1 j=1 (26) 6 CϕH (n). For E4 (n), we first calculate A(k, j)2 . We have A(k, j)2 = ρ(k − j)2 − 2ρ(k − j) e−θ(k−j) ρ(0) + ρ(0)2 e−2θ(k−j) + e−2θk ρ(j)2 −2 e−θ(k+j) ρ(j)ρ(k) + e−2θj ρ(k)2 + 2 e−θk ρ(k − j)ρ(j) − 2 e−θj ρ(k)ρ(k − j) −2 e−θ(k−j) ρ(0) e−θk ρ(j) + 2 e−θ(k−j) ρ(0)e−θj ρ(k). Hence, we need to study the rate of convergence of the following terms. We have n−1 n ∞ X 2X X ρ(k − j)2 − 2 ρ(i)2 n j=1 k=j+1 i=1 6 2 ∞ X n−1 i4H−4 + i=n 6 CψH (n), n−1 n ∞ X 1 X X −θ(k−j) e ρ(k − j) − e−θi ρ(i) 4ρ(0) n j=1 i=1 2 X 4H−3 i n i=1 6 4ρ(0) ∞ X j=n k=j+1 n−1 + 4ρ(0) X −θi 2H−1 e i n i=1 6 CϕH (n) 29 j 2H−2 e−θj and ∞ n−1 n 1 X X −2θ(k−j) X −2θi 6 Cn−1 . e − e 2ρ(0) n j=1 k=j+1 i=1 2 Moreover, it is clear that n−1 n  1 X X  −2θk e ρ(j)2 + e−θ(k+j) |ρ(j)||ρ(k)| + e−2θj ρ(k)2 6 Cn−1 n j=1 k=j+1 and n−1 n n−1 1 X X −θk C X 2H−2 e |ρ(k − j)||ρ(j)| 6 j 6 CϕH (n). n j=1 n j=1 k=j+1 In addition, n n−1 n−1 n 4 X X −θj C X −θj X e |ρ(k)||ρ(k − j)| 6 e (k − j)4H−4 n j=1 n j=1 k=j+1 k=j+1 C 6 n ∞ X 4H−4 i n−1 X e−θj 6 Cn−1 , j=1 i=1 n−1 n n−1 n 4 X 2H−2 X −θ(k−j) 4 X X −θ(k−j) −θk e |ρ(j)| e 6 j e n j=1 k=j+1 n j=1 k=j+1 6 CϕH (n), n−1 n n−1 n 4ρ(0) X X −θ(k−j) 4ρ(0) X −θj X −θ(k−j) −θj e |ρ(k)| e 6 e (k − j)2H−2 e n j=1 k=j+1 n j=1 k=j+1 n−1 ∞ 4ρ(0) X −θj X −θi 6 e e ρ(i) n j=1 i=1 6 Cn−1 . Thus, E4 (n) − 2 ∞ X i=1 ρ(i) − fH (θ) e−θi 30 2 6 CψH (n). (27) For E5 (n), we have n−1 n CX X (k − j)2H−2 (kj)H−1 |E5 (n)| 6 n j=1 k=j+1 n−j n−1 C X H−1 X 2H−2 6 j i (i + j)H−1 n j=1 i=1 !2 n C X 2H−2 j 6 CψH (n). 6 n j=1 (28) Finally, using D(k, j)2 6 |A(k, j)||D(k, j)| ∀k > j, we get |E6 (n)| 6 C|E5 (n)|. (29) Combining (24), (25), (26), (27), (28) and (29) the proof is completed. Let us now study the case when H = 3/4. Lemma 23 Let X be the solution of (14). If H = 34 , we have n X 9 C 1 (E[Xk Xj ])2 − 6 . 4 n log(n) j,k=1 32θ log(n) (30) k6=j Proof. Using similar argument as in above, it is straightforward to check that |E1 (n)| + |E2 (n)| + |E3 (n)| + |E5 (n)| + |E6 (n)| 6 C, and 9 1 6 C log(n)−1 E4 (n) − 4 log(n) 32θ which completes the proof. Lemma 24 Let X be the solution of (13), with the process G is a bifBm. If 0 < HK < 3/4, we have E[Vn (X)2 ] − 4 ∞ X  2 ρ(i) − (1 − 21−K ) e−θi ρ(0) − 23−2K ρ(0)2 6 CψHK (n). i=1 If HK = 3/4, we have 9 E[Vn (X)2 ] 6 C log(n)−1 . − log(n) 16θ4 31 Proof. We can write n n 2X 2 X 2 (E[Xk Xj ]) + (E[Xk2 ])2 , E[Vn (X) ] = n k,l=1 n k=1 2 k6=l where n n 1X 1X | (E[Xk2 ])2 − 22−2K ρ(0)2 | 6 |E[Xk2 ] − 21−K ρ(0)|(E[Xk2 ] + 21−K ρ(0)) n k=1 n k=1 n CX |E[Xk2 ] − 21−K ρ(0)| 6 n 6 C n k=1 n X k 2HK−2 k=1 6 CϕHK (n). Moreover, ∞ n X 1 X 2 (E[Xk Xj ]) − 2 (ρ(i) − (1 − 21−K ) e−θi ρ(0))2 6 CψHK (n). n j,k=1 i=1 k6=j Indeed, we have n 1 X (E[Xk Xj ])2 = E 1 (n) + E 2 (n) + E 3 (n) + E 4 (n) + E 5 (n) + E 6 (n) n k,j=1 where E 1 (n) := n−1 n 2X X a(k, j)2 , n j=1 k=j+1 E 2 (n) := n−1 n 4X X a(k, j)A(k, j) n j=1 k=j+1 n−1 n 4X X a(k, j)D(k, j), E 3 (n) := n j=1 k=j+1 n−1 n 2X X E 4 (n) := A(k, j)2 n j=1 k=j+1 n−1 n 4X X E 5 (n) := A(k, j)D(k, j), n j=1 k=j+1 n−1 n 2X X E 6 (n) := D(k, j)2 n j=1 k=j+1 32 with for 0 < s < t, a(s, t) := A(t, s) := D(t, s) := e−θ(t−s) E[Xs2 ], 2 1−K −θt −θs HK(2HK − 1) e 22−K K(K − 1) e−θt e−θs e Z tZ s Z t θv e s s Z s 0 eθu (u − v)2HK−2dudv, eθu eθv (uv)2H−1 (u2H + v 2H )K−2dudv. 0 Using similar arguments as in the subfractional case, we obtain ∞ X 23−2K ρ(0)2 e−2θ e−2θ 3−K −θi 3−K 2 + E 1 (n) − E (n) − 2 ρ(0) e ρ(i) − 2 ρ(0) 2 1 − e−2θ 1 − e−2θ i=1 + E 3 (n) + E 4 (n) − 2 6 CψHK (n) ∞ X i=1 (ρ(i) − ρ(0) e−θi )2 + E 5 (n) + E 6 (n) which completes the proof of the first inequality of Lemma 24. By using similar techniques as in above we also obtain for HK = 3/4, n X 9 1 1 (E[Xk Xj ])2 − 6 C log(n)−1 n log(n) j,k=1 32 θ4 k6=j which finishes the proof. References [1] Azmoodeh, E. and Morlanes, G. I. (2013). Drift parameter estimation for fractional Ornstein-Uhlenbeck process of the second kind. Statistics. DOI: 10.1080/02331888.2013.863888. [2] Azmoodeh, E. and Viitasaari, L. (2015). Parameter estimation based on discrete observations of fractional Ornstein-Uhlenbeck process of the second kind. Statist. Infer. Stoch. Proc. 18, no. 3, 205-227. [3] Belfadli, R., Es-Sebaiy, K. and Ouknine, Y. (2011). Parameter Estimation for Fractional Ornstein-Uhlenbeck Processes: Non-Ergodic Case. Frontiers in Science and Engineering (An International Journal Edited by Hassan II Academy of Science and Technology). 1, no. 1, 1-16. 33 [4] Biermé, H., Bonami, A., Nourdin, I. and Peccati, G. (2012). Optimal Berry-Esséen rates on the Wiener space: the barrier of third and fourth cumulants. ALEA 9, no. 2, 473-500. [5] Brouste, A. and Iacus, S. M. (2012). Parameter estimation for the discretely observed fractional Ornstein-Uhlenbeck process and the Yuima R package. Comput. Stat. 28, no. 4, 1529-1547. [6] Cheridito, P., Kawaguchi, H. and Maejima, M. (2003). Fractional Ornstein-Uhlenbeck processes, Electr. J. Prob. 8, 1-14. [7] Davydov, Y.A. and Martynova, G. V. (1987). Limit behavior of multiple stochastic integral. Statistics and control of random process. Preila, Nauka, Moscow, 55-57 (in Russian). [8] Dobrushin, R. L. and Major, P. (1979). Non-central limit theorems for nonlinear functionals of Gaussian fields. Z. Wahrsch. verw. Gebiete, 50, 27-52. [9] El Machkouri, M., Es-Sebaiy, K. and Ouknine, Y. (2015). Least squares estimator for non-ergodic OrnsteinUhlenbeck processes driven by Gaussian processes. Journal of the Korean Statistical Society, DOI: 10.1016/j.jkss.2015.12.001 (In press). [10] El Onsy, B., Es-Sebaiy, K. and Tudor, C. (2014). Statistical analysis of the non-ergodic fractional Ornstein-Uhlenbeck process of the second kind. Preprint. [11] El Onsy, B., Es-Sebaiy, K. and Viens, F. (2017). Parameter Estimation for OrnsteinUhlenbeck driven by fractional Ornstein-Uhlenbeck processes. Stochastics, Vol. 89, No. 2, 431-468. [12] Es-Sebaiy, K. and Viens, F. (2016). Optimal rates for parameter estimation of stationary Gaussian processes. Preprint. http://arxiv.org/pdf/1603.04542.pdf. [13] Fernique, X. Régularité des trajectoires des fonctions aléatoires gaussiennes. In Ecole d’Et´e de Probabilités de Saint-Flour IV-1974, Lecture Notes in Math. 480, 1-96. Springer V., 1975. [14] Hu, Y. and Nualart, D. (2010). Parameter estimation for fractional Ornstein-Uhlenbeck processes. Statist. Probab. Lett. 80, 1030-1038. [15] Hu, Y. and Song, J. (2013). Parameter estimation for fractional Ornstein-Uhlenbeck processes with discrete observations. F. Viens et al (eds), Malliavin Calculus and Stochastic Analysis: A Festschrift in Honor of David Nualart, 427-442, Springer. [16] Kloeden, P. and Neuenkirch, A. (2007). The pathwise convergence of approximation schemes for stochastic differential equations. LMS J. Comp. Math. 10, 235-253. 34 [17] Neufcourt, L. and Viens, F. (2014). A third-moment theorem and precise asymptotics for variations of stationary Gaussian sequences. In press in ALEA; preprint available at http://arxiv.org/abs/1603.00365 [18] Neuenkirch, A. and Tindel, S. (2014). A least square-type procedure for parameter estimation in stochastic differential equations with additive fractional noise.Statist. Infer. Stoch. Proc. 17, no. 1, 99-120. [19] Nourdin, I. and Peccati, G. (2015). The optimal fourth moment theorem. Proc. Amer. Math. Soc. 143, 3123-3133. [20] Nourdin, I. and Peccati, G. (2012). Normal approximations with Malliavin calculus : from Stein’s method to universality. Cambridge Tracts in Mathematics 192. Cambridge University Press, Cambridge. [21] Nualart, D. (2006). The Malliavin calculus and related topics. Springer-Verlag, Berlin. [22] Nualart, D. and Peccati, G. (2005). Central limit theorems for sequences of multiple stochastic integrals. Ann. Probab. 33, no. 1, 177-193. [23] Viens, F.; Vizcarra, A. Supremum concentration inequality and modulus of continuity for sub-nth chaos processes. J. Functional Analysis 248 (1) (2007), 1-26. [24] Young, L. C. (1936). An inequality of the Hölder type connected with Stieltjes integration. Acta Math. 251-282. 35
10math.ST
Density-aware Single Image De-raining using a Multi-stream Dense Network He Zhang Vishal M. Patel Department of Electrical and Computer Engineering Rutgers University, Piscataway, NJ 08854 arXiv:1802.07412v1 [cs.CV] 21 Feb 2018 {he.zhang92,vishal.m.patel}@rutgers.edu Abstract Single image rain streak removal is an extremely challenging problem due to the presence of non-uniform rain densities in images. We present a novel densityaware multi-stream densely connected convolutional neural network-based algorithm, called DID-MDN, for joint rain density estimation and de-raining. The proposed method enables the network itself to automatically determine the rain-density information and then efficiently remove the corresponding rain-streaks guided by the estimated raindensity label. To better characterize rain-streaks with different scales and shapes, a multi-stream densely connected de-raining network is proposed which efficiently leverages features from different scales. Furthermore, a new dataset containing images with rain-density labels is created and used to train the proposed density-aware network. Extensive experiments on synthetic and real datasets demonstrate that the proposed method achieves significant improvements over the recent state-of-the-art methods. In addition, an ablation study is performed to demonstrate the improvements obtained by different modules in the proposed method. Code can be found at: https://github.com/hezhangsprinter (a) (b) (c) (d) (e) (f) Figure 1: Image de-raining results. (a) Input rainy image. (b) Result from Fu et al. [6]. (c) DID-MDN. (d) Input rainy image. (e) Result from Li et al. [33]. (f) DID-MDN. Note that [6] tends to over de-rain the image while [33] tends to under de-rain the image. tively consider various shapes, scales and density of rain drops into their algorithms. State-of-the-art de-raining algorithms such as [33, 6] often tend to over de-rain or under de-rain the image if the rain condition present in the test image is not properly considered during training. For example, when a rainy image shown in Fig. 1(a) is de-rained using the method of Fu et al. [6], it tends to remove some important parts in the de-rained image such as the right arm of the person, as shown in Fig. 1(b). Similarly, when [33] is used to de-rain the image shown in Fig. 1(d), it tends to under de-rain the image and leaves some rain streaks in the output de-rained image. Hence, more adaptive and efficient methods, that can deal with different rain density levels present in the image, are needed. One possible solution to this problem is to build a very large training dataset with sufficient rain conditions containing various rain-density levels with different orientations and scales. This has been achieved by Fu et al. [6] and Yang et al.[33], where they synthesize a novel large-scale dataset consisting of rainy images with various conditions 1. Introduction In many applications such as drone-based video surveillance and self driving cars, one has to process images and videos containing undesirable artifacts such as rain, snow, and fog. Furthermore, the performance of many computer vision systems often degrades when they are presented with images containing some of these artifacts. Hence, it is important to develop algorithms that can automatically remove these artifacts. In this paper, we address the problem of rain streak removal from a single image. Various methods have been proposed in the literature to address this problem [17, 6, 35, 19, 2, 14, 9, 1, 36, 33, 5]. One of the main limitations of the existing single image de-raining methods is that they are designed to deal with certain types of rainy images and they do not effec1 and they train a single network based on this dataset for image de-raining. However, one drawback of this approach is that a single network may not be capable enough to learn all types of variations present in the training samples. It can be observed from Fig. 1 that both methods tend to either over de-rain or under de-rain results. Alternative solution to this problem is to learn a density-specific model for deraining. However, this solution lacks flexibility in practical de-raining as the density label information is needed for a given rainy image to determine which network to choose for de-raining. In order to address these issues, we propose a novel Density-aware Image De-raining method using a Multistream Dense Network (DID-MDN) that can automatically determine the rain-density information (i.e. heavy, medium or light) present in the input image (see Fig. 2). The proposed method consists of two main stages: rain-density classification and rain streak removal. To accurately estimate the rain-density level, a new residual-aware classifier that makes use of the residual component in the rainy image for density classification is proposed in this paper. The rain streak removal algorithm is based on a multi-stream densely-connected network that takes into account the distinct scale and shape information of rain streaks. Once the rain-density level is estimated, we fuse the estimated density information into our final multi-stream denselyconnected network to get the final de-rained output. Furthermore, to efficiently train the proposed network, a largescale dataset consisting of 12,000 images with different rain-density levels/labels (i.e. heavy, medium and light) is synthesized. Fig. 1(c) & (d) present sample results from our network, where one can clearly see that DID-MDN does not over de-rain or under de-rain the image and is able to provide better results as compared to [6] and [33]. This paper makes the following contributions: world) and comparisons are performed against several recent state-of-the-art approaches. Furthermore, an ablation study is conducted to demonstrate the effects of different modules in the proposed network. 2. Background and Related Work In this section, we briefly review several recent related works on single image de-raining and multi-scale feature aggregation. 2.1. Single Image De-raining Mathematically, a rainy image y can be modeled as a linear combination of a rain-streak component r with a clean background image x, as follows y = x + r. (1) In single image de-raining, given y the goal is to recover x. As can be observed from (1) that image de-raining is a highly ill-posed problem. Unlike video-based methods [23, 29, 25], which leverage temporal information in removing rain components, prior-based methods have been proposed in the literature to deal with this problem. These include sparse coding-based methods [14, 9, 41], lowrank representation-based methods [2, 35] and GMM-based (gaussian mixture model) methods [17]. One of the limitations of some of these prior-based methods is that they often tend to over-smooth the image details [14, 35]. Recently, due to the immense success of deep learning in both high-level and low-level vision tasks [8, 31, 38, 21, 32, 34], several CNN-based methods have also been proposed for image de-raining [3, 5, 33, 6]. In these methods, the idea is to learn a mapping between input rainy images and their corresponding ground truths using a CNN structure. 2.2. Multi-scale Feature Aggregation 1. A novel DID-MDN method which automatically determines the rain-density information and then efficiently removes the corresponding rain-streaks guided by the estimated rain-density label is proposed. 2. Based on the observation that residual can be used as a better feature representation in characterizing the raindensity information, a novel residual-aware classifier to efficiently determine the density-level of a given rainy image is proposed in this paper. 3. A new synthetic dataset consisting of 12,000 training images with rain-density labels and 1,200 test images is synthesized. To the best of our knowledge, this is the first dataset that contains the rain-density label information. Although the network is trained on our synthetic dataset, it generalizes well to real-world rainy images. 4. Extensive experiments are conducted on three highly challenging datasets (two synthetic and one real- It has been observed that combining convolutional features at different levels (scales) can lead to a better representation of an object in the image and its surrounding context [7, 39, 8, 11]. For instance, to efficiently leverage features obtained from different scales, the FCN (fully convolutional network) method [18] uses skip-connections and adds high-level prediction layers to intermediate layers to generate pixel-wise prediction results at multiple resolutions. Similarly, the U-Net architecture [24] consists of a contracting path to capture the context and a symmetric expanding path that enables the precise localization. The HED model [30] employs deeply supervised structures, and automatically learns rich hierarchical representations that are fused to resolve the challenging ambiguity in edge and object boundary detection. Multi-scale features have also been leveraged in various applications such as semantic segmentation [39], face-alignment [20], visual tracking [16] 2 Figure 2: An overview of the proposed DID-MDN method. The proposed network contains two modules: (a) residual-aware rain-density classifier, and (b) multi-stream densely-connected de-raining network. The goal of the residual-aware rain-density classifier is to determine the rain-density level given a rainy image. On the other hand, the multi-stream densely-connected de-raining network is designed to efficiently remove the rain streaks from the rainy images guided by the estimated rain-density information. crowd-counting [27], action recognition [42], depth estimation [4], single image dehazing [22, 37] and also in single image de-raining [33]. Similar to [33], we also leverage a multi-stream network to capture the rain-streak components with different scales and shapes. However, rather than using two convolutional layers with different dilation factors to combine features from different scales, we leverage the densely-connected block [11] as the building module and then we connect features from each block together for the final rain-streak removal. The ablation study demonstrates the effectiveness of our proposed network compared with the structure proposed in [33]. der de-rain the image. This is mainly due to the fact that a single network may not be sufficient enough to learn different rain-densities occurring in practice. We believe that incorporating density level information into the network can benefit the overall learning procedure and hence can guarantee better generalization to different rain conditions [23]. Similar observations have also been made in [23], where they use two different priors to characterize light rain and heavy rain, respectively. Unlike using two priors to characterize different rain-density conditions [23], the rain-density label estimated from a CNN classifier is used for guiding the de-raining process. To accurately estimate the density information given a rainy input image, a residual-aware raindensity classifier is proposed, where the residual information is leveraged to better represent the rain features. In addition, to train the classier, a large-scale synthetic dataset consisting of 12,000 rainy images with density labels is synthesized. Note that there are only three types of classes (i.e. labels) present in the dataset and they correspond to low, medium and high density. 3. Proposed Method The proposed DID-MDN architecture mainly consists of two modules: (a) residual-aware rain-density classifier, and (b) multi-stream densely connected de-raining network. The residual-aware rain-density classifier aims to determine the rain-density level given a rainy image. On the other hand, the multi-stream densely connected de-raining network is designed to efficiently remove the rain streaks from the rainy images guided by the estimated rain-density information. The entire network architecture of the proposed DID-MDN method is shown in Fig. 2. One common strategy in training a new classifier is to fine-tune a pre-defined model such as VGG-16 [26], Res-net [8] or Dense-net [11] on the newly introduced dataset. One of the fundamental reasons to leverage a fine-tune strategy for the new dataset is that discriminative features encoded in these pre-defined models can be beneficial in accelerating the training and it can also guarantee better generalization. However, we observed that directly fine-tuning such a ‘deep’ model on our task is not an efficient solution. This is 3.1. Residual-aware Rain-density Classifier As discussed above, even though some of the previous methods achieve significant improvements on the deraining performance, they often tend to over de-rain or un3 mainly due to the fact that high-level features (deeper part) of a CNN tend to pay more attention to localize the discriminative objects in the input image [40]. Hence, relatively small rain-streaks may not be localized well in these highlevel features. In other words, the rain-streak information may be lost in the high-level features and hence may degrade the overall classification performance. As a result, it is important to come up with a better feature representation to effectively characterize rain-streaks (i.e. rain-density). From (1), one can regard r = y − x as the residual component which can be used to characterize the rain-density. To estimate the residual component (r̂) from the observation y, a multi-stream dense-net (without the label fusion part) using the new dataset with heavy-density is trained. Then, the estimated residual is regarded as the input to train the final classifier. In this way, the residual estimation part can be regarded as the feature extraction procedure 1 , which is discussed in Section 3.2. The classification part is mainly composed of three convolutional layers (Conv) with kernel size 3 × 3, one average pooling (AP) layer with kernel size 9×9 and two fully-connected layers (FC). Details of the classifier are as follows: Conv(3,24)-Conv(24,64)-Conv(64,24)-APFC(127896,512)-FC(512,3), where (3,24) means that the input consists of 3 channels and the output consists of 24 channels. Note that the final layer consists of a set of 3 neurons indicating the rain-density class of the input image (i.e. low, medium, high). An ablation study, discussed in Section 4.3, is conducted to demonstrate the effectiveness of proposed residual-aware classifier as compared with the VGG-16 [26] model. (a) (b) Figure 3: Sample images containing rain-streaks with various scales and shapes.(a) contains smaller rain-streaks, (b) contains longer rain-streaks. images shown in Fig. 3, the rainy image in Fig. 3 (a) contains smaller rain-streaks, which can be captured by smallscale features (with smaller receptive fields), while the image in Fig. 3 (b) contains longer rain-streaks, which can be captured by large-scale features (with larger receptive fields). Hence, we believe that combining features from different scales can be a more efficient way to capture various rain streak components [10, 33]. Based on this observation and motivated by the success of using multi-scale features for single image de-raining [33], a more efficient multi-stream densely-connected network to estimate the rain-streak components is proposed, where each stream is built on the dense-block introduced in [11] with different kernel sizes (different receptive fields). These multi-stream blocks are denoted by Dense1 (7 × 7), Dense2 (5 × 5), and Dense3 (3 × 3), in yellow, green and blue blocks, respectively in Fig. 2. In addition, to further improve the information flow among different blocks and to leverage features from each dense-block in estimating the rain streak components, a modified connectivity is introduced, where all the features from each block are concatenated together for rain-streak estimation. Rather than leveraging only two convolutional layers in each stream [33], we create short paths among features from different scales to strengthen feature aggregation and to obtain better convergence. To demonstrate the effectiveness of our proposed multi-stream network compared with the multi-scale structure proposed in [33], an ablation study is conducted, which is described in Section 4. To leverage the rain-density information to guide the deraining process, the up-sampled label map 2 is concatenated with the rain streak features from all three streams. Then, the concatenated features are used to estimate the residual (r̂) rain-streak information. In addition, the residual is subtracted from the input rainy image to estimate the coarse de-rained image. Finally, to further refine the estimated Loss for the Residual-aware Classifier:. To efficiently train the classifier, a two-stage training protocol is leveraged. A residual feature extraction network is firstly trained to estimate the residual part of the given rainy image, then a classification sub-network is trained using the estimated residual as the input and is optimized via the ground truth labels (rain-density). Finally, the two stages (feature extraction and classification) are jointly optimized. The overall loss function used to train the residual-aware classier is as follows: L = LE,r + LC , (2) where LE,r indicates the per-pixel Euclidean-loss to estimate the residual component and LC indicates the crossentropy loss for rain-density classification. 3.2. Multi-stream Dense Network It is well-known that different rainy images contain rainstreaks with different scales and shapes. Considering the 2 For example, if the label is 1, then the corresponding up-sampled label-map is of the same dimension as the output features from each stream and all the pixel values of the label map are 1. 1 Classificaiton network can be regarded as two parts: 1.Feature extractor and 2. Classifer 4 3.3. Testing coarse de-rained image and make sure better details well preserved, another two convolutional layers with ReLU are adopted as the final refinement. There are six dense-blocks in each stream. Mathematically, each stream can be represented as sj = cat[DB1 , DB2 , ..., DB6 ], (3) where cat indicates concatenation, DBi , i = 1, · · · 6 denotes the output from the ith dense block, and sj , j = 1, 2, 3 denotes the jth stream. Furthermore, we adopt different transition layer combinations3 and kernel sizes in each stream. Details of each stream are as follows: Dense1: three transition-down layers, three transition-up layers and kernel size 7 × 7. Dense2: two transition-down layers, two no-sampling transition layers, two transition-up layers and kernel size 5 × 5. Dense3: one transition-down layer, four no-sampling transition layers, one transition-up layer and kernel size 3 × 3. Note that each dense-block is followed by a transition layer. Fig 4 presents an overview of the first stream, Dense1. During testing, the rain-density label information using the proposed residual-aware classifier is estimated. Then, the up-sampled label-map with the corresponding input image are fed into the multi-stream network to get the final de-rained image. 4. Experimental Results In this section, we present the experimental details and evaluation results on both synthetic and real-world datasets. De-raining performance on the synthetic data is evaluated in terms of PSNR and SSIM [28]. Performance of different methods on real-world images is evaluated visually since the ground truth images are not available. The proposed DID-MDN method is compared with the following recent state-of-the-art methods: (a) Discriminative sparse codingbased method (DSC) [19] (ICCV’15), (b) Gaussian mixture model (GMM) based method [17] (CVPR’16), (c) CNN method (CNN) [5] (TIP’17), (d) Joint Rain Detection and Removal (JORDER) method [33] (CVPR’17), (e) Deep detailed Network method (DDN) [6] (CVPR’17), and (f) Joint Bi-layer Optimization (JBO) method [41] (ICCV’17). 4.1. Synthetic Dataset Even though there exist several large-scale synthetic datasets [6, 36, 33], they lack the availability of the corresponding rain-density label information for each synthetic rainy image. Hence, we develop a new dataset, denoted as Train1, consisting of 12,000 images, where each image is assigned a label based on its corresponding rain-density level. There are three rain-density labels present in the dataset (e.g. light, medium and heavy). There are roughly 4,000 images per rain-density level in the dataset. Similarly, we also synthesize a new test set, denoted as Test1, which consists of a total of 1,200 images. It is ensured that each dataset contains rain streaks with different orientations and scales. Images are synthesized using Photoshop. We modify the noise level introduced in step 3 of 4 to generate different rain-density images, where light, medium and heavy rain conditions correspond to the noise levels 5% ∼ 35%, 35% ∼ 65%, and 65% ∼ 95%, respectively 5 . Sample synthesized images under these three conditions are shown in Fig 5. To better test the generalization capability of the proposed method, we also randomly sample 1,000 images from the synthetic dataset provided by Fu [6] as another testing set, denoted as Test2. Figure 4: Details of the first stream Dense1. Loss for the De-raining Network:. Motivated by the observation that CNN feature-based loss can better improve the semantic edge information [13, 15] and to further enhance the visual quality of the estimated de-rained image [36], we also leverage a weighted combination of pixelwise Euclidean loss and the feature-based loss. The loss for training the multi-stream densely connected network is as follows L = LE,r + LE,d + λF LF , (4) where LE,d represents the per-pixel Euclidean loss function to reconstruct the de-rained image and LF is the featurebased loss for the de-rained image, defined as 1 (5) kF (x̂)c,w,h − F (x)c,w,h k22 , LF = CW H where F represents a non-linear CNN transformation and x̂ is the recovered de-rained image. Here, we have assumed that the features are of size w × h with c channels. In our method, we compute the feature loss from the layer relu1 2 of the VGG-16 model [26]. 4 http://www.photoshopessentials.com/photoeffects/photoshopweather-effects-rain/ 5 The reason why we use three labels is that during our experiments, we found that having more than three rain-density levels does not significantly improve the performance. Hence, we only use three labels (heavy, medium and light) in the experiments. 3 The transition layer can function as up-sample transition, downsample transition or no-sampling transition [12]. 5 Table 1: Quantitative results evaluated in terms of average SSIM and PSNR (dB) (SSIM/PSNR). Input DSC [19] (ICCV’15) GMM [17] (CVPR’16) CNN [5] (TIP’17) JORDER [33] (CVPR’17) DDN [6] (CVPR’17) JBO [41] (ICCV’17) Test1 0.7781/21.15 0.7896/21.44 0.8352/22.75 0.8422/22.07 0.8622/24.32 0.8978/ 27.33 0.8522/23.05 0.9087/ 27.95 Test2 0.7695/19.31 0.7825/20.08 0.8105/20.66 0.8289/19.73 0.8405/22.26 0.8851/25.63 0.8356/22.45 0.9092/ 26.0745 • Yang-Multi [33]6 : Multi-stream network trained without the procedure of label fusion. • Multi-no-label: Multi-stream densely connected network trained without the procedure of label fusion. • DID-MDN (our): Multi-stream Densely-connected network trained with the procedure of estimated label fusion. Heavy Medium Light Figure 5: Samples synthetic images in three different conditions. Table 2: Quantitative results compared with three baseline configurations on Test1. Single Yang-Multi [33] Multi-no-label The average PSNR and SSIM results evaluated on Test1 are tabulated in Table 2. As shown in Fig. 6, even though the single stream network and Yang’s multi-stream network [33] are able to successfully remove the rain streak components, they both tend to over de-rain the image with the blurry output. The multi-stream network without label fusion is unable to accurately estimate the rain-density level and hence it tends to leave some rain streaks in the derained image (especially observed from the derained-part around the light). In contrast, the proposed multi-stream network with label fusion approach is capable of removing rain streaks while preserving the background details. Similar observations can be made using the quantitative results as shown in Table 2. DID-MDN PSNR (dB) 26.05 26.75 27.56 27.95 SSIM 0.8893 0.8901 0.9028 0.9087 Table 3: Accuracy of rain-density estimation evaluated on Test1. VGG-16 [26] Residual-aware 73.32 % 85.15 % Accuracy DID-MDN 4.2. Training Details During training, a 512 × 512 image is randomly cropped from the input image (or its horizontal flip) of size 586×586. Adam is used as optimization algorithm with a mini-batch size of 1. The learning rate starts from 0.001 and is divided by 10 after 20 epoch. The models are trained for up to 80×12000 iterations. We use a weight decay of 0.0001 and a momentum of 0.9. The entire network is trained using the Pytorch framework. During training, we set λF = 1. All the parameters are defined via crossvalidation using the validation set. 4.3.1 Results on Two Synthetic Datasets We compare quantitative and qualitative performance of different methods on the test images from the two synthetic datasets - Test1 and Test2. Quantitative results corresponding to different methods are tabulated in Table 1. It can be clearly observed that the proposed DID-MDN is able to achieve superior quantitative performance. To visually demonstrate the improvements obtained by the proposed method on the synthetic dataset, results on two sample images selected from Test2 and one sample chosen from our newly synthesized Test1 are presented in Figure 7. Note that we selectively sample images from all three conditions to show that our method performs well under different variations 7 . While the JORDER method [33] is able to remove some parts of the rain-streaks, it still tends to leave some rain-streaks in the de-rained images. Similar results are also observed from [41]. Even though the method 4.3. Ablation Study The first ablation study is conducted to demonstrate the effectiveness of the proposed residual-aware classifier compared to the VGG-16 [26] model. The two classifiers are trained using our synthesized training samples Train1 and tested on the Test1 set. The classification accuracy corresponding to both classifiers on Test1 is tabulated in Table 3. It can be observed that the proposed residual-aware classifier is more accurate than the VGG-16 model for predicting the rain-density levels. In the second ablation study, we demonstrate the effectiveness of different modules in our method by conducting the following experiments: 6 To better demonstrate the effectiveness of our proposed muli-stream network compared with the state-of-the-art multi-scale structure proposed in [33], we replace our multi-stream dense-net part with the multi-scale structured in [33] and keep all the other parts the same. 7 Due to space limitations and for better comparisons, we only show the results corresponding to the most recent state-of-the-art methods [33, 6, 41] in the main paper. More results corresponding to the other methods [19, 17, 5] can be found in Supplementary Material. • Single: A single-stream densely connected network (Dense2) without the procedure of label fusion. 6 PSNR: 16.47 SSIM: 0.51 PSNR: 22.87 SSIM: 0.8215 Input Single PSNR: 23.02 SSIM: 0.8213 PSNR: 23.47 SSIM: 0.8233 Yang-Multi [33] PSNR: 24.88 SSIM: 0.8623 Multi-no-label PSNR: Inf SSIM: 1 DID-MDN Ground Truth Figure 6: Results of ablation study on a synthetic image. PSNR: 17.27 SSIM: 0.8257 PSNR:21.89 SSIM: 0.9007 PSNR: 25.30 SSIM:0.9455 PSNR: 20.72 SSIM: 0.8885 PSNR: 25.95 SSIM: 0.9605 PSNR: Inf SSIM: 1 PSNR:19.31 SSIM: 0.7256 PSNR:22.28 SSIM: 0.8199 PSNR:26.88 SSIM:0.8814 PSNR: 21.42 SSIM:0.7878 PSNR: 29.88 SSIM:0.9252 PSNR: Inf SSIM:1 PSNR: 20.74 SSIM:0.7992 PSNR:24.20 SSIM:0.8502 PSNR:29.44 SSIM:0.9429 PSNR:25.32 SSIM: 0.8922 PSNR:29.84 SSIM:0.9482 PSNR: Inf SSIM:1 Input JORDER (CVPR’17) [33] DDN (CVPR’17) [6] JBO (ICCV’17) [41] DID-MDN Ground Truth Figure 7: Rain-streak removal results on sample images from the synthetic datasets Test1 and Test2. of Fu et al. [6] is able to remove the rain-streak, especially in the medium and light rain conditions, it tends to remove some important details as well, such as flower details, as shown in the second row and window structures as shown in the third row (Details can be better observed via zooming-in the figure). Overall, the proposed method is able to preserve better details while effectively removing the rain-streak components. 4.3.2 As before, previous methods either tend to under de-rain or over de-rain the images. In contrast, the proposed method achieves better results in terms of effectively removing rain streaks while preserving the image details. In addition, it can be observed that the proposed method is able to deal with different types of rain conditions, such as heavy rain shown in the second row of Fig 8 and medium rain shown in the fifth row of Fig 8. Furthermore, the proposed method can effectively deal with rain-streaks containing different shapes and scales such as small round rain streaks shown in the third row in Fig 8 and long-thin rain-streak in the second row in Fig 8. Overall, the results evaluated on real-world images captured from different rain conditions demonstrate the effectiveness and the robustness of the proposed DID- Results on Real-World Images The performance of the proposed method is also evaluated on many real-world images downloaded from the Internet and also real-world images published by the authors of [36, 6]. The de-raining results are shown in Fig 8. 7 Input JORDER (CVPR’17) [33] DDN (CVPR’17) [6] JBO (ICCV’17) [41] DID-MDN Figure 8: Rain-streak removal results on sample real-world images. MDN method. More results can be found in Supplementary Material. 4.3.3 work (DID-MDN) for jointly rain-density estimation and deraining. In comparison to existing approaches which attempt to solve the de-raining problem using a single network to learn to remove rain streaks with different densities (heavy, medium and light), we investigated the use of estimated rain-density label for guiding the synthesis of the derained image. To efficiently predict the rain-density label, a residual-aware rain-density classier is proposed in this paper. Detailed experiments and comparisons are performed on two synthetic and one real-world datasets to demonstrate that the proposed DID-MDN method significantly outperforms many recent state-of-the-art methods. Additionally, the proposed DID-MDN method is compared against baseline configurations to illustrate the performance gains obtained by each module. Running Time Comparisons Running time comparisons are shown in the table below. It can be observed that the testing time of the proposed DIDMDN is comparable to the DDN [6] method. On average, it takes about 0.3s to de-rain an image of size 512 × 512. Table 4: Running time (in seconds) for different methods averaged on 1000 images with size 512×512. 512X512 DSC GMM CNN (GPU) JORDER (GPU) DDN (GPU) JBO (CPU) DID-MDN (GPU) 189.3s 674.8s 2.8s 600.6s 0.3s 1.4s 0.2s 5. Conclusion References In this paper, we propose a novel density-aware image deraining method with multi-stream densely connected net- [1] D.-Y. Chen, C.-C. Chen, and L.-W. Kang. Visual depth guided color image rain streaks removal using sparse coding. 8 [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] Y. Li, R. T. Tan, X. Guo, J. Lu, and M. S. Brown. Rain streak removal using layer priors. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2736–2744, June 2016. [18] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3431–3440, 2015. [19] Y. Luo, Y. Xu, and H. Ji. Removing rain from a single image via discriminative sparse coding. In ICCV, pages 3397– 3405, 2015. [20] X. Peng, R. S. Feris, X. Wang, and D. N. Metaxas. A recurrent encoder-decoder network for sequential face alignment. In European Conference on Computer Vision, pages 38–56. Springer International Publishing, 2016. [21] X. Peng, X. Yu, K. Sohn, D. Metaxas, and M. Chandraker. Reconstruction for feature disentanglement in pose-invariant face recognition. In ICCV, 2017. [22] W. Ren, S. Liu, H. Zhang, J. Pan, X. Cao, and M.-H. Yang. Single image dehazing via multi-scale convolutional neural networks. In ECCV, pages 154–169. Springer, 2016. [23] W. Ren, J. Tian, Z. Han, A. Chan, and Y. Tang. Video desnowing and deraining based on matrix decomposition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4210–4219, 2017. [24] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234–241. Springer, 2015. [25] V. Santhaseelan and V. K. Asari. Utilizing local phase information to remove rain from video. International Journal of Computer Vision, 112(1):71–89, 2015. [26] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. [27] V. A. Sindagi and V. M. Patel. Generating high-quality crowd density maps using contextual pyramid cnns. In ICCV, 2017. [28] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600–612, 2004. [29] W. Wei, L. Yi, Q. Xie, Q. Zhao, D. Meng, and Z. Xu. Should we encode rain streaks in video as deterministic or stochastic? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2516–2525, 2017. [30] S. Xie and Z. Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395–1403, 2015. [31] T. Xu, P. Zhang, Q. Huang, H. Zhang, Z. Gan, X. Huang, and X. He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In CVPR, 2018. [32] J. Xue, H. Zhang, K. Dana, and K. Nishino. Differential angular imaging for material recognition. In CVPR, 2017. [33] W. Yang, R. T. Tan, J. Feng, J. Liu, Z. Guo, and S. Yan. Deep joint rain detection and removal from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1357–1366, 2017. IEEE transactions on circuits and systems for video technology, 24(8):1430–1455, 2014. Y.-L. Chen and C.-T. Hsu. A generalized low-rank appearance model for spatio-temporally correlated rain streaks. In IEEE ICCV, pages 1968–1975, 2013. D. Eigen, D. Krishnan, and R. Fergus. Restoring an image taken through a window covered with dirt or rain. In ICCV, pages 633–640, 2013. D. Eigen, C. Puhrsch, and R. Fergus. Depth map prediction from a single image using a multi-scale deep network. In NIPS, pages 2366–2374, 2014. X. Fu, J. Huang, X. Ding, Y. Liao, and J. Paisley. Clearing the skies: A deep network architecture for single-image rain removal. IEEE Transactions on Image Processing, 26(6):2944–2956, 2017. X. Fu, J. Huang, D. Zeng, Y. Huang, X. Ding, and J. Paisley. Removing rain from single images via a deep detail network. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1715–1723, July 2017. K. He, X. Zhang, S. Ren, and J. Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. In European Conference on Computer Vision, pages 346–361. Springer, 2014. K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016. D.-A. Huang, L.-W. Kang, Y.-C. F. Wang, and C.-W. Lin. Self-learning based image decomposition with applications to single image denoising. IEEE Transactions on multimedia, 16(1):83–93, 2014. D.-A. Huang, L.-W. Kang, M.-C. Yang, C.-W. Lin, and Y.-C. F. Wang. Context-aware single image rain removal. In Multimedia and Expo (ICME), 2012 IEEE International Conference on, pages 164–169. IEEE, 2012. G. Huang, Z. Liu, K. Q. Weinberger, and L. van der Maaten. Densely connected convolutional networks. arXiv preprint arXiv:1608.06993, 2016. S. Jégou, M. Drozdzal, D. Vazquez, A. Romero, and Y. Bengio. The one hundred layers tiramisu: Fully convolutional densenets for semantic segmentation. In Computer Vision and Pattern Recognition Workshops (CVPRW), 2017 IEEE Conference on, pages 1175–1183. IEEE, 2017. J. Johnson, A. Alahi, and L. Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694–711. Springer, 2016. L.-W. Kang, C.-W. Lin, and Y.-H. Fu. Automatic singleimage-based rain streaks removal via image decomposition. IEEE TIP, 21(4):1742–1755, 2012. C. Ledig, L. Theis, F. Huszár, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang, et al. Photo-realistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1–8, 2017. K. Li, Y. Kong, and Y. Fu. Multi-stream deep similarity learning networks for visual tracking. In IJCAI, 2017. 9 [34] H. Zhang and K. Dana. Multi-style generative network for real-time transfer. arXiv preprint arXiv:1703.06953, 2017. [35] H. Zhang and V. M. Patel. Convolutional sparse and lowrank coding-based rain streak removal. In Applications of Computer Vision (WACV), 2017 IEEE Winter Conference on, pages 1259–1267. IEEE, 2017. [36] H. Zhang, V. Sindagi, and V. M. Patel. Image de-raining using a conditional generative adversarial network. arXiv preprint arXiv:1701.05957, 2017. [37] H. Zhang, V. Sindagi, and V. M. Patel. Joint transmission map estimation and dehazing using deep networks. arXiv preprint arXiv:1708.00581, 2017. [38] Z. Zhang, Y. Xie, F. Xing, M. McGough, and L. Yang. Mdnet: A semantically and visually interpretable medical image diagnosis network. In CVPR, 2017. [39] H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia. Pyramid scene parsing network. In Proceedings of the IEEE International Conference on Computer Vision, pages 1–8, 2017. [40] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2921–2929, 2016. [41] L. Zhu, C.-W. Fu, D. Lischinski, and P.-A. Heng. Joint bilayer optimization for single-image rain streak removal. In Proceedings of the IEEE international conference on computer vision, pages 2526–2534, 2017. [42] Y. Zhu, Z. Lan, S. Newsam, and A. G. Hauptmann. Hidden two-stream convolutional networks for action recognition. arXiv preprint arXiv:1704.00389, 2017. 10
1cs.CV
Conversion Rate Optimization through Evolutionary Computation Risto Miikkulainen1,2 , Neil Iscoe1 , Aaron Shagrin1 , Ron Cordell1 , Sam Nazari1 , Cory Schoolland1 , Myles Brundage1 , Jonathan Epstein1 , Randy Dean1 , Gurmeet Lamba1 1 Sentient 2 The Technologies, Inc. University of Texas at Austin arXiv:1703.00556v4 [cs.HC] 30 Apr 2017 ABSTRACT Conversion optimization means designing a web interface so that as many users as possible take a desired action on it, such as register or purchase. Such design is usually done by hand, testing one change at a time through A/B testing, or a limited number of combinations through multivariate testing, making it possible to evaluate only a small fraction of designs in a vast design space. This paper describes Sentient Ascend, an automatic conversion optimization system that uses evolutionary optimization to create effective web interface designs. Ascend makes it possible to discover and utilize interactions between the design elements that are difficult to identify otherwise. Moreover, evaluation of design candidates is done in parallel online, i.e. with a large number of real users interacting with the system. A case study on an existing media site shows that significant improvements (i.e. over 43%) are possible beyond human design. Ascend can therefore be seen as an approach to massively multivariate conversion optimization, based on a massively parallel interactive evolution. KEYWORDS Conversion optimization, e-commerce, interactive evolution, online evolution, design 1 INTRODUCTION In e-commerce, designing web interfaces (i.e. web pages and interactions) that convert as many users as possible from casual browsers to paying customers is an important goal [1, 11]. While there are some well-known design principles, including simplicity and consistency, there are often also unexpected interactions between elements of the page that determine how well it converts. The same element, such as a headline, image, or testimonial, may work well in one context but not in others—it is often hard to predict the result, and even harder to decide how to improve a given page. An entire subfield of information technology has emerged in this area, called conversion rate optimization, or conversion science. The standard method is A/B testing, i.e. designing two different versions of the same page, showing them to different users, and collecting statistics on how well they each convert [8]. This process allows incorporating human knowledge about the domain and conversion optimization into the design, and then testing their effect. After observing the results, new designs can be compared and gradually improved. The A/B testing process is difficult and time-consuming: Only a very small fraction of page designs can be tested in this way, and subtle interactions in the design are likely to go unnoticed and unutilized. An alternative to A/B is multivariate testing, where all value combinations of a few elements are tested at once. While this process captures interactions between these elements, only a very small number of elements is usually included (e.g. 2-3); the rest of the design space remains unexplored. This paper describes a new technology for conversion optimization based on evolutionary computation. This technology is implemented in Ascend, a conversion optimization product by Sentient Technologies, deployed in numerous e-commerce websites of paying customers since September 2016 [13]. Ascend uses a customer-designed search space as a starting point. It consists of a list of elements on the web page that can be changed, and their possible alternative values, such as a header text, font, and color, background image, testimonial text, and content order. Ascend then automatically generates web-page candidates to be tested, and improves those candidates through evolutionary optimization. Because e-commerce sites often have high volume of traffic, fitness evaluations can be done live with a large number of real users in parallel. The evolutionary process in Ascend can thus be seen as a massively parallel version of interactive evolution, making it possible to optimize web designs in a few weeks. From the application point of view, Ascend is a novel method for massively multivariate optimization of web-page designs. Depending on the application, improvements of 20-200% over human design have been observed through this approach [13]. This paper describes the technology underlying Ascend, presents an example use case, and outlines future opportunities for evolutionary computation in optimizing e-commerce. 2 BACKGROUND With the explosive growth of e-commerce in recent years, entirely new areas of study have emerged. One of the main ones is conversion rate optimization, i.e. the study of how web interfaces should be designed so that they are as effective as possible in converting users from casual browsers to actual customers. Conversion means taking a desired action on the web interface such as making a purchase, registering for a marketing list, or clicking on other desired link in an email, website, or desktop, mobile, or social media application [1, 11]. Conversions are usually measured in number of clicks, but also in metrics such as resulting revenue or time spent on the site and rate of return to the site. Conversions are currently optimized in a labor-intensive manual process that requires significant expertise. The web design expert or marketer first creates designs that s/he believes to be effective. These designs are then tested in an A/B testing process, by directing user traffic to them, and measuring how well they convert. If the conversion rates are statistically significantly different, the better design is adopted. This design can then be improved further, using domain expertise to change it, in another few rounds of creation and testing. Conversion optimization is a fast-emerging component of ecommerce. In 2016, companies spent over $72 billion to drive customers to their websites [5]. Much of that investment does not result in sales: conversion rates are typically 2-4% (i.e. 2-4% of the users that come to the site convert within 30 days). In 2014, only 18% of the top 10,000 e-commerce sites did any conversion optimization; in January 2017, 30% of them did so [4]. The growth is largely due to available conversion optimization tools, such as Optimizely, Visual Website Optimizer, Mixpanel, and Adobe Target [4]. These tools make it possible to configure the designs easily, allocate users to them, record the results, and measure significance. This process has several limitations. First, while the tools make the task of designing effective web interfaces easier, the design is still done by human experts. The tools thus provide support for confirming the experts’ ideas, not helping them explore and discover novel designs. Second, since each step in the process requires statistical significance, only a few designs can be tested. Third, each improvement step amounts to one step in hillclimbing; such a process can get stuck in local maxima. Fourth, the process is aimed at reducing false positives and therefore increases false negatives, i.e. designs with good ideas may be overlooked. Fifth, while the tools provide support for multivariate testing, in practice only a few combinations can be tested (e.g. five possible values for two elements, or three possible values for three elements). As a result, it is difficult to discover and utilize interactions between design elements. Evolutionary optimization is well suited to address these limitations. Evolution is an efficient method for exploration; only weak statistical evidence is needed for progress; its stochastic nature avoids getting stuck in local maxima; good ideas will gradually become more prevalent. Most importantly, evolution searches for effective interactions. For instance, Ascend may find that the button needs to be green, but *only* when it is transparent, *and* the header is in small font, *and* the header text is aligned. Such interactions are very difficult to find using A/B testing, requiring human insight into the results. Evolution makes this discovery process automatic. With Ascend, it is thus possible to optimize conversions better and at a larger scale than before. Technically, Ascend is related to approaches to interactive evolution [12, 15] and crowdsourcing [2, 9] in that evaluations of candidates are done online by human users. The usual interactive evolution paradigm, however, employs a relatively small number of human evaluators, and their task is to select good candidates or evaluate the fitness of a pool of candidates explicitly. In contrast in Ascend, a massive number of human users are interacting with the candidates, and fitness is derived from their actions (i.e. convert or not) implicitly. 3 3.1 Defining the Search Space The starting point for Ascend is a search space defined by the web designer. Ascend can be configured to optimize a design of a single web-page, or a funnel consisting of multiple pages such as the landing page, selections, and a shopping cart. For each such space, the designer specifies the elements on that page and values that they can take. For instance in the landing page example of Figure 1, logo size, header image, button color, content order are such elements, and they can each take on 2-4 values. Ascend searches for good designs in the space of possible combinations of these values. This space is combinatorial, and can be very large, e.g. 1.1M in this example. Interestingly, it is exactly this combinatorial nature that makes web-page optimization a good application for evolution: Even though human designers have insight into what values to use, their combinations are difficult to predict, and need to be discovered by search process such as evolution. 3.2 Initializing Evolution A typical setup is that there is already a current design for the web interface, and the goal for Ascend is to improve over its performance. That is, the current design of the web interface is designated as the Control, and improvement is measured compared to that particular design. Because fitness is evaluated with real users, exploration incurs real cost to the customer. It is therefore important that the candidates perform reasonably well throughout evolution, and especially in the beginning. If the initial population is generated randomly, many web interfaces would perform poorly. Instead, the initial population is created using the Control as a starting point: The candidates are created by changing the value of one element systematically. In a small search space, the initial population thus consists of all candidates with one difference from the control; in a large search space, the population is a sample of the set of such candidates. With such an initialization, most of the candidates perform similarly to the control. The candidates also cover the search dimensions well, thus forming a good starting point for evolution. 3.3 Evolutionary Process Each page is represented as a genome, as shown for two example pages in Figure 2 (left side). The usual genetic operations such as crossover (re-combination of the elements in the two genomes; middle) and mutation (randomly changing one element in the offspring; right side) are then performed to create new candidates. In the current implementation, fitness-proportionate selection is used to generate offspring candidates from the current population. From the current population of n candidates, another n new candidates are generated in this way. Because evaluations are expensive, consuming traffic for which most customers have to pay, it is useful to minimize them during evolution. Each page needs to be tested only to the extent that it is possible to decide whether it is promising, i.e. whether it should serve as a parent in the next generation, or should be discarded. A process similar to age-layering [7, 14] is therefore used to allocate fitness evaluations. At each generation, each new candidate and each old candidate is evaluated with a small number (a maturity THE ASCEND METHOD Ascend consists of defining the space of possible web interfaces, initializing the population with a good coverage of that space, allocating traffic to candidates intelligently so that bad designs can be eliminated early, and testing candidates online in parallel. Each of these steps is described in more detail in this section. 2 Figure 2: Genetic Encoding and Operations on Web Interface Candidates. The pages are represented as concatenations of their element values with one-hot encoding. Crossover and mutation operate on these vectors as usual, creating new combinations of values. Figure 1: Elements and Values of an Example Web Page Design. In this example, 13 elements each have 2-4 possible values, resulting in 1.1M combinations. Alternative designs to be evaluated User_(j) User_(j+1) Interaction_ij Interaction_i(j+1) Design_i 11 00 00 11 00 11 00 11 00 11 00 11 User_(n) Learning Process for discovering optimal designs Best design Interaction_in Avg. return Figure 3: Overall Architecture of the Online Evolution System. The outcome of each interaction (i.e. whether the user converted or not) constitutes one evaluation of a design. Many such evaluations ij are run in parallel with different users j and averaged to estimate how good the design i is. After all designs have been evaluated, the adaptation process discards bad designs and generates more variations of the best designs. This process of generation, testing, and selection is repeated until a sufficiently good design has been found or the time allocated for the process has been spent. The best design found so far is output as the result of the learning process. The system thus discovers good designs for web interfaces through live online testing. age) of user interactions, such as 2000. The top n candidates are retained, and the bottom n discarded. In this manner, bad candidates are eliminated quickly. Good candidates receive progressively more evaluations, and the confidence in their fitness estimate increases. In this process, Ascend learns which combinations of elements are effective, and gradually focuses the search around the most promising designs. It is thus sufficient to test only a tiny fraction of the search space to find the best ones, i.e. thousands of pages instead of millions or billions. 3.4 learning, the population-based approach is particularly effective because the evaluations can be done in parallel. The entire population can be tested at once, as different users interact with the site simultaneously. It is also unnecessary to test each design to statistical significance; only weak statistical evidence is sufficient to proceed in the search. In this process, thousands of page designs can be tested in a short time, which is impossible through A/B or multivariate testing. Figure 3 shows the overall architecture of the system. A population of alternative designs (center) are adapted (right) based on evaluations with actual users (left). The population of designs (center) are evaluated with many users in parallel (left). The evolutionary process (right) generates new designs and outputs the best design in the end. The system also keeps track of which design has Online Evolution While in simple cases (where the space of possible designs is small) such optimization can potentially be carried out by simpler mechanisms such as systematic search, hill-climbing, or reinforcement 3 Figure 4: The control design and three best evolved designs. After 60 days of evolution with 599,008 user interactions, a design for the search widget was found that converted 46.6% better than the control (5.61% vs. 8.22%), as well as other good designs. Much of the improvement was based on discovering a combination of colors that draws attention to the widget and makes the call to action clear. been show to which user, so that they get to see the same design if they return within a certain time limit (e.g. the same day). 4 excellent result given that the control was a candidate that was already hand-optimized using state-of-the art tools. Unlike Control, the top candidates utilize bright background colors to draw attention to the widget. There is an important interaction between the background and the blue banner (whose color was fixed)—in the best two designs (in the middle) the background is distinct from the banner but not competing with it. Moreover, given the colored background, a white button with black text provided the most clear call for action. It is difficult to recognize such interactions ahead of time, yet evolution discovered them early on, and many of the later candidates built on them. Other factors such as an active call to action (i.e. “Get Started” and “Find my Program” rather than “Request Info”) amplified it further. At the time evolution was turned off, better designs were still being discovered, suggesting that a more prolonged evolution and a larger search space (e.g. including banner color and other choices) could have improved the results further. CASE STUDY As an example of how Ascend works, let us consider a case study on optimizing the web interface for a media site that connects users to online education programs. This experiment was run in September through November 2016 on the desktop traffic of the site. The initial design for this page is shown in the left side of Figure 4. It had been hand designed using standard tools such as Optimizely. Its conversion rate during the time of the experiment was found to be 5.61%, which is typical of such web interfaces. Based on this page, the web designers came up with nine elements, with two to nine values each, resulting in 381,024 potential combinations (Figure 5). While much larger search spaces are possible, this example represents a mid-size space common with many current sites. The initial population of 37 candidates was formed by systematically replacing each of the values in the control page with one of the alternative values, as described in section 3.2. Evolution was then run for 60 days, or four generations, altogether testing 111 candidates with 599,008 user interactions total. The estimated conversion rates of the candidates over this time are shown in Figure 6. The conversion rates of the top 20 candidates are shown in Figure 7. These figures show that evolution was successful in discovering significantly better candidates than control. As an independent verification, the three top candidates in Figure 4 were then subjected to an A/B test using Optimizely. In about 6500 user interactions, the best candidate was confirmed to increase the conversion rate by 43.5% with greater than 99% significance (and the other two by 37.1% and 28.2%)—which is an 5 FUTURE WORK Ascend has been applied to numerous web interfaces, and it has consistently improved conversion rates by 20-200% compared to hand designed controls [13]. The main limitation is often the human element: web designers, who are used to A/B and multivariate testing, often try to minimize the search space, i.e. the number of elements and values, as much as possible, thereby not giving evolution much space to explore and discover powerful solutions. Often the evolution discovers significant improvement in a couple of generations, and the designers are eager to adopt them right away, instead of letting evolution optimize the designs fully. Populationbased optimization requires different thinking; as designers become 4 Figure 5: A screenshot of the user interface for designing Ascend experiments, showing the elements and values in the learnhowtobecome.com case study. Nine elements with two to nine different values each result in 381,024 potential web page designs; the first value in each element is designated as the control. This is a mid-size problem typical of current web interface designs. more comfortable with it, we believe they will let evolution take its course, reaching more refined results. Currently Ascend delivers one best design, or a small number of good ones, in the end as the result, again in keeping with the A/B testing tradition. In many cases there are seasonal variations and other long-term changing trends, making the performance of good designs gradually decay. It is possible to counter this problem by running the optimization again every few months. However, a new paradigm of “always-on” would be more appropriate: Evolutionary optimization can be run continuously at a low volume, keeping up with changing trends (i.e. through dynamic evolutionary optimization; [3]). New designs can then be adopted periodically when their performance exceeds old designs significantly. Furthermore, currently Ascend optimizes a single design to be used with all future users of a mobile or desktop site. An interesting extension would be to take user segmentation [16] into account, and evolve different pages for different kinds of users. Moreover, such a mapping from user characterizations to page designs can be automatic: A mapping system such as a neural network can take user variables such as location, time, device, any past history with the site as inputs, and generate the vector of elements and their values as outputs. Neuroevolution [6, 10] can discover optimal such mappings, in effect evolve to discover a dynamic, continuous segmentation of the user space. Users will be shown designs that are likely to convert well based on experience with other users with similar characteristics, continuously and automatically. It will be possible to analyze such evolved neural networks and discover what variables are most predictive, characterize the main user segments, and thereby develop an in-depth understanding of the opportunity. Finally, the Ascend approach is not limited to optimizing conversions. Any outcome that can be measured, such as revenue or user retention, can be optimized. The approach can also be used in a different role, such as optimizing the amount of resources spent on attracting users, such as ad placement and selection, adword bidding, and email marketing. The approach can be seen as a fundamental step in bringing machine optimization into e-commerce, and demonstrating the value of evolutionary computation in real-world problems. 6 CONCLUSION Sentient Ascend demonstrates how interactive evolution can be scaled up by testing a large number of candidates in parallel on real users. It includes technology for keeping the cost of exploration reasonable, and for minimizing the number of evaluations needed. From the application point of view, Ascend is the first automated system for massively multivariate conversion optimization— replacing A/B with AI. It finds the subtle combinations of variables that lead to conversion increases. The web designer can spend more time trying ideas and less time doing statistics, giving them the freedom they need to make a difference. REFERENCES [1] Tim Ash, Rich Page, and Maura Ginty. 2012. Landing Page Optimization: The Definitie Guide to Testing and Tuning for Conversions (second ed.). Wiley, Hoboken, NJ. [2] Daren C. Brabham. 2013. Crowdsourcing. MIT Press, Cambridge, MA. [3] Jürgen Branke. 2002. Evolutionary Optimization in Dynamic Environments. Springer, Berlin. [4] Builtwith. 2017. A/B Testing Usage. (2017). https://trends.builtwith.com/ analytics/a-b-testing Retrieved 1/9/2017. [5] eMarketer. 2016. US Digital Ad Spending to Surpass TV this Year. (2016). https:// www.emarketer.com/Article/US-Digital-Ad-Spending-Surpass-TV-this-Year/ 1014469 Retrieved 2/1/2017. [6] Dario Floreano, Peter Dürr, and Claudio Mattiussi. 2008. Neuroevolution: From Architectures to Learning. Evolutionary Intelligence 1 (2008), 47–62. [7] Babak Hodjat and Hormoz Shahrzad. 2013. Introducing an Age-Varying Fitness Estimation Function. In Genetic Programming Theory and Practice X, Rick Riolo, Ekaterina Vladislavleva, Marylyn D Ritchie, and Jason H. Moore (Eds.). Springer, New York, 59–71. 5 Figure 6: Estimated Conversion Rates through the 60-day Online Evolution Run. Days are in the x-axis and the conversion rate on the y axis. The dark blue dots (on top) indicate the current best candidate, the light blue dots (in the middle) an average of all currently active candidates, and the orange dots (at the bottom) the estimated performance of the control design. The shaded areas display the 95% confidence intervals (from the binomial distribution with the observed mean). The dark blue peaks indicate the start of each new generation. Such peaks emerge because during the first few days, the new candidates have been evaluated only a small number of times, and some of them have very high estimated rates through random chance. Eventually they will be evaluated in a maturity age of 2000 user interactions, and the estimates become lower and the confidence intervals narrower. The elite candidates are tested across several generations as described in section 3.3, resulting in very narrow intervals towards the end. Estimated conversion rates of the best candidates in later generations are significantly higher than control, suggesting that evolution is effective in discovering better candidates. Interestingly, the active population average is also higher than control, indicating that the experiment did not incur any cost in performance. . [8] Ron Kohavi and Roger Longbotham. 2016. Online Controlled Experiments and A/B Tests. In Encyclopedia of Machine Learning and Data Mining, Claude Sammut and Geoffrey I. Webb (Eds.). Springer, New York. [9] Joel Lehman and Risto Miikkulainen. 2013. Boosting Interactive Evolution using Human Computation Markets. In Proceedings of the 2nd International Conference on the Theory and Practice of Natural Computation. Springer, Berlin. [10] Joel Lehman and Risto Miikkulainen. 2013. Neuroevolution. Scholarpedia 8, 6 (2013), 30977. http://nn.cs.utexas.edu/?lehman:scholarpedia13 [11] Khalid Salehd and Ayat Shukairy. 2011. Conversion Optimization: The Art and Science of Converting Prospects to Customers. O’Reilly Media, Inc., Sebastopol, CA. [12] Jimmy Secretan, Nicholas Beato, David B. D’Ambrosio, Adelein Rodriguez, Adam Campbell, J. T. Folsom-Kovarik, and Kenneth O. Stanley. 2011. Picbreeder: A Case Study in Collaborative Evolutionary Exploration of Design Space. Evolutionary Computation 19 (2011), 345–371. [13] Sentient Technologies 2017. It’s not A/B, I’s AI. (2017). http://www.sentient.ai/ ascend Retrieved 1/9/2017. [14] Hormoz Shahrzad, Babak Hodjat, and Risto Miikkulainen. 2016. Estimating the Advantage of Age-Layering in Evolutionary Algorithms. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO 2016). ACM, New York, NY, USA. [15] H. Takagi. 2001. Interactive Evolutionary Computation: Fusion of the Capacities of EC Optimization and Human Evaluation. Proc. IEEE 89, 9 (2001), 1275– 1296. http://ieeexplore.ieee.org/iel5/5/20546/00949485.pdf?tp=&arnumber= 949485&isnumber=20546 [16] Daniel Yankelovich and David Meer. 2006. Rediscovering Market Segmentation. Harvard Business Review 84, 2 (2006). 6 Figure 7: Estimated Conversion Rates of the Top 20 Candidates vs. Control. The candidates are identified in the x-axis with numbers corresponding to those in Figure 4. They are ordered according to their estimated conversion rate. All these candidates are significantly better than the control at the 95% level. The best one (on the right) is 46.6% better, as confirmed with an independent A/B test that showed 43.5% improvement. 7
9cs.NE
Using artificial intelligence for data reduction in mechanical engineering L. Mdlazi1, C.J. Stander1, P.S. Heyns 1, T. Marwala2 1 Dynamic Systems Group Department of Mechanical and Aeronautical Engineering, University of Pretoria Pretoria, 0002, South Africa E-mail: [email protected] 2 School of Electrical and Information Engineering, University of the Witwatersrand, Private Bag 3, Wits, 2050, South Africa E-mail: [email protected] Abstract In this paper artificial neural networks and support vector machines are used to reduce the amount of vibration data that is required to estimate the Time Domain Average of a gear vibration signal. Two models for estimating the time domain average of a gear vibration signal are proposed. The models are tested on data from an accelerated gear life test rig. Experimental results indicate that the required data for calculating the Time Domain Average of a gear vibration signal can be reduced by up to 75% when the proposed models are implemented. 1. Introduction Calculating the Time Domain Average (TDA) of a gear vibration signal by direct averaging using digital computers requires large amounts of data [17]. This requirement makes it difficult to develop online gearbox condition monitoring systems that utilize time domain averaging calculated by direct averaging to enhance diagnostic capability. This study presents a novel approach to estimating the TDA of a gear vibration signal, using less data than would be used when calculating the TDA by direct averaging. Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs) are used for estimating the TDA of a gear vibration signal. Two models are presented. The input data comprises rotationsynchronized gear vibration signals and the output is the TDA of the gear vibration signal. When Model 1 is used, the results indicate that the amount of gear vibration data required for calculating the TDA is reduced to 25 percent of the amount of data required when calculating the TDA by direct averaging. When Model 2 is used, the amount of data to be stored in the data acquisition system is reduced to less than 20 percent of the data that would be stored when calculating the TDA by direct averaging. The ANNs that are implemented are Multi-Layer Perceptrons (MLPs) and Radial Basis Functions (RBFs) [8]. Two parameters were selected to verify whether the TDA estimated by the models retains the original diagnostic capability of the TDA. These parameters are the kurtosis for impulses and the peak value for overall vibration. The computational time is also compared to determine the suitability of the proposed models in real-time analysis. 2. ANNs and SVMs ANNs and SVMs may be viewed as parameterized non-linear mapping of input data to the output data. Learning algorithms are viewed as methods for finding parameter values that look probable in the light of the data. The learning process takes place by training the ANNs or SVMs through supervised learning. Supervised learning is the case where the input data set and the output data set are both known, and ANNs or SVMs are used to approximate the functional mapping between the two data sets. 2.1 Multi-layer Peceptron A two-layered MLP architecture was used. This selection was made because of the universal approximation theorem, which states that a twolayered architecture is adequate for the MLP. The MLP provides a distributed representation with respect to the input space due to the cross-coupling between hidden units. The output of a two-layer perceptron can be expressed as the following equation: yk = f outer M j =1 w(ji1) xi + w(j10) + wk( 20) wkj( 2 ) f inner (1) where f outer and finner are activation functions, w(1) ji denotes a weight in the first layer, going from input i to hidden unit j, wk(2)0 (2) kj hidden unit k and w is the bias for the denotes a weight in the second layer. The parameter f inner was selected as hyperbolic tangent function “tanh” and f outer was selected as a linear function. The hyperbolic tangent function maps the interval [-∞,∞] onto the interval [-1,1] and the linear activation function maps the interval [-∞,∞] onto the interval [-∞,∞]. The maximum-likelihood approach was used for training the MLP network. The sum-of-squares-oferror and the weight decay regularization was used as cost functions [8, 9]. The weight decay penalizes large weights and ensures that the mapping function is smooth, avoiding an over-fitted mapping between the input data and the output data [9]. In this study, a regularization coefficient of 1.5 was found most suitable. The weights wi and biases in the hidden layers were varied using Scaled Conjugate Gradient (SCG) optimization until the cost function was minimized [10]. It was determined empirically that a two-layer MLP network with five hidden units was best suited to this application. 2.2 Radial basis functions The RBF network approximates functions by a combination of radial basis functions and a linear output layer. The RBF neural networks provide a smooth interpolating function for which the number of basis functions is determined by the complexity of the mapping to be represented, rather than by the data set. The RBF neural network mapping is given by yk ( x ) = M j =1 ϖ kjφ j ( x ) + ϖ k 0 (2) where ω k 0 are the biases, ω kj are the output layer weights, x is the d-dimensional input vector and φ j ( ⋅) is the jth basis function. The thin plate-spline basis function was used in this study [8]. The radial basis function is trained in two stages. In the first stage the input data set xn alone is used to determine the basis function parameters. After the first training stage, the basis functions are kept fixed and the second layer of weights is determined in the second training phase. Since the basis functions are considered fixed, the network is equivalent to a single-layer network that can be optimized by minimizing a suitable error function. The sum-of-square error function is also used to train RBFs. The error function is a quadratic function of the weights and its minimum can therefore be found in terms of the solution of a set of linear equations. For regression the basis function parameters are found by treating the basis function centers and widths, along with the secondlayer weights, as adaptive parameters to be determined by minimizing the error function. In this study it was determined empirically that a RBF network with five basis functions was most suitable. 2.3 Support vector machines SVMs were developed by Vapnik [9] and have gained much popularity in recent years. The SVM formulation embodies the Structural Risk Minimization (SRM) principle, which has been shown to be superior to the traditional Empirical Risk Minimization (ERM) principle employed by conventional neural networks [9, 11, 12]. SRM minimizes an upper limit on the expected risk, as opposed to the ERM that minimizes the error on the training data. It is this difference that gives SVMs a greater ability to generalize. When SVMs are applied to regression problems, loss functions that include a distance measure are used. The εinsensitive loss function [9] was selected for this study. The ε-insensitive loss function is defined by: Le ( y ) = 0 f (x) − y − ε for f ( x ) − y Otherwise . (3) In non-linear regression, non-linear mapping is used to map the data to a higher dimensional feature space where linear regression is performed. The kernel approach is employed to address the curse of dimensionality. The non-linear support vector regression solution, using the ε-insensitive loss function, is given by: max (α , α ∗ ) = max∗ ∗ α ,α αα i = 1 j =1 1 − 2 i =1 with constraints, l 0 ≤ α i ,α i* ≤ C , l il =1 (α i j l j j =1 α ∗ ( yi − ∈) − α i ( yi − ∈) (α * i (4) * j (5) − α i* ) = 0 Solving Equation (3) with the constraints in Equation (5) determines the Lagrange multipliers, α and α * and the regression function is given by i f (x ) = i =1 (α i − α i∗ )K (x i , x ) + b (6) − α i∗ )K (xi , x j ) (7) where w, x = SVs (α i FF Network Configuration Rotation synchronized gear vibration signals − α i )(α − α j )K (xi , x j ) i = 1, K , l Input 1 (αi − αi∗ )(K (xi , xr ) + K (xi , xs )). 2 i =1 Different kernels were investigated for mapping the data to a higher dimensional feature space where linear regression was performed. The exponential radial basis function kernel [11] with an order of 10 was found most suitable for this application. b =− 3. Proposed models Two different models are proposed. Model 1 maps the input space to the target using simple feedforward network configuration. The size of the input space is systematically reduced to find the optimal number of input vectors that can be used to estimate the target vector correctly. When the ANNs and SVMs are properly trained, Model 1 is capable of mapping the input space to the target, using less data than would otherwise be used when calculating the TDA by direct averaging. It was determined empirically that 40 rotationsynchronized gear vibration signals were suitable for predicting the TDA with Model 1. Consequently, the amount of data required for calculating the TDA is reduced to 25 percent when using 40 rotation-synchronized gear vibration signals, since 160 rotation-synchronized gear vibration signals were used in calculating the TDA by direct averaging. Figure 2 shows a schematic diagram of the proposed methodology. Output TDA of signal (ANN/SVMs) Figure 1. Schematic of proposed methodology Model 2 estimates the TDA of the input space in small sequential steps, analogous to taking a running average of the input space. This model consists of a number of feed-forward networks. Instead of using the network to estimate the TDA of the entire input data, Model 2 first sequentially estimates the average of subsections of the input data. The output of the first stage is used as input into the second network that estimates the TDA of the entire input data. The feed-forward networks are trained off-line to reduce computation time. In Model 2, data can be discarded immediately after use. This means that the model does not require large amounts of data to be stored in the data logger, even though all the data should be collected. In this study, 10 rotation-synchronized gear vibration signals were found most suitable for estimating the instantaneous average in the first stage of estimation. As a result, the amount of vibration data that is stored in the data logger was reduced to less than 20 percent of the amount of data that is stored in the data logger when calculating the TDA by direct averaging. 4. Estimation results Model 1 and Model 2 were used for estimating the TDA of the gear vibration signal so that the TDA estimated by the proposed models could be compared with the TDA calculated by direct averaging. The data that was used was from the accelerated gear life test rig [13, 14]. Comparisons were made in the time and the frequency domains. To quantify the accuracy of the TDA estimated by the models, the ‘fit’ parameter ηsim [15] defined by N ηsim = 100 N k =1 e( k ) k =1 [%] , (8) ydesired ( k ) was used. In Equation 8 e(k) is the simulation accuracy defined by Equation (9) and N is the number of data points used. e(k ) = y desired (k ) − y achieved (k ) (9) ydesired is the TDA signal calculated by direct averaging and yachieved is the TDA signal estimated by the models. The ‘fit’ parameter ηsim gives a single value for each simulation, therefore can be used to compare the performance of the different formulations over the entire gear life. A high value for the ‘fit’ parameter ηsim implies a bad fit whereas a low value implies a good fit. Through trial and error it was established that ηsim = 40% is a suitable upper cut-off point for simulation accuracy. For Model 1, 40 rotation-synchronized gear vibration signals from the first test with 1024 points per revolution were used for training the ANNs and 256 points per revolution were used to train SVMs. This resulted in training sets of dimensions 1×40×1024 and 1×40×256 respectively. Fifteen test data sets were measured through the life of the gear and used as validation sets. This resulted in validation data sets of dimensions 15×160×1024 15×160×256 for ANNs and SVMs respectively. For Model 2, the whole data set of 160 rotationsynchronized gear vibration signals from the first test was used for training the ANNs and SVMs. This resulted in a training set of dimensions 1×160×1024 for ANNs and 1×160×256 for SVMs. The rest of the data of dimensions 15×160×1024 and 15×160×256 were used as validation data for ANNs and SVMs respectively. Figure 2 shows the estimation results obtained when Model 1 with an MLP network was simulated using unseen validation data sets of 40 input signals during the running-in stages of gear life. The dotted line is the TDA estimated by Model 1, and the solid line is the TDA calculated by direct averaging. The first plot in Figure 2 is the time domain representation of the results; the second plot is the frequency domain representation. It is observed that both the time and frequency domain representations are almost exact fits. This shows that Model 1 with MLP networks retains the time and frequency domain properties of the original time domain averaging process when using gear vibration data from the accelerated gear life test rig. Similar performances were obtained throughout the life of the gear using both models with RBFs and SVMs, even though there were significant changes in the vibration signatures as the condition of the monitored gear deteriorated. The changes in the vibration signatures were due to changes in the meshing stiffness caused by cracks in the gear teeth. The good performance is due to the mapping and generalization capabilities of ANNs and SVMs. Figure 2. Model 1 with MLP estimation using data from a test conducted under constant load conditions Figure 3 shows the simulation accuracy ηsim plotted against the gear life for Model 1. It is observed that Model 1 with RBF network and Model 1 with SVMs give similarly performance. Their performance was slightly better than that of Model 1 with MLP networks. The performances of all three formulations are acceptable because ηsim is less than the cut-off value for all the formulations. this simulation was not optimal; consequently it did not generalize well to changes in the measured vibration as gear failure progressed. Similar results were obtained using Model 2. Figure 3. Model 1 Simulation accuracy ηsim vs. gear life under constant load conditions In addition to simply considering the goodness of fit, the diagnostic capabilities of the TDA estimated by the models was assessed using the peak value of the vibration Xmax during a given interval T and the Kurtosis. The peak value is used to monitor the overall magnitude of the vibration to distinguish between acceptable and unacceptable vibration levels and the kurtosis is useful for detecting the presence of an impulse within the vibration signal [16]. The peak value of the vibration Xmax and the kurtosis of the TDA calculated using direct averaging, were compared to the TDA estimate from the proposed models. Figure 4 is a plot of Xmax and kurtosis calculated from the TDA estimated by Model 1, superimposed on the Xmax and kurtosis calculated from the TDA obtained by using direct averaging. Figure 4 indicates that the kurtosis is an exact fit for all three formulations. This implies that the TDA predicted by Model 1 can be used to monitor the presence of impulses in the measured gear vibration signal. It is also observed that the kurtosis is very high during the early stages of gear life. This is characteristic of the running-in stages of the gear life, during which the vibration signature tends to be random. A similar trend is observed during the wear-out stage in which strong impulses are caused by the reduction in stiffness in cracked or broken gear teeth. Only the peak values obtained from Model 1 with MLP and SVM are close fits and can be used to monitor the amplitude of the overall vibration. Model 1 with RBF achieved an unsatisfactory performance because the RBF network selected in Figure 4. Comparison of kurtosis and peak values for the TDA calculated by direct averaging (+) and the TDA predicted by Model 1 (*) with MLP, RBF and SVMs To put the proposed models into perspective, their computation time was compared to that of the existing time domain averaging method using a Pentium 4 computer with a 1.60 GHz processor. The computation times are listed in Table 1. Table 1. Computation time in seconds Time parameter [s] Pre-processing Model 1 Training Model 1 Simulating Model 1 Pre-processing Model 2 Training Model 2 Simulating Model 2 TDA 1.011 0.75 1.011 - MLP 0.703 22.24 0.016 1.011 1.14 0.08 RBF 0.703 2.219 0.047 1.011 1.015 0.078 SVM 0.703 497.0 5.500 1.011 963.8 83.76 It is clear that Model 1 requires less pre-processing time than the TDA calculated by direct averaging. This is because Model 1 uses 25 percent of the vibration data while the original TDA process uses all of the vibration data. The required preprocessing time for Model 2 is equal to that of calculating the TDA by direct averaging because both models use the same amount of vibration data. When Model 1 and Model 2 are used, RBF and MLP give the best performance in terms of simulating time and SVMs give the poorest performance. The models are trained off-line therefore the training time would not influence the performance in real-time applications. When the models are used with SVMs their performance is poor. The poor performance with SVMs is due to the optimization problem. In SVMs optimization is a quadratic problem with 2N variables, where N is the number of data training points. Longer training times are needed because more operations are required in the process. This is much slower than the MLP and RBF neural networks in which only the weights and biases or the basis centers are obtained by minimizing the error functions. 5. Conclusion In this paper a novel approach that uses artificial neural networks and support vector machines to reduce the amount of data that is required to calculate the time domain average of a gear vibration signal is presented. Two models are proposed. Using Model 1 the data for calculating the Time Domain Average was reduced to 25 percent of the data required to calculate the Time Domain Average by direct averaging. Model 2 was found to be excellent at estimating the Time Domain Average but had the disadvantage of requiring more operations to execute. When using Model 2, less than 20 percent of the data are needed to be stored in the data logger at any given time. Furthermore, the suitability of the developed models for diagnostic purposes was assessed. It was observed that the performances of the Model 1 and Model 2 were similar over the entire life of the gear. The good performance of Model 2 can be attributed to the fact that Model 2 uses the whole data set for training and simulation, whereas Model 1 uses only a section of the data set. Using the whole data set during training and simulation exposes the formulations in the model to all the transient effects in the data, resulting in a more accurate estimate of the Time Domain Average. The performance of Model 1 relies on the generalization capabilities of the formulation used. 6. References [1] L. Hongxing, Z. Hongfu, J. Chengyu and Q. Liangheng, 2000 Mechanical Systems and Signal Processing 14(2), pp. 279-285, An improved algorithm for direct time domain averaging. [2] C.R. Trimble, 1968 Hewlett-Packard Journal 19(8), pp. 2-7. What is signal averaging? [3] S. Braun, 1975 Acustica (32), pp. 69-77, The extraction of periodic waveforms by time domain averaging. [4] S. Braun and B. Seth, 1980 Journal of Sound and Vibration 70(4), pp. 513-526 Analysis of repetitive mechanism signatures. [5] P. D. McFadden, 1987 Mechanical Systems and Signal Processing (1), pp. 83-95, A revised model for the extraction of periodic waveforms by time domain averaging. [6] P. D. McFadden, 1989 Mechanical Systems and Signal Processing (3), pp. 87-97, Interpolation techniques for time domain averaging of gear vibration. [7] B. Samanta, 2004 Mechanical Systems and Signal Processing (18), pp. 625-644, Gear fault detection using artificial neural networks and support vector machines. [8] C.M. Bishop, 1995 Neural networks for pattern recognition, Oxford: Clarendon Press. [9] S.R. Gunn, 1998 Technical report, University of Southampton, Department of Electrical and Computer Science, UK, Support vector machines for classification and regression. [10] S. Haykin, 1999 Neural networks, 2nd edition, New Jersey, USA: Prentice-Hall Inc. [11] V.N. Vapnik, 1999 IEEE Transactions on Neural Networks 10, pp. 988-1000, An overview of learning theory. [12] V.N. Vapnik, 1995 The nature of statistical learning theory, New York, USA: Springer-Verlag. [13] C.J. Stander and P.S Heyns, 2002 Proceedings of the 15th International Congress on Condition Monitoring and Diagnostic Engineering Management, Birmingham UK, 2-4 September 2002, pp. 220-230. Instantaneous shaft speed monitoring of gearboxes under fluctuating load conditions. [14] C.J. Stander and P.S Heyns, 2002 Mechanical Systems and Signal Processing 16(6) pp. 10051024. Using vibration monitoring for local fault detection on gears operating under fluctuating load conditions. [15] A.D. Raath, 1992 Structural dynamic response reconstruction in the time domain, Ph.D. thesis, Department of Mechanical and Aeronautical Engineering, University of Pretoria. [16] M. P. Norton, 1989 Fundamentals of noise and vibration analysis for engineers, New York: Cambridge University Press.
5cs.CE
arXiv:1412.3714v2 [cs.NE] 13 Dec 2014 Feature Weight Tuning for Recursive Neural Networks Jiwei Li Computer Science Department Stanford University Stanford, CA, USA 94305 [email protected] Abstract This paper addresses how a recursive neural network model can automatically leave out useless information and emphasize important evidence, in other words, to perform “weight tuning” for higher-level representation acquisition. We propose two models, Weighted Neural Network (WNN) and Binary-Expectation Neural Network (BENN), which automatically control how much one specific unit contributes to the higher-level representation. The proposed model can be viewed as incorporating a more powerful compositional function for embedding acquisition in recursive neural networks. Experimental results demonstrate the significant improvement over standard neural models. 1 Introduction Recursive neural network models [1] constitute one type of neural structure for obtaining higherlevel representations beyond word-level such as phrases or sentences. It works in a bottom-up fashion on tree structures (e.g., parse trees) in which long-term dependency can be to some extent captured. Figure 1 gives a brief illustration about how recursive neural models work to obtain the distributed representation for the short sentence “The movie is wonderful”. Suppose his and hwonderful are the embeddings for tokens is and wonderful. The representation for their parent node VP at second layer is given by: hVP = f(W · [his , hwonderful ] + b) (1) where W and b denote parameters involved in the convolutional function. f(·) is the activation function, usually tanh or sigmod or the rectifier linear function. For NLP tasks, the obtained embeddings could be further fed into task-specific machine learning models1 , through which parameters are to be optimized. Take sentiment analysis as an example, we could feed the aforementioned sentence embedding into a logistic regression model to classify it as either positive or negative. Embeddings are sometimes more capable of capturing latent semantic meanings or syntactic rules within the text than manually developed features, from which many NLP tasks would benefit (e.g., [2, 3]). Such a type of structure suffers some sorts of intrinsic drawbacks. Revisit Figure 1, common sense tells us that tokens like “the”, “movie” and “is” do not contribute much to the sentiment decision but word “wonderful” is the key part (and a good machine learning model should have the ability of learning these rules). Unfortunately, the intrinsic structure of recursive neural networks makes it less flexible to get rid of the influence from less sentiment-related tokens. If the keyword “wonderful” hides too deep in the parse tree, for example, as in the sentence “I studied Russia in Moscow, where 1 Of course, embeddings could also be optimized through the task-specific objective functions. 1 Figure 1: Illustration of Standard Recursive Neural Network for Sentence-level Representation Calculation. Model unigram SVM Recursive Neural Net Accuracy 0.743 0.730 Table 1: A brief comparison between SVM and standard neural network models for sentence-level sentiment classification using date set from [4]. Neural network models are trained with L2 regularization, using AdaGrad [5] with minibatches (for details about implementations of recursive networks, please see Section 2). Parameters are trained based on 5-fold cross validation on the training data. We report the best performance searching optimum regularization parameter, optimum batch size for mini-batches and convolutional function. Word embeddings are borrowed from Glove [6] with dimensionality of 300, which generates better performance than word2vect, SENNA [7] and RNNLM [8]. all my family think the winter is wonderful”, it will takes quite a few convolution steps before the keyword ‘wonderful” comes up to the surface, with the consequence that its influence on the final sentence representation could be very trivial. Such an issue, usually referred to as gradient vanishing [9]. is not specific for recursive models, but for most deep learning architectures. When we compare neural models with SVM, one notable weakness of big-of-word based SVM is its inability of considering how words are combined to form meanings (or order information in other words) [10]. But interestingly, such downside of SVM comes with the advantage of resilience in feature managing as the optimization is “flat-expanded”. Low weights will be assigned to lessinformative evidence, which could further be pushed to zero by regularization. Table 1 gives a brief comparison between unigram based SVM and neural network models for sentence-level sentiment prediction on Pang et al.’s dataset [4], and as can be seen, in this task, standard neural network models underperform SVM2 . Revisit the form of Equ.1, there are two straws we can grasp at to deal with the aforementioned problem: (1) expecting the learned feature embeddings for less useful words such as the3 exert very little influence (for example, a zero vector for the best) (2) expecting the compositional parameters W and b are extremely powerful. For the former, it is sometimes hard, as mostly we borrow (or initialize) word embeddings from those trained from large corpus (e.g., word2vec, RNNLM [8, 11], SENNA [7]), rather than training embeddings from task-specific objective functions as neural models can be easily over fitted given the small amount of training data4 . Regarding the latter issue, several alternative compositional functions have been proposed to enable more varieties in composition to cater. Recent proposed approaches include, for example, MatrixVector RNN [12], which represents every word as both a vector and a matrix, RNTN [2] which allows greater interactions between the input vectors, and the algorithm presented in [13] which 2 To note, results here are not comparable with Socher et al.’s work [2] which obtains state-of-art performance in sentiment classification, as here labels at sentence-level constitute only sort of supervision for both SVM and neural network models (for details, see footnote 7). 3 We just use this example for illustration. Practically, the might be a good sentiment indicator as it usually co-appears with superlatives. 4 There are cases, for example, [2], where task-specific word embeddings are learned. But it requires sufficient training data to avoid over fitting. For example, Socher et al.’s work labels every single node as positive/negative/neutral along parse trees (with a total number of more than 200,000 phrases). 2 associates different labels (e.g., POS tags, relation tags) with different sets of compositional parameters. These approaches to some extent enlarge the power of compositional functions. In this paper, we borrow the idea of “weight tuning” from feature based SVM and try to incorporate such idea into neural architectures. To achieve this goal, we propose two recursive neural architectures, Weighted Neural Network (WNN) and Binary-Expectation Neural Network (BENN). The major idea involved in the proposed approaches is to associate each node in the recursive network with additional parameters, indicating how important it is for final decision. For example, we would expect such type of a structure would dilute the influence of tokens like “the” and “movie” but magnifies the impact of tokens like “wonderful” and “great” in sentiment analysis tasks. Parameters associated with proposed models are automatically optimized through the objective function manifested by the data. The proposed model combines the capability of neural models to capture the local compositional meanings with weight tuning approach to reduce the influence of undesirable information at the same time, and yield better performances in a range of different NLP tasks when compared with standard neural models. The rest of this paper is organized as follows: Section 2 briefly describes the related work. The details of WNN and BENN are illustrated in Section 4 and experimental results are presented in Section 5, followed by a brief conclusion. 2 Related Work Distributed representations, calculated based on neural frameworks, are extended beyond tokenlevel, to represent N-grams [14], phrases [2], sentences (e.g., [3, 15]), discourse [16, 13], paragraphs [17] or documents [18]. Recursive and recurrent [19, 20] models constitute two types of commonly used frameworks for sentence-level embedding acquisition. Different variations of recurrent/recursive models are proposed to cater for different scenarios (e.g., [3, 2]). Other recently proposed approaches included sentence compositional approach proposed in [21], or paragraph/sentence vector [17] where representations are optimized through predicting words within the sentence. Neural network architecture sometimes requires a vector representation of each input token. Various deep learning architectures have been explored to learn these embeddings in an unsupervised manner from a large corpus [22, 23, 24, 25], which might have different generalization capabilities and are able to capture the semantic meanings depending on the specific task at hand. Both of the proposed architectures are in this work inspired by the long short-term memory (LSTM) model, first proposed by Hochreiter and Schmidhuber back in 1990s [26, 27] to process time sequence data where there are very long time lags of unknown size between important events5 . LSTM associates each time with a series of “gates” to determine whether the information from early timesequence should be forgotten [27] and when current information should be allowed to flow into or out of the memory. LSTM could partially address gradient vanishing problem in recurrent neural models and have been widely used in machine translation [28, 29] 3 “Weight Tuning” for Neural Network Let s denote a sequence of token s = {w1 , w2 , ..., wns }. It could be phrases, sentences etc. Each word w is associated with a specific vector embedding ew = {e1w , e2w , ..., eK w }, where K denotes the dimensionality of the word embedding. We wish to compute the vector representation for sentence s, denoted as hs = {h1s , h2s , ..., hK s } based on parse trees using recursive neural models. Parse tree for each sentence is obtained from Stanford Parser [30]. 3.1 WNN for Recursive Neural Network For any node C in the parse tree, it is associated with representation hC . The basic idea of WNN is to associate each node C with an additional weight variable MC , which is in range (0,1), to denote the importance of current node. Technically, MC is used to pushing the output representation of not-useful node towards the direction of 0 and retain relatively important information. 5 http://en.wikipedia.org/wiki/Long_short_term_memory 3 We expect that information regarding the importance of current node (e.g., whether it is relevant to positive/negative sentiment) is embedded in its representation hC . So we use a convolution function to enable this type of information to emerge to the surface from the following compositional functions: RC = f(WM · hC + bM ) (2) MC = sigmod(UTM · RC ) (3) where WM is a D × K dimensional matrix and bB is the 1 × K bias vector. RC is a K dimensional intermediate vector. Such implementation can be viewed as using a three-layer neural model with D latent neurons for an output projected to a [0,1] space. Let output(C) denote the output from node C to its parent. In WNN, output(C) would consider both current information, which is embedded in the embedding hC and its related importance MC . output(C) is therefore given by output(C) = MC ∗ hC (4) Recall the example in Figure 1, we have: output(the) = Mthe · hthe output(movie) = Mmovie · hmovie (5) If the model thinks not too much relevant information embedded in hC , the value of MC would be small, pushing the output vector towards 0. The representations for parents, for example VP and NP in Figure 1, are therefore computed as follows: hVP = tanh(WB · [output(is), output(wonderful)]) hNP = tanh(WB · [output(the), output(movie)]) (6) where WB denotes a K × 2K dimensional matrix and [output(is), output(wonderful)] denotes the concatenation of the two vectors. In an optimum situation, Mthe and Mmovie will take the values around 0, leading to the representation of node NP to an around-zero vector. Training WNN For illustration purpose, we use a binary classification task to show how to train WNN. To note, the described training approach applies to other situations (e.g., multi-class classification, regression) with minor adjustments. In a binary classification task, each sequence is associated with a gold-standard label ys . ys takes value of 1 if positive and 0 otherwise. Standardly, to determine the value of ys , we feed the representation hs into a logistic regression model: p(ys = 1) = sigmod(UT hs + b) (7) where UT is a 1 ∗ K vector and b denotes the bias. Then by adding the regularization part parameterized by Q, the loss function J(Θ) for the training dataset is given by: X J(Θ) = − log[p(ys = 1)ys · (1 − p(ys = 1))1−ys ] + Q θ2 (8) θ∈Θ Revisit the example in Figure 1, for any parameter θ to optimize, the calculation for gradient ∂J/∂θ is trivial once ∂[MVP · hVP ]/∂θ and ∂[MNP · hNP ]/∂θ are obtained, which are given by: ∂MVP · hVP ∂hVP ∂MVP = MVP + hVP ∂θ ∂θ ∂θ (9) To note, hVP is embraced in MVP . As all components in Equ 9 are continuous, the gradient can be efficiently obtained from standard backpropagation [31, 32]. 4 Figure 2: Illustration of WNN. 3.2 BENN for Recursive Neural Network BENN associates each node with a binary variable BC , which is sampled from a binary distribution parameterized by LC . LC is a scalar fixed to the range of [0,1], indicating the possibility that current node should pass information to its ancestors. LC is obtained in the similar ways as in WNN by using a convolution to project the current representation hC to a scalar lying within [0,1]. RC = f(WB · hC + bB ) (10) LC = sigmod(UTB · RC ) (11) BC ∼ binary(LC ) (12) For smoothing purpose, in BENN, current node C outputs the expectation of embedding hC to its parent, as given by: output(C) = E[hC ] (13) Take the case in Figure 1 as an example again, vector hNP will therefore follow the following distribution: p(hNP = tanh(WB [hthe , hmovie ])) = Lthe · Lmovie p(hNP = tanh(WB [0, hmovie ])) = (1 − Lthe ) · Lmovie (14) p(hNP = tanh(WB [hthe , 0])) = Lthe · (1 − Lmovie ) p(hNP = tanh(WB [0, 0])) = (1 − Lthe ) · (1 − Lmovie ) E[hNP ] can be further obtained based on such distribution X E[hNP ] = P(hNP = h) · h (15) h To note, for leaf nodes, E[hC ] = hC . Training BENN For training, we again use binary sentiment classification for illustration. For any sentence s with label ys , we have p(ys = 1) = sigmod(UT E[hs ] + b) (16) With respect to any given parameter θ, the derivative of E[hs ] is further given by: ∂E(hs ) ∂LNP · LVP · tanh(WB [E(hNP ), E(hVP )]) = ∂θ ∂θ ∂(1 − LNP ) · LVP · tanh(WB [0, E(hVP )]) + ∂θ (17) ∂LNP · (1 − LVP ) · tanh(WB [E(hNP ), 0] + ∂θ ∂(1 − LNP ) · (1 − LVP ) · tanh(WB [0, 0]) + ∂θ With all components being continuous, the gradient can be efficiently obtained from standard backpropagation. 5 Recursive SVM Model Standard (GLOVE) Standard (learned) MV-RNN RNTN Label-Specific WNN BENN unigram Accuracy 0.730 0.658 0.704 0.760 0.768 0.778 0.772 0.743 Table 2: Binary Sentiment Classification with Supervision only at Sentence Level. Word embeddings are initialized from 300 dimensional embeddings borrowed from GLOVE [6]. 4 Experiment We perform experiments to better understand the behavior of the proposed models compared with standard neural models (and other variations). To achieve this, we implement our model on problems that require fixed-length vector representations for phrases or sentences. 4.1 Sentiment Analysis Sentence-level Labels We first perform experiments on dateset from [4]. In this setting, binary labels at the top of sentence constitute the only resource of supervision (to note, it is different from settings described in [2]). All neural models adopt the same settings for fair comparison: L2 regularization, gradient decent based on AdaGrad with mini batch size of 25, tuned parameters for regularization on 5-fold cross validation. For standard neural models, we implement two settings: standard (GLOVE) where word embeddings are directly fixed to GLOVE and standard (learned) where word embeddings are treated as parameters to optimize in the framework. Additionally, we implemented some recent popular variations of recursive models with more sophisticatedly designed compositional functions, including: • MV-RNN (Matrix-Vector RNN): which was proposed in [12] which represents every node in a parse tree as both a vector and a matrix. Given the vector representation hC1 , matrix representation VC1 for child node C1 , hC2 and VC2 for child node C2 , the vector representation hp and matrix representation Vp for parent p are given by: hp = f(W1 [VC1 · hC2 , VC2 · hC1 ]) Vp = f(W1 [VC1 , VC2 ]) (18) We fix word vector embeddings using SENNA and treat matrix representations as parameters to optimize. • RNTN (Recursive Neural Tensor Network): proposed in [2]. Given hC1 and hC2 for children nodes, RNTN computes parent vector hp in the following way: hp = f([hC1 , hC2 ]T V[hC1 , hC2 ] + W[hC1 , hC2 ]) (19) • Label-specific: associate each of the sentence roles (i.e., VP, NP or NN) with a specific composition matrix. We report results in Table 2. As discussed earlier, standard neural models underperform the bag of word models. To note, for derivations of standard neural models such as Standard (learned) and MVRNN with many more parameters to learn, the performance is even worse due to over-fitting. WNN and BENN , although not significantly output bag of words SVM, generates better results, yielding significant improvement over standard neural models and existing revised versions. Figure 3 illustrates the automatic learned muted factor MC regarding different nodes in the parse tree based on recursive network. As we can observe , the model is capable of learning the proper weight of vocabularies, assigning larger weight values to important sentiment indicators (e.g., wonderful, silly and tedious) and suppressing the influence of less important ones. We attribute the better performance of proposed models over standard neural models to such automatic weight-tuning ability. 6 Figure 3: Visual illustration of automatic learning of weight MC associated with each node in WMM. To note, in this scenario, we are not claiming that we generate state-of-art results using the proposed model. More sophisticated bag-of-word models, for example, (e.g., [33]) can generate better performance that what the proposed models achieve. The point we wish to illustrate here is that the proposed models provide a promising perspective over standard neural models due to the “weight tuning” property. And in the cases where more detailed data is available to capture the compositionally, the proposed models hold promise to generate more compelling results, as we will illustrate in Socher et al’s setting for sentiment analysis. Socher et al’s settings We now consider Socher et al’s dataset [2] for sentiment analysis, where contains gold-standard labels at every phrase node in the parse tree. The task could be considered either as a 5-way fine-grained classification task where the labels are verynegative/negative/neutral/positive/very-positive or a 2-way coarse-way as positive/negative based on labeled dataset. We follow the experimental protocols described in [2] (word embeddings are treated as parameters to learn rather than fixed to externally borrowed embeddings). In this work we only consider labeling the full sentences. In addition to varieties of neural models mentioned in Socher et al’s work, we also report the performance of recently proposed paragraph vector model [17], which first obtains sentence embeddings in an unsupervised manner by predicting words within the context and then feeds the pre-obtained embeddings into a logistic regression model. paragraph vector achieves current the state-of-art performance regarding Socher et al’s dataset. Performances are reported in Table 3. As can be seen, the proposed approach slightly underperforms current state-of-art performance achieved by paragraph vector but outperforms all the other versions of recursive neural models, indicating the adding “weight tuning” parameters indeed leads to better compositionally. To note, when there is more comprehensive dataset which we can rely on to obtain the favorable task-specific word embeddings, compositionally plays an important role in deciding whether the review is positive or negative by harnessing local word order information. In that case, neural models exhibit its power in capturing local evidence from the composition, leading to significantly better performance than all bag-of-words based models (i.e., SVM and Bigram Naives Bayes). 4.2 Document-level Sentiment Analysis on IMDB dataset We move on to sentiment analysis at document level. We use the IMDB dataset proposed by Maas et al. [33]. The dataset consists of 100,000 movie reviews taken from IMDB and each movie review contains several sentences. We follow the experimental protocols described in [33]. We first train word vectors from word2vect using the 75,000 training documents. Next we train the compositional functions using the 25,000 labeled documents by keeping the word embedding fixed. We first obtain sentence-level representations using WNN/BENN (recursive). As each review contains multiple sentences, we convolute sentence representations to one single vector using 7 Model SVM Bigram Naives Recursive MV-RNN RNTN Paragraph Vector WNN BENN Fine-grained 0.407 0.419 0.432 0.444 0.457 0.487 0.482 0.475 Coarse-grained 0.794 0.831 0.824 0.829 0.854 0.878 0.865 0.870 Table 3: The performance of proposed approaches compared with other methods on Stanford sentiment treebank dataset. Baseline performances are reported from [2, 17]. Model SVM-unigram SVM-bigram recursive+recurrent WNN BENN Precision 0.869 0.892 0.870 0.902 0.910 Table 4: The performance of proposed model compared to other approaches on binary classification on IMDB dataset. Results for baselines are reported from [33]. To note, the reported results here underperform current state-of-the-art performances. Paragraph vectors [17] reported an accuracy of 92.58 in terms of IMDB dataset. WNN/BENN recurrent network. We cross validate parameters using the labeled documents and test the models on the 25,000 testing reviews. The results of our approach and other baselines are reported in Table 5. As can be seen, for long documents, bag-of-words (both unigram and diagram) perform quite well and it is difficult to beat. Standard neural models again do not generate competent results compared with bag of word models in this task. But by incorporating weighted tuning mechanism, we got much better performance, roughly 5% when compared against standard neural models. Although WNN and BENN still underperform current state-of-art model Paragraph Vector [17], they produces better performance than bag-of-word models. 4.3 Sentence Representations for Coherence Evaluation Sentiment analysis forces more on the semantic perspective of meaning. Next we turn to a more syntactic oriented task, where we obtain sentence-level representations based on the proposed model to decide the coherence of a given sequence of sentences. We use corpora widely employed for coherence prediction [34, 35]. One contains reports on airplane accidents from the National Transportation Safety Board and the other contains reports about earthquakes from the Associated Press. Standardly, we use pairs of articles, one containing the original document order which is assumed to be coherent and used as positive examples, and the other a random permutation of the sentences from the same document, which are treated as not-coherent examples. We follow the protocols introduced in [34, 36, 37] by considering a window approach and feeding the concatenation of representations of adjacent sentences into a logistic regression model, to be classified as either coherent or non-coherent. In test time, we assume that the model makes a right decision if the original document gets a score higher than the one with random permutations. Current state-of-art performance regarding this task is obtained by using standard recursive network as described in [37]. Table 5 illustrates the performance of different models. Entity-grid model [34] generates state-of-art performance among all non-neural network models. As can be seen, neural models perform pretty well in this task when compared against existing feature based algorithm. From the reported results, better sentence representations are obtained by incorporating “weighted tuning” properties, pushing the state of art of this task to the accuracy of 0.936. 8 Model WNN-recursive BENN-recursive recursive Entity-Grid Accuracy 0.930 0.936 0.920 0.888 Table 5: Comparison of Different Coherence models. Reported baseline results are reprinted from [34]. 5 Conclusion In this paper, we propose two revised versions of neural models, WNN and BENN for obtaining higher level feature representations for a sequence of tokens. The proposed framework automatically incorporates the concept of “weight tuning” of SVM into the DL architectures which lead to better higher-level representations and generate better performance against standard neural models in multiple tasks. While it still underperforms bag-of-word models in some cases, and the newly proposed paragraph vector approach, it provides as an alternative to existing recursive neural models for representation learning. To note, while we limit our attentions to recursive models in this work, the idea of weight tuning in WNN and BENN, that associates nodes in neural models with additional weighed variables is a general one and can be extended to many other deep learning models with minor adjustment. References [1] Ronald J Williams and David Zipser. A learning algorithm for continually running fully recurrent neural networks. Neural computation, 1(2):270–280, 1989. [2] Richard Socher, Alex Perelygin, Jean Y Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1631–1642. Citeseer, 2013. [3] Ozan Irsoy and Claire Cardie. Bidirectional recursive neural networks for token-level labeling with structure. arXiv preprint arXiv:1312.0493, 2013. [4] Bo Pang, Lillian Lee, and Shivakumar Vaithyanathan. Thumbs up?: sentiment classification using machine learning techniques. In Proceedings of the ACL-02 conference on Empirical methods in natural language processing-Volume 10, pages 79–86. Association for Computational Linguistics, 2002. [5] John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. The Journal of Machine Learning Research, 12:2121–2159, 2011. [6] RichardSocher JeffreyPennington and ChristopherD Manning. Glove: Global vectors for word representation. 2014. [7] Ronan Collobert, Jason Weston, Léon Bottou, Michael Karlen, Koray Kavukcuoglu, and Pavel Kuksa. Natural language processing (almost) from scratch. The Journal of Machine Learning Research, 12:2493–2537, 2011. [8] Tomas Mikolov, Stefan Kombrink, Anoop Deoras, Lukar Burget, and J Cernocky. Rnnlmrecurrent neural network language modeling toolkit. In Proc. of the 2011 ASRU Workshop, pages 196–201, 2011. [9] Yoshua Bengio, Patrice Simard, and Paolo Frasconi. Learning long-term dependencies with gradient descent is difficult. Neural Networks, IEEE Transactions on, 5(2):157–166, 1994. [10] Raymond J Mooney and Razvan C Bunescu. Subsequence kernels for relation extraction. In Advances in neural information processing systems, pages 171–178, 2005. [11] Tomas Mikolov, Martin Karafiát, Lukas Burget, Jan Cernockỳ, and Sanjeev Khudanpur. Recurrent neural network based language model. In INTERSPEECH, pages 1045–1048, 2010. 9 [12] Richard Socher, Brody Huval, Christopher D Manning, and Andrew Y Ng. Semantic compositionality through recursive matrix-vector spaces. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1201–1211. Association for Computational Linguistics, 2012. [13] Jiwei Li, Rumeng Li, and Eduard Hovy. Recursive deep models for discourse parsing. 2014. [14] Richard Socher, Jeffrey Pennington, Eric H Huang, Andrew Y Ng, and Christopher D Manning. Semi-supervised recursive autoencoders for predicting sentiment distributions. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 151–161. Association for Computational Linguistics, 2011. [15] Phil Blunsom, Edward Grefenstette, Nal Kalchbrenner, et al. A convolutional neural network for modelling sentences. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics. Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, 2014. [16] Yangfeng Ji and Jacob Eisenstein. Representation learning for text-level discourse parsing. 2014. [17] Quoc V Le and Tomas Mikolov. Distributed representations of sentences and documents. arXiv preprint arXiv:1405.4053, 2014. [18] Misha Denil, Alban Demiraj, Nal Kalchbrenner, Phil Blunsom, and Nando de Freitas. Modelling, visualising and summarising documents with a single convolutional neural network. arXiv preprint arXiv:1406.3830, 2014. [19] Mike Schuster and Kuldip K Paliwal. Bidirectional recurrent neural networks. Signal Processing, IEEE Transactions on, 45(11):2673–2681, 1997. [20] Ilya Sutskever, James Martens, and Geoffrey E Hinton. Generating text with recurrent neural networks. In Proceedings of the 28th International Conference on Machine Learning (ICML11), pages 1017–1024, 2011. [21] Nal Kalchbrenner, Edward Grefenstette, and Phil Blunsom. A convolutional neural network for modelling sentences. Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, June 2014. [22] Yoshua Bengio, Holger Schwenk, Jean-Sébastien Senécal, Fréderic Morin, and Jean-Luc Gauvain. Neural probabilistic language models. In Innovations in Machine Learning, pages 137– 186. Springer, 2006. [23] Ronan Collobert and Jason Weston. A unified architecture for natural language processing: Deep neural networks with multitask learning. In Proceedings of the 25th international conference on Machine learning, pages 160–167. ACM, 2008. [24] Andriy Mnih and Geoffrey Hinton. Three new graphical models for statistical language modelling. In Proceedings of the 24th international conference on Machine learning, pages 641– 648. ACM, 2007. [25] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013. [26] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997. [27] Felix A Gers, Jürgen Schmidhuber, and Fred Cummins. Learning to forget: Continual prediction with lstm. Neural computation, 12(10):2451–2471, 2000. [28] Martin Sundermeyer, Tamer Alkhouli, Joern Wuebker, and Hermann Ney. Translation modeling with bidirectional recurrent neural networks. In Proceedings of the Conference on Empirical Methods on Natural Language Processing, October, 2014. [29] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. [30] Richard Socher, John Bauer, Christopher D Manning, and Andrew Y Ng. Parsing with compositional vector grammars. In In Proceedings of the ACL conference. Citeseer, 2013. 10 [31] Christoph Goller and Andreas Kuchler. Learning task-dependent distributed representations by backpropagation through structure. In Neural Networks, 1996., IEEE International Conference on, volume 1, pages 347–352. IEEE, 1996. [32] Richard Socher, Christopher D Manning, and Andrew Y Ng. Learning continuous phrase representations and syntactic parsing with recursive neural networks. In Proceedings of the NIPS-2010 Deep Learning and Unsupervised Feature Learning Workshop, pages 1–9, 2010. [33] Andrew L Maas, Raymond E Daly, Peter T Pham, Dan Huang, Andrew Y Ng, and Christopher Potts. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies-Volume 1, pages 142–150. Association for Computational Linguistics, 2011. [34] Regina Barzilay and Lillian Lee. Catching the drift: Probabilistic content models, with applications to generation and summarization. In HLT-NAACL, pages 113–120, 2004. [35] Regina Barzilay and Mirella Lapata. Modeling local coherence: An entity-based approach. Computational Linguistics, 34(1):1–34, 2008. [36] Annie Louis and Ani Nenkova. A coherence model based on syntactic patterns. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1157–1168. Association for Computational Linguistics, 2012. [37] Jiwei Li and Eduard Hovy. A model of coherence based on distributed sentence representation. 11
9cs.NE
The Nearest Neighbor Information Estimator is Adaptively Near Minimax Rate-Optimal arXiv:1711.08824v2 [stat.ML] 22 Feb 2018 Jiantao Jiao∗, Weihao Gao†, Yanjun Han∗ February 23, 2018 Abstract We analyze the Kozachenko–Leonenko (KL) nearest neighbor estimator for the differential entropy. We obtain the first uniform upper bound on its performance over Hölder balls on a torus without assuming any conditions on how close the density could be from zero. Accompanying a new minimax lower bound over the Hölder ball, we show that the KL estimator is achieving the minimax rates up to logarithmic factors without cognizance of the smoothness parameter s of the Hölder ball for s ∈ (0, 2] and arbitrary dimension d, rendering it the first estimator that provably satisfies this property. 1 Introduction Information theoretic measures such as information entropy, Kullback-Leibler divergence and mutual information quantify the amount of information among random variables. They have many applications in modern machine learning tasks, such as classification [PLD05], clustering [MNL12, VSG14, CABE+ 15, LM17] and feature selection [Bat94, Fle04]. Information theoretic measures and their variants can also be applied in several data science domains such as causal inference [GKOV16], sociology [RRF+ 11] and computational biology [KSM+ 14]. Estimating information theoretic measures from data is a crucial sub-routine in the aforementioned applications and has attracted many interest in statistics community. In this paper, we study the problem of estimating Shannon differential entropy, which is the basis of estimating other information theoretic measures for continuous random variables. Suppose we observe n independent identically distributed random variables X = {X1 , . . . , Xn } drawn from density function f where Xi ∈ Rd . We consider the problem of estimating the differential entropy R h(f ) = − f (x) ln f (x)dx , (1) from the empirical observations X. The fundamental limit of estimating the differential entropy is given by the minimax risk ∗ 1/2  , R(n, F) = inf ĥ supf ∈F E(ĥ(X) − h(f ))2 (2) Department of Electrical Engineering, Stanford University, email: {jiantao,yjhan}@stanford.edu Department of Electrical and Computer Engineering, Coordinated Science Laboratory, University of Illinois at Urbana-Champaign, email: [email protected] † 1 where the infimum is taken over all estimators ĥ that is a function of the empirical data X. F denotes a (nonparametric) class of density functions. The problem of differential entropy estimation has been investigated extensively in the literature. As discussed in [BDGVdM97], there exist two main approaches, where one is based on kernel density estimators [KKP+ 15], and the other is based on the nearest neighbor methods [TVdM96, SRH12, SP16, DF17, BSY16], which is pioneered by the work of [KL87]. The problem of differential entropy estimation lies in the general problem of estimating nonparametric functionals. Unlike the parametric counterparts, the problem of estimating nonparametric functionals is challenging even for smooth functionals. Initial efforts have focused on inference of linear, quadratic, and cubic functionals in Gaussian white noise and density models and have laid the foundation for the ensuing research. We do not attempt to survey the extensive literature in this area, but instead refer to the interested reader to, e.g., [HM87, BR88, DN90, Fan91, BM95, KP96, Lau96, Nem00, CL03, CL05, TLRvdV08] and the references therein. For non-smooth functionals such as entropy, there is some recent progress [LNS99, HJ+ 17, HJMW17] on designing theoretically minimax optimal estimators, while these estimators typically require the knowledge of the smoothness parameters, and the practical performances of these estimators are not yet known. The nearest neighbor differential entropy estimator, or Kozachenko-Leonenko (KL) estimator is computed in the following way. Let Ri be the distance between Xi and its nearest neighbor among {X1 , . . . , Xi−1 , Xi+1 , . . . , Xn }. Let B(x, ρ) denote the closed ℓ2 ball centered at x of radius ρ and λ be the Lebesgue measure on Rd . The KL differential entropy estimator is defined as d Pn (3) ĥKL n (X) = γ + ln n + ln Vd + n i=1 ln Ri 1 Pn = γ + n i=1 ln (nλ(B(Xi , Ri ))) , (4) R ∞ −t where γ = − 0 e ln tdt = 0.5772156 . . . is the Euler–Mascheroni constant and Vd = π d/2 /Γ(1 + d/2) be the Lebesgue measure of the unit ball in Rd . There exist an intuitive explanation behind the construction of the KL differential entropy estimator. Writing informally, we have n h(f ) = Ef [− ln f (X)] ≈ n 1X 1X − ln f (Xi ) ≈ − ln fˆ(Xi ), n n i=1 (5) i=1 where the first approximation is based on the law of large numbers, and in the second approximation we have replaced f by a nearest neighbor density estimator fˆ. The nearest neighbor density estimator fˆ(Xi ) follows from the intuition that 1 fˆ(Xi )λ(B(Xi , Ri )) ≈ . n (6) Here the final additive bias correction term γ follows from a detailed analysis of the bias of the KL estimator, which will become apparent later. There exist extensive literature on the analysis of the KL differential entropy estimator, which we refer to [BD15] for a recent survey. One of the major difficulties in analyzing the KL estimator is that the nearest neighbor density estimator exhibits a huge bias when the density is small. Indeed, it was shown in [MR79] that the bias of the nearest neighbor density estimator in fact does not vanish even when n → ∞ and deteriorates as f (x) gets close to zero. In the literature, a large collection of work assume that the density is uniformly bounded away from zero [Hal84, Joe89, VE92, KKP+ 15, SRH12], while others put various assumptions quantifying on average how close the density is to zero [HM93, Lev78, TVdM96, EHHG09, GOV16, SP16, DF17]. In this paper, we focus on removing assumptions on how close the density is to zero. 2 1.1 Main Contribution Let Hds (L; [0, 1]d ) be the Hölder ball in the unit cube (formally defined later in Definition 2) and s ∈ (0, 2] is the Hölder smoothness parameter. Then, the worst case risk of the nearest neighbor differential entropy estimator over Hds (L; [0, 1]d ) is controlled by the following theorem. Theorem 1. Let X = {X1 , . . . , Xn } be i.i.d. samples from density function f . Then, for 0 < s ≤ 2, the KL differential entropy estimator ĥKL n in (3) satisfies sup f ∈Hsd (L;[0,1]d ) Ef  2 ĥKL (X) − h(f ) n !1 2   s 1 ≤ C n− s+d (ln n) + n− 2 . (7) where C is a constant depends only on s, L and d. The KL estimator is in fact nearly minimax up to logarithmic factors: we quote the following theorem from [HJ+ 17]. Theorem 2. Let X = {X1 , . . . , Xn } be i.i.d. samples from density function f . Then, there exists a constant L0 depending on s, d only such that for all L ≥ L0 , s > 0, inf sup ĥ f ∈Hsd (L;[0,1]d )  2 Ef ĥ(X) − h(f ) !1 2   s+2d s 1 ≥ c n− s+d (ln n)− s+d + n− 2 . (8) where c is a constant depends only on s, L and d. Remark 1. We emphasize that one cannot remove the condition L ≥ L0 in Theorem 2. Indeed, if the Hölder ball has a too small width, then the density itself is bounded away from zero, which 4s makes the differential entropy a smooth functional, with minimax rates n− 4s+d +n−1/2 [RLTvdV08, RLM+ 16, MNR17]. Theorem 1 and 2 imply that the KL estimator achieves the minimax rates up to logarithmic factors without knowing s for all s ∈ (0, 2], which implies that it is near minimax rate-optimal (within logarithmic factors) when the dimension d ≤ 2. We cannot expect the vanilla version of the KL estimator to adapt to higher order of smoothness since the nearest neighbor density estimator can be viewed as a variable width kernel density estimator with the box kernel, and it is well known in the literature (see, e.g., [Tsy08, Chapter 1]) that any positive kernel cannot exploit the smoothness s > 2. We refer to [HJ+ 17] for a more detailed discussion on this difficulty. The significance of our work is multi-folded: • We obtain the first uniform upper bound on the performance of the KL differential entropy estimator over Hölder balls without assuming how close the density could be from zero. We emphasize that assuming conditions of this type, such as the density is bounded away from zero, could make the problem significantly easier. For example, if the density f is assumed to satisfy f (x) ≥ c for some constant c > 0, then the differential entropy becomes a smooth functional and consequently, the general technique for estimating smooth nonparametric functionals [RLTvdV08, RLM+ 16, MNR17] can be directly applied here to achieve the minimax 4s rates n− 4s+d + n−1/2 . The main technical tools that enabled us to remove the conditions on how close the density could be from zero are the Besicovith covering lemma and the generalized Hardy–Littlewood maximal inequality. In a recent paper [DF17], an assumption based 3 on Hardy–Littlewood maximal function was considered instead of simply assuming the density is away from zero. Our paper is different from [DF17] since we completely removed the assumption on maximal function and make the result potentially easier to generalize to other problems. • We show that the KL entropy estimator nearly achieves the minimax rates without knowing the smoothness parameter s. In the functional estimation literature, designing estimators that can be theoretically proved to adapt to unknown levels of smoothness is usually achieved using the Lepski method [Lep92, GN08, MTTR15, MTR16, HJMW17], which is not known to be performing well in general in practice. On the other hand, a simple plug-in approach can achieves the rate of n−s/(s+d) , but only when s is known [HJ+ 17]. The KL estimator is well known to exhibit excellent empirical performance, but existing theory has not yet demonstrated its near-“optimality” when the smoothness parameter s is not known. Our work makes a step towards closing this gap and provides a theoretical explanation for the wide usage of the KL estimator in practice. 1.2 Notations For non-negative sequences aγ , bγ , we use the notation aγ .α bγ to denote that there exists a a universal constant C that only depends on α such that supγ bγγ ≤ C, and aγ &α bγ is equivalent to bγ .α aγ . Notation aγ ≍α bγ is equivalent to aγ .α bγ and bγ .α aγ . We write aγ . bγ if the constant is universal and does not depend on any parameters. Notation aγ ≫ bγ means a that lim inf γ bγγ = ∞, and aγ ≪ bγ is equivalent to bγ ≫ aγ . We write a ∧ b = min{a, b} and a ∨ b = max{a, b}. 1.3 Organizations The rest of the paper is organized as follows. • In Section 2, we formally present the definition of Hölder ball in the unit cube [0, 1]d . • In Section 3, we prove that the upper bound of the ℓ2 error is O(n−s/(s+d) (ln n) + n−1/2 ) by consider the bias and variance separately. We refer to [HJ+ 17] for the proof of the lower bound. • In Section 4, we discuss some related work and future directions. 2 Definition of Hölder Ball In order to define the Hölder ball in the unit cube [0, 1]d , we first review the definition of Hölder ball in Rd . Definition 1 (Hölder ball in Rd ). The Hölder ball Hds (L; Rd ) is specified by the parameters s > 0 (order of smoothness), d ∈ Z+ (dimension of the argument) and L > 0 (smoothness constant) and is as follows. A positive real s can be uniquely represented as s = k + α, (9) where k is a nonnegative integer and 0 < α ≤ 1. By definition, Hds (L; Rd ) is comprised of all k times continuously differentiable functions f : Rd 7→ R, 4 (10) with Hölder continuous, with exponent α and constant L, derivatives of order k: kD k f (x)[δ1 , . . . , δd ] − D k f (x′ )[δ1 , . . . , δd ]k ≤ Lkx − x′ kα kδkk , ∀x, x′ ∈ Rd , δ ∈ Rd . (11) Here k · k is the Euclidean norm on Rd , and D k f (x)[δ1 , . . . , δk ] is the k-th differential of f taken as a point x along the directions δ1 , . . . , δk : D k f (x)[δ1 , . . . , δk ] = ∂k ∂t1 ...∂tk f (x + t1 δ1 + . . . + tk δk ). (12) t1 =t2 =...=tk =0 In this paper, we consider functions that lie in Hölder balls in [0, 1]d . The Hölder ball in the compact set [0, 1]d is defined as follows. Definition 2 (Hölder ball in the unit cube). A function f : [0, 1]d 7→ R is said to belong to the Hölder ball Hds (L; [0, 1]d ) if and only if there exists another function f1 ∈ Hds (L; Rd ) such that f (x) = f1 (x), x ∈ [0, 1], (13) and f1 (x) is a 1-periodic function in each variable. Here Hds (L; [0, 1]d ) is introduced in Definition 1. In other words, ∀x ∈ Rd , 1 ≤ j ≤ d, f1 (x + ej ) = f1 (x), (14) where {ej : 1 ≤ j ≤ d} is the standard basis in Rd . Definition 2 has appeared in the literature [KKPW14]. It is motivated by the observations that sliding window kernel methods usually can not deal with the boundary effects without additional assumptions [KA05]. Indeed, near the boundary the sliding window kernel density estimator may have a significantly larger bias than that of the interior points. In the nonparametric statistics literature, it is usually assumed that the density has its value and all the derivatives vanishing at the boundary, and our assumptions are in fact weaker than that. 3 Proof of upper bound In this section, we will prove that   2  12 s 1 KL .s,L,d n− s+d (ln n) + n− 2 , E ĥn (X) − h(f ) (15) for any f ∈ Hds (L; [0, 1]d ) and s ∈ (0, 2]. The proof consists two parts: (i) the upper bound of the bias is O(n−s/(s+d) (ln n)); (ii) the upper bound of the variance is O(n−1 ). We show the proof of the bias in the main text and the proof of the variance is in Appendix A. Firstly, we introduce the following notation Z 1 µ(B(x, t)) f (u)du . (16) = ft (x) = λ(B(x, t)) Vd td u:|u−x|≤t Here µ is the probability measure specified by density function f on the torus, λ is the Lebesgue measure on Rd , and Vd = π d/2 /Γ(1 + d/2) is the Lebesgue measure of the unit ball in d-dimensional Euclidean space. Hence ft (x) is the average density of a neighborhood near x. We first state two main lemmas about ft (x) which will be used later in the proof. 5 Lemma 1. If f ∈ Hds (L; [0, 1]d ) for some 0 < s ≤ 2, then for any x ∈ [0, 1]d and t > 0, we have | ft (x) − f (x) | ≤ dLts , s+d (17) where ft (x) is defined in (16). Lemma 2. If f ∈ Hds (L; [0, 1]d ) for some 0 < s ≤ 2 and f (x) ≥ 0 for all x ∈ [0, 1]d , then for any x and any t > 0, we have  s/(s+d)   f (x) .s,L,d max ft (x), ft (x)Vd td , (18) where ft (x) is defined in (16). Furthermore, we can show that f (x) .s,L,d 1. We relegate the proof of Lemma 1 and Lemma 2 to Appendix. B. Now we investigate the bias of The following argument reduces the bias analysis of ĥKL n (X) to a function analytic problem. For notation simplicity, we introduce a new random variable X which is i.i.d. with {X1 , . . . , Xn } d and study ĥKL n+1 ({X1 , . . . , Xn , X}). For every x ∈ R , denote R(x) = min1≤i≤n,m∈Zd km + Xi − xk be the nearest neighbor distance of x on the torus. Then, ĥKL n (X). E[ĥKL n+1 ({X1 , . . . , Xn , X})] − h(f ) (19) = γ + E [ ln ( (n + 1)λ(B(X, R(X))) )] + E [ln f (X)]    f (X)λ(B(X, R(X))) + ( E [ ln ((n + 1)µ(B(X, R(X))) ) ] + γ ) = E ln µ(B(X, R(X)))   f (X) = E ln + ( E [ ln ((n + 1)µ(B(X, R(X))) ) ] + γ ) . fR(X) (X) (20) (21) (22) We first show that the second term E [ln ((n + 1)µ(B(X, R(X))))] + γ can be universally controlled regardless of the smoothness of f . Indeed, the random variable µ(B(X, R(X))) ∼ Beta(1, n) and it was shown in [BD15, Theorem 7.2] that there exists a universal constant C > 0 such that E [ln ((n + 1)µ(B(X, R(X))))] + γ Now, it left to show that for 0 < s ≤ 2,   f (X) E ln fR(X) (X) ≤ C . n s .s,L,d n− s+d ln(n). (23) (24) i h s f (X) .s,L,d n− s+d and We split our analysis into two parts. Section 3.1 shows that E ln R(X) f (X) i h s f (X) − s+d . n ln(n), which completes the proof. section 3.2 shows that E ln R(X) s,L,d f (X) 3.1 h i fR(X) (X) Upper bound on E ln f (X) By the fact that ln y ≤ y − 1 for any y > 0, we have     fR(X) (X) − f (X) fR(X) (X) ≤ E E ln f (X) f (X) Z  E[fR(x) (x)] − f (x) dx. = [0,1]d ∩{x:f (x)6=0} 6 (25) (26) Here the expectation is taken with respect to the randomness in R(x) = min1≤i≤n,m∈Zd km + Xi − xk, x ∈ Rd . Define function g(x; f, n) as   1 d g(x; f, n) = sup u ≥ 0 : Vd u fu (x) ≤ , (27) n then for any x ∈ [0, 1]d , we can split E[fR(x) (x)] − f (x) into three terms as E[fR(x) (x)] − f (x) = E[(fR(x) (x) − f (x))1(R(x) ≤ n−1/(s+d) )] (28) + E[(fR(x) (x) − f (x))1(n−1/(s+d) < R(x) ≤ g(x; f, n))] (29) + E[(fR(x) (x) − f (x))1(R(x) > g(x; f, n) ∨ n−1/(s+d) )] (30) = C1 + C2 + C3 . (31) Now we handle three terms separately. Our goal is to show that for every x ∈ [0, 1], Ci .s,L,d n−s/(s+d) for i ∈ {1, 2, 3}. Then, taking the integral with respect to x leads to the desired bound. 1. Term C1 : whenever R(x) ≤ n−1/(s+d) , by Lemma 1, we have |fR(x) (x) − f (x)| ≤ dLR(x)s .s,L,d n−s/(s+d) , s+d (32) which implies that h C1 ≤ E fR(x) (x) − f (x) i 1(R(x) ≤ n−1/(s+d) ) .s,L,d n−s/(s+d) . (33) 2. Term C2 : whenever R(x) satisfies that n−1/(s+d) < R(x) ≤ g(x; f, n), by definition of g(x; f, n), we have Vd R(x)d fR( x) (x) ≤ n1 , which implies that fR(x) (x) ≤ 1 1 .s,L,d n−s/(s+d) . ≤ d −d/(s+d) nVd R(x) nVd n (34) It follows from Lemma 2 that in this case s/(s+d)  f (x) .s,L,d fR(x) (x) ∨ fR(x) (x)Vd R(x)d ≤ n−s/(s+d) ∨ n−s/(s+d) = n−s/(s+d) . (35) (36) Hence, we have C2 = ≤ h  i E (fR(x) (x) − f (x))1 n−1/(s+d) < R(x) ≤ g(x; f, n) h  i E (fR(x) (x) + f (x))1 n−1/(s+d) < R(x) ≤ g(x; f, n) .s,L,d n−s/(s+d) . (37) (38) (39) 3. Term C3 : we have i  h C3 ≤ E (fR(x) (x) + f (x))1 R(x) > g(x; f, n) ∨ n−1/(s+d) . (40) fR(x) (x) .s,L,d Vd R(x)d fR(x) (x)nd/(s+d) , (41) For any x such that R(x) > n−1/(s+d) , we have 7 and by Lemma 2, f (x) .s,L,d fR(x) (x) ∨ (Vd R(x)d fR(x) (x))s/(s+d) d ≤ s/(s+d) fR(x) (x) + (Vd R(x) fR(x) (x)) (42) . (43) Hence, f (x) + fR(x) (x) .s,L,d 2fR(x) (x) + (Vd R(x)d fR(x) (x))s/(s+d) .s,L,d Vd R(x)d fR(x) (x)nd/(s+d) + (Vd R(x)d fR(x) (x))s/(s+d) d .s,L,d Vd R(x) fR(x) (x)n d/(s+d) , (44) (45) (46) where in the last step we have used the fact that Vd R(x)d fR(x) (x) > n−1 since R(x) > g(x; f, n). Finally, we have C3 .s,L,d nd/(s+d) E[(Vd R(x)d fR(x) (x))1(R(x) > g(x; f, n))] h  i = nd/(s+d) E (Vd R(x)d fR(x) (x))1 Vd R(x)d fR(x) (x) > 1/n . (47) (48) Note that Vd R(x)d fR(x) (x) ∼ Beta(1, n), and if Y ∼ Beta(1, n), we have E[Y 2 ] = 1 1 n + . 2. 2 2 (n + 1) (n + 2) (n + 1) n Notice that E[Y 1 (Y > 1/n)] ≤ nE[Y 2 ]. Hence, we have i h C3 .s,L,d nd/(s+d) n E (Vd R(x)d fR(x) (x))2 .s,L,d 3.2 nd/(s+d) n = n−s/(s+d) . n2 (49) (50) (51) h i f (X) Upper bound on E ln fR(X) (X) By splitting the term into two parts, we have # "Z   f (x) f (X) dx f (x) ln E ln = E fR(X) (X) fR(x) (x) [0,1]d ∩{x:f (x)6=0} Z  f (x) −s/(s+d) f (x) ln = E 1(fR(x) (x) > n )dx fR(x) (x) A Z  f (x) −s/(s+d) f (x) ln + E 1(fR(x) (x) ≤ n )dx fR(x) (x) A = C4 + C5 . here we denote A = C4 ≤ = + ≤ [0, 1]d (52) (53) (54) (55) ∩ {x : f (x) 6= 0} for simplicity of notation. For the term C4 , we have    f (x) − fR(x) (x) −s/(s+d) 1(fR(x) (x) > n )dx (56) f (x) E fR(x) (x) A "Z # (f (x) − fR(x) (x))2 E 1(fR(x) (x) > n−s/(s+d) )dx (57) fR(x) (x) A Z   −s/(s+d) f (x) − fR(x) (x) 1(fR(x) (x) > n E )dx (58) A   Z Z  2 s/(s+d) f (x) − fR(x) (x) dx . (59) f (x) − fR(x) (x) dx + E n E Z A A 8 h f i (X) In the proof of upper bound of E ln R(X) , we have shown that E[fR(x) (x) − f (x)] .s,L,d f (X) i h f (X) , we have n−s/(s+d) for any x ∈ A. Similarly as in the proof of upper bound of E ln R(X) f (X)   E (fR(x) (x) − f (x))2 .s,L,d n−2s/(s+d) for every x ∈ A. Therefore, we have C4 .s,L,d ns/(s+d) n−2s/(s+d) + n−s/(s+d) . n−s/(s+d) . (60) Now we consider C5 . We conjecture that C5 .s,L,d n−s/(s+d) in this case, but we were not able to prove it. Below we prove that C5 .s,L,d n−s/(s+d) ln(n). Define function M (x) = sup t>0 1 . ft (x) (61) Since fR(x) (x) ≤ n−s/(s+d) , we have M (x) = supt>0 (1/ft (x)) ≥ 1/fR(x) (x) ≥ ns/(s+d) . Denote ln+ (y) = max{ln(y), 0} for any y > 0, therefore, we have that   Z  f (x) + −s/(s+d) f (x) ln 1(fR(x) (x) ≤ n )dx (62) C5 ≤ E fR(x) (x) A Z    f (x) + s/(s+d) f (x) ln ≤ E 1(M (x) ≥ n )dx (63) fR(x) (x) A    Z 1 + 1(M (x) ≥ ns/(s+d) )dx (64) f (x)E ln ≤ (n + 1)Vd R(x)d fR(x) (x) A Z  i h (65) f (x)E ln+ (n + 1)Vd R(x)d f (x) 1(M (x) ≥ ns/(s+d) )dx + A = C51 + C52 , (66) where the last inequality uses the fact ln+ (xy) ≤ ln+ x + ln+ y for all x, y > 0. As for C51 , since Vd R(x)d fR(x) (x) ∼ Beta(1, n), and for Y ∼ Beta(1, n), the density of Y is n(1 − y)n−1 , therefore  + E ln  1 (n + 1)Y  = = ≤ ≤ ≤ 1 n+1   1 ln n(1 − x)n−1 dx (n + 1)x 0    Z 1 1 ln ((n + 1)x) n(1 − x)n−1 dx E ln + 1 (n + 1)Y n+1    Z 1 1 E ln + ln(n + 1) n(1 − x)n−1 dx 1 (n + 1)Y n+1    1 1 n ) E ln + ln(n + 1)(1 − (n + 1)Y n+1 ψ(n + 1) − ψ(1) − ln(n + 1) + ln(n + 1) Z (67) (68) (69) (70) (71) = ψ(n + 1) − ψ(1) . ln(n), (72) i h  1 = ψ(n + 1) − ψ(1) − ln(n + 1) where in the last inequality we used the fact that E ln (n+1)Y where ψ(·) is the digamma function. Hence, Z f (x)1(M (x) ≥ ns/(s+d) )dx. C51 .s,L,d ln(n + 1) (73) A Now we introduce the following lemma. 9 Lemma 3. Let µ1 , µ2 be two Borel measures that are finite on the bounded Borel sets of Rd . Then, for all t > 0 and any Borel set A ⊂ Rd , )! (   Cd µ2 (B(x, ρ)) µ2 (AD ). (74) >t ≤ x ∈ A : sup µ1 t 0<ρ≤D µ1 (B(x, ρ)) Here Cd > 0 is a constant that depends only on the dimension d and AD = {x : ∃y ∈ A, |y − x| ≤ D}. (75) Applying the second part of Lemma 3 with µ2 being the Lebesgue measure and µ1 being the measure specified by f (x) on the torus, we can view the function M (x) as M (x) = µ2 (B(x, ρ)) . 0<ρ≤1/2 µ1 (B(x, ρ)) sup (76) Taking A = [0, 1]d ∩ {x : f (x) 6= 0}, t = ns/(s+d) , then µ2 (A 1 ) ≤ 2d , so we know that 2 Z f (x)1(M (x) ≥ ns/(s+d) )dx   ln(n) · µ1 x ∈ [0, 1]d , f (x) 6= 0, M (x) ≥ ns/(s+d) C51 .s,L,d ln(n) · (77) A = ≤ ln(n) · Cd n−s/(s+d) µ2 (A 1 ) .s,L,d n−s/(s+d) ln(n). 2 (78) (79) Now we deal with C52 . Recall that in Lemma 2, we know that f (x) .s,L,d 1 for any x, and R(x) ≤ 1, so ln+ ((n + 1)Vd R(x)d f (x)) .s,L,d ln(n). Therefore, Z (80) f (x)1(M (x) ≥ ns/(s+d) )dx C52 .s,L,d ln(n) · .s,L,d n A −s/(s+d) ln(n). (81) Therefore, we have provedhthat C5 ≤iC51 + C52 .s,L,d n−s/(s+d) ln(n), which completes the proof of the upper bound on E ln f f (X)(X) . R(X) 3.3 Proof of Lemma 3 We first introduce the Besicovitch covering lemma, which plays a crucial role in the analysis of nearest neighbor methods. Lemma 4. [EG15, Theorem 1.27][Besicovitch covering lemma] Let A ⊂ Rd , and suppose that {Bx }x∈A is a collection of balls such that Bx = B(x, rx ), rx > 0. Assume that A is bounded or that supx∈A rx < ∞. Then there exist an at most countable collection of balls {Bj } and a constant Cd depending only on the dimension d such that [ X (82) A⊂ Bj , and χBj (x) ≤ Cd . j j Here χB (x) = 1(x ∈ B). 10 Now we are ready to prove the lemma. Let   µ2 (B(x, ρ)) . M (x) = sup 0<ρ≤D µ1 (B(x, ρ)) (83) Let Ot = {x ∈ A : M (x) > t}. Hence, for all x ∈ Ot , there exists Bx = B(x, rx ) such that µ2 (Bx ) > tµ1 (Bx ), 0 < rx ≤ D. It follows from the Besicovitch lemma applying to the set Ot that there exists a set E ⊂ Ot , which has at most countable cardinality, such that [ X Ot ⊂ Bj , and χBj (x) ≤ Cd . (84) j∈E j∈E Let AD = {x : ∃y ∈ A, |y − x| ≤ D}, therefore Bj ⊂ AD for every j. Then, 1X µ2 (Bj ) t j∈E j∈E Z Z X 1 1X Cd χBj dµ2 = = χBj dµ2 ≤ µ2 (AD ). t t t AD AD µ1 (Ot ) ≤ X µ1 (Bj ) < j∈E 4 (85) j∈E Discussions It is an tempting question to ask whether one can close the logarithmic gap between Theorem 1 and 2. We believe that neither the upper bound nor the lower bound are tight. In fact, we s conjecture that the upper bound in Theorems 1 could be improved to n− s+d + n−1/2 , while the lower bound could be improved to (n ln n)− s+d + n−1/2 . We believe that the logarithmic gap is inherent in the KL estimator, since the general phenomenon of effective sample size enlargement demonstrated in [VV11, JVHW15, WY16, HJW16, JHW16, JVHW17] shows that the performance of the minimax rate-optimal estimators with n samples is usually that of the plug-in estimator with n ln n samples, while the KL estimator could be viewed as a plug-in kernel density estimator with variable width. A kernel density estimator baseds differential entropy estimator was constructed in [HJ+ 17] which achieves upper bound (n ln n)− s+d + n−1/2 over Hds (L; [0, 1]d ). One future interesting direction is to analyze the minimax risk of nearest neighbor KullbackLeibler divergence estimator using the Rsimilar technique. The Kullback-Leibler divergence, or relative entropy, defined as DKL (p||q) = p(x) ln(p(x)/q(x))dx, is used as an asymmetric distance of two distributions and a training objective for machine learning. Nearest neighbor estimators of divergence from samples have been proposed in [WKV09]. Recently, the minimax rate-optimal estimator of divergence has been studied in [HJW16, BZLV16]. Analyzing the nearest neighbor divergence estimator and its minimax optimality is a possible future direction. Another related problem is the minimax risk for nearest neighbor mutual information estimator. Mutual information, defined as I(X; Y ) = DKL (PXY ||PX PY ), is widely used in many machine learning and data science applications. Mutual information can be written as I(X; Y ) = h(X) + h(Y ) − h(X, Y ) if the distribution is purely continuous, therefore, one can apply differential entropy estimator three times to obtain an estimator for mutual information. However, the stateof-the-art nearest neighbor estimator is the KSG estimator proposed in [KSG04], by adaptively choose the hyperparameter k to cancel the distance terms. The convergence rate of KSG estimator has been studied recently in [GOV17], but closing the gap of the upper bound and lower bound is still an open problem. Also, KSG estimator is proved to be consistent even if the underlying distribution is not purely continuous nor purely discrete [GKOV17], but the convergence rate in this case is still missing. 11 Acknowledgement We are grateful to Gérard Biau, Shashank Singh, and Terence Tao for insightful discussions. A Proof of variance In this section, we will prove that   1 Var ĥKL (X) .d . n n (86) The prove utilizes Efron-Stein inequality to control the variance. To do so, for any i = 2, . . . , n, define Ri′ = min j6=i,j6=1,m∈Zd kXj − Xi − mk , (87) be the distance between Xi and its nearest neighbor among all the samples except X1 . Define n ĥKL n−1 (X) = γ + 1X ln(nλ(B(Xi , Ri′ ))) . n (88) i=2 Let τi ∈ {1, . . . , n} be the index of the nearest neighbor of Xi . Note that Ri′ = Ri if τi 6= 1 and i 6= 1.PAs shown in [BD15, Section 7.2], τi is independent of Ri or Ri′ with P (τi = 1) = 1/(n − 1), and ni=2 τi ≤ βd for a constant βd only depends on d. Therefore, follow the analysis in [BD15, Section 7.2], we have   Var ĥKL (X) n  2  KL KL ≤ 2nE ĥn (X) − ĥn−1 (X) (89) = 2nE + n X h1 ln(nλ(B(X1 , R1 ))) n2 i=2 ≤  2 i 1{τi = 1} ln(nλ(B(Xi , Ri ))) − ln(nλ(B(Xi , Ri′ ))) ! n X 2 h E 1+2 1{τi = 1} ln2 (nλ(B(X1 , R1 ))) n (90) i=2 + n X 1{τi = 1} ln (nλ(B(Xi , Ri ))) + 2 i=2 ≤ +  2 + 4βd   2 E ln (nλ(B(X1 , R1 ))) + n n X i=2 = n X i=2 n X i=2   P (τi = 1)E ln2 (nλ(B(Xi , Ri′ ))) i 1{τi = 1} ln2 (nλ(B(Xi , Ri′ ))) (91)   P (τi = 1)E ln2 (nλ(B(Xi , Ri )))  2 + 4βd  2  4 + 8βd  2 E ln (nλ(B(X1 , R1 ))) + E ln (nλ(B(X1 , R1′ ))) . n n (92) (93) Here the first inequality uses Efron-Stein inequality and the second inequality uses Cauchy-Schwarz    inequality. Now we prove that E ln2 (nλ(B(X1 , R1 ))) .d 1 and E ln2 (nλ(B(X1 , R1′ ))) .d 1. 12 Using Cauchy-Schwarz inequality, we have      2   2  2 λ(B(X1 , R1 )) ) + E ln (nµ(B(X1 , R1 ))) , (94) E ln (nλ(B(X1 , R1 ))) ≤ 2 E ln ( µ(B(X1 , R1 ))    ′  2    ′ 2 λ(B(X1 , R1 )) E ln (nλ(B(X1 , R1 ))) ≤ 3 E ln ( ) + E ln2 ((n − 1)µ(B(X1 , R1′ ))) ′ µ(B(X1 , R1 )) n  2 + ln ( ) . (95) n−1 ′ Since n − 1), therefore we know that both   2 µ(B(X1 , R1 )) ∼ Beta(1, n) 2and µ(B(X1 , R1 )) ∼′ Beta(1, 2 E ln (nµ(B(X1 , R1 ))) and E ln ((n − 1)µ(B(X1 , R1 ))) equal to certain h constants. lni (n/(n−1)) λ(B(X1 ,R1 )) ) .d 1 and is smaller than ln2 2 for n ≥ 2. So we only need to prove that E ln2 ( µ(B(X 1 ,R1 )) i h ′ )) λ(B(X ,R E ln2 ( µ(B(X11 ,R1′ )) ) .d 1. Recall that we have defined the maximal function as follows, 1 M (x) = λ(B(x, r)) . µ(B(x, r)) 0≤r≤1/2 (96) m(x) = µ(B(x, r)) . 0≤r≤1/2 λ(B(x, r)) (97) sup Similarly, we define sup Therefore,     λ(B(X1 , R1 )) ) ≤ E max{ln2 (M (x)), ln2 (m(x))} E ln2 ( µ(B(X1 , R1 ))   ≤ E ln2 (M (x) + 1) + ln2 (m(x) + 1)     = E ln2 (M (x) + 1) + E ln2 (m(x) + 1) . Similarly this inequality holds if we replace R1 by R1′ . By Lemma 3, we have Z   ln2 (M (x) + 1)dµ(x) E ln2 (M (x) + 1) = d [0,1] Z ∞ n o µ x ∈ [0, 1]d : ln2 (M (x) + 1) > t dt = Z0 ∞ n o √ µ x ∈ [0, 1]d : M (x) > e t − 1 dt = Z0 ∞ 1 √ dt .d 1. .d t e −1 0 For E[ln2 (m(x) + 1)], we rewrite the term as Z   f (x) ln2 (m(x) + 1)dx E ln2 (m(x) + 1) = d [0,1] Z ∞ n o λ( x ∈ [0, 1]d : f (x) ln2 (m(x) + 1) > t dt . = (98) (99) (100) (101) (102) (103) (104) (105) (106) t=0 Notice that {f (x) ln2 (m(x) + 1) > t} implies either m(x) > t2 or f (x) > t/ ln2 (t2 + 1). Moreover, since ln t − 2 ln ln(t2 + 1) 1 1 ln(t2 ) ln ln(t2 + 1) = lim − 2 lim = , 2 2 2 t→∞ t→∞ t→∞ ln(t + 1) 2 ln(t + 1) ln(t + 1) 2 lim 13 (107) so there exists a t0 such that ln t−2 ln ln(t2 +1) ln(t2 +1) > 1/3 and t/ ln2 (t2 + 1) > 1 for every t > t0 . Then for every t > t0 , f (x) > t/ ln2 (t2 + 1) implies that f (x) ln2 (f (x)) = t t(ln t − 2 ln ln(t2 + 1))2 > . 2 2 9 ln (t + 1) (108) For every t < t0 , f (x) > t/ ln2 (t2 + 1) implies f (x) > t/ ln2 (t20 + 1) = t/C for some constant C. Therefore, {f (x) ln2 (m(x) + 1) > t} implies at least one of the following three statement holds: (i) {m(x) > t2 }, (ii) {f (x) ln2 f (x) > t/9} or (iii) {f (x) > t/C}. Therefore, Z ∞ n o λ( x ∈ [0, 1]d : f (x) ln2 (m(x) + 1) > t dt (109) t=0 Z ∞ n Z ∞ n o o λ( x ∈ [0, 1]d : f (x) ln2 f (x) > t/9 dt λ( x ∈ [0, 1]d : m(x) > t2 dt + ≤ t=0 Zt=0 n o ∞ λ( x ∈ [0, 1]d : f (x) > t/C dt (110) + t=0 Z ∞ Z Z 1 1 1 .d + f (x) ln2 f (x)dx + f (x)dx (111) 2 9 C t=0 t .d 1. (112) Hence, the proof is completed. B Proof of lemmas In this section we provide the proofs of the lemmas used in this paper. B.1 Proof of Lemma 1 We consider the cases s ∈ (0, 1] and s ∈ (1, 2] separately. For s ∈ (0, 1], following the definition of Hölder smoothness, we have, Z 1 (113) f (u)du − f (x) | ft (x) − f (x) | = Vd td u:||u−x||≤t Z 1 |f (u) − f (x)|du (114) ≤ Vd td u:ku−xk≤t Z 1 ≤ Lku − xks du . (115) Vd td u:ku−xk≤t By denoting ρ = ku − xk and considering θ ∈ S d−1 on the unit d-dimensional sphere, we rewrite the above integral using polar coordinate system and obtain, Z t Z 1 | ft (x) − f (x) | = Lρs ρd−1 dρdθ (116) Vd td ρ=0 θ∈S d−1 Z t 1 = dVd Lρs+d−1 dρ (117) Vd td ρ=0 = dLts dVd Lts+d = . (s + d)Vd td s+d 14 (118) Now we consider the case s ∈ (1, 2]. Now we rewrite the difference as Z 1 f (u)du − f (x) | ft (x) − f (x) | = Vd td u:ku−xk≤t Z 1 = ( f (x + v) + f (x − v) ) dv − f (x) 2Vd td v:kvk≤t Z 1 ≤ f (x + v) + f (x − v) − 2f (x) dv . 2Vd td v:kvk≤t (119) (120) (121) For fixed v, we bound |f (x + v) + f (x − v) − 2f (x)| using the Gradient Theorem and the definition of Hölder smoothness as follows, |f (x + v) + f (x − v) − 2f (x)| = = = = ≤ Z ( f (x + v) − f (x) ) + ( f (x − v) − f (x) ) Z −1 Z 1 ∇f (x + αv) · d(x + αv) ∇f (x + αv) · d(x + αv) + α=0 α=0 Z 1 Z 1 ( ∇f (x − αv) · v ) dα ( ∇f (x + αv) · v ) dα − α=0 α=0 Z 1 ( ∇f (x + αv) − ∇f (x − αv) ) · vdα (123) (124) (125) (126) α=0 1 k ∇f (x + αv) − ∇f (x − αv)kkvkdα α=0 Z 1 Lk2αvks−1 kvkdα α=0 Z 1 Lkvks 2s−1 s . = Lkvk (2α)s−1 dα = s 0 ≤ (122) Plug it into (121) and using the similar method in the s ∈ (0, 1] case, we have Z 1 Lkvks 2s−1 | ft (x) − f (x) | ≤ dv d 2Vd t v:kvk≤t s Z t Z 1 Lρs 2s−1 d−1 = ρ dρdθ 2Vd td ρ=0 θ∈S d−1 s Z t dVd Lρs+d−1 2s−1 1 dρ = s 2Vd td ρ=0 = dLts 1 dVd L2s−1 ts+d ≤ , 2Vd td s s+d s+d (127) (128) (129) (130) (131) (132) (133) where the last inequality uses the fact that s ∈ (1, 2]. B.2 Proof of Lemma 2 We consider the following two cases. If f (x) ≥ 2dLts /(s + d), then by Lemma 1, we have f (x) ≤ ft (x) + f (x) dLts ≤ ft (x) + . s+d 2 15 (134) Hence, f (x) ≤ 2ft (x) in this case. If f (x) < 2dLts /(s+d), then define t0 = (f (x)(s+d)/2dL)1/s < t. By the nonnegativity of f , we have = ≥ = = f (x)Vd td Z Zt f (x)dx = ft0 (x)Vd td0 f (x)dx ≥ B(x,t0 ) B(x,t)   dLts0 Vd td0 f (x) − s+d     f (x)(s + d) (s+d)/s dL f (x)(s + d) d/s − Vd f (x)Vd 2dL s+d 2dL d/s    s+d 2−d/s − 2−(s+d)/s . f (x)(s+d)/s Vd dL (135) (136) (137) (138) (139) Therefore, we have f (x) .s,L,d (ft (x)Vd td )s/(s+d) in this case. We obtain the desired statement by combining the two cases. Furthermore, by taking t = 1/2, we have Vd td ft (x) < 1, so ft (x) .s,L,d 1. By applying this lemma immediately we obtain f (x) .s,L,d 1. References [Bat94] R. Battiti. Using mutual information for selecting features in supervised neural net learning. Neural Networks, IEEE Transactions on, 5(4):537–550, 1994. [BD15] Gérard Biau and Luc Devroye. Lectures on the nearest neighbor method. Springer, 2015. [BDGVdM97] Jan Beirlant, Edward J Dudewicz, László Györfi, and Edward C Van der Meulen. Nonparametric entropy estimation: An overview. International Journal of Mathematical and Statistical Sciences, 6(1):17–39, 1997. [BM95] Lucien Birgé and Pascal Massart. Estimation of integral functionals of a density. The Annals of Statistics, pages 11–29, 1995. [BR88] Peter J Bickel and Yaacov Ritov. Estimating integrated squared density derivatives: sharp best order of convergence estimates. Sankhyā: The Indian Journal of Statistics, Series A, pages 381–393, 1988. [BSY16] Thomas B Berrett, Richard J Samworth, and Ming Yuan. Efficient multivariate entropy estimation via k-nearest neighbour distances. arXiv preprint arXiv:1606.00304, 2016. [BZLV16] Yuheng Bu, Shaofeng Zou, Yingbin Liang, and Venugopal V Veeravalli. Estimation of KL divergence between large-alphabet distributions. In 2016 IEEE International Symposium on Information Theory (ISIT), pages 1118–1122. IEEE, 2016. [CABE+ 15] C. Chan, A. Al-Bashabsheh, J. B. Ebrahimi, T. Kaced, and T. Liu. Multivariate mutual information inspired by secret-key agreement. Proceedings of the IEEE, 103(10):1883–1913, 2015. [CL03] T Tony Cai and Mark G Low. A note on nonparametric estimation of linear functionals. Annals of statistics, pages 1140–1153, 2003. 16 [CL05] T Tony Cai and Mark G Low. Nonquadratic estimators of a quadratic functional. The Annals of Statistics, pages 2930–2956, 2005. [DF17] Sylvain Delattre and Nicolas Fournier. On the kozachenko–leonenko entropy estimator. Journal of Statistical Planning and Inference, 185:69–93, 2017. [DN90] David L Donoho and Michael Nussbaum. Minimax quadratic estimation of a quadratic functional. Journal of Complexity, 6(3):290–323, 1990. [EG15] Lawrence Craig Evans and Ronald F Gariepy. Measure theory and fine properties of functions. CRC press, 2015. [EHHG09] Fidah El Haje Hussein and Yu Golubev. On entropy estimation by m-spacing method. Journal of Mathematical Sciences, 163(3):290–309, 2009. [Fan91] Jianqing Fan. On the estimation of quadratic functionals. The Annals of Statistics, pages 1273–1294, 1991. [Fle04] F. Fleuret. Fast binary feature selection with conditional mutual information. The Journal of Machine Learning Research, 5:1531–1555, 2004. [GKOV16] Weihao Gao, Sreeram Kannan, Sewoong Oh, and Pramod Viswanath. Conditional dependence via shannon capacity: Axioms, estimators and applications. In International Conference on Machine Learning, pages 2780–2789, 2016. [GKOV17] Weihao Gao, Sreeram Kannan, Sewoong Oh, and Pramod Viswanath. Estimating mutual information for discrete-continuous mixtures. In Advances in Neural Information Processing Systems, pages 5988–5999, 2017. [GN08] Evarist Giné and Richard Nickl. A simple adaptive estimator of the integrated square of a density. Bernoulli, pages 47–61, 2008. [GOV16] Weihao Gao, Sewoong Oh, and Pramod Viswanath. Breaking the bandwidth barrier: Geometrical adaptive entropy estimation. In Advances in Neural Information Processing Systems, pages 2460–2468, 2016. [GOV17] Weihao Gao, Sewoong Oh, and Pramod Viswanath. Demystifying fixed k-nearest neighbor information estimators. In Information Theory (ISIT), 2017 IEEE International Symposium on, pages 1267–1271. IEEE, 2017. [Hal84] Peter Hall. Limit theorems for sums of general functions of m-spacings. In Mathematical Proceedings of the Cambridge Philosophical Society, volume 96, pages 517–532. Cambridge University Press, 1984. [HJ+ 17] Yanjun Han, Jiantao Jiao, , Tsachy Weissman, and Yihong Wu. Optimal rates of entropy estimation over lipschitz balls. arXiv preprint arXiv:1711.02141, 2017. [HJMW17] Yanjun Han, Jiantao Jiao, Rajarshi Mukherjee, and Tsachy Weissman. On estimation of lr -norms in gaussian white noise models. arXiv preprint arXiv:1710.03863, 2017. [HJW16] Yanjun Han, Jiantao Jiao, and Tsachy Weissman. Minimax rate-optimal estimation of divergences between discrete distributions. arXiv preprint arXiv:1605.09124, 2016. 17 [HM87] Peter Hall and James Stephen Marron. Estimation of integrated squared density derivatives. Statistics & Probability Letters, 6(2):109–115, 1987. [HM93] Peter Hall and Sally C Morton. On the estimation of entropy. Annals of the Institute of Statistical Mathematics, 45(1):69–88, 1993. [JHW16] Jiantao Jiao, Yanjun Han, and Tsachy Weissman. Minimax estimation of the L1 distance. In 2016 IEEE International Symposium on Information Theory (ISIT), pages 750–754. IEEE, 2016. [Joe89] Harry Joe. Estimation of entropy and other functionals of a multivariate density. Annals of the Institute of Statistical Mathematics, 41(4):683–697, 1989. [JVHW15] Jiantao Jiao, Kartik Venkat, Yanjun Han, and Tsachy Weissman. Minimax estimation of functionals of discrete distributions. Information Theory, IEEE Transactions on, 61(5):2835–2885, 2015. [JVHW17] Jiantao Jiao, Kartik Venkat, Yanjun Han, and Tsachy Weissman. Maximum likelihood estimation of functionals of discrete distributions. IEEE Transactions on Information Theory, 63(10):6774–6798, 2017. [KA05] Rhoana J Karunamuni and Tom Alberts. On boundary correction in kernel density estimation. Statistical Methodology, 2(3):191–212, 2005. [KKP+ 15] Kirthevasan Kandasamy, Akshay Krishnamurthy, Barnabas Poczos, Larry Wasserman, et al. Nonparametric von Mises estimators for entropies, divergences and mutual informations. In Advances in Neural Information Processing Systems, pages 397–405, 2015. [KKPW14] Akshay Krishnamurthy, Kirthevasan Kandasamy, Barnabas Poczos, and Larry Wasserman. Nonparametric estimation of Rényi divergence and friends. In International Conference on Machine Learning, pages 919–927, 2014. [KL87] LF Kozachenko and Nikolai N Leonenko. Sample estimate of the entropy of a random vector. Problemy Peredachi Informatsii, 23(2):9–16, 1987. [KP96] Gérard Kerkyacharian and Dominique Picard. Estimating nonquadratic functionals of a density using haar wavelets. The Annals of Statistics, 24(2):485–507, 1996. [KSG04] Alexander Kraskov, Harald Stögbauer, and Peter Grassberger. Estimating mutual information. Physical Review E, 69(6):066138, 2004. [KSM+ 14] Smita Krishnaswamy, Matthew H Spitzer, Michael Mingueneau, Sean C Bendall, Oren Litvin, Erica Stone, Dana Pe’er, and Garry P Nolan. Conditional densitybased analysis of t cell signaling in single-cell data. Science, 346(6213):1250689, 2014. [Lau96] Béatrice Laurent. Efficient estimation of integral functionals of a density. The Annals of Statistics, 24(2):659–681, 1996. [Lep92] Oleg V Lepski. On problems of adaptive estimation in white gaussian noise. Topics in nonparametric estimation, 12:87–106, 1992. 18 [Lev78] Boris Ya Levit. Asymptotically efficient estimation of nonlinear functionals. Problemy Peredachi Informatsii, 14(3):65–72, 1978. [LM17] Pan Li and Olgica Milenkovic. Inhomogoenous hypergraph clustering with applications. arXiv preprint arXiv:1709.01249, 2017. [LNS99] Oleg Lepski, Arkady Nemirovski, and Vladimir Spokoiny. On estimation of the Lr norm of a regression function. Probability theory and related fields, 113(2):221–253, 1999. [MNL12] A. C. Müller, S. Nowozin, and C. H. Lampert. Information theoretic clustering using minimum spanning trees. Springer, 2012. [MNR17] Rajarshi Mukherjee, Whitney K Newey, and James M Robins. Semiparametric efficient empirical higher order influence function estimators. arXiv preprint arXiv:1705.07577, 2017. [MR79] YP Mack and Murray Rosenblatt. Multivariate k-nearest neighbor density estimates. Journal of Multivariate Analysis, 9(1):1–15, 1979. [MTR16] Rajarshi Mukherjee, Eric Tchetgen Tchetgen, and James Robins. On adaptive estimation of nonparametric functionals. arXiv preprint arXiv:1608.01364, 2016. [MTTR15] Rajarshi Mukherjee, Eric Tchetgen Tchetgen, and James Robins. Lepski’s method and adaptive estimation of nonlinear integral functionals of density. arXiv preprint arXiv:1508.00249, 2015. [Nem00] Arkadi Nemirovski. Topics in non-parametric. Ecole dEté de Probabilités de SaintFlour, 28:85, 2000. [PLD05] H. Peng, F. Long, and C. Ding. Feature selection based on mutual information criteria of max-dependency, max-relevance, and min-redundancy. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 27(8):1226–1238, 2005. [RLM+ 16] James Robins, Lingling Li, Rajarshi Mukherjee, Eric Tchetgen Tchetgen, and Aad van der Vaart. Higher order estimating equations for high-dimensional models. The Annals of Statistics (To Appear), 2016. [RLTvdV08] James Robins, Lingling Li, Eric Tchetgen, and Aad van der Vaart. Higher order influence functions and minimax estimation of nonlinear functionals. In Probability and Statistics: Essays in Honor of David A. Freedman, pages 335–421. Institute of Mathematical Statistics, 2008. [RRF+ 11] David N Reshef, Yakir A Reshef, Hilary K Finucane, Sharon R Grossman, Gilean McVean, Peter J Turnbaugh, Eric S Lander, Michael Mitzenmacher, and Pardis C Sabeti. Detecting novel associations in large data sets. science, 334(6062):1518–1524, 2011. [SP16] Shashank Singh and Barnabás Póczos. Finite-sample analysis of fixed-k nearest neighbor density functional estimators. In Advances in Neural Information Processing Systems, pages 1217–1225, 2016. 19 [SRH12] Kumar Sricharan, Raviv Raich, and Alfred O Hero. Estimation of nonlinear functionals of densities with confidence. IEEE Transactions on Information Theory, 58(7):4135–4159, 2012. [TLRvdV08] Eric Tchetgen, Lingling Li, James Robins, and Aad van der Vaart. Minimax estimation of the integral of a power of a density. Statistics & Probability Letters, 78(18):3307–3311, 2008. [Tsy08] A. Tsybakov. Introduction to Nonparametric Estimation. Springer-Verlag, 2008. [TVdM96] Alexandre B Tsybakov and EC Van der Meulen. Root-n consistent estimators of entropy for densities with unbounded support. Scandinavian Journal of Statistics, pages 75–83, 1996. [VE92] Bert Van Es. Estimating functionals related to a density by a class of statistics based on spacings. Scandinavian Journal of Statistics, pages 61–72, 1992. [VSG14] G. Ver Steeg and A. Galstyan. Maximally informative hierarchical representations of high-dimensional data. stat, 1050:27, 2014. [VV11] Gregory Valiant and Paul Valiant. The power of linear estimators. In Foundations of Computer Science (FOCS), 2011 IEEE 52nd Annual Symposium on, pages 403–412. IEEE, 2011. [WKV09] Qing Wang, Sanjeev R Kulkarni, and Sergio Verdú. Divergence estimation for multidimensional densities via k-nearest-neighbor distances. Information Theory, IEEE Transactions on, 55(5):2392–2405, 2009. [WY16] Yihong Wu and Pengkun Yang. Minimax rates of entropy estimation on large alphabets via best polynomial approximation. IEEE Transactions on Information Theory, 62(6):3702–3720, 2016. 20
7cs.IT
arXiv:1710.09490v1 [cs.CV] 25 Oct 2017 ZOU, LI, HOIEM: COMPLETE 3D SCENE PARSING FROM SINGLE RGBD IMAGE 1 Complete 3D Scene Parsing from Single RGBD Image Chuhang Zou http://web.engr.illinois.edu/~czou4/ Zhizhong Li http://zli115.web.engr.illinois.edu Department of Computer Science University of Illinois at Urbana-Champaign USA Derek Hoiem http://dhoiem.cs.illinois.edu 1 Abstract Inferring the location, shape, and class of each object in a single image is an important task in computer vision. In this paper, we aim to predict the full 3D parse of both visible and occluded portions of the scene from one RGBD image. We parse the scene by modeling objects as detailed CAD models with class labels and layouts as 3D planes. Such an interpretation is useful for visual reasoning and robotics, but difficult to produce due to the high degree of occlusion and the diversity of object classes. We follow the recent approaches that retrieve shape candidates for each RGBD region proposal, transfer and align associated 3D models to compose a scene that is consistent with observations. We propose to use support inference to aid interpretation and propose a retrieval scheme that uses convolutional neural networks (CNNs) to classify regions and retrieve objects with similar shapes. We demonstrate the performance of our method compared with the state-of-the-art on our new NYUd v2 dataset annotations which are semi-automatically labelled with detailed 3D shapes for all the objects. Introduction In this paper, we aim to predict the complete 3D models of indoor objects and layout surfaces from single RGBD images. The prediction is represented by detailed CAD objects with class labels and 3D planes of layouts. This interpretation is useful for robotics, graphics, and human activity interpretation, but difficult to produce due to the high degree of occlusion and the diversity of object classes. Our approach. We propose an approach to recover a 3D model of room layout and objects from an RGBD image. A major challenge is how to cope with the huge diversity of layouts and objects. Rather than restricting to a parametric model and a few detectable object classes, as in previous single-view reconstruction work, our models represent every layout surface and object with a 3D mesh that approximates the original depth image under projection. We take a data-driven approach that proposes a set of potential object regions, matches each region to a similar region in training images, and transfers and aligns the associated labeled 3D models while encouraging their agreement with observations. During the matching step, we use CNNs to retrieve objects of similar class and shape and further incorporate c 2017. The copyright of this document resides with its authors. It may be distributed unchanged freely in print or electronic forms. 2 ZOU, LI, HOIEM: COMPLETE 3D SCENE PARSING FROM SINGLE RGBD IMAGE ✔ ✗ Support CNNs RGB&depth CNNs Feature extraction Region Proposals Input RGB +Depth Classification CNNs Shape Retrieval CNNs ... ✔ Shape Candidates Scene Composition Figure 1: Overview of our approach. Given one RGBD image, our approach is to perform complete 3D parsing. We generate layout proposals and object proposals, predict each proposal’s support height and class and retrieve a similar object shape. We then select a subset of layout and object shapes to compose the scene with complete 3D interpretation. support estimation to aid interpretation. We hypothesize, and confirm in experiments, that support height information will help most for interpreting occluded objects because the full extent of an occluded object can be inferred from support height. The subset of proposed 3D objects and layouts that best represent the overall scene is then selected by our optimization method based on consistency with observed depth, coverage, and constraints on occupied space. The flexibility of our models is enabled through our approach (Fig. 1) to propose a large number of likely layout surfaces and objects and then compose a complete scene out of a subset of those proposals while accounting for occlusion, image appearance, depth, and layout consistency. Detailed 3D labeling. Our approach requires a dataset with labeled 3D shape for region matching, shape retrieval, and evaluation. We make use of the NYUd v2 dataset [18] which consists of 1449 indoor scene RGBD images, with each image segmented and labeled with object instances and categories. Each segmented object also has a corresponding annotated 3D model, provided by Guo and Hoiem [4]. The 3D labeling provides groundtruth 3D scene representation with layout surfaces as 3D planar regions, furniture as CAD exemplars, and other objects as coarser polygonal shapes. However, the polygonal shapes are too coarse to enable comparison of object shapes. Therefore, we extend the labeling by Guo and Hoiem with more detailed 3D annotations in the object scale. Annotations are labeled automatically and are adjusted manually as described in Sec. 3. We evaluate our method on our newly annotated groundtruth. We measure success according to accuracy of depth prediction of complete layout surfaces, voxel occupancy accuracy, and semantic segmentation performance. We evaluate our method on our newly annotated NYUd v2 dataset. Experiments on object retrieval demonstrate better region classification and shape estimation compared with the state-of-the-art. Our performance of scene composition shows better semantic segmentation results and competitive 3D estimation results compared with the state-of-the-art. Our contributions are: 1. We refine the NYUd v2 dataset with detailed 3D shape annotations for all the objects in each image. The labelling process is performed by a semi-automatic approach and can be utilized for labelling other RGBD scene datasets. 2. We apply support inference to aid region classification in images. 3. We use CNNs to classify regions and retrieve objects with similar shapes; 4. We demonstrate better performance in full 3D scene parsing from single RGBD images compared with the state-of-the-art. ZOU, LI, HOIEM: COMPLETE 3D SCENE PARSING FROM SINGLE RGBD IMAGE 3 This paper is an extension of our previous work [5] (available only on arxiv) that predicts full 3D scene parsing from an RGBD image. Our main new contributions are refinement of the NYUd v2 dataset with detailed 3D shape annotations, use of CNNs to classify regions and retrieve object models with similar shapes to a region, and use of support inference to aid region classification. We also provide more detailed discussion and conduct more extensive experiments, demonstrating qualitative and quantitative improvement. 2 Related work The most relavent work to our paper is Guo et al [5] which predicts the complete 3D models of indoor scene from single RGBD image as introduced above. Both our approach and Guo et al. recover complete models from a limited viewpoint, which is in contrast to the multiview whole-room 3D context interpretation method by Zhang et al. [24] that advocate making use of 360◦ full-view panoramas. We introduce other related topics as follows. Inferring shape from single RGB (D) Image. Within RGB images, Lim et al. [12, 13] find furniture instances and Aubry et al. [1] recognize chairs using HOG-based part detectors. In RGBD images, Song and Xiao [19] search for main furniture by sliding 3D shapes and enumerating all possible poses. Similarly, Gupta et al. [8] fit shape models of 6 classes with poses to improve object detection. Our approach finds an approximate shape for any object and layout in the scene. We take an exemplar-based approach, apply region-to-region retrieval to transfer similar 3D shape from training region to query region. Semantic segmentation for single RGBD Images. Silberman et al. [18] use both image and depth cues to jointly segment the objects into 4 categories and infer support relations. Gupta et al. [6] apply both generic and class-specific features to assign 40-class region labels. Their following work encodes both RGB and depth descriptor (HHA) [7]: height above ground, angle with gravity and the horizontal disparity into the CNNs structure for better region classification. Long et al. [15] introduce a fully convolutional network structure for learning better features. Our method make use of the RGB and HHA features to categorize objects into a larger variety of 81 classes to distinguish infrequent objects in the scenes, rather than restricting to a parametric model and a few detectable objects. Though semantic segmentation is not the main purpose of our method, we can infer region labels by projecting the 3D scene models to the 2D images. Support height estimation. Guo and Hoiem [4] localize the height and full extent of support surfaces from one RGBD image. In addition, object height priors have shown to be crucial geometric cues for better object detection in both 2D [10, 21] and 3D [8, 14, 19]. Deng et al. [3] applies height above ground to distinguish objects. We propose to use object’s support height to aid region class interpretation, which helps in classification of occluded regions by distinguishing objects that appear at different height levels: e.g. chair should be on the floor and alarm clock should be on the table. 3 Detailed 3D annotations for indoor scenes We conduct our experiments on the NYUdv2 dataset [18], which provides complete 3D labeling of both objects and layouts of 1449 RGB-D indoor images. Each object and layout has a 2D segment labeling and a corresponding annotated 3D model, provided by Guo and Hoiem [4]. The 3D annotations use 30 models to represent 6 categories of furniture 4 ZOU, LI, HOIEM: COMPLETE 3D SCENE PARSING FROM SINGLE RGBD IMAGE Input RGB (D) Annotations by Guo and Hoiem Our detailed annotations Figure 2: Samples of our semi-automatic 3D object annotations in NYUd v2 dataset. Images from first to third row: input RGB (D) image, 3D annotations by Guo and Hoiem [4], our refined 3D annotations. Our annotations of the scene is much detailed in object shape scale. that are most common and use extruded polygons to label all other objects. These models provide a good approximation of object extent but are often poor representations of object shape, as shown in Fig 2. Therefore, we extend the NYUd v2 dataset by replacing the extruded polygons with CAD models collected from ShapeNet [2] and ModelNet [22]. To align name-space between datasets, we manually map all model class labels to the 633-class 3D object labels in NYUd v2 dataset. The shape retrieval and alignment process is performed automatically and then adjusted manually, as follows. Coarse alignment. For each groundtruth 2D region ri in the NYUd v2 dataset, we retrieve model set M = {Mi } from our collected models that have the same class label as ri . We also include the region’s original coarse 3D annotation by Guo and Hoiem [4] in the model set M, so that we can preserve the original labeling if no provided CAD models are better fit in depth. We initialize each Mi ’s 3D location as the world coordinate center of the 3D annotation labeled by Guo and Hoiem. We resize Mi to have the same height as the 3D annotation. Fine alignment. Next, we align each retrieved 3D object model Mi to fit the available depth map of the corresponding 2D region ri in the target scene. The initial alignment is often not in the correct scale and orientation; e.g., a region of a left-facing chair often resembles a right-facing chair and needs to be rotated. We found that using Iterative Closest Point to solve for all parameters did not yield good results. Instead, we enumerate 16 equally-spaced orientations from -180 to 180 from top-down view and allows 2 minor scale revision ratio as {1.0, 0.9}. We perform ICP to solve for translation initialized using scale and rotation, and pick the best ICP result based on the following cost function: FittingCost(Mi , Ti ) = Cdepth ∑ ˆ j; Mi , Ti )| |Id ( j) − d( j∈ri ∩s(Mi ,Ti ) + Cmissing ∑ j∈ri ∩¬s(Mi ,Ti ) +Cocc ∑ ˆ j; Mi , Ti ) − Id ( j), 0) max(d( (1) j∈¬ri ∩s(Mi ,Ti ) where Ti represents scale, rotation, and translation, s(.) is the mask of the rendered aligned ˆ j) means the rendered depth at object, Id ( j) denotes the observed depth at pixel j and d( j. The first term encourages depth similarity to the groundtruth RGBD region; the second ZOU, LI, HOIEM: COMPLETE 3D SCENE PARSING FROM SINGLE RGBD IMAGE 5 penalizes pixels in the proposed region that are not rendered, and the third term penalizes pixels in the rendered model that are closer than the observed depth image (so the model does not stick out into space known to be empty). Based on the fitting cost of Eq. 1, our algorithm picks the model Mi with the best translation, orientation, and scale Ti . We set the term weights Cdepth , Cmissing , Cocc as 1.0, 0.9, 0.5 based on a grid search in the validation set. For efficiency, we first obtain the top 5 models based on the fitting cost, each maximized only over the 16 initial orientations before ICP. For each of these models, we then solve for the best translation Ti for each scale and rotation based on Eq. 1 and finally select the aligned model with the lowest fitting cost. Post-processing. Automatic fitting may fail due to high occlusion or missing depth values. We manually conduct a post-processing check and refine badfitting models, which affects roughly 10% of models. Using a GUI, an annotator checks the automatically produced shape for each region. If the result is not satisfactory, the user compares to other top model fits, and if none of those are good matches, then the fitting optimization based on Eq. 1 is applied to the original polygonal 3D labeling. This helps to ensure that our detailed shape annotations are a strict imFigure 3: Cumulative relative depth provement over the original course annotations. Validation. Figure 3 reports the cumulative rela- error of our detailed 3D annotations tive error of the rendered depth of our detailed 3D an- and the 3D annotations by Guo and notations compared with groundtruth depth in NYUd Hoiem [4] in NYUd v2 dataset. v2 dataset. The relative error rD is computed as: rD = |d p − dˆp | 1 ∑ ∑ |SI | I∈SI p∈I dp (2) where SI = {I1 , I2 , . . . , IN } is all the RGBD images in the dataset; p represents a pixel in each image I; d p is the groundtruth depth of pixel p from sensor; and dˆp is the rendered depth of the 3D label annotation at pixel p. For comparison, we report the rD of the 3D annotations by Guo and Hoiem [4]. Our annotations have more points with low relative depth error, and achieve a better modeling of depth for each image. 4 Generating object candidates Given an RGBD image as input, we aim to find a set of layout and object models that fit RGB and depth observations and provide a likely explanation for the unobserved portion of the scene. To produce object candidate regions, we use the method for RGBD images by Gupta et al. [7] and extract top ranked 2000 region proposals for each image. Our experiments show that this region retrieval is more effective than the method based on Prims algorithm [16] used in our previous work [5]. Likely object categories and 3D shapes are then assigned to each candidate region. 6 ZOU, LI, HOIEM: COMPLETE 3D SCENE PARSING FROM SINGLE RGBD IMAGE Proposed height Region proposal depth Support from below height stack Floor height depth height Region Proposal’s Bottom to Floor 32×32×4 Height Conv 3×3, 8 filters, stride 1 Max Pool Max Pool Conv 3×3, 3×3, 3×3, stride 2 16 filters, stride 2 stride 1 FC 512 FC 512 ✔ FC 2 Figure 4: The CNN for predicting a candidate object’s support height. We perform ReLU between the convolutional layer and the max pooling layer. Local response normalization is performed before the first fully connected (FC) layer. We add dropout with 0.5 before each FC layer during training. 4.1 Predicting region’s support height and class We train and use CNN networks to predict the object category and support height of each region, as shown in Fig. 4 and Fig. 5. The support height is used as a feature for the object classification. We also train and use a CNN with a Siamese network design to find the most similar 3D shape of a training object, based on region and depth features. Support height prediction. We predict support height for each object with the aim of better predicting class and position. We first find candidate support heights using the method of Guo and Hoiem [4] and use a CNN to estimate which is most likely for a given object region based on crops of the depth maps and height maps of the region proposal and region that extends from bottom of the region proposal to the estimated floor position. The support height network also predicts whether the object is supported from below or behind. To create the feature vector, we subtract the candidate support height from the height cropped images, re-size all four crops to 32 × 32 patches, and concatenate them, as illustrated in Fig. 4. In the test set, we identify the closest candidate support height with 92% ac0.96/1.0 (shape similarity) curacy, with an average distance error of bed cosine 0.18m. As a feature for classification, we softmax use the support height relative to the camera FC 512 FC 82 height, which leads to slightly better perforFC 1024 FC 1024 mance than using support height relative to the estimated ground likely because dataset images are taken from consistent heights but estimated ground height may be mistaken. Categorization. Our classification netRegion Proposal RGBD Exemplar RGBD work gets input of the region proposal’s support height and type, along with CNN features from both RGB and depth. The Figure 5: The CNNs for region classification network predicts the probability for each (left) and similar shape retrieval (right). We class as shown in Fig. 5. To model the var- perform ReLU and dropout with 0.5 after the ious classes of shapes in indoor scene, we first FC layer for both of the networks during classify the regions into the 78 most com- training. mon classes, which have at least 10 training Support CNNs RGB CNNs Depth CNNs RGB CNNs Depth CNNs ZOU, LI, HOIEM: COMPLETE 3D SCENE PARSING FROM SINGLE RGBD IMAGE 7 samples. Less common objects are classified as “other prop", “other furniture" and “other structure" based on rules by Silberman et al. [18]. In addition, we identify a region proposal that is not representative for an object shape (e.g. a piece of chair leg region when the whole chair region is visible) as a “bad region" class. This leads to our 78 + 3 + 1 = 82-class classifier. The input support height and type are directly predicted by our support height prediction network. To create our classification features, we copy the two predicted support values 100 times each (a useful trick to reduce sensitivity to local optima for important low-dimensional features) and concatenate them to the region proposal’s RGB and HHA features from Gupta et al. [7] in both the 2D bounding box and masked region as in [9]. Experiments show that using the predicted support type and the support height improves the classification accuracy by about 1%, with larger improvement for occluded objects. 4.2 Predicting region’s 3D shape Using a Siamese network (Fig. 5), we learn a region-to-region similarity measure that predicts 3D shape similarity of the corresponding objects. The network embeds the RGB and HHA features used in our classification network into a space where cosine distance correlates to shape similarity, as in [23]. In training, we use surface-to-surface distance [17] between mesh pairs as the ground truth similarity and train the network to penalize errors in shape similarity orderings. Each region pair’s shape similarity score is compared with the next pair’s among the randomly sampled batch in the current epoch and penalized only if the ordering disagrees with the groundtruth similarity. We attempted sharing embedding weights with the classification network but observed a 1% drop in classification performance. We also found predicted class probability to be unhelpful for predicting shape similarity. Candidate region selection. We apply the above retrieval scheme to each of the 2000 region proposals in each image, obtaining shape similarity rank compared with all the training samples and 81-object class and non-object class probability for each region proposal. In order to reduce the number of retrieved candidates before the scene interpretation in Sec. 5, we first reduce the number of region proposals using non-maximal suppression based on non-object class probability and threshold on the non-object class probability. We set the threshold to obtain 190 region proposals for each image, on average. We select the two most probable classes for each remaining region proposal and select five most similar shapes for each class, leading to 10 shape candidates for each region proposal. Then, we further refine these 10 retrieved shapes. We align each shape candidate to the target scene by translating the model using the offset between the depth point mass centers of the region proposal and the retrieved region. We then perform ICP, using a grid of initial values for rotation and scale and pick the best one for each shape based on the fitting energy in Eq. 1. We set the term weight: Cmissing as 0.6, Cdepth as 1.0, Cocc as 0.9 based on a grid search on the validation set. We tried using the estimated support height for each region proposal for aligning the related 3D shape models but observed a worse performance in the scene composition result. This is because a relatively small error in object’s support height estimation can cause a larger error in fitting. Finally, we select the two most promising shape candidates based on the following energy function, El (mi ) =w f E f itting (mi ,ti ) + wc log P(ci |ri ) + wb log P(bi |ri ) (3) 8 ZOU, LI, HOIEM: COMPLETE 3D SCENE PARSING FROM SINGLE RGBD IMAGE E f itting is the fitting energy defined in Eq. 1 that we used for alignment. P(ci |ri ) and P(bi |ri ) are the softmax class probability and the non-object class probability output by our classification network for the regon proposal ri . We normalize P(ci |ri ) to sum to 1, in order not to penalize the non-object class twice in the energy function. We set the term weights w f = 1.0, wc = −1500, wb = 1300 using a grid search. Note that E f itting is on the scale of the number of pixels in the region. 4.3 Training details We first find meta-parameters on the validation set after training classifiers on the training set. Then, we retrain on both training and validation in order to report results on the test set. We train our networks with the region proposals that have the highest (and at least 0.5) 2D intersection-over-union (IoU) with each groundtruth region in the train set. We train the support height prediction network with the groundtruth support type for the region proposal and set the groundtruth support height as the closest support height candidate that is within 0.15 meters from the related 3D annotation’s bottom. For training regions that are supported from behind, we do not penalize the support height estimation, since our support height candidates are for vertical support. For training the classification network, we also include the non-object class region proposals that have < 0.3 IoU with the groundtruth regions. We randomly sample the same number of the non-object regions as the total number of the object regions during training. To avoid unbalanced weights for different classes, we sample from the dataset the same number of training regions for each class in each epoch. When training the shape similarity network, we translate each 3D model to origin and resize to 200 × 200 × 200-voxel cuboid before computing the surface-to-surface distance. We use ADAM [11] to train each network with the hyper-parameter of β1 = 0.9, β2 = 0.999. The learning rate for each network is: support height prediction 0.0008, classification 0.003 and Siamese network 0.0001. 5 Scene composition with candidate 3D shape Given a set of retrieved candidate 3D shapes and the layout candidates from Guo et al. [5], we select a subset of the candidates that closely reproduce the original depth image when rendered, correspond to minimally overlapping 3D aligned models. The composition is hard because of the high degree of occlusions and large amount of objects in the scene. We apply the same method proposed by Guo et al. [5] to perform scene composition, this leads to our final scene interpretation. 6 6.1 Experiments Experiments setting We evaluate our method on our detailed 3D annotated NYUd v2 dataset. We also report the result of the state-of-the-art by Guo et al. [5] on our new annotations. Same as Gupta et al. [7], we tune the parameters on the validation set while training on the train set; we report our result on the test set by training on both the train and validation set. 9 ZOU, LI, HOIEM: COMPLETE 3D SCENE PARSING FROM SINGLE RGBD IMAGE Table 1: Our 81-class classification accuracy on groundtruth 2D regions in the test set. We compare two methods: our classification network with/without estimating support height. We compute the average accuracy for each class, average precision based on the predicted probability and the accuracy averaged over instances. The classification networks are trained and evaluated 10 times and the means and standard deviations (reflecting variation due to randomness in learning) are reported. 15 common object class results are also listed. Bold numbers signify better performance. Method avg per class avg precision avg over instance picture chair cabinet pillow bottle books paper table box window door sofa bag lamp clothes w/o support height w/ support height 43.7± 0.3 44.7± 0.3 37.7 ± 0.1 39.7±0.1 40.8± 0.3 42.7±0.2 57.5 57.5 46.1 53.1 39.5 44.35 66.5 69.0 55.6 54.9 30.0 33.5 40.7 43.5 36.7 39.5 10.6 14.1 61.4 62.4 54.9 57.8 63.0 65.9 14.9 15.1 64.6 65.9 25.3 26.9 Table 3: Quantitative evaluation for our retrieval method compared with Guo et al’s method. Method Guo et al. [5] Ours 6.2 avg class accuracy (%) top 1 top 2 top 3 top 1 11.97 41.56 0.134 0.191 16.36 54.47 19.65 62.07 avg 3D IoU top 2 top 3 0.177 0.231 0.200 0.249 avg surface distance (m) top 1 top 2 top 3 0.033 0.026 0.029 0.024 0.027 0.023 Evaluation of region classification and shape retrieval Classification. We report the region classification Table 2: Classification accuracy accuracy on the groundtruth 2D regions in the test set under different occlusion ratios of by our classification network, as shown in Table 1. groundtruth regions in test set Overall, including support height prediction will imOcclusion Ratio < 0.5 > 0.5 prove the classification results. For certain classes, w/ support height 45.8 38.5 objects often appear on the floor level (e.g. chair, w/o support height 44.2 35.7 desk) will have better classification accuracy given object’s height, while objects often appear on several heights (e.g. picture) do not have this benefit. Table. 2 shows the per instance classification accuracy under different occlusion ratios of the groundtruth regions in the test set. The improvement in classification accuracy is larger for highly occluded area, which conforms with our claim that estimating object’s support height will help classify occluded regions. Retrieval. We evaluate our candidate shape retrieval method compared with the stateof-the-art by Guo et al. [5] given groundtruth 2D regions in the dataset, as shown in Table 3. Since we use groundtuth regions, we do not perform candidate selection in Sec. 4.2 for this evaluation. We evaluate 1) top N retrieved class accuracy and 2) top N retrieved shape similarity, based on the shape intersect over union (IoU) and the surface-to-surface distance. In our experiment, we set N = {1, 2, 3}. To avoid rotation ambiguity in shape similarity measurement, we rotate each retrieved object to find the best shape similarity score with the groundtruth 3D shape. Our retrieval method outperforms the state-of-the-art under all the evaluation criteria. 6.3 Evaluation of scene composition 84-class semantic segmentation. We evaluate the 84-class semantic segmentation (81 object classes and 3 layout classes: wall, ceiling, floor) on the rendered 2D image of our scene composition result. We compare our method with Guo et al. [5] with both automatically generated region proposals and groundtruth regions. Table 4 shows the average class accuracy (avacc), average class accuracy weighted by frequency (fwavacc) and the average pixel ac- 10 ZOU, LI, HOIEM: COMPLETE 3D SCENE PARSING FROM SINGLE RGBD IMAGE Table 4: Results of 84-class semantic segmentation on both automatic region proposals and groundtruth regions. (b) Each class’s w/ support accuracy minus w/o support accuracy with automatic region proposals (a) Overall results Method Mean coverage weighted Mean coverage unweighted Guo et al. [5] MCG+CCA Ours w/o support Ours 48.12 50.43 50.43 50.23 29.75 32.91 34.34 33.92 Ours w/ GT support Ours w/ GT-region 50.52 52.39 34.32 35.61 Detailed groundtruth labeling 65.42 49.15 Table 5: Results of 84-class instance segmentation on both automatic region proposals and groundtruth regions. Method Mean coverage weighted Mean coverage unweighted Guo et al. [5] MCG+CCA Ours w/o support Ours 48.12 50.43 50.43 50.23 29.75 32.91 34.34 33.92 Ours w/ GT support Ours w/ GT-region 50.52 52.39 34.32 35.61 Detailed groundtruth labeling 65.42 49.15 curacy (pixacc). Our method has a better semantic segmentation result. We report our result with/without support height prediction and the result given the groundtruth support height as the upper bound. We see improvements, compared to Guo et al., due to both classification method (CNN vs. CCA) and region proposal method (MCG by Gupta et al. vs. prim’s by Guo et al.). We also report the 40-class semantic segmentation compared with the state-ofthe-art [15]: 47.7 pixacc / 18.3 avacc / 33.8 fwavacc vs. their 65.4/34.0/49.5. Note that our method focuses on predicting 3D geometry rather than semantic segmentation. 84-class instance segmentation. We report the instance segmentation result with the experiment setting same as semantic segmentation. The evaluation follows the protocol in RMRC [20]. Our method outperforms the state-of-the-art slightly. Note that the lower results for MeanCovU is caused by the large diversity of classes with less frequency. 3D estimation. We evaluate 3D estimation as in Guo et al. [5] with layout pixel-wise and depth prediction and freespace and occupancy estimation. As shown in Fig. 6, our method has competitive results in 3D estimation compared with the state-of-the-art. Qualitative results. Sample qualitative results are shown in Fig. 6 and Fig. 7. Our method has better region classification and shape estimation property, which results in a slightly better scene composition result. Failure cases are caused by bad pruning in region proposals (last two column, Fig. 6) and confusion between similar class (top right, Fig. 7). 7 Conclusions In this paper, we predict the full 3D parse of both visible and occluded portions of the scene from one RGBD image. We propose to use support inference to aid interpretation and proTable 6: Results of 3D layout, freespace and occupancy estimation. Method Guo et al. [5] Ours Layout Pixel Error overall visible occluded 10.9 10.6 14.0 13.6 4.8 4.8 Layout Depth Error overall visible occluded 0.166 0.15 0.204 0.181 0.074 0.074 Freespace precision recall 0.954 0.954 0.914 0.919 precision 0.504 0.478 Occupancy recall precision-ε 0.380 0.397 0.751 0.741 recall-ε 0.646 0.710 11 ZOU, LI, HOIEM: COMPLETE 3D SCENE PARSING FROM SINGLE RGBD IMAGE Input RGB(D) GT Guo et al. Ours Input RGB(D) GT Guo et al. Ours Figure 6: Qualitative results on scene composition with automatic region proposals. We randomly sample images from the top 25% (first four rows), medium 50% (row 6-9) and worst 25% (last four rows) based on 84-class semantic segmentation accuracy. 12 ZOU, LI, HOIEM: COMPLETE 3D SCENE PARSING FROM SINGLE RGBD IMAGE Input RGB(D) GT Guo et al. Ours Input RGB(D) GT Guo et al. Ours Figure 7: Qualitative results on scene composition given groundtruth 2D labeling as region proposals. We randomly sample images from the top 25% (first two rows), medium 50% (row 3-4) and worst 25% (last two rows) based on 84-class semantic segmentation accuracy. pose a retrieval scheme that uses CNNs to classify regions and find objects with similar shapes. Experiments demonstrate better performance of our method in semantic segmentation and instance segmentation and competitive results in 3D scene estimation. Acknowledgements This research is supported in part by ONR MURI grant N000141010934 and ONR MURI grant N000141612007. We thank David Forsyth for insightful comments and discussion. References [1] M. Aubry, D. Maturana, A. A. Efros, B. C. Russell, and J. Sivic. Seeing 3d chairs: Exemplar part-based 2d-3d alignment using a large dataset of cad models. In CVPR, 2014. [2] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. [3] Zhuo Deng, Sinisa Todorovic, and Longin Jan Latecki. Semantic segmentation of rgbd images with mutex constraints. In Proceedings of the IEEE International Conference on Computer Vision, pages 1733–1741, 2015. ZOU, LI, HOIEM: COMPLETE 3D SCENE PARSING FROM SINGLE RGBD IMAGE 13 [4] Ruiqi Guo and Derek Hoiem. Support surface prediction in indoor scenes. In ICCV, 2013. [5] Ruiqi Guo, Chuhang Zou, and Derek Hoiem. Predicting complete 3d models of indoor scenes. arXiv preprint arXiv:1504.02437, 2015. [6] Saurabh Gupta, Pablo Arbelaez, and Jitendra Malik. Perceptual organization and recognition of indoor scenes from RGB-D images. In CVPR, 2013. [7] Saurabh Gupta, Ross Girshick, Pablo Arbeláez, and Jitendra Malik. Learning rich features from rgb-d images for object detection and segmentation. In European Conference on Computer Vision, pages 345–360. Springer, 2014. [8] Saurabh Gupta, Pablo Arbeláez, Ross Girshick, and Jitendra Malik. Aligning 3d models to rgb-d images of cluttered scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4731–4740, 2015. [9] Bharath Hariharan, Pablo Arbeláez, Ross Girshick, and Jitendra Malik. Simultaneous detection and segmentation. In Computer vision–ECCV 2014, pages 297–312. Springer, 2014. [10] Derek Hoiem, Alexei A Efros, and Martial Hebert. Putting objects in perspective. International Journal of Computer Vision, 80(1):3–15, 2008. [11] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. [12] Joseph J Lim, Hamed Pirsiavash, and Antonio Torralba. Parsing IKEA objects: Fine pose estimation. In ICCV, 2013. [13] Joseph J Lim, Aditya Khosla, and Antonio Torralba. Fpm: Fine pose parts-based model with 3d cad models. In European Conference on Computer Vision, pages 478–493. Springer, 2014. [14] Dahua Lin, Sanja Fidler, and Raquel Urtasun. Holistic scene understanding for 3d object detection with rgbd cameras. In ICCV, 2013. [15] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3431–3440, 2015. [16] Santiago Manen, Matthieu Guillaumin, and Luc Van Gool. Prime object proposals with randomized prim’s algorithm. In Proceedings of the IEEE International Conference on Computer Vision, pages 2536–2543, 2013. [17] Jason Rock, Tanmay Gupta, Justin Thorsen, JunYoung Gwak, Daeyun Shin, and Derek Hoiem. Completing 3d object shape from one depth image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2484–2493, 2015. [18] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. Computer Vision–ECCV 2012, pages 746–760, 2012. 14 ZOU, LI, HOIEM: COMPLETE 3D SCENE PARSING FROM SINGLE RGBD IMAGE [19] Shuran Song and Jianxiong Xiao. Sliding shapes for 3d object detection in depth images. In European conference on computer vision, pages 634–651. Springer, 2014. [20] R Urtasun, R Fergus, D Hoiem, A Torralba, A Geiger, P Lenz, N Silberman, J Xiao, and S Fidler. Reconstruction meets recognition challenge, 2013. [21] Stefan Walk, Konrad Schindler, and Bernt Schiele. Disparity statistics for pedestrian detection: Combining appearance, motion and stereo. In Computer Vision–ECCV 2010, pages 182–195. Springer, 2010. [22] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaoou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1912–1920, 2015. [23] Wen-tau Yih, Kristina Toutanova, John C Platt, and Christopher Meek. Learning discriminative projections for text similarity measures. In Proceedings of the Fifteenth Conference on Computational Natural Language Learning, pages 247–256. Association for Computational Linguistics, 2011. [24] Yinda Zhang, Shuran Song, Ping Tan, and Jianxiong Xiao. Panocontext: A wholeroom 3d context model for panoramic scene understanding. In European Conference on Computer Vision, pages 668–686. Springer, 2014.
1cs.CV
Hashing Over Predicted Future Frames for Informed Exploration of Deep Reinforcement Learning arXiv:1707.00524v1 [cs.LG] 3 Jul 2017 Haiyan Yin Sinno Jialin Pan Nanyang Technological University, Singaproe {haiyanyin, sinnopan}@ntu.edu.sg Abstract In reinforcement learning (RL) tasks, an efficient exploration mechanism should be able to encourage an agent to take actions that lead to less frequent states which may yield higher accumulative future return. However, both knowing about the future and evaluating the frequentness of states are non-trivial tasks, especially for deep RL domains, where a state is represented by high-dimensional image frames. In this paper, we propose a novel informed exploration framework for deep RL tasks, where we build the capability for a RL agent to predict over the future transitions and evaluate the frequentness for the predicted future frames in a meaningful manner. To this end, we train a deep prediction model to generate future frames given a state-action pair, and a convolutional autoencoder model to generate deep features for conducting hashing over the seen frames. In addition, to utilize the counts derived from the seen frames to evaluate the frequentness for the predicted frames, we tackle the challenge of making the hash codes for the predicted future frames to match with their corresponding seen frames. In this way, we could derive a reliable metric for evaluating the novelty of the future direction pointed by each action, and hence inform the agent to explore the least frequent one. We use Atari 2600 games as the testing environment and demonstrate that the proposed framework achieves significant performance gain over a state-of-the-art informed exploration approach in most of the domains. 1 Introduction Reinforcement learning (RL) involves an agent progressively interacting with an initially unknown environment, in order to learn an optimal policy with the objective of maximizing the cumulative rewards collected from the environment [22]. Throughout the learning process, the RL agent alternates between two primal behaviors: exploration - to try out novel states that could potentially lead to high future rewards; and exploitation - to perform greedily according to the learned knowledge [22]. In the past, exploitation of learned knowledge has been well studied, while how to efficiently explore through the state space is still remained as a critical challenge, especially for deep RL domains. In deep RL domains, a state is represented by low-level sensory inputs only, such as image pixels [11, 14, 9], which is often high-dimensional or/and continuous. Thus, the state space for deep RL is huge and often intractable for searching. When performing exploration through such a huge state space, most existing deep RL works (e.g., [25, 3, 24, 19, 18]) adopt a simple exploration heuristic, -greedy strategy, where the RL agent takes a random action with a probability of , e.g., via uniform sampling for discrete-action domains [22, 11] or corrupting action with i.i.d. Gaussian noise for continuous-action domains [10]. In such a way, the agent explores the state space without conscious, i.e., without incorporating any meaningful knowledge about the environment. Such exploration heuristic turns out to work well in simple problem domains but fails to handle more challenging 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA. domains, such as those with extremely sparse rewards that lead to exponentially increasing state space, and those with a large action space. Unlike the unconscious exploratory behavior for agents using -greedy strategy, when human beings are intending to explore an unfamiliar task domain, one often actively applies domain knowledge for the task, accounts for the state space that has been less frequently visited, and intentionally tries out actions that lead to novel states. In this work, we aim to mimic such exploratory behaviors to improve upon the -greedy strategy with random action selection, and come up with a more efficient informed exploration framework for deep RL agents. On one hand, we develop agent’s knowledge on the environment and make it able to predict the future trajectories. On the other hand, we integrate the developed knowledge with hashing techniques over the high-dimensional state space in order to make the agent be able to realistically evaluate the novelty for the predicted future trajectories. Specifically, in our proposed informed exploration framework, first, we train an action-conditional prediction model to predict future frames given a state-action pair. Second, to perform hashing over the high-dimensional state space seen by the agent, we train a deep convolutional autoencoder to generate high-level features for the state and apply locality-sensitive hashing (LSH) [5] on the high-level state features to generate binary codes to represent each state. However, the learned hashing function is counting over the actually seen states, while we need to query the counts for the predicted future frames to compute their novelty. Hence, we introduce an additional training phase for the autoencoder to match the hash codes for the predicted frames with that of their corresponding ground-truth frames (i.e., the actually seen frames). In this way, we are able to utilize the environment knowledge and hashing techniques over the high-dimensional states to generate a reliable novelty evaluation metric for the future direction pointed by each action given a state. 2 Related Work Recently, works on enhancing the exploration behavior of the deep RL agent have demonstrated great potential in improving the performance of various deep RL task domains. In [12], asynchronous training techniques is adopted and multiple agents are created to perform gradient-based learning to update the model parameters separately in the Atari 2600 domain. In [16], an ensemble of Q-functions is trained to reduce the bias for the values being approximated and increases the exploration depth in the Atari 2600 domain. In [6], an exploration strategy based on maximizing the information gain of the agent’s belief about the environment dynamics is adopted for tasks with continuous states and actions. In [1], the KL-divergence between the probability obtained from a learned dynamics model and that of the actual trajectory is used to measure the surprise over the experience. In [3, 23], the novelty of a state is measured based on count-based mechanisms, and reward shaping is performed by adding a reward bonus term on the Q-value which is computed based on the counts. For all those approaches, the exploration strategy is incorporated either in the function approximation or during the optimization process, and the agent still needs to randomly choose action to explore without relying on any knowledge about the model. In this work, we aim to conduct informed exploration, i.e., to utilize the model-based knowledge to derive a deterministic choice of action for the agent to explore. Exploration with Deep Prediction Models Recent works aiming to incentivize exploration via deep prediction models have shown promising results for deep RL domains. In [21], an autoencoder model is trained jointly with the policy model, and the reconstruction error from the autoencoder is used to determine the rareness of a state. In [17], a pixelCNN is trained jointly with the policy model as a density model for the state. The prediction gain of a state is measured as the difference of the state density given by the pixelCNN after and before observing that state. In both approaches, the novelty of a state is measured by the loss or output of another model, which is not exact statistics. In our work, we use the counts derived from hashing over the state space to reliably infer the novelty of a state. In [15], which is a mostly related work to ours, an action-conditional prediction model is trained to predict the future frame given a state-action pair. Then they compute the gaussian kernal distance between a predicted future frame and a set of history frames, and inform the agent to take the action that leads to the frame which is most dissimilar compared to a window of recent frames. In our work, we use the same action-conditional architecture as in [15] to construct the prediction model. However, we involve a hashing mechanism to count over the state space to determine the novelty of a state based on the counting statistics. We enable the model to query the exact counting 2 statistics for the predicted frames, and compute the novelty for each action direction based on the novelty evaluated on multi-step prediction. Hashing for Deep RL Domain Running RL algorithms on the discretized features yields faster learning and promising performance. It is shown that the latent features learned by autoencoder trained in an unsupervised manner is of great promise to efficiently discretize the high-dimensional state space [4]. In [23], the state space for Atari 2600 domain is first discretized using the latent features derived from an autoencoder model. Then hashing is performed to encourage exploration by computing a reward bonus term in a form like MBIE-EB [20]. Our work also introduce hashing over the state space based on latent features trained from deep autoencoder model, but the exploration mechanism is significantly different from [23]. First, in our work, we count over the actually seen image frames, but query the hash for the predicted frames. Second, as the reward bonus is added to the function approximation target in their approach, the counts for future states influence their previous states backwards via bellman equation, whereas in our work, the count is used for informed exploration, which does not have direct influence on the approximated Q-values. 3 Methodology 3.1 Notations In this paper, we consider a discounted finite-horizon Markov Decision Process (MDP) with discrete actions. Formally, it is defined by a tuple (S, A, P, R, γ), where S is a set of states which could be high-dimentional or continuous, A is a set of actions, P is a state transition probability distribution with P(s0 |s, a) specifying the probability for transiting to state s0 after issuing action a at state s, R is a reward function mapping each state-action pair to a reward in R, and γ ∈ [0, 1] is a discount factor. The goal of the RL agent is to learn a policy π that maximizes the expected cumulative future PT rewards given the policy: Eπ [ t=0 γ t R(st , at )]. In the context of deep RL, at each step t, a RL agent receives a state observation St ∈ Rr×m×n , where r is the number of consequent frames to represent a state, and m × n is the dimension for each frame. The agent selects an action at ∈ A among all the l possible choices, and receives a reward rt ∈ R. 3.2 Informed Exploration Framework We propose an informed exploration framework to mimic the exploratory behavior of human beings under an unfamiliar task domain. Generally, the RL agent no longer randomly selects an action to explore without incorporating any domain knowledge. Instead, we aim to let the agent intentionally select the action that leads to the least frequent future states and thus explore the state space in an informed and deterministic manner. To this end, we build the capability of the RL agent on performing the following two tasks: 1) predicting over future transitions, and 2) evaluating the visiting frequency for those predicted future frames. Figure 1: Deep neural network architectures adopted for informed exploration. Left: actionconditional prediction model for predicting over future transition frames; right: autoencoder model for conducting hashing over the state space. 3.2.1 Learning Transition Model with Prediction Network The architecture for action-conditional prediction is shown in Figure 1 (left). To be specific, we train a deep prediction model f : (St , at ) → St+1 to make the agent able to predict over the future 3 transitions given a state-action pair. The state input St is a set of r recent image frames, and the action input at is represented by at ∈ Rl , which is a one-hot vector, where l is the number of actions for the task domain. To predict a new state, the model predicts one single frame at a time, denoted as s̄ ∈ Rm×n . The new state St+1 is formed by concatenating the predicted new frame with the most recent r−1 frames. We adopt the action-conditional transformation as proposed in [15] to form a joint feature for the state input and the action input. Specifically, the state input is first passed through three stacked convolutional layers to form a feature vector hst ∈ Rh . Then the state feature hst and the one-hot action feature at perform a linear transformation by multiplying with their corresponding transformation matrix Wts ∈ Rk×h and Wta ∈ Rk×l . After the linear transformation, both features are shaped with the same dimensionality. Then the features for the state and action after the linear transformation performs a multiplicative interaction to form a joint feature as follows, ht = Wts hst Wta hat . Afterwards, the joint feature ht is passed through stacked deconvolutional layers and a sigmoid layer to form the final prediction output. To predict over multiple future steps, the prediction model progressively composes the new state using its prediction result to predict the next-step transition. 3.2.2 Hashing over the State Space with Autoencoder and LSH To evaluate the novelty of a state, we adopt a hashing model to count over the state space. We first train an autoencoder model on frames, g : s ∈ Rm×n → ŝ ∈ Rm×n , in an unsupervised manner with the reconstruction loss as follows [8], n Lrec (st ) = − m  1 XX log p(ŝtij ) , mn j=1 i=1 (1) where ŝtij is the reconstructed pixel at the i-th row and the j-th column. The architecture for the autoencoder model is shown in Figure 1 (right). To be specific, each convolutional layer is followed by a Rectifier Linear Unit (ReLU) [13] layer and a max pooling layer with a kernel size 2 × 2. To discretize the state space, we hash over the last frame st of each state. We adopt the output of the last ReLU layer from the encoder as the high-level state features, and denote by φ(·) the corresponding feature map that generates the high-level feature vector zt ∈ Rd of a state, i.e., φ(st ) = zt . To further discretize the state feature, locality-sensitive hashing (LSH) [5] is adopted upon zt . To this end, a projection matrix A ∈ Rp×d is randomly initialized with i.i.d. entries drawn from a standard Gaussian N (0, 1). Then by projecting feature z through A, the sign of the outputs form a binary code, c ∈ Rp . With the introduced discretization scheme, we are able to count over the state space for the problem domain. During the RL process, a hash table H is created. The count for a state St , denoted by ψt , can be stored, queried and updated from the hash table. Overall, the process for counting over a state St is expressed in the following formulas. zt = φ(st ), ct = sgn(Azt ), and ψt = H(ct ). 3.2.3 (2) Matching the Prediction with Reality To derive the novelty for the predicted frames while updating the hash table with seen frames, we need to match the predictions with realities, i.e., make the hash codes for the predicted frames to be the same as their corresponding ground-truth seen frames in training. To this end, we introduce an additional training phase for the autoencoder model g(·). To make the hash codes to be the same, the derived feature vectors of the predicted frames and the ground-truth seen fames through φ(·) need to be close to each other. We introduce an additional loss function of a pair of a ground-truth seen frame and a predicted frame (st , s̄t ) as follows, Lmat (st , s̄t ) = kφ(st ) − φ(s̄t )k2 (3) Finally, by combing (1) and (3), we define the following overall loss function, L(st , s̄t ; θ) = Lrec (st ) + Lrec (s̄t ) + λLmat (st , s̄t ), (4) where θ is the parameter for the autoencoder. Note that even though the prediction model could generate almost identical frames, training the autoencoder with only the reconstruction loss may lead to distinct state codes in all the task domains 4 (details will be shown in Section 4). Therefore, the effort for matching the codes is necessary. However, matching the state code while guaranteeing a satisfying reconstruction behavior is extremely challenging. Fine tuning an autoencoder fully trained with Lrec by the code matching loss Lmat would fast disrupt the reconstruction behavior before the code loss could decrease to the expected level. Training the autoencoder from scratch with both Lrec and Lmat is also difficult, as Lmat is initially very low and Lrec is very high. The network can hardly find a direction to consistently decrease Lrec with such an imbalance. Therefore, in this work, we propose to train the autoencoder for two phases, where the first phase uses Lrec to train until convergence, and the second phase uses the composed loss function L as proposed in (4) to address the requirement for matching the prediction with reality. 3.2.4 Computing Novelty for States Once the prediction model f (·) and the autoencoder model g(·) are both trained, the agent could perform informed exploration in the following manner. At each step, the agent performs exploration with a probability less than  and perform greedy action selection otherwise. Given the state St , when performing exploration, the agent predicts through the future trajectories with length r for all the possible actions aj ∈ A. Formally, the novelty score for an action aj given state St , denoted by ρ(aj |St ), is computed as q X β i−1 p , (5) ρ(aj |St ) = ψt+i + 0.01 i=1 where ψt+i is the count for the future state St+i derived from (2) for the predicted frames {s̄t+j }ij=1 , q is the predefined prediction length, and β is a real-valued discount rate. After evaluating the novelty for all the possible actions, the agent selects the one with the highest novelty score to explore. Overall, the policy for the RL agent with the proposed informed exploration strategy is defined as: ( arg min[Q(St , a)] p ≥ ε a at = arg max[ρ(a|St )] p < ε a where Q(St , a) is the Q-value function, and p is a random value sampled from Uniform (0,1). 4 Experiments In the empirical evaluation, we use the Arcade Learning Environment (ALE) [2], which consists of Atari 2600 video games as the testing domain. We choose 5 representative games that require significant exploration to learn the policy: Breakout, Freeway, Frostbite, Ms-Pacman and Q-bert. Among the five games, Frostbite has a large action space that consists of a full set of 18 actions, Breakout has a state distribution that changes significantly as the ability of the policy network changes, and all the others have sparse rewards. For all the tasks, we use the state representation that concatenates 4 consequent image frames of size 84 × 84. 4.1 Evaluation on Prediction Model The architecture of the prediction model is identical to the one shown in Figure 1 (left). To train the prediction model, we create a training dataset which consists of 500,000 transition records generated by a fully trained DQN agent performing under -greedy with uniform random action selection, where  is set equal to 0.3. During training, we adopt Adam [7] as the optimization algorithm and use a learning rate of 10−3 and a mini-batch size of 100. Moreover, we discount the gradient scale by multiplying the gradient value by 10−2 . The state input is normalized by dividing the pixel values by 255. We show the pixel prediction loss in mean square error (MSE) for multi-step future prediction in Table 1. For all the task domains, the prediction errors are within a small scale. The prediction error increases with the increase of the prediction length. We demonstrate that the trained prediction models are able to generate realistic future frames which are visualized to be very close to the ground-truth frames from the results shown in Figure 2. 5 Game 1-step 3-step 5-step 10-step Breakout Freeway Frostbite Ms-Pacman Q-bert 1.114e-05 2.856e-05 7.230e-05 1.413e-04 5.300e-05 3.611e-04 0.939e-05 2.401e-04 4.353e-04 1.570e-04 4.471e-04 1.424e-04 5.142e-04 6.913e-04 2.688e-04 5.296e-04 2.479e-04 1.800e-03 1.226e-03 4.552e-04 Table 1: The prediction loss in MSE for the trained prediction models. 4.2 Evaluation on Hashing with Autoencoder and LSH The architecture for the autoencoder model is identical to that shown in Figure 1 (right). The autoencoder is trained on a dataset collected in an identical manner as that for the prediction model. It is trained under two phases. In the first phase, it is trained with only reconstruction loss. We use Adam as optimization algorithm, 10−3 as learning rate, a mini-batch size of 100 and discount the gradient by multiplying 10−2 . In the second phase, we train the autoencoder based on the loss in (4). In the second phase, we use Adam as optimization algorithm, 10−4 as learning rate, a mini-batch size of 100, and λ value of 0.01. We discount the gradient by multiplying the gradient value by 5 × 10−3 . Figure 2: The prediction and reconstruction result for each task domain. For each task, we present 2 sets of frames, where each set consists of four frames. The 2 sets of frames for each domain are put side-by-side in a row. The four frames for each set are organized as follows: (1) the ground-truth frame seen by the agent; (2) the predicted frame by the prediction model; (3) the reconstruction of autoencoder trained only with reconstruction loss; (4) the reconstruction of autoencoder trained after the second phase (i.e., trained with both reconstruction loss and code matching loss). Overall, it is extremely challenging to match the state codes for the predicted frames and their corresponding seen frames while maintaining a satisfying reconstruction performance. We demonstrate this in Figure 3 (a) by showing the code loss, which is measured in terms of the number of mismatch in binary codes between a pair of predicted frame and its corresponding ground-truth frame. The presented result is derived by averaging over 10,000 pairs of codes. First, the result shows that without the second phase, it is impossible to perform hashing with autoencoder trained only by the reconstruction loss, since the average code losses are above 1 in all the domains and with distinct hash codes, the count values returned from querying the hash table are meaningless. Second, the result shows that after the training of the second phase, the code loss is significantly reduced. 6 0Breakout Freeway Frostbite Pacman Q-bert 0.008 Phase-1 0.007 Phase-2 0.006 0.005 0.004 0.003 0.002 0.001 0.000Breakout Freeway Frostbite Pacman Q-bert (a) Code Loss (b) Reconstruction Loss Phase-1 Phase-2 8 Reconstruction Loss (MSE) Code Error (Num. Bits) 10 6 4 2 Figure 3: Comparison of the code loss and reconstruction loss (MSE) for the autoencoder model after the training of phase 1 and phase 2. We also show the reconstruction errors measured in terms of MSE after the training of the two phases for each domain in Figure 3 (b). By incorporating the code matching loss, the reconstruction behavior for the autoencoder receives slightly negative effect. A comparison of frame reconstruction effect after the training of the two phases are shown in Figure 2. It is shown that after training to match the state codes, the reconstructed frames are slightly blurred, but still able to reflect the essential features in each problem domain except for Q-bert. Moreover, we use Breakout as an illustrative example to demonstrate that the presented hashing framework can generate meaningful hash codes for predicted future frames (Figure 4). For a given ground-truth frame, we show the predicted frames with length 5 for taking each action. It can be found that the trajectories of ball positions are predicted to be the same regardless of the action choice, and different actions lead to different board positions. For the hash codes, as the first two actions, no-op and fire, lead to little change in frames, almost all the future frames are hashed into the same code. The last two actions, right and left, lead to significant changes in board position, so the codes for future frames are much more distinct from each other. The result also shows that the hash code is less sensitive to ball position compared to the board position in this game domain. Figure 4: Left: the predicted future trajectories for each action in Breakout. In each row, the first frame is the ground-truth frame and the following five frames are the predicted future trajectories with length 5. In each row, the agent takes one of the following actions (constantly): (1) no-op; (2) fire; (3) right; (4) left. Right: the hash codes for the frames in the same row ordered in a top-down manner. To save the space, four binary codes are grouped into one hex code, i.e., in a range of 0 to 15, and the color map is normalized linearly by the hex value. 7 4.3 Evaluation on Informed Exploration Framework To evaluate the efficiency of the proposed informed exploration framework, we integrate it into the DQN algorithm [11]. We compare it with two baselines: (1) DQN that performs -greedy with uniform random choice of action for exploration, denoted by DQN-Random; (2) the state-of-the-art informed exploration approach proposed in [15], denoted by DQN-Informed. The proposed informed exploration approach in this paper is denoted by DQN-Informed-Hash. Note that both DQN-Informed and the proposed approach adopt DQN as the base algorithm. We run the experiment by following the standard setting used in [11]. For all the methods, we train the agent over 50 million frames and evaluate the RL agent over 100 episodes of game play. We use q = 3 as the length for hashing over future frames. We report the result in Table 2. 1 2 Model Breakout Freeway Frostbite Ms-Pacman Q-bert DQN-Random DQN-Informed DQN-Informed-Hash 401.2 (26.9)1 0.93 (1.8)2 451.93 (45.79) 30.9 (0.2) 32.2 (0.2) 33.92 (0.14) 328.3 (250.5)1 1287.01 (314.3)2 1812.10 (433.64) 2281 (53) 2522 (57) 3526.60 (398.62) 3876 (106) 8238 (498) 8827.83 (531.64) Scores are obtained from [11]. Scores are obtained by implementing DQN-Informed by us in tasks not used in [15]. Table 2: Performance score for the proposed approach and baseline RL approaches. The performance scores for DQN-Random and DQN-Informed are obtained from [15] except for game Breakout and Frostbite, which are not included in their work. Among all the test domains, DQN-Informed-Hash outperforms the baseline approaches, and there are significant performance gains observed in each domain. Note that in Breakout, the agent fails to progress with DQN-Informed and always scores almost 0. It may due to that the kernel-based pixel distance evaluation metric used in DQN-Informed encourages the agent to explore states that is dissimilar from the recent history, which is insufficient to let the agent explore. Note that DQN-Informed-Hash demonstrates the superior performance with a deterministic exploration mechanism. It indicates that counting over the predicted future frames could provide a meaningful direction for exploration. 5 Conclusion In this paper, we propose an informed exploration framework for deep RL tasks with discrete action space. By incorporating a deep convolutional prediction model over future transitions and a hashing mechanism based on a deep autoencoder model and LSH, we enable the agent to predict over the future trajectories and intuitively evaluate the novelty for each future action direction based on the hashing result. The empirical result on Atari 2600 domain shows that the proposed informed exploration framework could efficiently encourage exploration in several challenging deep RL domains. References [1] Joshua Achiam and Shankar Sastry. Surprise-based intrinsic motivation for deep reinforcement learning. 2017. [2] M. G. Bellemare, Y. Naddaf, J. Veness, and M. Bowling. The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research, 47:253–279, jun 2013. [3] Marc Bellemare, Sriram Srinivasan, Georg Ostrovski, Tom Schaul, David Saxton, and Remi Munos. Unifying count-based exploration and intrinsic motivation. In Advances in Neural Information Processing Systems, pages 1471–1479, 2016. [4] Charles Blundell, Benigno Uria, Alexander Pritzel, Yazhe Li, Avraham Ruderman, Joel Z Leibo, Jack Rae, Daan Wierstra, and Demis Hassabis. Model-free episodic control. arXiv preprint arXiv:1606.04460, 2016. [5] Moses S Charikar. Similarity estimation techniques from rounding algorithms. In Proceedings of the thiry-fourth annual ACM symposium on Theory of computing, pages 380–388. ACM, 2002. [6] Rein Houthooft, Xi Chen, Yan Duan, John Schulman, Filip De Turck, and Pieter Abbeel. Vime: Variational information maximizing exploration. In Advances in Neural Information Processing Systems, pages 1109– 1117, 2016. [7] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 8 [8] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. [9] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. JMLR, 17(39):1–40, 2016. [10] Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015. [11] V. Mnih, K. Kavukcuoglu, D. Silver, A. a Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, S. Petersen, A. Sadik C. Beattie, I. Antonoglou, D. Kumaran H. King, D. Wierstra, S. Legg, and D. Hassabis. Human-level control through deep reinforcement learning. Nature, 2015. [12] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy P Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International Conference on Machine Learning, 2016. [13] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th international conference on machine learning (ICML-10), pages 807–814, 2010. [14] Junhyuk Oh, Valliappa Chockalingam, Satinder Singh, and Honglak Lee. Control of memory, active perception, and action in minecraft. ICML, 2016. [15] Junhyuk Oh, Xiaoxiao Guo, Honglak Lee, Richard L Lewis, and Satinder Singh. Action-conditional video prediction using deep networks in atari games. In Advances in Neural Information Processing Systems 28, pages 2845–2853. Curran Associates, Inc., 2015. [16] Ian Osband, Charles Blundell, Alexander Pritzel, and Benjamin Van Roy. Deep exploration via bootstrapped dqn. In Advances In Neural Information Processing Systems, pages 4026–4034, 2016. [17] Georg Ostrovski, Marc G Bellemare, Aaron van den Oord, and Remi Munos. Count-based exploration with neural density models. arXiv preprint arXiv:1703.01310, 2017. [18] Emilio Parisotto, Jimmy Ba, and Ruslan Salakhutdinov. Actor-mimic deep multitask and transfer reinforcement learning. In ICLR, 2016. [19] Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. In ICLR, 2016. [20] Satinder P Singh, Andrew G Barto, and Nuttapong Chentanez. Intrinsically motivated reinforcement learning. In NIPS, volume 17, pages 1281–1288, 2004. [21] Bradly C Stadie, Sergey Levine, and Pieter Abbeel. Incentivizing exploration in reinforcement learning with deep predictive models. arXiv preprint arXiv:1507.00814, 2015. [22] Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction, volume 1. MIT press Cambridge, 1998. [23] Haoran Tang, Rein Houthooft, Davis Foote, Adam Stooke, Xi Chen, Yan Duan, John Schulman, Filip De Turck, and Pieter Abbeel. # exploration: A study of count-based exploration for deep reinforcement learning. arXiv preprint arXiv:1611.04717, 2016. [24] Hado Van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double q-learning. In AAAI, pages 2094–2100, 2016. [25] Ziyu Wang, Tom Schaul, Matteo Hessel, Hado van Hasselt, Marc Lanctot, and Nando de Freitas. Dueling network architectures for deep reinforcement learning. In ICML, pages 1995–2003, 2016. 9 Appendix A. Informed Exploration Algorithm Algorithm 1: Informed exploration via hashing over predicted future frames. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 Train the prediction model f: S r×m×n , a l → s̄ m×n , the autoencoder model g: s m×n → ŝ m×n . Initialize Ap×d with i.i.d. entries drawn from standard Gaussian N (0, 1). Initialize an empty hash table H(.) ≡ 0, and a Q-function Q: S r×m×n → q l for each episode do while not terminal do Receive a state observation St from the environment. Compute the value t . Draw r ∼ Uniform(0,1) if r < t then . Informed exploration. for k=1,...,l do Let novelty ρ(ak |St ) = 0. Let at = ak . for i = 0 → q-1 do Predict s̄t+i+1 = f (St+i , at+i ). Compute latent feature zt+i+1 = φ (s̄t+i+1 ). Compute hash code ct+i+1 = sgn(A · zt+i+1 ). Derive the count ψt+i+1 = H(ct+i+1 ). Update novelty ρ(ak |St ) based on ψt+i+1 . Update St+i+1 [1:r−1,:,:] = St+i [2:r,:,:] , St+i+1 [r,:,:] = s̄t+i+1 Let at+i+1 = argmaxa Q(St+i+1 ). end end Let at = argmaxa ρ(a|St ). . Deterministically explore the least frequent action. else Let at = argmaxa Q(St ) . Greedy. end Take action at , and observe reward rt . Compute hash code for observed state ct = sgn(A · φ(st )) . Hashing. Update hash table H(ct ) ← H(ct ) + 1 Update Q-function on sampled transitions frequently. . Q-learning. end end 10 Appendix B. Result for Multi-step Prediction over Future Frames In this section, we show the multi-step prediction result over future frames in each task domain. We demonstrate the predicted future frames with a prediction length of 6. For each task, we illustrate 2 sets of prediction results, where each set consists of 2 rows. The first frame from the first row corresponds to the ground-truth frame st , and the following 6 frames in the first row correspond to the ground-truth frames from st+1 to st+6 . The frames in the second row correspond to the predicted future frames from s̄t+1 to s̄t+6 . Figure 5: Multi-step prediction result for Breakout. 11 Figure 6: Multi-step prediction result for Freeway. Figure 7: Multi-step prediction result for Frostbite. 12 Figure 8: Multi-step prediction result for Ms_pacman. Figure 9: Multi-step prediction result for Qbert. 13 Appendix C. Result for Hashing over Image Frames We demonstrate the effect of hashing over image frames in Figure 10. For each domain, we show a set of image frames that are hashed into the same hash code in a row. Overall, by adopting hashing-based counting approach, we expect that the hashing mechanism could consist of certain degree of generalization ability, so that similar frames could be hashed into the same hash code, while dissimilar frames could be distinguished. To this end, using features derived from autoencoder layer is apparently better than using raw pixels, since the image frames are too distinct at pixel level. However, when using autoencoder features, it is extremely hard to control the degree of generalization achieved by the output of the feature layer. Moreover, the binarilization approach via LSH and the additional training phase of autoencoder to match the codes for predicted frames and the real seen frames makes it even harder to ensure that the degree of generalization for the hash code is within an expected level. From the results shown in Figuer 10, we observe the hashing mechanism could identify some significant game-related features. For instance, in Breakout, we find the hash code is much more sensitive to the shape of bricks and board positions than the ball position; in Freeway, the intensity of cars and the position of the cars matters; in Frostbite, the hash codes distinguish well on the layout of platform pattern. However, we also find the current limitation of representing multiple patterns by one code, and the challenge of distinguishing frames based on some desired game-related feature, such as the ball position for Breakout, which may fail to be captured by the hashing mechanism without intelligently tune the hashing mechanism based on the desired purpose. 14 Result for hashing over image frames in each domain. Each row corresponds to the frames that lead to the same hash code in that corresponding game domain. Figure 10: 15
2cs.AI
Noncentralized Cryptocurrency wtih No Blockchain by Qian Xiaochao [email protected] btc: 18Bf718HZNboFvEqQuhJCTdKJtM8YxHPZW Sept. 2016, final Abstract. We give an explicit definition of decentralization and show you that decentralization is almost impossible for the current stage and Bitcoin is the first truly noncentralized currency in the currency history. We propose a new framework of noncentralized cryptocurrency system with an assumption of the existence of a weak adversary for a bank alliance. It abandons the mining process and blockchain, and removes history transactions from data synchronization. We propose a consensus algorithm named ―Converged Consensus‖ for a noncentralized cryptocurrency system. Keywords: Cryptocurrency,Converged Consensus, Bitcoin, Noncentralization Introduction Bitcoin is a peer to peer distributed digital currency system whose implementation is mainly based on cryptography. As a decentralized cryptocurrency, it has attracted lots of attention and has been widely adopted over the whole world. It is proposed by Satoshi Nakamoto in 2008 and created in 2009. Its goal is to build a practical decentralized cryptocurrency system. To achieve this goal, it adopts a proof of work mechanism to distribute the coins and validate the transactions. This mechanism is implemented by a sophisticated design of block and blockchain. The system adopts a longest blockchain principle to represent the consensus of users about the general ledger. In the Bitcoin system every transaction is stored in the blockchain. Over time, the size of the blockchain grows very fast. Ordinary users hate to verify transactions and synchronize the whole blockchain and the trading between Bitcoin and some fiat currency is invietable, so that they have to rely on some exchange hub which leads to additional safety dependence which severely compromises the safety of Bitcoins. The Mt.Gox event broke out against such a background. Many innocent users lost their Bitcoins. In addition, the intensive mining process results in a colossal waste of electricity and computing resources. 1 The proof of work mechanism makes the Bitcoin system expose itself to any potential superiority computing power attacker who may even hold no Bitcoin. There is a bad news about the Bitcoin system in Jun 2014, a major mining pool ghash.io had ever controlled more than 50% of computing power. This is a dangerous sign and is unacceptable for an alleged decentralized cryptocurrency system. Various of phenomenon express that the feasibility of a decentralized cryptocurrency is questionable. The main content of this paper is divided to two parts. We first try to show you that decentralization is almost impossible for the current stage in Part I and we propose a new framework of noncentralized cryptocurrency system for a bank alliance in Part II. Noncentralization Definition Centralization: A cryptocurrency system which is totally controlled by one single entity. Definition Decentralization: A cryptocurrency system which has only one safety dependency that is the private key. Definition Noncentralization: A cryptocurrency system which is neither centralized nor decentralized. Simply speaking, Decentralization means no trust; Noncentralization means partial trust and Centralization means fully trust. Mining Centralization The longest chain principle of Bitcoin provides a winner-take-all competition mechanism. For any miner, the best strategy is to try its best to get stronger or to merge with other miners to get itself stronger. As long as there are more than one miner (a mining pool is deemed as a miner) such fierce competition won’t cease. Thus the mining system has strong incentive to get centralized. The equilibrium result is that there is only one miner left. Although there are still more than one major mining pools at present, the weak pools are going to disappear because they are not able to get enough rewards to cover the cost for running a mining pool. The number of mining pools is going down. Even though the number won’t reduce to one, how could we know the major mining pools are not controlled by some single underground big boss? This is called Sybil attack which means mining pools with different names are merely sybils of some underground big boss. According to previous definition of Decentralization and Noncentralization, Bitcoin is obviously not decentralized as some fans claimed but running as a noncentralized system. As can be seen, essentially, Satoshi’s competitive blockchain solution to decentralization is centralization which is a trivial solution. As a decentralized cryptocurrency, the Bitcoin system is clearly failed, however, a failed experiment can provide us much valuable information. 2 First Noncentralized Currency Although mining centralization is inevitable, Bitcoin is still a truly noncentralized cryptocurrency system. In fact, the mining center will control most part of the Bitcoin system but not the whole system. In a typical centralized currency system, the central bank has super power of transferring money from any account to any other account, while the mining center doesn’t has such super power and it’s possible for stakeholders to fork the Bitcoin system whenever they find the mining center is no longer trustworthy. So Bitcoin is not totally controlled by the mining center. According to our previous definition, Bitcoin system is a noncentralized cryptocurrency and it is the first truly noncentralized currency in the currency history. Virtual General Adversary (VGA) For any decentralized system we assume there is one single Virtual General Adversary which is trying to attack the system. "Virtual General" means it includes any factor (including communication problem and some abstract attack behaviors) which may lead to the failure of decentralization. Virtually, the adversary can do whatever it can do to attack the system. We consider 4 major attack behaviors. ○ 1 maliciously construct messages, ○ 2 maliciously schedule messages, ○ 3 sybil attack ○ 4 centralization The Bitcoin system manages to build up strong resistance to ○ 1 , because constructing a block with high difficulty value is very expensive, nevertheless, it has very weak resistance to ○ 2 (e.g. a mining pool may withhold blocks it discovered) ○ 3 (we don’t know whether there is a underground boss)○ 4 (there is strong incentive to get centralized). These mean that the blockchain mechanism merely partially meets the challenge ○ 1 . To overcome any one of these 4 attack behaviors is a big challenge, let alone to overcome all of them simultaneously to build a truly decentralized cryptocurrency. Decentralization is too good to be true for the current stage. Probably, it belongs to the far future say the 22nd century, by when some truly revolutionary breakthrough of communication technique or computational theory may have occurred. To take the second best, we try to propose a new framework of noncentralized cryptocurrency system with an assumption of the existence of a weak adversary, which may be practical for the current stage. When we talk about an adversary we consider an adversary quadruple Q:= <R,S,I,C> where R is the basic Resources (e.g. computing power, money) the adversary possesses, S is the ability the adversary Scheduling messages, I is the ability the adversary gathering Information of the system, C is the attraction degree of the system getting Centralized. 3 It’s hard to accurately quantify the value of Q and if we simply assume a super high value of Q then we always get negative result. Hence, we assume a weak virtual general adversary with a low Q value for our new system. How weak exactly? Honestly, we have no idea. Just like the Bitcoin system we need to do experiments to gather information and finetune the parameters. We suppose the system is found by a bank alliance which includes N member banks all over the whole world. The banks are highly trusted, not completely trusted though. That accords with the assumption of a weak adversary. We name the new system X-Coin. Every expected event in the system has a timeout value, some of which are set globally the others are set locally. Balanceview Since we don't need mining anymore, we replace the concept of block and blockchain with normal data list. Definition Balanceview is a data list compounding a balance record list and a bank list. Definition Baseview is a Balanceview based on which some important actions are taken. balance record [ pubkey, balance ] balanceview [ bank list height, //sequence number of a Balanceview increasing continuously from 0 baseview hash, // refers to the Baseview of this Balanceview package hash, // refers to the package which is actually calculated [record1, record2, record3], ] transaction [ height, baseview hash, unique code, sender pubkey, receiver pubkey, volume, transaction fee, sender signature 4 ] transaction list [ [tx1, tx2, tx3... ], bank signature ] An end user sets a max transaction fee he will pay, for every new transaction and sends it to a selected bank. The bank will pick up the transaction if it provides enough transaction fee. Every bank verifies and collects transactions and makes a transaction list and broadcasts the list to other banks. Definition agent An agent is a bank which chooses to collect transaction lists from banks and make a transaction package and broadcasts the package to be verified and granted by banks. A bank grants a package which shares the same Baseview by signing the package and sends a granter item back to the agent. Every bank freely sets a granting rule to select out packages which it wants to grant. Definition Package-51 An agent collects granter items from other banks. Once it collects over 51% of grants, it combines the transactions package and a granter items list to form a pakcage-51 and broadcasts this package-51 to announce a new candidate Balanceview. A new Balanceview can be calculated according to a package-51 and the last Baseview. The total transaction fees are divided by the ratio: p for the agent, q for the deputy bank, r equally divided to all banks. granter item [ granter's pubkey, granter's signature // signs a package ] granter item list [ [item1, item2, item3... ], ] transaction package [ height, baseview hash, pubkey of source node, transaction list, 5 signature, ] package-51 [ package, granter item list, signature of agent ] Consensus It's possible that more than one package based on one Baseview get 51% of grants, consequently, more than one package-51 is broadcasted. Then the consensus of the new Balanceview can be split. This circumstance is analogous to a blockchain forks. According to the longest chain principle, the Bitcoin system guarentee the termination of a consensus process, which means whenever you see a longer valid blockchain you decide on that blockchain. But the consistency is not 100% gaurenteed. Because multiple valid blocks at one height is possible and the virtual general adversary may mailously schedule the spreading of blocks. To prevent such uncertainty, a checkpoint mechanism and a confirmation delay of 6 blocks is introduced. Only decentralization guarantees both termination and consistency. Bitcoin is obviously not decentralized. Taking the consensus standard of Bitcoin as a reference, X-Coin guarantees the termination and consistency with high probability. Our main idea of consensus algorithm is roughly as simple as following. Suppose there are 100 red balls and 100 black balls in a bag, one randomly selects 5 balls, if red balls is more than black balls then one balck ball will turn to red ball, vise versa. He does this time and time again, eventually colors of all 200 balls will converge to one single color, ether all black or all red. Suppose there are 100 balls in a bag, if one consecutively 20 times randomly selects 1 ball are all red balls. Then the probability of most of 100 balls are red balls is extremly high. In practice, we present a Balanceview Indirectly denoted by ibv to save the space. When necessary, we reconstruct the Balanceview according to an ibv. indirect presented Balanceview [ height, hash of package-51, // refers to a package-51 which is actually calculated ] When a bank receives the first valid package-51 at its current height, it makes an ibv of this new Balanceview as its current ibv. Converged Consensus Algorithm: 6 while no consecutive identical k ibv_update observed randomly sample n ibv // timeout gives a NULL ibv ibv_update = the most frequent ibv if ibv_update.height < current height discard //we should never go backward decide on ibv_update if ibv_update == NULL or Baseview collision detected then alert Termination: Every honest node eventually decides, with probability 1. Since the adversary has limited ability of scheduling the messages, termination will be reached, sooner or later. A bank reconstructs Baseview according to ibv_update. A bank’s current Balanceview is volatile as the bank keeps updating. A Baseview is rather stable but still not permanent because the algorithm doesn’t guarantee completely consistent. Note that the distribution of ibv gets more and more consistent as the sampling and updating go and the speed of the converging process gets faster and faster due to the positive feedback effect which is a good merit for practical application. To avoid double spending, referring to Bitcoin’s strategy, we may delay the confirmation of transactions by several Baseviews. A Baseview collision means some bank decides on different ibv at one height. If a bank detects a Baseview collision (theoretically possible, but very rare) at one height or it decides on a NULL ibv then a human intervention is needed. Meta data & meta-transaction We can adjust the major parameters of X-Coin system by a meta-transaction which requires 75% of grants to take effect. We can adjust the ratio <p=0%, q=70%, r=30%> for dividing transaction fees which controls the intensity of competition among banks. We shouldn’t give strong incentive to banks choosing to be an agent to increase the initial consistency as much as possible. So we set p=0% or a very small positive number. The ratio q controls the competition intensity among the banks serving end users. The ratio r provides the basic support to all banks to make the public contribution such as verification, granting and keeping sampling. According to the Moore’s Law, the cost of making public contribution will be trivial for a commercial bank and the banks are no longer a completely selfish entity hidden in the dark corner of the internet which can only be handled by Decentralization. We are now Noncentralization. The banks are legal public companies. They care about their public reputation and image. X-Coin system has weak resistance to a small number of dishonest banks. We can adjust the sampling and termination parameters according to conditions of the environment. We can even inflate the money supply by a meta-transaction. If necessary, we can switch the system to a leader election mode, in which, we specifies 7 a fixed bank as the agent to increase the consistency and efficiency or to fix some emergent problems. We can make a statistic tool as a monitor model for a standard client for banks to monitor each other' behavior and a bank’s extremely abnormal behaviors will be reported. We can use meta-transaction to remove a highly suspicious bank from the bank list as long as 75% of banks are convinced. Banks have sufficient incentive and social pressure to behaving normally and have little incentive to behave badly. Conclusion We agree that maintaining a trustable financial system is very expensive. Satoshi’s original idea is to trade expensive electricity and computing resources for Dencentralization. If the strategy does work then it worth the cost. Unfortunately, we have shown you that Bitcoin is actually running as a Noncentralized system and has obvious trend to evolve to a centralized system. So it’s unnecessary to waste that much electricity and computing resources to do meaningless hashing. As a decentralized system Bitcoin is obviously failed. For Decentralization the blockchain mechanism doesn’t work and for Noncentralization the blockchain mechanism is unnecessary. The blockchain mechanism has been proven by the reality to be a trivial and wasteful design. We assert the blockchain mechanism will be a flash in the pan. The most valuable information we get from the failed Bitcoin experiment is that Bitcoin has shown us that a noncentralized cryptocurrency system is possible. We have shown you that Decentralization is almost impossible in the current stage. Researchers should not claim a Decentralization and we should not believe in a claimed one. It’s all Noncentralization. We have one question for those who claim a Decentralization. Whether the system has only one safety dependency? If the answer is no, what are other unknown unclear undefined safety dependencies? How could they ensure the safety without even knowing all safety dependencies? Actually, they couldn't give an accurate quantitified definition of the adversary. They don't know what their adversary exactly is and how strong their adversary exactly is. So, their calculations on the safety analysis are very questionable. Decentralization is simply too good to be true. It's not the time for Decentralization. We suggest give the noncentralized X-Coin system based on Converged Consensus a try. Although it is extremely hard to build a truly decentralized cryptocurrency system, we human beings do have gotten one truly decentralized currency system. It is the gold system which has only one safety dependency that is YOU. Your gold is safe as long as you keep your gold well so the only safety dependency is YOU—the owner of gold. We think of the gold system as a NATUAL Decentralized currency system. The gold system itself is a special Balanceview which represents the NATUAL consensus of the distribution of fortune of all mankind. Finally, we have a conjecture that the timeline of currency of mankind should be like this: Natural Decentralization (gold)  Artificial Centralization (fiat currency)  8 Artificial Noncentralization (Bitcoin, X-Coin)  Artificial Decentralization (X-Gold, our ultimate goal) References 1. Satoshi Nakamoto, Bitcoin: A Peer-to-Peer Electronic Cash System, 2008 2. Michael J. Fischer, Nancy A. Lynch, and Michael S. Paterson. Impossibility of distributed consensus with one faulty process. Journal of the ACM, 32(2):374–382, April 1985. 3. M. Pease, R. Shostak and L. Lamport. Reaching agreement in the presence of faults. Journal of the ACM., 27:228–234, 1980. 4. L. Lamport, R. Shostak and M. Pease. The Byzantine Generals Problem. ACM Transactions on Programming Languages and Systems 4 (3): 382–401. July 1982. 5. http://en.wikipedia.org/wiki/Absorbing_Markov_chain Revision 15 March 2014 17:30 UTC Appendix Solution to the ball problem: The changing of the state of all balls forms a Markov chain. Obviously, it’s an absorbing Markov chain. Suppose its transition matrix P have t transient states (two colors) and r absorbing states (single color). Then: Q R P= 0 Ir where Q is a t-by-t matrix, R is a nonzero t-by-r matrix, 0 is an r-by-t zero matrix, and Ir is the r-by-r identity matrix. Thus, Q describes the probability of transitioning from some transient state to another while R describes the probability of transitioning from some transient state to some absorbing state. n steps transition matrix: Q R n Qn R′ Pn = = 0 Ir 0 Ir The probability that the process will be absorbed is 1. Because limn→∞ Qn = 0 9
5cs.CE
1 On Perfect Privacy and Maximal Correlation Borzoo Rassouli and Deniz Gündüz Abstract The problem of private data disclosure is studied from an information theoretic perspective. Considering a pair of arXiv:1712.08500v1 [cs.IT] 22 Dec 2017 correlated random variables (X, Y ), where Y denotes the observed data while X denotes the private latent variables, the following problem is addressed: What is the maximum information that can be revealed about Y , while disclosing no information about X? Assuming that a Markov kernel maps Y to the revealed information U , it is shown that the maximum mutual information between Y and U , i.e., I(Y ; U ), can be obtained as the solution of a standard linear program, when X and U are required to be independent, called perfect privacy. This solution is shown to be greater than or equal to the non-private information about X carried by Y . Maximal information disclosure under perfect privacy is is shown to be the solution of a linear program also when the utility is measured by the reduction in the mean square error, E[(Y − U )2 ], or the probability of error, Pr{Y 6= U }. For jointly Gaussian (X, Y ), it is shown that perfect privacy is not possible if the kernel is applied to only Y ; whereas perfect privacy can be achieved if the mapping is from both X and Y ; that is, if the private latent variables can also be observed at the encoder. Next, measuring the utility and privacy by I(Y ; U ) and I(X; U ), respectively, the slope of the optimal utility-privacy trade-off curve is studied when I(X; U ) = 0. Finally, through a similar but independent analysis, an alternative characterization of the maximal correlation between two random variables is provided. Index Terms Privacy, perfect privacy, non-private information, maximal correlation, mutual information. I. I NTRODUCTION With the explosion of machine learning algorithms, and their applications in many areas of science, technology, and governance, data is becoming an extremely valuable asset. However, with the growing power of machine learning algorithms in learning individual behavioral patterns from diverse data sources, privacy is becoming a major concern, calling for strict regulations on data ownership and distribution. On the other hand, many recent examples of de-anonymization attacks on publicly available anonymized data ([1], [2]) show that regulation on its own will not be sufficient to limit access to private data. An alternative approach, also considered in this paper, is to process the data at the time of release, such that no private information is leaked, called perfect privacy. Assuming that the joint distribution of the observed data and the latent variables that should be kept private is known, an information-theoretic study is carried out in this paper to characterize the fundamental limits on perfect privacy. Consider a situation in which Alice wants to release some useful information about herself to Bob, represented by random variable Y , and she receives some utility from this disclosure of information. This may represent some Borzoo Rassouli and Deniz Gündüz are with the Information Processing and Communications Lab at the Department of Electrical and Electronics Engineering, Imperial College London, United Kingdom. emails: {b.rassouli12, d.gunduz}@imperial.ac.uk. 2 data measured by a health monitoring system [3], her smart meter measurements [4], or the sequence of a portion of her DNA to detect potential illnesses [5]. At the same time, she wishes to conceal from Bob some private information which depends on Y , represented by X. To this end, instead of letting Bob have a direct access to Y , a privacy-preserving mapping is applied, whereby a distorted version of Y , denoted by U , is revealed to Bob. In this context, privacy and utility are competing goals: The more distorted version of Y is revealed by the privacy mapping, the less information can Bob infer about X, while the less utility can be obtained. This trade-off is the very result of the dependencies between X and Y . An extreme point of this trade-off is the scenario termed as perfect privacy, which refers to the situation where nothing is allowed to be inferred about X by Bob through the disclosure of U . This condition is modelled by the statistical independence of X and U . The concern of privacy and the design of privacy-preserving mappings has been the focus of a broad area of research, e.g., [6]–[9], while the information-theoretic view of privacy has gained increasing attention more recently. In [10], a general statistical inference framework is proposed to capture the loss of privacy in legitimate transactions of data. In [11], the privacy-utility trade-off under the self-information cost function (log-loss) is considered and called the privacy funnel, which is closely related to the information bottleneck introduced in [12]. In [13], sharp bounds on the optimal privacy-utility trade-off for the privacy funnel are derived, and an alternative characterization of the perfect privacy condition (see [14]) is proposed. Measuring both the privacy and the utility in terms of mutual information, perfect privacy is fully characterized in [15] for the binary case. Furthermore, a new quantity is introduced to capture the amount of private information about the latent variable X carried by the observable data Y . We study the information theoretic perfect privacy in this paper, and our main contributions can be summarized as follows: • Adopting mutual information as the utility measure, i.e., I(Y ; U ), we show that the maximum utility under perfect privacy is the solution to a standard linear program (LP). We obtain similar results when other measures of utility, e.g., the minimum mean-square error or the probability of error, are considered. • We show that when (X, Y ) is a jointly Gaussian pair with non-zero correlation coefficient, for the privacy mapping pU |Y , perfect privacy is not feasible. In other words, U is independent of X if and only if it is also independent of Y , i.e., maximum privacy is obtained at the expense of zero utility. This, however, is not the case when the mapping is of the form pU |X,Y ; that is, when the encoder has access to the private latent variables as well as the data. • Denoting the maximum I(Y ; U ) under perfect privacy by g0 (X, Y ), we characterize the relationship between the non-private information about X carried by Y , DX (Y ) as defined in [15], and g0 (X, Y ). • Considering mutual information as both the privacy and the utility measure, the optimal utility-privacy trade-off curve, characterized by the supremum of I(Y ; U ) over pU |Y vs. I(X; U ), is not a straightforward problem. Instead, we investigate the slope of this curve when I(X; U ) = 0. This linear approximation to the trade-off curve provides the maximum utility rate when a small amount of private data leakage is allowed. We obtain this slope when perfect privacy is not feasible, i.e., g0 (X, Y ) = 0, and propose a lower bound on it when perfect privacy is feasible, i.e., g0 (X, Y ) > 0. 3 • As a by-product of this slope analysis, we provide an alternative characterization of the maximal correlation between two random variables [16]–[18]. Notations. Random variables are denoted by capital letters, their realizations by lower case letters, and their alphabets by capital letters in calligraphic font. Matrices and vectors are denoted by bold capital and bold lower case letters, respectively. For integers m ≤ n, we have the discrete interval [m : n] , {m, m + 1, . . . , n}, and the tuple (am , am+1 , . . . , an ) is written in short as a[m:n] . For an integer n ≥ 1, 1n denotes an n-dimensional all-one column vector. For a random variable X ∈ X , with finite |X |, the probability simplex P(X ) is the standard (|X | − 1)-simplex given by  P(X ) =  v ∈ R|X | 1T|X | · v = 1, vi ≥ 0, ∀i ∈ [1 : |X |] . Furthermore, to each probability mass function (pmf) on X, denoted by pX (·), corresponds a matrix PX = diag(pX ), where pX is a probability vector in P(X ), whose i-th element is pX (xi ) (i ∈ [1 : |X |]). For a pair of random variables (X, Y ) with joint pmf pX,Y , PX,Y is an |X| × |Y | matrix with (i, j)-th entry equal to pX,Y (i, j). Likewise, PX|Y is an |X| × |Y | matrix with (i, j)-th entry equal to pX|Y (i|j). FY (·) denotes the cumulative distribution function (CDF) of random variable Y , and if it admits a density, its probability density function (pdf) is denoted by fY (·). For 0 ≤ t ≤ 1, Hb (t) , −t log2 t − (1 − t) log2 (1 − t) denotes the binary entropy function. The unit-step function is denoted by s(·). Throughout the paper, for a random variable Y with the corresponding probability vector pY , H(Y ) and H(pY ) are written interchangeably, and so are the quantities D(pY (·)||qY (·)) and D(pY ||qY ). II. S YSTEM MODEL AND PRELIMINARIES Consider a pair of random variables (X, Y ) ∈ X ×Y (|X |, |Y| < ∞) distributed according to the joint distribution pX,Y . We assume that pY (y) > 0, ∀y ∈ Y and pX (x) > 0, ∀x ∈ X , since otherwise the supports Y or/and X could have been modified accordingly. This equivalently means that the probability vectors pY and pX are in the interior of their corresponding probability simplices, i.e. P(Y) and P(X ). Let X denote the private/sensitive data that the user wants to conceal and Y denote the useful data the user wishes to disclose. Assume that the privacy mapping/data release mechanism takes Y as input and maps it to the released data denoted by U . In this scenario, X − Y − U form a Markov chain, and the privacy mapping is captured by the conditional distribution pU |Y . Let g (X, Y ) be defined1 as [15] g (X, Y ) , sup pU |Y : X−Y −U I(X;U )≤ I(Y ; U ). (1) In other words, when mutual information is adopted as a measure of both utility and privacy, (1) gives the best utility that can be obtained by privacy mappings which keep the sensitive data (X) private within a certain level of . 1 This is done with an abuse of notation, as g (X, Y ) should be written as g (pX,Y ). 4 Proposition 1. It is sufficient to have |U| ≤ |Y| + 1. Also, we can write g (X, Y ) = max pU |Y : X−Y −U I(X;U )≤ I(Y ; U ) = max pU |Y : X−Y −U I(X;U )= I(Y ; U ). (2) Proof. The proof is provided in Appendix A. Later, we show that it is sufficient to restrict our attention to |U| ≤ |Y|, when  = 0. III. P ERFECT P RIVACY Definition. For a pair of random variables (X, Y ), we say that perfect privacy is feasible if there exists a random variable U that satisfies the following conditions: 1) X − Y − U forms a Markov chain, 2) X ⊥ ⊥ U , i.e., X and U are independent, 3) Y 6⊥ ⊥ U , i.e., Y and U are not independent. From the above definition, we can say that perfect privacy being feasible is equivalent to having g0 (X, Y ) > 0. Proposition 2. Perfect privacy is feasible if and only if   dim Null(PX|Y ) 6= 0. (3) Proof. In [14, Theorem 4], the authors showed that for a given pair of random variables (X, Y ) ∈ X ×Y (|X |, |Y| < ∞), there exists a random variable U satisfying the conditions of perfect privacy if and only if the columns of PX|Y are linearly dependent. Equivalently, there must exist a non-zero vector v, such that PX|Y v = 0, which is equivalent to (3). Proposition 3. For the null space of PX|Y , we have z ∈ Null(PX|Y ) =⇒ 1T|Y| · z = 0. Therefore, for any z ∈ Null(PX|Y ), there exists a positive real number α, such that pY + αz ∈ P(Y). Proof. We have 1T|Y| · z = 1T|X | PX|Y z = 0, (4) (5) where (4) follows from the fact that 1T|X | PX|Y = 1T|Y| , and (5) from the assumption that z belongs to Null(PX|Y ). The last claim of the proposition is due to the fact that pY is in the interior of P(Y). Theorem 1. For a pair of random variables (X, Y ) ∈ X × Y (|X |, |Y| < ∞), g0 (X, Y ) is the solution to a standard linear program (LP) as given in (13). Proof. Let PX|Y be an |X | × |Y| matrix with (i, j)-th entry equal to pX|Y (i|j). 5 From the singular value decomposition2 of PX|Y , we have PX|Y = UΣVT , where the matrix of right eigenvectors is h V = v1 v2 ... v|Y| i |Y|×|Y| . (6) (3) is equivalent to having the null space of PX|Y written as Null(PX|Y ) = Span{vm , vm+1 , . . . , v|Y| }, for some m ≤ |Y|. (7) The random variables X and U are independent if and only if pX (·) = pX|U (·|u), ∀u ∈ U ⇐⇒ pX = pX|u , ∀u ∈ U. (8) Furthermore, if X − Y − U form a Markov chain, (8) is equivalent to PX|Y (pY − pY |u ) = 0, ∀u ∈ U ⇐⇒ (pY − pY |u ) ∈ Null(PX|Y ), ∀u ∈ U. From the column vectors in (6) and the definition of index m afterwards, construct the matrix A as h iT A , v1 v2 . . . vm−1 . (9) From (7), we can write (pY − pY |u ) ∈ Null(PX|Y ), ∀u ∈ U ⇐⇒ A(pY − pY |u ) = 0, ∀u ∈ U. Therefore, for the triplet (X, Y, U ), if X − Y − U forms a Markov chain and X ⊥ ⊥ U , we must have pY |u ∈ S, ∀u ∈ U, where S is a convex polytope defined as   S , x ∈ R|Y| Ax = ApY , x ≥ 0 . (10) Note that any element of S is a probability vector according to Proposition 3. On the other hand, for any pair (Y, U ), for which pY |u ∈ S, ∀u ∈ U, we can simply have X − Y − U and X⊥ ⊥ U . Therefore, we can write X − Y − U, X ⊥ ⊥ U ⇐⇒ pY |u ∈ S, ∀u ∈ U. (11) This leads us to g0 (X, Y ) = = max pU |Y : X−Y −U I(X;U )=0 max pU |Y : pY |u ∈S, ∀u∈U = H(Y ) − 2 We I(Y ; U ) I(Y ; U ) min pU (·),pY |u ∈S, ∀u∈U : P u pU (u)pY |u =pY H(Y |U ), assume, without loss of generality, that the singular values are arranged in a descending order. (12) 6 where in (12), since the minimization is over pY |u rather than pU |Y , a constraint was added to preserve the marginal distribution pY .  Proposition 4. In minimizing H(Y |U ) over pY |u ∈ S, ∀u ∈ U  p (u)p = p Y , it is sufficient to Y |u u U P consider only |Y| extreme points of S. Proof. The proof is provided in Appendix B. From Proposition 4, the problem in (12) can be divided into two phases: in phase one, the extreme points of set S are identified, while in phase two, proper weights over these extreme points are obtained to minimize the objective function. For the first phase, we proceed as follows. The extreme points of S are the basic feasible solutions (see [19], [20]) of it, i.e., the basic feasible solutions of the set   |Y| x ∈ R Ax = b , x ≥ 0 , where b = ApY . The procedure of finding the extreme points of S is as follows. Pick a set B ⊂ [1 : |Y|] of indices that correspond to m − 1 linearly independent columns of matrix A in (9). Since matrix A(m−1)×|Y| is full rank (note that its  |Y| rows are mutually orthonormal and rank(A) = m − 1), there are at most m−1 ways of choosing m − 1 linearly independent columns of A. Let AB be an (m − 1) × (m − 1) matrix whose columns are the columns of A indexed iT h by the indices in B. Also, for any x ∈ S, let x̃ = xTB xTN , where xB and xN are (m − 1)-dimensional and (|Y|−m+1)-dimensional vectors whose elements are the elements of x indexed by the indices in B and [1 : |Y|]\B, respectively. For any basic feasible solution x∗ , there exists a set B ⊂ [1 : |Y|] of indices that correspond to a set of linearly iT h independent columns of A, such that the corresponding vector of x∗ , i.e. x̃∗ = x∗B T x∗N T , satisfies the following x∗N = 0, x∗B = A−1 B b, x∗B ≥ 0. On the other hand, for any set  B ⊂ [1 : |Y|] of indices that correspond to a set of linearly independent columns of A−1 b  B  is the corresponding vector of a basic feasible solution. Hence, the extreme points A, if A−1 b ≥ 0, then B 0  |Y| of S are obtained as mentioned above, and their number is at most m−1 . For the second phase, we proceed as follows. Assume that the extreme points of S, found in the previous phase, are denoted by p1 , p2 , . . . , pK . Then (12) is equivalent to h i g0 (X, Y ) = H(Y )− min H(p1 ) H(p2 ) . . . H(pK ) · w w≥0 h i s.t. p1 p2 . . . pK w = pY , where w is a K-dimensional weight vector, and it can be verified that the constraint (13) PK i=1 wi = 1 is satisfied if the constraint in (13) is met. The problem in (13) is a standard linear program (LP), which can be efficiently solved. 7 The following example clarifies the optimization procedure in the proof of Theorem 1. Example 1. Consider the pair (X, Y ) ∈ [1 : 2] × [1 : 4] joint distribution is specified by the following matrix:   0.15 0.2 0.0625 0.05 . PX,Y =  0.35 0.05 0.0625 0.075 This results in   1 2 1 4  pY =  1 , 8   PX|Y  0.3 = 0.7 0.8 0.5 0.2 0.5  0.4 . 0.6 1 8   Since |Y| > |X |, we have dim Null(PX|Y ) 6= 0; and therefore, g0 (X, Y ) > 0. The singular value decomposition of PX|Y is PX|Y  −0.7071 = −0.7071  −0.7071 1.4142  0.7071 0 0 0.5292  −0.5   0 0 −0.5   0 0 −0.5  −0.5  −0.5394   −0.3154 −0.0876 ,  0.8452 −0.1889  −0.1135 0.8159 −0.4163 0.5345 −0.8018 0 0.2673 where it is obvious that columns 3 and 4 of the matrix of the right eigenvectors span the null space of PX|Y . Hence, the matrix A in (9) is given by  −0.5 A= 0.5345 −0.5 −0.5 −0.8018 0 −0.5  . 0.2673 For the first phase, i.e., finding the extreme points of S, it is clear that there are 4 2  possible ways of choosing 2 linearly independent columns of A. Hence, the index set B can be {1, 2}, {1, 3}, {1, 4}, {2, 3}, {2, 4} or {3, 4}. From xB = A−1 B b, we get x{1,2} x{2,3}       0.675 0.1875 −0.625  , x{1,3} =   , x{1,4} =   = 0.325 0.8125 1.625       −0.125 0.1563 0.625 , x{2,4} =   , x{3,4} =  . = 1.125 0.8437 0.375 It is obvious that x{1,4} and x{2,3} are not feasible, since they do not satisfy xB ≥ 0. Therefore, the extreme points of S are obtained as         0.675 0.1875 0 0                 0.325  0  0.1563  0   , p2 =   , p3 =   , p4 =  . p1 =           0  0.8125  0  0.625         0 0 0.8437 0.375 8 Now, for the second phase, the standard LP in (13) is written as h i min 0.9097 0.6962 0.6254 0.9544 .w w≥0     1 0.675 0.1875 0 0 2   1     0.325 0 0.1563 0   w = 4 , S.t.  1   8  0 0.8125 0 0.625     1 0 0 0.8437 0.375 8 where the minimum value of the objective function is 0.8437 bits, which is achieved by   0.698       0.1538 ∗ .  w =  0.1481   0 Therefore, g0 (X, Y ) = H(Y ) − 0.8437 = 0.9063 (note that H(Y |X) = 1.3061), U = {u1 , u2 , u3 }, pU = h iT 0.698 0.1538 0.1481 and pY |ui = pi , ∀i ∈ [1 : 3]. Finally, p∗U |Y corresponds to the matrix P∗U |Y given as   0.9423 0.9074 0 0     P∗U |Y = 0.0577 0 1 0 .   0 0.0926 0 1 Remark 1. It can be verified that in the degenerate case of X ⊥ ⊥ Y , we have Null(PX|Y ) = Span{v1 , v2 , . . . , v|Y| }, or equivalently, S = P(Y). In this case, the extreme points of S have zero entropy. Therefore, the minimum value of H(Y |U ) is zero, and U = {u1 , u2 , . . . , u|Y| } with pU (ui ) = pY (yi ), ∀i ∈ [1 : |Y|] and pY |ui = ei , which is the ith extreme point of P(Y). As a result, g0 (X, Y ) = H(Y ), which is also consistent with the fact that U = Y is independent of X and maximizes I(Y ; U ). A. MMSE under perfect privacy Assume that instead of I(Y ; U ), the goal is to minimize E[(Y − U )2 ] under the perfect privacy constraint. This can be formulated as follows: min E[(Y − U )2 ], pU |Y : X−Y −U X⊥ ⊥U (14) where the expectation is according to the joint distribution pY,U . Obviously, an upperbound for (14) is Var[Y ], as one could choose U = E[Y ]. In what follows, we show that (14) has a similar solution to that of g0 (X, Y ). The only difference is that the realizations of U , i.e., the particular values of the elements in U, are irrelevant for the 9 solution of g0 (X, Y ), since only their mass probabilities have a role when evaluating I(Y ; U ), while the objective function in (14) takes into account both the pmf and the realizations of U . We can write   EU,Y [(Y − U )2 ] = EU EY |U [(Y − U )2 |U ]     ≥ EU EY |U (Y − E[Y |U ])2 |U Z = Var[Y |U = u]dF (u), (15) (16)  where (15) is a classical result from MMSE estimation [21]; and in (16), we have used Var[Y |U = u] = EY |U  2 U = u . Therefore, from (11) and (16), we have E[Y |U = u] min: E[(Y − U )2 ] ≥ p U |Y X−Y −U X⊥ ⊥U Y− Z min FU (·), pY |u ∈S: p dF (u)=pY U Y |u Var[Y |U = u]dF (u), (17) R where the equality holds if and only if E[Y |U = u] = u, ∀u ∈ U. Proposition 5. Var[Y |U = u] is a strictly concave function of pY |u . Proof. The proof is provided in Appendix C. From the concavity of Var[Y |U = u] in Proposition 5, we can apply the reasoning in the proof of Proposition 4, and conclude that in (17), it is sufficient to consider only the extreme points of S. Hence, the problem has two phases, where in phase one, the extreme points of S are found. For the second phase, denoting the extreme points of S by p1 , p2 , . . . , pK , (17) boils down to a standard linear program as follows. h i min Var1 Var2 . . . VarK · w w≥0 h i S.t. p1 p2 . . . pK w = pY , (18) where Vari (∀i ∈ [1 : K]) denotes Var[Y |U = u] under pi , i.e., when pY |u = pi . Finally, once the LP in (18) is solved, the realizations of the random variable U are set to equalize the expectations of Y under the corresponding distributions of those extreme points (pi ) of S with non-zero mass probability. For example, the problem in (14) for the pair (X, Y ) given in Example 1 is h i min 0.2194 0.6094 0.52754 0.2344 · w w≥0     1 0.675 0.1875 0 0   2  1  0.325   0 0.1563 0   w = 4 , s.t.    1  0 8 0.8125 0 0.625     1 0 0 0.8437 0.375 8 (19) 10 where the extreme points of S, i.e., pi (i ∈ [1 : 4]), are already known from Example 1. The minimum value of the standard LP in (24) is 0.2406, which is achieved by   0.7407       0 ∗ .  w =  0.0593   0.2 h Therefore, MMSE= 0.2406 (≤ Var[Y ] = 1.1094), pU = 0.7407 0.0593 iT and pY |u1 = p1 , pY |ui = 0.2 pi+1 , i = 2, 3. By setting u = E[Y |U = u], we obtain U = {u1 , u2 , u3 } = {1.325, 3.6874, 3.375}. B. Minimum probability of error under perfect privacy The objective of the optimization can be the error probability as min Pr{Y 6= U }. (20) pU |Y : X−Y −U X⊥ ⊥U Obviously, an upper bound for (20) is 1 − maxy pY (y) as one could choose U = arg maxy pY (y). For an arbitrary joint distribution on (Y, U ), we can write Pr{Y 6= U } = 1 − Pr{Y = U } Z =1− Pr{Y = u|U = u}dFU (u) ZU ≥1− max pY |U (y|u)dFU (u), U y where (21) holds with equality when u = arg maxy pY |U (y|u). Then, Z max pY |U (y|u)dFU (u). min: Pr{Y 6= U } ≥ 1 − max p U |Y X−Y −U X⊥ ⊥U FU (·), pY |u ∈S: p dF (u)=pY U Y |u R U y (21) (22) It can be verified that maxy pY (y) is convex in pY (·). Hence, following the same reasoning in the proof of Proposition 4, it is sufficient to consider only the extreme points of S in the optimization in (22). Therefore, the problem has two phases: in phase one, the extreme points of S are identified. For the second phase, denoting the extreme points of S by p1 , p2 , . . . , pK , the problem boils down to a standard linear program as follows: h i 1− max pm1 pm2 . . . pmK · w w≥0 h i s.t. p1 p2 . . . pK w = pY , (23) where pmi is the maximum element of the vector pi , i ∈ [1 : K]. Once the LP is solved, the optimal conditionals pY |u and the optimal mass probabilities of U are obtained. Finally, the realizations of the random variable U are 11 set to equalize u = arg maxy pY |u (y|u). For example, the problem in (20) for the pair (X, Y ) given in Example 1 is 1− max w≥0 h 0.675  0.675   0.325 s.t.    0  0 0.8125 0.1875 0 0.8125 0 i 0.625 · w    1 0 0 2  1    0.1563 0   w = 4 , 1  8 0 0.625    1 0.8437 0.375 8 0.8437 (24) where the extreme points of S, i.e. pi (i ∈ [1 : 4]), are already known from Example 1. The minimum probability of error is obtained as 0.2789 (≤ 1 − maxy pY (y) = 12 ) achieved by   0.698       0.1538 . w∗ =    0.1482   0 h iT Hence, pU = 0.698 0.1538 0.1482 with U = {u1 , u2 , u3 } = {1, 3, 4}, and pY |ui = pi , i ∈ [1 : 3]. Thus far, we have investigated the constraint of perfect privacy when |X |, |Y| < ∞. The next theorem and its succeeding example consider two cases in which at least one of |X | and |Y| is infinite. The following theorem shows that perfect privacy is not feasible for the (correlated) jointly Gaussian pair. Theorem 2. Let (X, Y ) ∼ N (µ, Σ) be a pair of jointly Gaussian random variables, where     2 µX σX ρσX σY , µ =  ,Σ =  µY ρσX σY σY2 (25) in which ρ 6= 0, since otherwise X ⊥ ⊥ Y . We have g0 (X, Y ) = 0 for the above pair. Proof. If there exists a random variable U such that X − Y − U form a Markov chain and X ⊥ ⊥ U , we must have FX (·) = FX|U (·|u), ∀u ∈ U, and hence, fX (·) = fX|U (·|u), ∀u ∈ U, since X has a density. Equivalently, we must have Z fX (·) = fX|Y (·|y)dFY |U (y|u), ∀u ∈ U. (26) Also, to have g0 (X, Y ) > 0, there must exists at least u1 , u2 ∈ U, such that FY |U (·|u1 ) 6= FY |U (·|u2 ). (27) In what follows we show that if (26) holds, (27) cannot be satisfied; and therefore, perfect privacy is not feasible for a jointly Gaussian (X, Y ) pair. It is known that X conditioned on {Y = y} is also Gaussian, given by   ρσX 2 2 X|{Y = y} ∼ N (y − µY ) + µX , (1 − ρ )σX . | {z } σ | Y {z } σ2 αy+β (28) 12 From (26), (28), and for u1 , u2 ∈ U, we have Z fX (x) = Z = or, equivalently, Z (x−αy−β)2 2σ 2 e− √ (x−αy−β)2 2σ 2 e− √ 2πσ 2 (x−αy−β)2 2σ 2 e− √ 2πσ 2 dFY |U (y|u1 ) dFY |U (y|u2 ), ∀x ∈ R,   d FY |U (y|u1 ) − FY |U (y|u2 ) = 0, ∀x ∈ R. 2πσ 2 (29) Multiplying both sides of (29) by ejωx , and taking the integral with respect to x, we obtain 2  Z − (x−αy−β)   Z 2σ 2 e jωx √ e d FY |U (y|u1 ) − FY |U (y|u2 ) dx = 0. 2πσ 2 By Fubini’s theorem3 , we can write (x−αy−β)2    Z Z − 2σ 2 jωx e √ e dx d FY |U (y|u1 ) − FY |U (y|u2 ) = 0. 2πσ 2 After some manipulations, we get Z e jωαy   d FY |U (y|u1 ) − FY |U (y|u2 ) = 0. (30) Since ρ 6= 0, from (28), we have α 6= 0. Hence, the LHS of (30) is a Fourier transform. Due to the invertiblity R of the Fourier transform, i.e. ejωt dg(t) = 0 ⇐⇒ dg(t) = 0, we must have FY |U (·|u1 ) = FY |U (·|u2 ). Therefore, (27) does not hold and perfect privacy is not feasible for the (correlated) jointly Gaussian pair (X, Y ). In the following example, we consider |X | = 2 and |Y| = ∞. We observe that we can have bounded I(X; Y ) and h(Y ), while an unbounded g0 (X, Y ) without even revealing Y undistorted. This renders the usage of mutual information as a measure of dependence counterintuitive for continuous alphabets. This is related to the fact that differential entropy cannot be interpreted as a measure of the information content in a random variable, as it can take negative values. Example 2. Let X = {x0 , x1 } with pX (x0 ) = p. Let Y |{X = x0 } ∼ Uniform[0, 1] and Y |{X = x1 } ∼ Uniform[0, 2]. It can be verified that the probability density  1+p     2 1−p fY (y) = 2     0 function (pdf) of Y is y ∈ [0, 1] y ∈ [1, 2] , (31) y 6∈ [0, 2] and the conditional pmf of X conditioned on Y is given by pX|Y (x0 |y) = 3 Note that R   2p 1+p y ∈ [0, 1]  0 y 6∈ [0, 1] |fX|U (x|u1 ) − fX|U (x|u2 )|dx ≤ R , pX|Y (x1 |y) =      1−p 1+p 1 y ∈ [1, 2] .     0 y 6∈ [0, 2] [|fX|U (x|u1 )| + |fX|U (x|u2 )|]dx = 2 < +∞. y ∈ [0, 1] 13 Since the support of Y is the interval [0, 2], the support of Y |{U = u} must be a subset of [0, 2], ∀u ∈ U. Also, the independence of X and U in X − Y − U implies  Z  pX|U (x0 |u) = p, ∀u ∈ U =⇒ Pr 0 ≤ Y ≤ 1 U = u = 1 dFY |U (y|u) = 0 1+p , ∀u ∈ U. 2 Finally, in order to preserve the pdf of Y in (31), the conditional CDF FY |U (·|u) must satisfy the following Z FY |U (·|u)dFU (u) = FY (·), (32) U where FY (·) is the CDF corresponding to the pdf in (31). Let F be the set of all CDFs defined on [0, 2] and F be defined as   Z 1 1+p dFY (y) = . F , FY ∈ F 2 0 We can write g0 (X, Y ) = sup I(Y ; U ). (33) FU (·), FY |U (·|u)∈F, ∀u∈U , R F (·|u)dFU (u)=FY (·) U Y |U In what follows, we show that the supremum in (33) is unbounded. Let M be an arbitrary positive integer. Construct 1 , ∀i ∈ [1 : M ]. Also, let the joint distribution pY,U as follows. Let U = {u1 , u2 , . . . , uM }, with pU (ui ) = M  1+p i   M y ∈ [ i−1  M , M]  2 1−p i fY |U (y|ui ) = , ∀i ∈ [1 : M ]. y ∈ [1 + i−1 2 M M ,1 + M]     0 otherwise It can be verified that FY |U (·|ui ) ∈ F, ∀i ∈ [1 : M ], and (32) is also satisfied. We have h(Y ) = Hb ( 1+p 2 ) and h(Y |U ) = Hb ( 1+p 2 ) − log2 M . Hence, with this construction, we have I(Y ; U ) = log2 M. Since this is true for any positive integer M , we conclude that g0 = +∞, while I(X; Y ) = Hb ( 1+p 2 ) − 1 + p is finite. Therefore, letting U be a quantized version of Y results in an unbounded g0 . When we consider MMSE as the utility function, we have to solve the following problem Z Var[Y |U = u]dF (u). min FU (·), FY |U (·|u)∈F, ∀u∈U , R (·|u)dFU (u)=FY (·) F U Y |U Since Var[Y |U = u] is a strictly concave function of FY |U (·|u), it is sufficient to consider the optimization over the extreme points of F, which are the distributions concentrated at two mass points; one in the interval 1+p 2 , and the other one in 1−p = 1+p 2 s(y − y1 (u)) + 2 s(y 1−p 2 . [0, 1] with mass probability the interval [1, 2] with mass probability words, we have FY |U (y|u) − y2 (u)), where s(·) denotes the unit step function, and y1 (u) ∈ [0, 1], y2 (u) ∈ [1, 2] denote the two aforementioned mass points as a function of u. A simple analysis shows that Var[Y |U = u] =  2 1 − p2 y1 (u) − y2 (u) . 4 In other 14 Therefore, Z Var[Y |U = u]dF (u) min FU (·), FY |U (·|u)∈F, ∀u∈U , R F (·|u)dFU (u)=FY (·) U Y |U = min R Z  1 − p2 4 Z min FU (·), y1 (u)∈[0,1],y2 (u)∈[1,2],∀u∈U 1−p [ 1+p 2 s(y−y1 (u))+ 2 s(y−y2 (u))]dFU (u)=FY U 2 = 1−p ∀u∈U , 4 min 1 Z Z FU (·), FY |U (·|u)∈F, R F (·|u)dFU (u)=FY (·) U Y |U 0 2 y1 (u) − y2 (u) FU (·), y1 (u)∈[0,1],y2 (u)∈[1,2],∀u∈U 1−p [ 1+p 2 s(y−y1 (u))+ 2 s(y−y2 (u))]dFU (u)=FY (y) U ≥ R 1 − p2 4 dF (u) 2 [y1 (u) − y2 (u)]dF (u) (34) (y) 2y dFY |U (y|u) − 1+p Z 1 2  2 2y dFY |U (y|u) dF (u) 1−p (35)  2 1 − p2 = E[Y |0 ≤ Y ≤ 1] − E[Y |1 ≤ Y ≤ 2] 4 1 − p2 1 3 2 ( − ) 4 2 2 1 − p2 = , 4 = (36) R1 where (34) is due to the convexity of x2 and Jensen’s inequality, and (35) is from the fact that 0 ydFY |U (y|u) = R2 1+p 1−p 2 y1 (u) and 1 ydFY |U (y|u) = 2 y2 (u).   3−p In order to achieve the minimum in (36), we proceed as follows. Let U ∼ Uniform 1−p , and 2 , 2 FY |U (y|u) = 1−p 1−p 1+p 1+p s(y − u + )+ s(y − u + ), 2 2 2 2 which means that Y |{U = u} has two mass points at u − and 1−p 2 , 1−p 2 ∈ [0, 1] and u + 1+p 2 ∈ [1, 2] with probabilities (37) 1+p 2 respectively. It can then be verified that (37) satisfies (32). Finally, the following equalities hold. 1 − p2 4   1−p 3−p E[Y |U = u] = u, ∀u ∈ , . 2 2 Var[Y |U = u] = Therefore, we have MMSE = 1−p2 4 < 1 3 − p2 4 = Var[Y ]. When we consider the probability of error as the utility function, we have to solve the following problem 1− max FU (·), FY |U (·|u)∈F, ∀u∈U , R F (·|u)dFU (u)=FY (·) U Y |U Pr{Y = U }. It can be verified that similarly to the analysis when MMSE is the utility function, we can restrict our attention to the extreme points of F and obtain the minimum error probability of and FY |U (y|u) = 1+p 2 s(y − u) + 1−p 2 s(y 1−p 2 , which is achieved by U ∼ Uniform[0, 1] − u − 1). IV. N ON - PRIVATE INFORMATION VS . g0 (X, Y ) For a pair of random variables (X, Y ) ∈ X × Y (|X |, |Y| < ∞), the private information about X carried by Y is defined in [15] as CX (Y ) , min W :X→W →Y, H(W |Y )=0 H(W ). (38) 15 Since H(W |Y ) = 0 implies that W is a deterministic function of Y , (38) means that among all the functions of Y that make X and Y conditionally independent, we want to find the one with the lowest entropy. It can be verified that I(X; Y ) ≤ CX (Y ) ≤ H(Y ), where the first inequality is due to the data processing inequality applied on the Markov chain X − W − Y , i.e., I(W ; Y ) ≥ I(X; Y ), and the second inequality is a direct result of the fact that W = Y satisfies the constraints in (38). The non-private information about X carried by Y is defined in [15] as DX (Y ) , H(Y ) − CX (Y ). (39) Let T X : Y → P(X ) be a mapping from Y to the probability simplex on X defined by y → pX|Y (·|y). It was shown in [15, Theorem 3] that the minimizer in (38) is W ∗ = T X (Y ); and hence, DX (Y ) = H(Y ) − H(T X (Y )). (40) Furthermore, it was proved in [15, Lemma 5] that CX (Y ) = H(Y ), i.e., DX (Y ) = 0, if and only if there do not exist y1 , y2 ∈ Y such that pX|Y (·|y1 ) = pX|Y (·|y2 ). In [15], g0 (X, Y ) and DX (Y ) were loosely connected to each other, as the latter represents roughly the amount of information contained in Y and not correlated with X. Three examples were provided, where in two of them g0 (X, Y ) = DX (Y ), while in the last one g0 (X, Y ) > DX (Y ). Finally a question was raised regarding the condition on the joint distribution pX,Y under which g0 (X, Y ) = DX (Y ) holds. In what follows, we characterize the relation between DX (Y ) and g0 (X, Y ). Let PX|Y denote the matrix corresponding to the conditional distribution pX|Y . From [15, Lemma 5], if DX (Y ) > 0, we must have at least y1 , y2 ∈ Y such that their corresponding columns in PX|Y are equal. Let E1 ⊂ [1 : |Y|] be a set of (at least two) indices corresponding to the columns in PX|Y that are equal, i.e., pX|yi = pX|yj , ∀i, j ∈ E1 , i 6= j, and pX|yi 6= pX|yk , ∀i ∈ E1 , ∀k ∈ [1 : |Y|]\E1 . We can generalize this definition if the matrix PX|Y has several subsets of identical columns. Hence, the corresponding index sets are denoted by Em , ∀m ∈ [1 : B], for some B ≥ 1. In other words, pX|yi = pX|yj , ∀i, j ∈ Em , i 6= j, and pX|yi 6= pX|yk , ∀i ∈ Em , ∀k ∈ [1 : |Y|]\Em , ∀m ∈ [1 : B]. Let G , PB i=1 |Ei |. For the |X | × |Y|-dimensional matrix PX|Y that has the index sets Em , ∀m ∈ [1 : B], we construct a corresponding |X | × (|Y| − G + B)-dimensional matrix P̂X|Y from PX|Y by eliminating all the columns in each Em , except one. For example,  0.3 0.3 0.4   PX|Y = 0.2 0.2 0.5  0.5 0.5 0.1 we have the following pair    0.5 0.4 0.3 0.4 0.5       0.5 0.5 and P̂X|Y = 0.2 0.5 0.5 ,    0 0.1 0.5 0.1 0 where B = 2, G = 4, E1 = {1, 2}, and E2 = {3, 5}. (41) 16 Theorem 3. For a pair of random variables (X, Y ) ∈ X × Y (|X |, |Y| < ∞) distributed according to pX,Y , we have g0 (X, Y ) ≥ DX (Y ), (42) where the equality holds if and only if either of the following holds: 1) Perfect privacy is not feasible, i.e. dim(Null(PX|Y )) = 0, 2) Perfect privacy is feasible, and dim(Null(P̂X|Y )) = 0. In other words,   |Y| − |X | ≤ G − B, and σi (P̂X|Y ) 6= 0, ∀i ∈ 1 : |Y| − G + B . Proof. It is obvious that when there exist no y1 , y2 ∈ Y such that pX|Y (·|y1 ) = pX|Y (·|y2 ), we have DX (Y ) = 0, and (42) holds from the non-negativity of g0 . Assume that there exist index sets Em , ∀m ∈ [1 : B], corresponding to equal columns of PX|Y , as defined before. Since pX|yi = pX|yj , ∀i, j ∈ Em , ∀m ∈ [1 : B], we have T X (yi ) = T X (yj ), ∀i, j ∈ Em , ∀m ∈ [1 : B]. Hence, T X (Y ) is a random variable whose support has the cardinality |Y| − G + B and whose mass probabilities are the elements of the following set X    X X pY (yi ), pY (yi ), . . . , E pY (yi ) ∪ pY (yi ) i 6∈ ∪B m=1 m . i∈E1 Let S0 be a set of QB i=1 i∈E2 (43) i∈EB |Ei | probability vectors on the simplex P(Y) given by   B Y 0 Ei , S = sm[1:B] ∀m[1:B] ∈ (44) i=1 where the tuple (m1 , m2 , . . . , mB ) is written in short as m[1:B] and the probability vectors sm[1:B] are defined element-wise as  P    t∈Ei pY (yt ) k = mi , i ∈ [1 : B] B  Y sm[1:B] (k) = , ∀k ∈ [1 : |Y|], ∀m ∈ Ei . pY (yk ) k∈ 6 ∪B E [1:B] i=1 i   i=1   0 otherwise (45) Proposition 6. For the set S0 in (44) and the set S in (10), we have S0 ⊆ S, and H(s) = H(T X (Y )), ∀s ∈ S0 . Furthermore, the probability vector pY can be written as a convex combination of the points in S0 , i.e. pY = X m[1:B] ∈ where αm[1:B] ≥ 0, ∀m[1:B] ∈ QB i=1 αm[1:B] sm[1:B] , QB i=1 Ei Ei and X Q m[1:B] ∈ B i=1 Ei Proof. The proof is provided in Appendix D. αm[1:B] = 1. (46) 17 h For example, assume that in the example in (41), we have pY = 0.1 pY = 1 9 s1,3 + 2 9 s1,5 + 2 9 s2,3 + 0.15 0.25 0.3 iT . We can write 4 9 s2,5 ,  s1,3 0.2 0.3 where   0.3   0   0                   0.3   0.3   0   0                  = 0.45 , s1,5 =  0  , s2,3 = 0.45 , s2,5 =  0  .                 0.25 0.25 0.25 0.25         0.45 0 0.45 0 Finally, we can write g0 (X, Y ) = H(Y ) − H(Y |U ) (47) αm[1:B] H(sm[1:B] ) (48) αm[1:B] H(T X (Y )) (49) min FU (·), pY |u ∈S, ∀u∈U : R p dF (u)=pY U Y |u ≥ H(Y ) − X Q m[1:B] ∈ B i=1 Ei = H(Y ) − X Q m[1:B] ∈ B i=1 Ei = H(Y ) − H(T X (Y )) = DX (Y ), (50) where (47) is from (12); (48) is justified as follows. According to Proposition 6, S0 ⊆ S, and pY is preserved from (46). Hence, the vectors in S0 belong to the constraint of the minimization in (47), and the inequality follows. (49) is from Proposition 6, and (50) is due to (40). This proves (42). For the proof of the necessary and sufficient condition of equality in (42), first, we prove the second direction, i.e. the sufficient condition. If perfect privacy is not feasible, we have g0 (X, Y ) = 0, and the equality is immediate. Assume that perfect privacy is feasible and there exist index sets Em , ∀m ∈ [1 : B], corresponding to equal columns of PX|Y , as defined before. Proposition 7. If dim(Null(P̂X|Y )) = 0, the extreme points of the convex polytope S, defined in (10), are the elements of S0 . If dim(Null(P̂X|Y )) 6= 0, none of the elements in S0 is an extreme point of S. Proof. The proof is provided in Appendix E . Now, if dim(Null(P̂X|Y )) = 0, from Proposition 7, we can say that for any vector s that is an extreme point of S, we have H(s) = H(T X (Y )), which means, from Proposition 4, that min FU (·), pY |u ∈S, ∀u∈U : R p dF (u)=pY U Y |u H(Y |U ) = H(T X (Y )). This is equivalent to g0 (X, Y ) = DX (Y ), from (12) and (40). For the first direction, i.e. the necessary condition of equality in (42), assume that g0 (X, Y ) = DX (Y ). If perfect privacy is not feasible, there is nothing to prove, i.e. g0 (X, Y ) = DX (Y ) = 0. If perfect privacy is feasible, i.e. g0 (X, Y ) > 0, we must have DX (Y ) > 0, which from [15], means that there must exist index sets Em , ∀m ∈ [1 : B] , corresponding to equal columns of PX|Y . We prove that dim(Null(P̂X|Y )) = 0 by contradiction. Assume that 18 dim(Null(P̂X|Y )) 6= 0. From Proposition 7, we conclude that none of the elements in S0 is an extreme point of S. In other words, for any s in S0 , we can find the triplet (s0 , s00 , β), such that s = βs0 + (1 − β)s00 , where s0 , s00 ∈ S (s0 6= s00 ) and β ∈ (0, 1). Therefore, H(T X (Y )) = X αm[1:B] H(sm[1:B] ) Q m[1:B] ∈ B i=1 Ei = X   αm[1:B] H βm[1:B] s0m[1:B] + (1 − βm[1:B] )s00m[1:B] Q m[1:B] ∈ B i=1 Ei > X Q m[1:B] ∈ B i=1 Ei ≥ X βm[1:B] αm[1:B] H(s0m[1:B] ) + min FU (·), pY |u ∈S, ∀u∈U : R p dF (u)=pY U Y |u m[1:B] ∈ (1 − βm[1:B] )αm[1:B] H(s00m[1:B] ) (51) QB i=1 Ei H(Y |U ), (52) where (51) is due to the strict concavity of the entropy; (52) comes from the fact that s0m[1:B] and s00m[1:B] with QB corresponding mass probabilities βm[1:B] αm[1:B] and (1 − βm[1:B] )αm[1:B] , ∀m[1:B] ∈ i=1 Ei , belong to the constraints of minimization in (52). This results in g0 (X, Y ) > DX (Y ), which is a contradiction. Hence, we must have dim(Null(P̂X|Y )) = 0. V. F ULL DATA OBSERVATION VS . OUTPUT PERTURBATION Thus far, we have assumed that the privacy mapping/data release mechanism takes Y as input and maps it to the released data denoted by U , where X − Y − U form a Markov chain and the privacy mapping is captured by the conditional distribution pU |Y . In a more general scenario, the privacy mapping can take a noisy version W of (X, Y ) as input, as in [22]. In this case, the privacy mapping is denoted by pU |W , and (X, Y ) → W → U form a Markov chain, where the triplet (X, Y, W ) ∈ X × Y × W (X , Y, W < ∞) is distributed according to some given joint distribution pX,Y,W . In this model, perfect privacy is feasible for the triplet (X, Y, W ) if there exists a privacy mapping pU |W whose output (U ) depends on the useful data (Y ), while being independent of the private data (X); that is, I(Y ; U ) > 0 and I(X; U ) = 0 as before. Proposition 8. Perfect privacy is feasible for (X, Y, W ) if and only if   dim Null(PX|W )\Null(PY |W ) 6= 0. (53) Proof. The proof follows similarly to that of Proposition 2 by noting that both X − (X, Y ) − W − U and Y − (X, Y ) − W − U form Markov chains. In other words, there must exist a vector in P(W), such that a change in pW along that vector changes pY , while keeps pX unchanged. It can be verified that for the general scenario of (X, Y ) − W − U , where the mapping is denoted by pU |W , perfect privacy can be obtained through a similar LP as in Theorem 1 with the following modifications: The convex 19 polytope S is modified as the set of probability vectors x in P(W), such that (pW − x) ∈ Null(PX|W ); denoting the the extreme points of S by p1 , p2 , . . . , pK , (13) changes to h max I(Y ; U ) = H(Y )− min H(PY |W p1 ) H(PY |W p2 ) . . . pU |W : w≥0 i H(PY |W pK ) · w (X,Y )−W −U X⊥ ⊥U h s.t. p1 p2 ... i pK w = pW . The special cases of full data observation and output perturbation ([22]) refer to scenarios where the privacy mapping has direct access to both the private and useful data (W = (X, Y )) and only the useful data (W = Y ), respectively. With these definitions, Sections II to IV consider the particular case of output perturbation. In what follows, we consider the full data observation scenario briefly. Proposition 9. If Y is not a deterministic function of X, perfect privacy is always feasible in the full data observation model. Proof. If Y is not a deterministic function of X, there must exist x1 ∈ X and y1 , y2 ∈ Y (y1 6= y2 ) such that pX,Y (x1 , y1 ) > 0 and pX,Y (x1 , y2 ) > 0. Let U = {u1 , u2 } and pU (u1 ) = 21 . Choose a sufficiently small  > 0 and let    p (x , y ) +    X,Y 1 1 pX,Y |U (x, y|u1 ) = pX,Y (x1 , y2 ) −      p (x, y) X,Y (x, y) = (x1 , y1 ) (x, y) = (x1 , y2 ) , otherwise pX,Y |U (x, y|u2 ) = 2pX,Y (x, y) − pX,Y |U (x, y|u1 ), ∀(x, y) ∈ X × Y. (54) It can be verified that pX,Y is preserved in pX,Y,U . Also, pX|U (·|u) = pX (·), ∀u ∈ U, and pY |U (y1 |u1 ) 6= pY (y1 ), where the former indicates that X ⊥ ⊥ U , and the latter shows that Y 6⊥ ⊥ U. Considering the output perturbation model, Theorem 2 proved that perfect privacy is not feasible for the (correlated) jointly Gaussian pair. The following theorem states the opposite for the full data model. Theorem 5. For a jointly Gaussian pair (X, Y ) with correlation coefficient ρ (6= 0), perfect privacy is feasible for the full data observation model, and we have sup I(Y ; U ) ≥ − log ρ. (55) FU |X,Y :X⊥ ⊥U 2 Proof. Denoting the variances of X and Y by σX and σY2 , respectively, it is already known that we can write Y = p ρσY X + σY 1 − ρ2 N, σX (56) 20 where N ∼ N (0, 1) is independent of X. By letting U = √1 σY 1−ρ2 (Y − ρσY σX X), i.e. U = N , we have X ⊥ ⊥ U, and I(Y ; U ) = h(Y ) − h(Y |U )   p 1 ρσY log2 2πeσY2 − h X + σY 1 − ρ2 N N 2 σX 1 ρσ Y = log2 2πeσY2 − h( X) 2 σX 1 1 = log2 2πeσY2 − log2 2πeρ2 σY2 2 2 (58) = − log2 ρ, (59) = (57) where (56) is used in (57), and (58) follows from the fact that X ⊥ ⊥ N. VI. M AXIMAL CORRELATION Consider a pair of random variables (X, Y ) ∈ X × Y distributed according to pX,Y , with |X |, |Y| < ∞. Let F̃ denote the set of all real-valued functions of X, and define   F , f (·) ∈ F̃ E[f (X)] = 0, E[f 2 (X)] = 1 . Let G̃ and G be defined similarly for the random variable Y . The maximal correlation of (X, Y ) is defined as ([16], [17], [18]): ρm (X; Y ) = max E[f (X)g(Y )]. f ∈F ,g∈G If F (and/or G) is empty4 , then ρm is defined to be zero. An alternative characterization of the maximal correlation is given by Witsenhausen in [23]5 as follows. Let the matrix Q be defined as −1 −1 Q , PX 2 PX,Y PY 2 , (60) with singular values σ1 ≥ σ2 ≥ · · · . It is shown in [23] that σ1 = 1, and the maximal correlation of (X, Y ), i.e., ρm (X; Y ), is equal to the second largest singular value of matrix Q, i.e., σ2 . In what follows, we propose an alternative characterization of the maximal correlation, which also helps us interprete the other singular values of matrix Q. The following preliminaries from [25] are needed in the sequel. A. Preliminaries Assume that R is an n-by-n real symmetric matrix, and c is an n-dimensional vector satisfying kck2 = 1. Assume that we are interested in finding the stationary values of xT Rx, 4 When 5 For X (and/or Y ) is constant almost surely. other characterizations, see [24]. (61) 21 subject to the constraints cT x = 0, kxk2 = 1. (62) Letting λ and µ be the Lagrange multipliers, we have L(x, λ, µ) = xT Rx − λ(xT x − 1) + 2µxT c. Differentiating the Lagrangian with respect to x, we obtain Rx − λx + µc = 0, (63) which results in µ = −cT Rx, after multiplying both sides by cT and noting that kck2 = 1. By substituting this value of µ in (63), we are led to PRx = λx, where P = I − ccT . Since P is a projection matrix, i.e. P2 = P, the stationary values of xT Rx are the singular values of the matrix PR that occur at the corresponding eigenvectors. Finally, assume that the vector c in the constraints is replaced with an n × r matrix C with r ≤ n. Also, assume that the columns of matrix C are orthonormal. It can be verified that the results remain the same after having P modified as P = I − CCT . B. Alternative characterization of ρm (X; Y ) Consider a pair of random variables (X, Y ) ∈ X × Y (|X |, |Y| < ∞) distributed according to PX,Y , with the marginals pX and pY . The matrix PX|Y can be viewed as a channel with input Y and output X. When the input of this channel is distributed according to qY , the output is distributed according to qX = PX|Y qY . Let r : P(Y)\{pY } → [0, 1] be defined as r(qY ) , D(qX ||pX ) , D(qY ||pY ) (64) We write qY → pY when kqY − pY k2 → 0 and qY 6= pY . We are interested in finding the stationary values of r(qY ) when qY → pY . Theorem 6. The stationary values of (64), when qY → pY , are the squared singular values of matrix Q , −1 PX 2 PX,Y −1 PY 2 , and in particular, ρ2m (X; Y ) = lim η→0 sup qY : 0<kqY −pY k2 ≤η r(qY ). Proof. Having qY → pY , we can write qY = pY + , 1T|Y| ·  = 0, kk2 → 0,  6= 0, where  is an auxiliary vector. From the relationship qX = PX|Y qY , we have r(qY ) = D(pX + PX|Y ||pX ) . D(pY + ||pY ) (65) 22 Assume that p0 and p are two probability vectors in the interior of P(Y). Let p0 (·) and p(·) denote their corresponding probability mass functions. We can write the Taylor series expansion of the relative entropy as 1 D(p0 + ||p) = D(p0 ||p) + T · ∇D|p0 + T ∇2 D|p0  + · · · , 2 (66) where h 0 (y1 ) ∇D|p0 = log pp(y +1 1) h 1 ∇2 D|p0 = diag p (y ) 0 log p0 (y2 ) p(y2 ) 1 p0 (y2 ) 1 +1 ... ... 1 p0 (y|Y| log i , ) p0 (y|Y| ) p(y|Y| ) +1 iT , are the gradient and the Hessian of D(·||p) at p0 , respectively, and the higher order terms of  are denoted by dots in (66). Therefore, (65) boils down to r(qY ) = = D(pX ||pX ) + T PTX|Y 1|X | + T PTX|Y P−1 X PX|Y  + . . . D(pY ||pY ) + T .1|Y| + T P−1 Y  + ... −1 T PTX|Y PX PX|Y  + . . . T P−1 Y  + ... , (67) where we have used the facts that D(p||p) = 0, PTX|Y 1|X | = 1|Y| and T · 1|Y| = 0. When kk2 → 0, the higher order terms of  in (67), shown with dots, can be ignored. Hence, we are interested in finding the stationary values of T PTX|Y P−1 X PX|Y  T P−1 Y  , (68) when 1T|Y| ·  = 0,  6= 0. Note that the condition kk2 → 0 is redundant as the norm kk2 cancels out from both the numerator and the denominator of (68). We can equivalently write (68) as 1 1 2 vT PY2 PTX|Y P−1 X PX|Y PY v vT · v −1 , 1 where v , PY 2 , v 6= 0, cT .v = 0 with c = PY2 1|Y| , and it is obvious that kck2 = 1. Without loss of generality, we assume that kvk2 = 1. Therefore, we are led to finding the stationary values of vT Rv, (69) 1 1 2 where R = PY2 PTX|Y P−1 X PX|Y PY , subject to the constraints cT · v = 0, kvk2 = 1. (70) Note that R is a |Y|-by-|Y| real symmetric matrix, and c is a |Y|-dimensional vector satisfying kck2 = 1. Therefore, (69) is the same problem as in (61) whose stationary values are the eigenvalues of the matrix (I − ccT )R, which occur at their corresponding eigenvectors. We have 1 1 2 R = PY2 PTX|Y P−1 X PX|Y PY =  T   −1 −1 −1 −1 PX 2 PX,Y PY 2 PX 2 PX,Y PY 2 = QT Q, 23 where Q is defined in (60). Also, c is the eigenvector of R corresponding to the eigenvalue of 1, which follows from: 1 1 2 Rc = PY2 PTX|Y P−1 X PX|Y PY c 1 = PY2 PTX|Y P−1 X PX|Y PY 1|Y| 1 = PY2 PTX|Y P−1 X PX|Y pY 1 = PY2 PTX|Y P−1 X pX 1 = PY2 PTX|Y 1|X | 1 = PY2 1|Y| = c. Therefore, the eigenvalues of the matrix (I − ccT )R are λ1 = σ22 ≥ λ2 = σ32 ≥ · · · and 0, where σi ’s are the singular values of matrix Q, and hence, λ1 = ρ2m . This leads us to the following equality for the maximal correlation ρ2m (X, Y ) = lim sup qY →pY qY 6=pY D(qX ||pX ) . D(qY ||pY ) The other eigenvalues of (I − ccT )R (or equivalently, the other singular values of matrix Q, except the largest one) can be interpreted in a similar way. Assume that v1 is the maximizer of (69), i.e., v1 is the eigenvector of (I − ccT )R that corresponds to the eigenvalue λ1 = ρ2m . Equivalently, when qY → pY (qY 6= pY ), the ratio 1 in (64) is maximized if qY converges to pY in the direction of 1 = PY2 v1 . If besides the constraints in (70), we also impose the constraint that v should be orthogonal to v1 , i.e., replacing c by matrix C whose first and second columns are, respectively, c and v1 , the maximum of (69) would be λ2 = σ32 , achieved by its corresponding 1 eigenvector v2 . Equivalently, when qY → pY (qY 6= pY ) and (qY − pY ) ⊥ PY2 v1 , the ratio in (64) is maximized 1 if qY converges to pY in the direction of 2 = PY2 v2 . This procedure can be continued to cover all the singular values of matrix Q, from the second largest to the smallest. Remark 2. A natural question arises whether the largest singular value, which is one, has a similar interpretation. If in (70), the constraint cT · v = 0 is omitted, the maximum of (69) would be 1, which occurs at v = c. The constraint cT ·v = 0 is due to 1T|Y| · = 0, which in turn results from the fact that qY is a probability vector. Therefore, if the definition of relative entropy is extended to the vectors with positive elements, when qY → pY (qY 6= pY ) and qY can be any vector with positive elements, the ratio in (64) is maximized if qY converges to pY in the 1 direction of 0 = PY2 c = pY . In Sections III and IV of this paper, the problem of perfect privacy, i.e., the quantity g0 (X, Y ), is studied. Although the optimal utility-privacy trade-off curve is not straightforward analytically, it is of interest to check the behaviour of this curve when a small amount of private data leakage is allowed. To this end, we study the slope of this curve in the following section. 24 VII. T HE SLOPE OF THE UTILITY- PRIVACY TRADE - OFF REGION In this section, we consider the trade-off region g (X, Y ) vs.  as defined in (1). We are interested in evaluating the rate of increase in the utility when I(X; U ) ≈ 0. In other words, we focus on the slope of this trade-off curve at (0, g0 ). For the case when g0 = 0, this slope is obtained, while for the case g0 > 0, a lower bound is proposed. Let V ∗ ∈ [1, +∞] be defined as V ∗ , sup qY : qY 6=pY D(qY ||pY ) , D(qX ||pX ) (71) with the convention that if for some qY (6= pY ), we have qX = pX , then V ∗ = +∞. Proposition 10. We have g0 (X; Y ) = 0 if and only if V ∗ < +∞. Proof. The proof is provided in Appendix G. A. The slope at (0,0) Proposition 11. We have lim →0 g (X; Y ) = V ∗.  (72) Proof. When perfect privacy is feasible, it is already known that g0 (X, Y ) > 0 and the slope is infinity at the origin. This, along with Proposition 10, proves (72) when g0 (X, Y ) > 0. In what follows, we consider the situation where perfect privacy is not feasible, i.e. V ∗ < +∞. From the definition in (71), for a fixed δ > 0, ∃ q∗Y 6= pY : V ∗ − δ < D(q∗Y ||pY ) ≤ V ∗, D(q∗X ||pX ) where q∗X is induced by q∗Y , i.e. q∗X = PX|Y q∗Y . Let U = {u1 , u2 }. For sufficiently small ζ > 0, let pU (u1 ) = ζ, pY |u1 = q∗Y , pY |u2 = 1 (pY − ζq∗Y ), 1−ζ where this sufficiently small ζ makes pY |u2 a probability vector. We have P pU (u)D(pY |u ||pY ) I(Y ; U ) = P u∈U I(X; U ) u∈U pU (u)D(pX|u ||pX )   1 (1 − ζ)D 1−ζ (pY − ζq∗Y ) pY + ζD(q∗Y ||pY )   = . 1 ∗ (1 − ζ)D 1−ζ (pX − ζqX ) pX + ζD(q∗X ||pX ) (73) We are interested in inspecting the behaviour of (73), when ζ → 0. From the Taylor series expansion, we have     ζ 1 ∗ ∗ (pY − ζqY ) pY = D pY + (pY − qY ) pY D 1−ζ 1−ζ =0+0+ ζ2 (pY − q∗Y )T P−1 (pY − q∗Y ) + . . . , 2(1 − ζ)2 | {zY } Ky (74) 25 where Ky is a constant and the dots denote the higher order terms in the Taylor series expansion. A similar expansion can be written for the first term in the denominator of (73) with the corresponding constant denoted by Kx . Hence, we have I(Y ; U ) lim = lim ζ→0 I(X; U ) ζ→0 = ζ2 2(1−ζ) Ky ζ2 2(1−ζ) Kx + ζD(q∗Y ||pY ) + ζD(q∗X ||pX ) D(q∗Y ||pY ) D(q∗X ||pX ) (75) (76) > V ∗ − δ. (77) Since δ is chosen arbitrarily, we can write lim →0 g (X; Y ) ≥ V ∗.  (78) On the other hand, lim →0 I(Y ; U ) g (X; Y ) ≤ sup  I(X; U) U :X−Y −U I(X;U )>0 P = sup U :X−Y −U I(X;U )>0 ≤ sup qY 6=pY u∈U :pY |u 6=pY pU (u)D(pY |u ||pY ) u∈U :pY |u 6=pY pU (u)D(pX|u ||pX ) P (79) D(qY ||pY ) D(qX ||pX ) = V ∗, (80) where the assumption of pY |u 6= pY in the summations of (79) causes no loss of generality, as it only excludes the zero terms. From (78) and (80), (72) is proved. Remark 3. In the general observation model of Section V, i.e., (X, Y ) − W − U , where the privacy-preserving mapping is pU |W , it can be similarly verified that the slope at (0, 0) of the optimal utility-privacy trade-off curve, characterized by I(Y ; U ) vs. I(X; U ), is given by sup qW : qW 6∈Null(PY |W ) D(PY |W qW ||pY ) , D(PX|W qW ||pX ) (81) where the same convention of the definition in (71) holds; the probability vectors pW , pX , pY , and the matrices PY |W , PX|W are all derived from the joint distribution pX,Y,W . Remark 4. Let X ∼ Bernoulli(px ) (px ∈ (0, 12 ]) and Y is connected to X through a binary symmetric channel (BSC) with crossover probability α ∈ (0, 12 ]. Hence, Y is distributed as Y ∼ Bernoulli(py ) with py = px ∗ α, where p ∗ q = p(1 − q) + q(1 − p) (p, q ∈ [0, 1]) denotes the binary convolution. From [15, Lemma 1], g (X, Y ) ≤ H(Y ) − Hb (α ∗ Hb−1 (H(X) − )), (82) 26 where Hb (t) = −t log t − (1 − t) log(1 − t) is the binary entropy function, and Hb−1 : [0, 1] → [0, 21 ] is its inverse. From (82), we have H(Y ) − Hb (α ∗ Hb−1 (H(X) − )) g (X; Y ) ≤ lim →0 →0     d −1 = H(Y ) − Hb (α ∗ Hb (H(X) − )) d lim = (1 − 2α) log log 1−py py 1−px px (83) =0 , (84) where (83) is from the application of L’Hospital’s rule. According to [15], perfect privacy is not feasible for this (X, Y ), and we have g0 (X, Y ) = 0. Hence, g (X, Y ) D(qY ||pY ) = sup →0  qY 6=pY D(qX ||pX ) 1 = X ||pX ) inf qY 6=pY D(q D(qY ||pY ) lim (85) (86) 1 ≥ inf qY →pY qY 6=pY (87) D(qX ||pX ) D(qY ||pY ) 1 ≥ 2 , ρm (88) where (85) is from Proposition 11; (86) is permissible, since the ratios involved are bounded away from zero and infinity; (87) is a direct result of adding constraint to the infimum, and (88) is from the analysis after (64). Note that, in the specific case of this example where |Y| = 2, the inequality in (88) can be replaced by equality, as the standard 1-simplex has only one dimension. In other words, when qY → pY , the infimum and the supremum of the ratio are the same, i.e., 1 ρ2m . From (88) and (84), we must have 1−p log py y 1 ≤ (1 − 2α) . x ρ2m log 1−p p (89) x −1 −1 By a simple calculation of ρm , which is the second largest singular value of the matrix PX 2 PX,Y PY 2 , we get ρ2m = α2 + (1 − 2α)(px + py − 2px py ) − 1. py (1 − py ) (90) It is obvious that for a fixed px , when α → 21 ; we have py → 12 , and therefore, from (90), ρm → 0. This is intuitive, since having α → 12 , the pair (X, Y ) moves towards independence; and therefore, any correlation between them vanishes. As a result, the left hand side (LHS) of (89) becomes unbounded, while its right hand side (RHS) tends to zero. For example, with px = 0.6 and α = 0.45, the LHS is approximately 104.12, while its RHS is 0.0099, which makes (89) invalid. The reason for this phenomenon is that the upperbound in (82) does not hold in general, which is due to a subtle error in employing Mrs. Gerber’s lemma in [15, Lemma 1], in which, the conditional entropy H(Y |U ) is bounded below as H(Y |U ) = H(X ⊕ N |U ) ≥ Hb (α ∗ Hb−1 (H(X|U ))), (91) 27 where the crossover probability α is captured in the additive binary noise N ∼ Bernoulli(α), and (91) is due to Mrs. Gerber’s lemma. Then, H(X|U ) is replaced with H(X) − , since I(X; U ) = , to obtain the bound in (82). However, in the statement of Mrs. Gerber’s lemma [26], N must be independent of the pair (X, U ), while this is not necessarily the case here. Assume that, in the Markov chain X − Y − U , U is obtained by passing Y through a Z channel6 . Then, N is not independent of the pair (X, U ). Actually, it can be verified that for one realization of U , N becomes a deterministic function of X. Therefore, the application of Mrs. Gerber’s lemma is not permissible. Remark 5. It can be readily verified from Theorem 6 and Proposition 11 that 1 = lim δ→0 V∗ ρ2m (X; Y ) = lim δ→0 where in (92), we use the convention 1 ∞ sup U :X−Y −U EU [D (pX|U (·|U )||pX (·))]≤δ sup U :X−Y −U maxu D (pX|U (·|u)||pX (·))≤δ I(X; U ) , I(Y ; U ) (92) I(X; U ) , I(Y ; U ) (93)  = 0; in (92), the term EU [D pX|U (·|U )||pX (·) ] is equal to I(X; U ), and it is written like this to be comparable with the constraint of supremization in (93), i.e., an average constraint in (92) versus a per-realization constraint in (93). B. The slope at (0, g0 ) In the previous subsection this slope was obtained when g0 = 0. In what follows, we consider the case g0 > 0, and we are interested in finding g (X; Y ) − g0 . →0  lim (94) In the sequel, we propose a lower bound for (94). Assume that g0 , as obtained through an LP formulation in Theorem 1, is achieved by U 0 ∈ U 0 = {u01 , u02 , . . . , u0|U 0 | }, pY |u0 , ∀u0 ∈ U 0 , where the vectors pY |u0 , ∀u0 ∈ U 0 belong to the extreme points of the set S, defined in (10). Define ψ(u0 ) , sup qY : 0<D(qY ||pY |u0 )<+∞ D(qY ||pY |u0 ) , ∀u0 ∈ U 0 , D(qX ||pX ) (95) and if for some u0 , there is no qY for which 0 < D(qY ||pY |u0 ) < +∞ (which happens exactly when pY |u0 is a corner point of the probability simplex), then let ψ(u0 ) = 0. Proposition 12. We have ψ(u0 ) < +∞, ∀u0 ∈ U 0 . Proof. The proof is provided in Appendix H. Let L , max ψ(u0 ), 0 0 (96) u ∈U 6 The Z channel has binary input and output alphabets, and conditional pmf p(0|0) = 1, p(1|1) = p(0|1) = 1 . 2 28 and denote a/the maximizer of (96) by u0j for some j ∈ [1 : |U 0 |]. From (95) and (96), for a fixed δ, we have ∃ qY 6= pY |u0j , D(qY ||pY |u0j ) < +∞ : L − δ < D(qY ||pY |u0j ) D(qX ||pX ) ≤L (97) Construct the pair (Y, U ) as follows. Let U = {u1 , u2 , . . . , u|U 0 | , ûj }, and pU (ui ) = pU 0 (u0i ), pY |ui = pY |u0i , ∀i ∈ [1 : |U 0 |], i 6= j pU (uj ) = pU 0 (u0j ), pU (ûj ) = (1 − )pU 0 (u0j ), pY |uj = qY , pY |ûj = (98) 1 (pY |u0j − qY ). 1− (99) Note that for sufficiently small  > 0, pY |ûj is a probability vector, since we have D(qY ||pY |u0j ) < +∞. In other words, for any entry of the vector pY |u0j that is zero (note that it is an extreme point of S), the corresponding entry in qY is also zero. Finally, it can be verified that from (98) and (99), the marginal probability vector pY is also preserved. With this construction, we can have the Markov chain X − Y − U and P P 0 0 I(Y ; U ) − g0 u∈U pU (u)D(pY |u ||pY ) − u0 ∈U 0 pU (u )D(pY |u0 ||pY ) P = I(X; U ) u∈U pU (u)D(pX|u ||pX ) P 0 0 0 u∈{uj ,ûj } pU (u)D(pY |u ||pY ) − pU (uj )D(pY |uj ||pY ) P = u∈{uj ,ûj } pU (u)D(pX|u ||pX )   1 D(qY ||pY ) + (1 − )D 1− (pY |u0j − qY ) pY − D(pY |u0j ||pY )   , = 1 D(qX ||pX ) + (1 − )D 1− (pX − qX ) pX (100) (101) where the numerator in (100) is from (98); the denominator in (100) is from the fact that pX|ui = pX|u0i , ∀i ∈ [1 : |U 0 |], i 6= j and pX|u0 = PX|Y pY |u0 = pX , ∀u0 ∈ U 0 ; (101) is due to (99). We can write the Taylor series expansion for the second term in the numerator of (101) as     1  D (pY |u0j − qY ) pY = D pY |u0j + (pY |u0j − qY ) pY 1− 1−  log pY |U 0 (y1 |u0j ) pY (y1 ) pY |U 0 (y2 |u0j ) pY (y2 ) +1        log + 1   = D(pY |u0j ||pY ) + (pY |u0j − qY )T .    1−   ...   pY |U 0 (y|Y| |u0j ) log pY (y|Y| ) + 1 + 2 0 (pY |u0j − qY )T P−1 Y |u0j (pY |uj − qY ) + . . . , 2(1 − )2 | {z } Ky where PY |u0j = diag(pY |u0j ). Replacing the above in (101), the numerator becomes   0 p 0 (y1 |uj ) + 1 log Y |U pY (y1 )     pY |U 0 (y2 |u0j )   log + 1 2 p (y ) T  2 Y + D(qY ||pY ) + (1 − )D(pY |u0j ||pY ) + (pY |u0j − qY ) .   2(1 − ) Ky − D(pY |u0j ||pY ) + . . .   ...   pY |U 0 (y|Y| |u0j ) log pY (y|Y| ) + 1 which, after some manipulations, becomes equal to D(qY ||pY |u0j ) + 2 Ky + . . . . 2(1 − ) 29 Following similar steps, an expansion for the second term in the denominator of (101) can be obtained. Letting  → 0, and ignoring the higher order terms (denoted by the dots), we get 2  Ky D(qY ||pY |u0j ) + 2(1−) I(Y ; U ) − g0 = lim 2  →0 D(qX ||pX ) + I(X; U ) 2(1−) Kx = D(qY ||pY |u0j ) D(qX ||pX ) > L − δ, (102) where (102) is due to (97). Since δ > 0 was chosen arbitrarily, we get lim →0 g (X; Y ) − g0 ≥ L.  (103) On the other hand, construct the pair (Y, U ) as follows. Let U = {u1 , u2 , . . . , u|U 0 | , ũ1 , ũ2 , . . . , ũ|Y| }, and pU (ui ) = (1 − )pU 0 (u0i ), pY |ui = pY |u0i , ∀i ∈ [1 : |U 0 |] pU (ũi ) = pY (yi ), pY |ũi = ei , ∀i ∈ [1 : |Y|], where ei is an |Y|-dimensional vector denoting the ith extreme point of the probability simplex, i.e. e1 = h iT h iT 1 0 0 . . . , e2 = 0 1 0 . . . , and so on. It can be verified that the marginal probability vector pY is preserved. With this construction, we can have the Markov chain X − Y − U and P|Y| (1 − )g0 + i=1 pU (ũi )D(pY |ũi ||pY ) − g0 I(Y ; U ) − g0 = P|Y| I(X; U ) i=1 pU (ũi )D(pX|ũi ||pX ) P|Y|  pY (yi )D(ei ||pY ) − g0 = Pi=1 |Y|  i=1 pY (yi )D(pX|yi ||pX ) = H(Y ) − g0 , I(X; Y ) (104) where we have used the facts that D(ei ||pY ) = − log pY (yi ) and pX|ũi = PX|Y ei = pX|yi . Therefore, combining (103) and (104), we have   H(Y ) − g0 g (X; Y ) − g0 lim ≥ max L, . →0  I(X; Y ) Example 3. Assume X = {1, 2}, Y = {1, 2, 3}, and    0.2   0.5   pY = 0.5 , PX|Y =    0.5 0.3 0.3 0.6 0.7 0.4   =⇒ pX   0.43 . = 0.57 We have H(Y ) = 1.4855, I(X; Y ) = 0.0539, g0 (X; Y ) = 0.5147, where g0 is achieved, through an LP formulation, by U 0 ∈ {u01 .u02 } with pU 0 (u01 ) = 0.3077 and     0.65 0         pY |u01 = 0.35 , pY |u02 = 0.5667 .     0 0.4333 (105) 30 Fig. 1: A pictorial representation of Example 3. We can simply obtain ψ(·) as ψ(u0 ) = sup qY 6=pY |u0 D(qY ||pY |u0 )<+∞  D(qY ||pY |u0 )  43.52 u0 = u01 = ,  15.86 u0 = u0 D(qX ||pX ) 2 where ψ(u01 ) and ψ(u02 ) are achieved at qY = e1 and qY = e2 , respectively. This results in L = 43.52, and from (105), we have g (X; Y ) − g0 ≥ max {43.52, 18.011} = 43.52. →0  lim Figure 1 illustrates this example, where the triangle represents the probability simplex with the corner points ei , i ∈ [1 : 3]. v3 is the third right eigenvector of PX|Y , whose span is the null space of PX|Y . The line segment passing through pY in the direction of v3 is the convex polytope S with the two extreme points pY |u01 and pY |u02 . VIII. C ONCLUSIONS This paper addresses the problem of perfect privacy, where the goal is to find the maximum utility obtained through a (distorted) disclosure of available data Y , while guaranteeing maximum privacy for the private latent variable X. It is shown that this problem boils down to a standard linear program when the utility is measured by the mutual information between Y and its disclosed (distorted) version U . Similar results are obtained for other utility measures, in particular mean-square error and probability of error. It is shown that when the private variable and the observed data form a jointly Gaussian pair, utility can be obtained only at the expense of privacy when the data release mechanism has access only to the observed data Y . On the other hand, it is shown that, when 31 the privacy mapping has direct access to both the data Y and the latent variable X, perfect privacy is feasible. Measuring both the utility and privacy by mutual information, we have then investigated the slope of the optimal utility-privacy trade-off curve when the revealed and private data are independent, i.e., I(X; U ) = 0. Finally, we have proposed an alternative characterization of the maximal correlation between two random variables. A PPENDIX A Let U be an arbitrary set. Let P be the set of probability mass functions (pmf) on Y. Let r : P → R|Y|+1 be a vector-valued mapping defined element-wise as ri (pY |U (·|u)) = pY |U (yi |u), i ∈ [1 : |Y| − 1], r|Y| (pY |U (·|u)) = H(Y |U = u), r|Y|+1 (pY |U (·|u)) = H(X|U = u). (106) Since P corresponds to the standard (|Y|−1)-simplex, which is a closed and bounded subset of R|Y| , it is compact. Also, r is a continuous mapping from P to R|Y|+1 . Therefore, from the support lemma [26], for every U ∼ F (u) defined on U, there exists a random variable U 0 ∼ p(u0 ) with |U 0 | ≤ |Y| + 1 and a collection of conditional pmfs pY |U 0 (·|u0 ) ∈ P indexed by u0 ∈ U 0 , such that Z X ri (p(y|u))dF (u) = ri (p(y|u0 ))p(u0 ), i ∈ [1 : |Y| + 1]. U u0 ∈U 0 This means that for an arbitrary U with X − Y − U , the terms pY (·), I(Y ; U ) and I(X; U ) are preserved if U is 0 replaced withU 0 . Since we can simply have X − Y − U  , there is no loss  of optimality in considering |U|≤ |Y| + 1. Let P , pU |Y (·|·) U ∈ U, Y ∈ Y, |U| ≤ |Y| + 1 and Py , pU |Y (·|y) U ∈ U, |U| ≤ |Y| + 1 , ∀y ∈ Y. The set Py is the standard (|U| − 1)-simplex, and therefore compact (since |U| ≤ |Y| + 1 < ∞). The set 0 P  = ∪y∈Y Py is a finite (|Y| < ∞) union of these compact sets, which is still compact. Finally, the set P = pU |Y (·|·) ∈ P X − Y − U, I(X; U ) ≤  is a closed subset of P (due to the continuity of mutual information and closedness of the interval [0, ]), and therefore, it is also compact. Since I(Y ; U ) is a continuous mapping over P 0 , its supremum is achieved; and therefore, it is a maximum. This proves the first equality in (2). The second equality in (2) follows from the convexity of the objective function on P 0 , and the maximum occurs at an extreme point of P 0 , for which I(X; U ) = . A PPENDIX B It is already known that when g0 > 0, we have pY |u ∈ S, ∀u ∈ U. The reasoning in Appendix A can be modified as follows. P can be replaced with S, and the mapping r is modified as r : S → R|Y| , where the last constraint in (106) is removed, as for any element in S, we have H(X|U = u) = H(X). Therefore, we obtain the sufficiency of |U| ≤ |Y|. 32 Now, assume that the minimum in (12) is achieved by K(≤ |Y|) points in S. We prove that all of these K points must belong to the extreme points of S. Let p be an arbitrary point among these K points. p can be written as7 p= |Y| X αi pi , (107) i=1 where αi ≥ 0 (∀i ∈ [1 : |Y|]) and P|Y| i=1 αi = 1; points pi (∀i ∈ [1 : |Y|]) belong to the extreme points of S and pi 6= pj (i 6= j). From the concavity of entropy, we have H(p) ≥ |Y| X αi H(pi ), (108) i=1 where the equality holds if and only if all of the αi s but one are zero. From the definition of an extreme point, if p is not an extreme point of S, it can be written as in (107) with at least two non-zero αi s, which makes the inequality in (108) strict. However, this violates the assumption that the K points achieve the minimum. Hence, all of the K points of the minimizer must belong to the set of extreme points of S. A PPENDIX C Let pY |u be given as pY |u = λpY |u1 + (1 − λ)pY |u2 , where λ ∈ [0, 1]. It is obvious that for an arbitrary function b(·), E[b(Y )|U = u] = λE[b(Y )|U = u1 ] + (1 − λ)E[b(Y )|U = u2 ]. (109) Therefore,   2 U =u Var[Y |U = u] = E Y − E[Y |U = u]  2 = E[Y 2 |U = u] − E[Y |U = u] 2 = λE[Y |U = u1 ] + (1 − λ)E[Y |U = u2 ] − λE[Y |U = u1 ] + (1 − λ)E[Y |U = u2 ] (110) 2 2   2  2 ≥ λE[Y 2 |U = u1 ] + (1 − λ)E[Y 2 |U = u2 ] − λ E[Y |U = u1 ] − (1 − λ) E[Y |U = u2 ] (111)  = λE 2   2  Y − E[Y |U = u1 ] U = u1 + (1 − λ)E Y − E[Y |U = u2 ] U = u2 = λVar[Y |U = u1 ] + (1 − λ)Var[Y |U = u2 ], where (110) is due to (109); and (111) is due to the (strict) convexity of x2 . 7 The convex polytope S is an at most (|Y| − 1)-dimensional convex subset of R|Y| . Therefore, any point in S can be written as a convex combination of at most |Y| extreme points of S. 33 A PPENDIX D QB From the construction in (45), we have ∀m[1:B] ∈ i=1 Ei , |Y| X PX|Y sm[1:B] = pX|yk sm[1:B] (k) k=1 X = X pX|yk sm[1:B] (k) + pX|yk sm[1:B] (k) k6∈∪B i=1 Ei k∈∪B i=1 Ei ! X = pX|ymk X pY (yt ) X X = = X pX|yk pY (yk ) k6∈∪B i=1 Ei pX|yk pY (yk ) + X pX|yk pY (yk ) k6∈∪B i=1 Ei k∈∪B i=1 Ei |Y| X pX|yk pY (yk ) k6∈∪B i=1 Ei pX|yt pY (yt ) + k∈[1:B] t∈Ek X + t∈Ek k∈[1:B] = X pX|yk pY (yk ) k=1 = pX . Hence, s ∈ S, ∀s ∈ S0 , which means that S0 ⊆ S. From (45), the non-zero entries of any probability vector s ∈ S0 are the same as the elements of the set in (43). Therefore, H(s) = H(T X (Y )), ∀s ∈ S0 . Finally, let the set of |Y|-dimensional probability vectors {sm1 }m1 ∈E1 on Y be defined element-wise as    pY (yk ) ∀k 6∈ E1   P sm1 (yk ) = , ∀m1 ∈ E1 , ∀k ∈ [1 : |Y|]. k = m1 j∈E1 pY (yj )     0 k 6= m1 , k ∈ E1 By induction, define    s (y ) ∀k 6∈ En  n  m[1:n−1] k Y P sm[1:n] (yk ) = Ei , ∀n ∈ [2 : B], ∀k ∈ [1 : |Y|], , ∀m ∈ p (y ) k = m [1:n] Y j n j∈En   i=1   0 k 6= mn , k ∈ En (112) (113) where it can be verified that (113) and (45) are equivalent for n = B. By constructions in (112) and (113), we can, respectively, write pY = X m1 ∈E1 and X sm[1:n−1] = pY (ym1 ) sm1 , k∈E1 pY (yk ) P n−1 Y pY (ymn ) sm[1:n] , ∀m[1:n−1] ∈ Ei , ∀n ∈ [2 : B]. k∈En pY (yk ) i=1 P mn ∈En (114) (115) Therefore, pY can be written from (114) and (115) as pY = X pY (m1 )pY (m2 ) . . . pY (mB ) P P sm[1:B] . p k∈E1 Y (yk ) k∈E2 pY (yk ) . . . k∈EB pY (yk ) P m[1:B] ∈ QB i=1 Ei (116) 34 By letting B Y pY (m1 )pY (m2 ) . . . pY (mB ) P P , ∀m[1:B] ∈ Ei , k∈E1 pY (yk ) k∈E2 pY (yk ) . . . k∈EB pY (yk ) i=1 αm[1:B] = P we are led to (46). A PPENDIX E   Without loss of generality, by an appropriate labelling of the elements in Y, we can assume that E1 = 1 : |E1 | ,   E2 = |E1 | + 1 : |E1 | + |E2 | , and so on. Let the common column vector corresponding to Em be denoted by pm , i.e. pm , pX|yi , ∀i ∈ Em , ∀m ∈ [1 : B]. We can write   PX|Y = p1 , . . . , p1 , p2 , . . . , p2 , . . . , pB , . . . , pB , pX|yG+1 , pX|yG+2 , . . . , pX|y|Y| , | {z } | {z } | {z } |E1 | times |E2 | times |EB | times and P̂X|Y   = p1 , p2 , . . . , pB , pX|yG+1 , pX|yG+2 , . . . , pX|y|Y| . Define the vectors   e1m 0Pm−1 |Ei |  i=1     2  1  , em  =    −1   0(|Y|−Pm−1 |Ei |−2) i=1 where  Pm−1 i=1 0Pm−1 |Ei |    i=1   Pm−1 0   1   i=1 |Ei |    1  2       |Em |−1 1|Em |−1  |Em |−1 1  = , . . . , e =  m   , ∀m ∈ [1 : B], 2     −1       −1   0(|Y|−Pm |E |) i i=1 0(|Y|−Pm−1 |Ei |−3) i=1 (117) |Ei | = ∅ when m = 1. Proposition 13. Let   N , Span eim ∀i ∈ [1 : |Em | − 1], ∀m ∈ [1 : B] . We have N ⊆ Null(PX|Y ), (118) where (118) holds with equality if and only if dim(Null(P̂X|Y )) = 0. Proof. The proof is provided in Appendix F. If dim(Null(P̂X|Y )) = 0, any element in S0 is an extreme point of S. The reasoning is as follows. Note that S0 ⊆ S. Hence, it remains to show that no point of S0 can be written as a convex combination of two different points of S. Pick an arbitrary point s in S0 . It can be verified that no  > 0 and e ∈ N exist such that both s + e and s − e remain a probability vector. This is due to having a negative element in either or both of them. From Proposition 13, we have N = Null(PX|Y ) which in turn means that s cannot be written as a convex combination of two different points of S. Therefore, the elements in S0 belong to the extreme points of S. On the other hand, when dim(Null(P̂X|Y )) = 0, assume that there exists s∗ 6∈ S0 that is an extreme point of S. We show that this leads to a contradiction. Firstly, note that among the elements of s∗ that correspond to 35 Em , ∀m ∈ [1 : B], there must be at most one non-zero element. This is justified as follows. Assume that s∗i and s∗j (i 6= j) are two non-zero elements of s∗ , where i, j ∈ Em for some m ∈ [1 : B]. Construct the |Y|-dimensional vector f where fi = −fj = 1, and the remaining terms are zero. Obviously, f ∈ Null(PX|Y ), as PX|Y f = 0. Let  = min{s∗i , s∗j }. It is obvious that s∗ can be written as a convex combination of the vectors s∗ + f and s∗ − f , where both are in S. However, this contradicts the assumption of s∗ being an extreme point of S. Hence, among the elements of s∗ that correspond to Em , ∀m ∈ [1 : B], at most one element is non-zero. As a result, we can find ∗ ∗ 0 ∗ a point s ∈ S0 whose positions of its non-zero elements in ∪B i=1 Ei matches those of s . Since s 6∈ S , s must differ with this s in at least one element. Assume that for some m ∈ [1 : B], there exists j ∈ Em such that s∗j 6= sj . Then, the elements of ∆s = s∗ − s that correspond to Em are all zero, except ∆sj = s∗j − sj 6= 0. It can then be verified that ∆s cannot be written as a linear combination of the vectors in N, as no linear combination of the vectors eim , ∀i ∈ [1 : |Em | − 1] can produce a vector whose all the elements corresponding to Em are zero except one. Since dim(Null(P̂X|Y )) = 0, we have from Proposition 13 that N = Null(PX|Y ), which in turn means that s∗ − s 6∈ Null(PX|Y ). Therefore, s∗ 6∈ S, which is a contradiction. If s∗j = sj , ∀j ∈ ∪B i=1 Ei , then we must have s∗j 6= sj for some j ∈ [G + 1 : |Y|]. Still, ∆s cannot be written as a linear combination of the vectors in N, as for any vector n ∈ N, we have nk = 0, ∀k ∈ [G + 1 : |Y|]. This leads us to s∗ 6∈ S, which is again a contradiction. Therefore, we conclude that when dim(Null(P̂X|Y )) = 0, the extreme points of S are the elements of S0 . If dim(Null(P̂X|Y )) 6= 0, from Proposition 13, there must exist a non-zero vector v such that v ∈ Null(PX|Y ) and v 6∈ N. Pick an arbitrary point of S0 . In order to make the analysis simple, let the picked vector be ŝ0 , which is  P p (y ) Y i   i∈E1   0 |E1 |−1    P    i∈E2 pY (yi )      0 |E2 |−1     ..     .  ŝ0 , s1,|E1 |+1,|E1 |+|E2 |+1,...,G−|EB |+1 =   P  i∈EB pY (yi )      0|EB |−1       pY (yG+1 )      ..   .   pY (y|Y| ) From v, we can construct a non-zero vector ṽ ∈ Null(PX|Y ), as done in (119). Then, it is obvious that for sufficiently small  > 0, ŝ0 can be written as a convex combination of ŝ0 + ṽ and ŝ0 − ṽ, where both are in S. This shows that ŝ0 cannot be an extreme point of S. A similar approach8 can be applied to show that the other 8 The only difference is in constructing a vector ṽ, such that when a point of S0 is perturbed along the direction of ṽ, it still lies in S. This can be done by noting that it is sufficient to construct a ṽ whose position of zero elements in [1 : G] matches that of the arbitrary point from S. In a similar way that ṽ(∈ Null(PX|Y )) was constructed from v in (119), by using other orthogonal vectors in N, instead of eim , a new ṽ can be constructed whose position of zero elements in [1 : G] matches that of the arbitrary point from S. 36 points of S0 do not belong to the set of extreme points of S. Hence, from dim(Null(P̂X|Y )) 6= 0, we conclude that none of the elements in S0 is an extreme point of S. A PPENDIX F The fact that N ⊆ Null(PX|Y ) can be verified by observing that PX|Y eim = 0, ∀i ∈ [1 : |Em | − 1], ∀m ∈ [1 : B]. If dim(Null(P̂X|Y )) = 0, we must have N = Null(PX|Y ). If this is not true, from (118), there must exist a non-zero vector v such that v ∈ Null(P ) and v 6∈ N. Let v denote the ith element in v. We can write i X|Y P i∈E1 v+ B |EX m |−1 X m=1 i=1 i i αm em vi      0  |E1 |−1   P    i∈E2 vi     0   |E2 |−1    ..     .  = ṽ,  = P   i∈EB vi       0|EB |−1       vG+1      ..   .   v|Y| (119) i , ∀i ∈ [1 : |Em | − 1], ∀m ∈ [1 : B] are obtained uniquely, as where it can be verified that the coefficients αm the vectors eim are mutually orthogonal. Since v, eim ∈ Null(PX|Y ), ∀i ∈ [1 : |Em | − 1], ∀m ∈ [1 : B], we have ṽ ∈ Null(PX|Y ). Also, note that ṽ is a non-zero vector, since otherwise (119) would result in v ∈ N. Finally, from the structure of ṽ and P̂X|Y , we observe that P̂X|Y ṽ0 = PX|Y ṽ = 0, where ṽ0 is obtained from eliminating the zero vectors of ṽ, denoted by 0|Ei |−1 , ∀i ∈ [1 : B], in (119). Since ṽ is a non-zero vector, so must be ṽ0 . Hence, dim(Null(P̂X|Y )) 6= 0, which is a contradiction. Therefore, we must have N = Null(PX|Y ). If N = Null(PX|Y ), we must have dim(Null(P̂X|Y )) = 0. If this is not true, there exists a non-zero vector r̃0 such that P̂X|Y r̃0 = 0, and correspondingly a non-zero vector r̃ such that PX|Y r̃ = 0, where the relation between r̃0 and r̃ is similar to that between ṽ0 and ṽ in the previous paragraph. Therefore, we have r̃ ∈ Null(PX|Y ). However, it can be verified that due to the structure of the vectors eim , i.e. the positions of the zero and non-zero elements in [1 : G], r̃ cannot be written as a linear combination of the vectors eim . This results in r̃ 6∈ N, which is a contradiction, as we assumed N = Null(PX|Y ). This proves that dim(Null(P̂X|Y )) = 0. A PPENDIX G When g0 (X, Y ) = 0, there is no qY 6= pY , such that qX = pX , since otherwise we could have constructed a random variable U ∈ {u1 , u2 }, and a sufficiently small α > 0, such that pU (u1 ) = α, pY |u1 = qY , pY |u2 = 1 (pY − αqY ), 1−α where the sufficiently small α makes pY |u2 still a probability vector. With this construction, it can be verified that X − Y − U, X ⊥ ⊥ U and Y 6⊥ ⊥ U which contradicts g0 (X, Y ) = 0. Hence, the only way to have V ∗ > M, ∀M ∈ R 37 is through the existence of a sequence of distributions, i.e. {qnY }n , where qnY = 6  pY , ∀n and qnX → pX . Since perfect privacy is not feasible, we must have |Y| ≤ |X | and σi (PX|Y ) 6= 0, ∀i ∈ 1 : min{|X |, |Y|} . This means that in order to have qnX → pX , we must have qnY → pY . We know that when qY → pY (qY 6= pY ), the ratio in (64) is bounded below by the minimum eigenvalue of the matrix QT Q. If for an arbitrary non-zero vector e, we have Qe = 0, then we must have −1 −1 −1 1 Qe = PX 2 PX,Y PY 2 e = PX 2 PX|Y PY2 e = 0, | {z } e0  which is not possible, since e0 is a non-zero vector, and so is PX|Y e0 due to the fact that σi (PX|Y ) 6= 0, ∀i ∈ 1 :  min{|X |, |Y|} and |Y| ≤ |X |, i.e. the null space of PX|Y is only the all-zero vector. Therefore, the minimum eigenvalue of the matrix QT Q is bounded away from zero. Equivalently, the inverse of (64) is bounded above by the inverse of the minimum eigenvalue of QT Q. Hence, V ∗ < +∞. The proof of the second direction is immediate, since having g0 (X, Y ) > 0 leads to the existence of qY 6= pY , such that qX = pX , which in turn violates V ∗ < +∞. A PPENDIX H Firstly, note that for any point qY that satisfies 0 < D(qY ||pY |u0 ) < +∞, we have qY 6∈ S (i.e., qX 6= pX ), where S is defined in (10), since otherwise from the fact that D(qY ||pY |u0 ) < +∞, for sufficiently small , we can make the probability vector q0Y = 1 0 1− (pY |u − qY ) which also belongs to S, as PX|Y q0Y = pX . However, this violates the fact that pY |u0 is an extreme point of S, since it can be written as a convex combination of two points of S, i.e. qY and q0Y (qY 6= q0Y ). Alternatively, we can say that for any qY that satisfies 0 < D(qY ||pY |u0 ) < +∞, we have (qY − pY |u0 ) 6∈ Null(PX|Y ). Hence, the only way to have ψ(u0 ) possibly unbounded is through the existence of a sequence of distributions, i.e. {qnY }n , where 0 < D(qnY ||pY |u0 ) < +∞, ∀n and qnX → pX , which requires qnY converging to a point of S. Let Iu0 denote the set of indices corresponding to the zero elements of pY |u0 . Let T denote the set of probability vectors p, such that D(p||pY |u0 ) < +∞. In other words, T is the set of probability vectors whose elements corresponding to the indices in Iu0 are zero. Since T is a closed set, we conclude that if qnY (∈ T) converges to p0 (a point of S), p0 must also be in T, i.e. it satisfies D(p0 ||pY |u0 ) < +∞. If D(p0 ||pY |u0 ) > 0, from what mentioned before, we have (p0 − pY |u0 ) 6∈ Null(PX|Y ), which contradicts the fact that p0 ∈ S, hence, we must have p0 = pY |u0 . Therefore, it suffices to consider the following problem lim inf qY :qY →pY |u0 0<D(qY ||pY |u0 )<+∞ D(qX ||pX ) . D(qY ||pY |u0 ) Similarly to the analysis after (64), the above becomes equal to the minimum eigenvalue of Q̃T Q̃, where −1 1 Q̃ = PX 2 P̃X|Y PY2 |u0 , (120) and P̃X|Y is an |X |×(|Y|−|Iu0 |)-dimensional matrix obtained by eliminating the columns of PX|Y that correspond to the indices in Iu0 ; PY |u0 is a (|Y| − |Iu0 |) × (|Y| − |Iu0 |) diagonal matrix whose diagonal elements are the corresponding non-zero elements of pY |u0 . 38 In what follows, we show that the minimum eigenvalue of the matrix Q̃T Q̃ is bounded away from zero, since 1 otherwise there must exists a non-zero (|Y|−|Iu0 |)-dimensional vector ẽ, such that Q̃ẽ = 0. Let ẽ0 , PY2 |u0 ẽ. From −1 1 (120), since PY2 |u0 and PX 2 are full rank matrices, we must have ẽ0 ∈ Null(P̃X|Y ), and therefore according to Proposition 3, 1T|Y|−|I u0 | .ẽ0 = 0. Construct the |Y|- dimensional vector e0 as follows. Let its elements corresponding to the indices in Iu0 be zero, and its other terms be equal to the elements of ẽ0 . It is obvious that e0 ∈ Null(PX|Y ), since the elements of e0 corresponding to the columns of PX|Y that are not in P̃X|Y are zero and we have ẽ0 ∈ Null(P̃X|Y ). From Proposition 3, having e0 ∈ Null(PX|Y ) results in 1T|Y| .e0 = 0. For sufficiently small  > 0, let qY = pY |u0 + e0 . Since  6= 0 and e0 6= 0, we have D(qY ||pY |u0 ) > 0. Moreover, since the elements in qY corresponding to the indices in Iu0 are zero, we have D(qY ||pY |u0 ) < +∞. Therefore, from the reasoning at the beginning of this Appendix, we have qY 6∈ S. However, since e0 ∈ Null(PX|Y ) and PX|Y pY |u0 = pX , we have qY ∈ S, which is a contradiction. Therefore, the minimum eigenvalue of the matrix Q̃T Q̃ is bounded away from zero. This in turn means that the inverse of ψ(u0 ) is bounded away from zero, and therefore, ψ(u0 ) < +∞, ∀u0 . R EFERENCES [1] A. Narayanan and V. Shmatikov, “Robust de-anonymization of large sparse datasets,” in IEEE Symp. on Security and Privacy (SP), 2008, pp. 111–125. [2] X. Ding, L. Zhang, and W. Zhiguo, “A brief survey on de-anonymization attacks in online social networks,” in International Conf. on Computational Aspects of Social Networks (CASoN), 2010, pp. 611–615. [3] S. Kumar, W. Nilsen, M. Pavel, and M. Srivastava, “Mobile health: Revolutionizing healthcare through transdisciplinary research,” Computer, vol. 46, pp. 28–35, 2013. [4] J. Gomez-Vilardebo and D. Gündüz, “Smart meter privacy for multiple users in the presence of an alternative energy source,” IEEE Trans. on Information Forensics and Security, pp. 132–141, 2015. [5] A. Motahari, G. Bresler, and D. Tse, “Information theory of DNA shotgun sequencing,” IEEE Trans. on Information Theory, vol. 59, no. 10, pp. 6273–6289, Oct. 2013. [6] C. Dwork, F. McSherry, K. Nissim, and A. Smith, “Calibrating noise to sensitivity in private data analysis,” Theory of Cryptography, Springer, pp. 265–284, 2006. [7] L. Sweeney, “k-anonymity: A model for protecting privacy,” Intl. Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, vol. 10, no. 5, pp. 557–570, 2002. [8] A. Machanavajjhala, D. Kifer, J. Gehrke, and M. Venkitasubramaniam, “l-diversity: Privacy beyond k-anonymity,” ACM Trans. on Knowledge Discovery from Data, vol. 1. [9] N. Li, T. Li, and S. Venkatasubramanian, “t-closeness: Privacy beyond k-anonymity and l-diversity,” IEEE Intl. Conf. on Data Eng., 2007. [10] F. Calmon and N. Fawaz, “Privacy against statistical inference,” in 50th Annual Allerton Conference, Illinois, USA, Oct. 2012, pp. 1401– 1407. [11] A. Makhdoumi, S. Salamatian, N. Fawaz, and M. Médard, “From the information bottleneck to the privacy funnel,” in IEEE Information Theory Workshop (ITW), 2014, pp. 501–505. [12] N. Tishby, F. Pereira, and W. Bialek, “The information bottleneck method,” arXiv preprint physics/0004057, 2000. [13] F. Calmon, A. Makhdoumi, and M. Médard, “Fundamental limits of perfect privacy,” in IEEE Int. Symp. Inf. Theory (ISIT), 2015, pp. 1796–1800. [14] T. Berger and R. Yeung, “Multiterminal source encoding with encoder breakdown,” IEEE Trans. Inf. Theory, pp. 237–244, 1989. [15] S. Asoodeh, F. Alajaji, and T. Linder, “Notes on information-theoretic privacy,” in 52nd Annual Allerton Conference, Illinois, USA, Oct. 2014, pp. 1272–1278. [16] H. Hirschfeld, “A connection between correlation and contingency,” in Proc. Cambridge Philosophical Soc. 31, 1935, pp. 520–524. [17] H. Gebelein, “Das statistische problem der korrelation als variations- und eigenwert-problem und sein zusammenhang mit der ausgle´ angew. Math. und Mech. 21, pp. 364–379, 1941. ichungsrechnung,,” Zeitschrift f’ur 39 [18] A. Rényi, “On measures of dependence,” Acta Math. Hung., vol. 10, pp. 539–550, 1959. [19] D. Bertsimas and J. N. Tsitsiklis, Introduction to linear optimization. [20] K. G. Murty, Linear Programming. Athena Scientic, 1997. John Wiley and Sons, 1983. [21] B. C. Levy, Principles of Signal Detection and Parameter Estimation. Springer, 2008. [22] Y. tradeoffs Wang, Y. Basciftci, and P. Ishwar, “Privacy-utility under constrained data release mechanisms,” https://arxiv.org/pdf/1710.09295.pdf, Oct. 2017. [23] H. Witsenhausen, “On sequences of pairs of dependent random variables,” SIAM Journal on Applied Mathematics, vol. 28, no. 1, pp. 100–113, Jan. 1975. [24] V. Anantharam, A. Gohari, S. Kamath, and C. Nair, “On maximal correlation, hypercontractivity, and the data processing inequality studied by Erkip and Cover,” https://arxiv.org/pdf/1304.6133.pdf, Apr. 2013. [25] G. Golub, “Some modified matrix eigenvalue problems,” SIAM Review, vol. 15, no. 2, pp. 318–334, Apr. 1973. [26] A. E. Gamal and Y.-H. Kim, Network Information Theory. Cambridge University Press, 2012.
7cs.IT
Deep Interactive Evolution Philip Bontrager1 , Wending Lin3 , Julian Togelius1 and Sebastian Risi2 arXiv:1801.08230v1 [cs.NE] 24 Jan 2018 1 New York University New York, United States [email protected], [email protected] 2 IT University of Copenhagen Copenhagen, Denmark [email protected] 3 Beijing University of Posts and Telecommunications Beijing, China [email protected] Abstract. This paper describes an approach that combines generative adversarial networks (GANs) with interactive evolutionary computation (IEC). While GANs can be trained to produce lifelike images, they are normally sampled randomly from the learned distribution, providing limited control over the resulting output. On the other hand, interactive evolution has shown promise in creating various artifacts such as images, music and 3D objects, but traditionally relies on a hand-designed evolvable representation of the target domain. The main insight in this paper is that a GAN trained on a specific target domain can act as a compact and robust genotype-to-phenotype mapping (i.e. most produced phenotypes do resemble valid domain artifacts). Once such a GAN is trained, the latent vector given as input to the GAN’s generator network can be put under evolutionary control, allowing controllable and high-quality image generation. In this paper, we demonstrate the advantage of this novel approach through a user study in which participants were able to evolve images that strongly resemble specific target images. 1 Introduction This paper addresses the question how to generate artifacts, included but not limited to images, through indirect interaction with computer software. This means that instead of effecting change directly to the artifact, for example through drawing a picture with virtual brushes in a traditional image editing software, the user communicates with the software through for example evaluative feedback and suggestions. Such interactive or AI-assisted creation has a number of potential uses. Perhaps most obviously, combining the generative capacity of computer software with human input can allow humans to create artifacts (such as images) of a quality and richness they could not create on their own using direct methods because they do not possess the requisite technical skills (such as drawing). But such methods could also be used to explore or enforce particular aesthetics, allowing users to create art in some other aesthetic than they would normally adhere to. Use cases for such AI-assisted creation methods could include creating avatars for games and virtual environments, creating facial composites for help in criminal cases, and product design, customization or search. Interactive evolution is a particular form of AI-assisted creation where the human user functions as the fitness function in an evolutionary algorithm. The system serves up a number of artifacts to the human every generation, and the human responds by indicating which of these artifact(s) they prefer. Through crossover and/or mutation, the system then generates another generation of artifacts, from which the user selects their favorite artifact(s), and so on [19]. While the conceptual appeal of interactive evolution is strong, there are several problems hampering the practical usefulness of the technique. One issue is that a large number of evaluations may be necessary to find the desired artifact, leading to an issue known as user fatigue. It might also not be possible to find the desired artifact through optimizing for traits that seem to be correlated with, or found in, the desired artifact; some studies suggest that such “goal-directed evolution” often fails [17,21]. Part of the problem here might be the underlying representation and genotype-to-phenotype mapping employed in the interactive evolution system, which might not be conducive to search, or which might thread the wrong balance between generality and domainspecificity. One potential solution is to find better artifact representations, which can map between genotype and phenotype within a particular domain while preserving the property that most points in search space correspond to reasonable artifacts. We suggest that this could be done by using the generator parts of Generative Adversarial Networks (GANs), which are trained to produce images (or other artifacts) from a particular domain. We hypothesize that using a generative representation acquired through adversarial training can give us a searchable domain-specific space. We call this approach deep interactive evolution. In the rest of this paper, we discuss previous work in interactive evolution and generative adversarial networks and present our specific approach. We also show the results of letting a set of users use deep interactive evolution using a generator trained on a shoe dataset, and one trained on a face dataset. 2 Background This section reviews interactive evolutionary and generative adversarial networks, which are fundamental for the combined approach presented in this paper. 2.1 Interactive Evolutionary Computation (IEC) In interactive evolutionary computation (IEC) the traditional objective fitness function is replaced by a human selecting the candidates for the next generation. IEC has traditionally been used in optimization tasks of subjective criteria or in open-ended domains where the objective is undefined [14,19,20]. As surveyed by Takagi [19], IEC has been applied to several domains including image generation [18], games [9,16], music [10], industrial design, and data mining. Evolutionary art and music often hinges on human evaluation of content due to the subjective nature of aesthetics and beauty. Most IEC approaches follow the paradigm of the original Blind Watchmaker from Dawkins [4], in which artifacts are evolved by users through an interactive interface that allows them to iteratively select from a set of candidate designs. However, a problem with this IEC approach is user fatigue, and many attempts have been made to minimize it [3,11,14]. User fatigue is a problem because evolution takes time and humans tend to suffer from fatigue after evaluating relatively few generations. One approach to limit user fatigue is collaborative interactive evolution, in which users can continue evolution from promising starting point generated by other users. Picbreeder4 is an example of such a system, which allows users to collaboratively evolve images online, building on the intermediate results published by others [17]. Another approach is to seed the initial population of IEC with meaningful, high-quality individuals [23]. In this paper we introduce a novel approach that should suffer less from user fatigue by restricting the space of possible artifacts to a certain class of images (e.g. faces, shoes, etc.); the hypothesis is that this way the user needs to spend less time exploring the space of possible solutions. 2.2 Generative Adversarial Networks (GANs) Algorithm 1 Generative Adversarial Network 1: Initialize θ and ω 2: for t training iterations do 3: for k discriminator updates do 4: x ← Minibatch sample of real data 5: z ← Minibatch of latent variables where z ∼ p(z) 6: L ← Loss(Dω (x), Real) + Loss(Dω (Gθ (z)), Generated) 7: ω ← gradient-based update w.r.t L 8: z ← Minibatch of latent variables where z ∼ p(z) 9: L ← Loss(Dω (Gθ (z)), Real) 10: θ ← gradient-based update w.r.t L The space of potential images in our approach is represented by a Generative and Adversarial Network (GAN). GANs [7] are a new class of deep learning algorithms that have extended the reach of deep learning methods to unsupervised domains, which do not require labeled training data. The basic idea behind GANs is that two networks compete against each other: one network is generative and the other network is discriminative. The generative network tries to create synthetic content (e.g. images of faces) that is indistinguishable to the discriminative network from real data (e.g. images of real faces). Both the generator and discriminator are trained in turns, with the generator becoming better at forging content, and the discriminator becoming better at telling real from synthetic content. 4 http://picbreeder.org/ In more detail, a generator network (called G) conditioned on latent variables, generates images to convince a separate discriminator network (called D) that its generated images are authentic ones from a real-world training set. Algorithm 1 outlines the basic steps of training a generator using this adversarial technique. Typically the discriminator is only updated once for every time the generator is updated, so k is 1. A lot of research on GANs is now focused on what loss functions allow for the most stable training. The latent variables are random variables that are typically sampled from a normal distribution. A generator is trained on a fixed number of latent variables and the stochasticity forces the generator to generalize. 3 Approach: Deep Interactive Evolution Trained Generative Network 1. Transposed Convolution 2. Tanh Output Images 128 x 128 x 3 User Input: 64 x 64 x 128 1. Select Images 2. Set Variation Preference 32 x 32 x 256 16 x 16 x 512 Apply Mutation: 8x8x 1024 • 50% probability 1. Keep Selected chromosomes • Standard 2. Uniform Crossover of selections random numbers deviation set by the user 4x4 x 1024 Next Generation: • Normal Latent Variables 1. Transposed Convolution Kernel: 4 Stride: 2 Padding: 1 2. Batch Normalization 3. Relu 3. Random Samples Input Latent Variables Fig. 1. This figure illustrates the four steps of DeepIE. First, latent variables are passed through a trained image generator to produce images. Next, the user selects the images based on their interests. Third, new sets of latent variables are chosen based on the selected image. Last, the latent variables are mutated with the intensity set by the user. Figure 1 depicts our Deep Interactive Evolution (DeepIE) approach. The main differentiating factor between DeepIE and other interactive evolution techniques is the employed generator. The content generator is trained over a dataset to constrain and enhance what is being evolved. In the implementation in this paper, we trained a nonspecialized network over 2D images. In general, a number of goals can be optimized for during the training process. For example, for generating art, a network that specializes in creative output can be used. Recently Elgamma et. al. proposed CAN, a variation to the GAN training that prioritizes creative output [5]. In other work, Kim et. al. train a network to map latent variables to discrete outputs, such as text [12]. The generator works by transforming a number of latent variables into content. In our work, the generator has 20 latent variables as input and the content is images from various domains. Typically the latent variables are chosen at random to produce random, pleasing, images. In DeepIE, we initially sample the latent variables randomly and then evolve them to get the generator to produce the user’s optimal image. The randomly sampled latent variables are used to generate images. The best-generated images are then selected by the user. Then the latent variables that produced the images are evolved to create a set of new variables. There are many ways this can be approached. The user could simply have the ability to select images or the complexity could be increased and the user could be allowed to provide more information such as the ability to rank the images or score them. Depending on the interface, an evolutionary strategy can be chosen to take advantage of the provided information. It is important to use crossover as it allows the user the ability to guide the process through picking images that can be later combined. The user then determines the amount of mutation to apply which can be implemented in a number of ways. This setup is very general. To create images in a specific domain, the person just needs a sufficient amount of data to train the generator. Once the generator is automatically trained on a domain, the user can evolve the inputs exactly the way they have always done. To use DeepIE on content outside of images, the generator architecture might need to be modified but the algorithm still remains the same. 3.1 Experimental Setup In this work, we implement a vanilla version of DeepIE. This allows us to observe the effectiveness of the technique in the most general sense. Here we use a fairly simple setup for the evolutionary process. Also, the interface is kept minimalistic to try to keep the focus off of how different interface design could affect the usability of the system. More involved interfaces could allow more information to be shared with the evolutionary algorithm allowing us to tune the process for different people and domains. This can provide more information but also may increase user fatigue. Testing this is beyond the scope of this paper. Generator The recent advances in neural network image generators are what enable this new take on interactive evolutionary computation to work. There are a number of techniques for designing and training a data generator. These include GANs, variational autoencoders, and autoregressive techniques, as well as a few more techniques [6]. For our setup, we decided to implement a GAN. GANs have seen a lot of recent success in generating realistic looking, relatively high resolution, images. At the time of publishing this paper, one of the state of the art techniques for training a GAN on images is the Wasserstein GAN with Gradient Penalty (WGAN-GP) [8]. This process uses a Wasserstein distance function as a loss for the discriminator and gradient penalties to keep the gradients in an acceptable range for the loss function. The generator has been shown to produce even better results if the discriminator is made to classify the data as well. In this work, we trained our network with the default setup for WGAN-GP without classification to keep our setup as general as possible. This training method is typically combined with a deep convolutional network architecture (DCGAN) [15]. Our implementation of the network is shown in Figure 1 Algorithm 2 Deep Interactive Evolution. Defaults m, n ← 20, µ ← 0, σ 2 ← 1, p ← .5, f oreign ← 2 1: Gθ ← trainGAN(data) 2: interf ace ← Interface() 3: Z ← m by n matrix where zi,j ∼ N (µ, σ 2 ) . we use WGAN-GP [8] . population size m with n latent variables each 4: repeat 5: images ← Gθ (Z) 6: interface.display(images) 7: wait until interface.buttonNextPressed() is True 8: indices ← interface.getSelectedImages() 9: selection ← Zindicies 10: s2 ← interface.getMutationParameter() 11: ∆ ← m − length(selection) 12: x ← max(0, ∆ − f oreign) 13: cross ← x by n matrix where crossi ∼ mutate(unif orm(selection), s2 ) 14: x ← max(f oreign, ∆) 15: new ← x by n matrix where newi,j ∼ N (µ, σ 2 ) 16: selection ← apply mutate with s2 to each selectioni 17: Z ← selection + cross + new 18: until user stops 19: function UNIFORM(population) 20: a ← random individual from population 21: b ← random individual from population 22: mask ← vector of length n where maski ∼ Bernoulli(.5) return mask · a + (1 − mask) · b 23: function MUTATE(individual, std) 24: mutation ← Bernoulli(p) 25: noise ← vector of length n wherenoisei ∼ N (µ, std) return individual + mutation · noise on the left end. The generator network is made of repeating modules that consist of a transposed convolution, a batch normalization layer, and a ReLU nonlinearity function. The discriminator is the same architecture upside-down, each module consists of a convolutional layer and a LeakyReLU nonlinearity. The convolutional layer has a kernel of size 3 and stride of 2 for downsampling. To get good consistency from the batch normalization layers during training and evaluation, its better to train with large batch sizes. To minimize the number of variables to evolve, our network only has 20 latent variables. This means the input to G is a random normal vector of size 20. Berthelot et. al. did tests in their work on Boundary Equilibrium GANs and found that changing the number of latent variables did not have a noticeable impact on the diversity of the network [2]. Evolution The implemented evolutionary process consists of mutation and crossover as the two primary operators for optimization. Mutation allows for local variation of the user’s selections. Crossover is what really allows the user to guide the process. The user can select images that do not match what they are looking for but contain a feature they want in their final design. Crossover allows the possibility of that feature being added to the current best design. To keep the number of generations low, it is important to have frequent mutations to quickly sample a design space. We have 50% probability of mutation. For mutation, we generate a random normal vector of the same size as the latent variable vector and add it to the latent variables. The standard deviation is set by the user. At the low end, it is set to 0 which results in no mutation. At the high end, it is 1 which is the same magnitude as the latent variables themselves. We use uniform crossover to combine selected parents. A new chromosome is created by randomly sampling from the variables of two parents, childi = B RandomChoice(parentA i , parenti ) for each variable i. Since the variables are not independent in their effect on the generator, this form of crossover can lead to some unexpected results. Testing does show some independence though as features of two images will combine in practice. Since the user only selects the images they want to keep, the system has no way of evaluating the relative quality or ranking of the selected images. The selected images are kept. The population is rebuilt through random crossover and two new foreign chromosomes. The new chromosomes are added to continually provide the user with new features to select. Our population size is 20 to keep from being overwhelming. We are aware that using these mutation and crossover techniques result in latent variables that do not match the prior distribution of variables that the network, G, was trained on. This is likely the cause of increased image artifacts the longer the images are evolved. Other options for mutation and crossover could involve interpolating between vectors along the hypersphere to maintain the expected distribution of inputs for G. Data We decided to produce 2D images as there is a lot of recent advances in using networks to generate images. We looked for datasets that represent domains that are interesting to work in. In our initial setup, we use three datasets: CelebA face dataset [13], UT Zappos50K shoes [22], and 3D Chairs images [1]. A separate network is trained for each domain. To add a new domain, one just has to apply a new dataset. In theory, this setup should translate to 3D data with very few alterations. Interface Figure 2 shows a screenshot of our implementation of DeepIE. There is a dropdown menu at the top left that allows the user to select the domain to design in. They are then presented with the entire generation. The user simply selects which images will go to the next generation. At the bottom is a slider that allows the user to set the standard deviation of the noise that is added during mutation. The user only has to know that further right on the slider increases variability and further left decreases it. Fig. 2. Web interface for our DeepIE framework. 3.2 User Testing To get feedback on our implementation of DeepIE, we had volunteers try the interface. We gave them two main tasks: 1. Reproduce the image of a shoe the was created in our system. 2. Reproduce the image of a face of their choosing. The intention is to see how much control the user has over the evolution process. While a major use case for this tool is to help someone communicate an image that only exists in their head, for the purpose of testing we required a target image to be chosen in advance. The volunteers were asked to familiarize themselves with the system by figuring out how to evolve a chair. Once they were comfortable, they were asked to select a shoe image from a list of generated images. They then were asked to use a minimum of 10 generations to try and recreate that image. The volunteers were given more freedom during the second task. They could choose any public domain face image, or their own, to recreate. Once again they were asked to use a minimum of 10 generations. After they completed these tasks we asked them to rate their ability to do each of the two tasks on a scale from 1 through 5. We also asked them to describe their strategies for creating the image as well as their experience. After they finished answering these questions we showed them a randomized mix of all the shoes they selected, and of all the faces they selected and asked them to pick the best from each domain. In this way, we could get a crude measure of whether the images were improving over time. 4 Results 1 2 3 4 5 10 - Target Fig. 3. This demonstrates the authors’ attempt to evolve an image that looks like Nicolas Cage. Each column contains the selected images for the given generation. In this example three images were selected at each step (it is possible to select less or more images). The last column shows the best image after 10 generations. The output of our trained network is visible in the images we provided. Due to difficulties in hosting a PyTorch enabled website, we are not currently publicly hosting our site. Figure 3 shows the process of evolving an image of Nicholas Cage through selection. Each column represents the selected images for the marked generation. In the example given, each selection is made based on a particular feature it possesses; i.e. face shape, hair, pose, etc. Over time, two separate tracks that are being evolved will merge in an acceptable way and they can be merged and the original tracks can be let go. This can be seen between generation 3 and 4 in Figure 3 where the top and middle image from generation 3 created the middle image in generation 4. 4.1 User Testing To test the approach in more depth, 16 volunteers tried the system as has been previously described. The outcomes of their efforts can be seen in Figure 4. The target image for each user appears on the left of their respective domain. For each user, we looked at two different outcomes: the final evolved image, and the user’s favorite evolved image. Ideally, these would be the same image but due to the stochastic nature of the optimization process, the best image is sometimes lost. Other times the image does not really improve for the last few generations, making it hard to discern the best generation. With the shoes, in particular, we observed that the image could be recreated by many users within only a few generations. Since we asked the users to go for 10 generations, it resulted in a lot of images that were essentially the same. The results in Figure 4 suggest that there is a wide diversity in how well people were able to recreate the target. The shoes tended to be pretty close, though people are less picky about the distinction between two similar looking shoes. For the face recreation tasks, people usually were only able to focus on a few characteristics. Quantitative Analysis As was mentioned earlier, we measured how far into the evolutionary process the best image occurred. Taking this value over the total number of iterations that a user searched provides a ratio of where in the evolutionary process the best image was found. Figure 5 is a histogram of the distribution of where the best image was found for both shoes and faces. Shoes are represented in orange and faces in blue. It’s clear right away that the best faces appeared much closer the final iteration than the best shoes. Looking at a paired-samples t-test, there is a significant difference in the ratios for faces (µ = .79, σ = .14) and shoes (µ = .56, σ = .16) with a p = 0.001. This confirms what was observed from the volunteers, where they said they found the shoe they were looking for in the first few generations, but then continued on to complete the required 10 generations. For this reason, the best shoe image is scattered all over the different generations. As can be seen in figure 4, there is often little difference between the best image and final image for shoes. This implies that the search converged early and the shoes mostly looked the same from there on out. The user’s ”best” pick at the end was mostly a random guess between very similar looking shoes. The data for faces, on the other hand, tells a different story. The individual ratios are clustered close to 1. This is a good sign that the image of the face was improving throughout the process. The best image was selected at a later time, randomized with all the selected images, and the user still chose one of the later designs. With an average number of generations of 12.7, most of the users stuck to the requested 10 generations so we are not able to measure for how long the faces would continue to improve. It is important to note this progress is based on the user’s reporting of which is the best image. This is by no means an objective measure. This measure of improvement measures that the user is able to better express what they want to express. This is an important metric since this is a communication tool, but the metric shouldn’t be confused as an indicator of image quality. Experience Based on self-reported numbers, users felt that they got much closer to reproducing the shoes than they did to the face. This could be predicted from figure 4. On average users reported 2.2 out of 5 for their ability to reproduce faces and 3.8 out of 5 for their ability to reproduce shoes, both with a standard deviation of 1. What users wrote about their experience was split between amusement and frustration. Some people embrace the chaotic nature of the system while others were frustrated with not having enough control. Part of the frustration could be caused by the nature of the task we assigned them. By providing the image, we gave them a very rigid goal which is different from the intended use case. Of course, our current implementation has limitations in the quality of images that it can produce and that has to be acceptable to the user. Faces Target Final Shoes Best Target Final Best 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Fig. 4. Evolved Images from the user study. Target images are the intended target for a row. Final images are the last selected image. Best images are the images that the user selected as best from a compilation of all their selections over the entire evolutionary process. Best Image Iteration Histogram 6 5 Subjects 4 3 2 1 0 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Best Image Iteration / Total Iterations 0.9 Fig. 5. This histogram depicts at which point during the evolutionary process the best image was found for each of the 16 participants. Data for the shoe domain is shown in orange while the results in the face domain are shown in blue. Strategies By reviewing the user’s feedback on the strategies they used, we were able to put together three main strategies. These are the selection strategies people deployed to try and reach their target image. The three primary strategies are as follows; 1. Collect Distinct Traits 2. Select Best Likeness 3. Hierarchical Trait Selection Collect Distinct Traits was by far the most popular strategy with over half of the users subscribing to this strategy. This strategy involves breaking down the target into distinct parts and then identifying the parts in the population provided. For a face, this can include hairstyle, gender, face profile, etc. The user of this strategy selects images that are the best representation of these traits. Once traits start to merge into single images, then the individual traits can be ignored. This can be followed until all the desired traits are combined and then the focus moves to finding the best variation of this image. Select Best Likeness is a simpler strategy. The user looks for the image or images that are closest to the final result. User 2 in 4 used this strategy. This strategy is the closest to using a distance-based loss function as the evaluation technique. The main difference is that users reported focusing on the overall look of the image instead of the details which most loss functions are very bad at. A quarter of our subjects reported using this strategy. Half of them implemented it by only selecting one image per generation and relying on the mutation slider. The final strategy, Hierarchical Trait Selection, is a variation of the first strategy. Instead of maintaining separate threads of different traits, the user focuses on the first desired trait that they see. They then cultivate its diversity until they find an image that has another desired trait. At this point, they have two desired traits and are ready to repeat the process for adding a third. This was the least common strategy with only two people using it. User 16, is one of the users who used this strategy. We did not expect these distinct strategies to emerge, as such, we have no experiment that compares the effectiveness of the different strategies. We also don’t measure how practice affects the outcome. These strategies could also just be the result of trying to recreate faces. They do seem like fundamental methods for designing complex systems through a series of selections. In the deploying of these strategies, users found specific things they wanted from the interface that is not there. The two most popular requests were for some form of elitism and for the ability to re-roll random samples. Elitism would allow people to keep their best images and then feel comfortable to explore more risky strategies. Re-rolling, or generating a new batch of random samples, would allow the users to quickly search for new features without affecting the features they already found. These emergent strategies and requests show that there are lots of ways to optimize the interface that can be explored. It also demonstrates that even with one of the simplest interfaces there is a certain depth to designing an image through selection. 5 Discussion This work allows interactive evolutionary computation to be brought to many new domains. Existing techniques are either very general, but very difficult to corral towards a specific problem domain, or they require a hand build model of the domain of the problem domain. By building the domain model using data, there are many new domains that IEC can be applied too. Obviously, there are still many things that cannot be modeled well with these machine learning techniques, but there is a lot of active research in the field and the applications keep growing. Aside from the ability to learns domains directly from data, DeepIEC allows the user to control the direction of the evolutionary process. Despite the complex mapping from latent variables to output, GAN’s are trained to have a smooth output. This means two very similar latent codes will very likely produce similar outputs. Also, the midpoint between two latent codes, not too far apart, is likely to map to images that share qualities of both original images. The way that variation and combinations in the phenotype carry over to the genotype allows the evolutionary process be more transparent. This, in turn, makes it possible to direct the process. The testing that we did on our implementation of DeepIE focused on the ability for the user to guide the population toward a predefined target. This ability is great for communicating an idea or design that you otherwise might not have the ability to create. This allows someone without the ability to draw to better express themselves through the medium of pictures. But we did not look, in this work, at our tool’s ability to be used for creative exploration. DeepIE shares a key strength that IEC has, which is its ability to help a user explore a creative space. The user can simply select the most interesting suggestion and see where it takes them. In fact, some of the volunteers for our system were tempted to do that at times instead of pursuing their target. By coupling this ability to help a user explore a space with DeepIE’s ability to focus on a domain, we get a tool that can help a designer come up with new ideas. This is an area we think that DeepIE could have an impact. While this is a positive for exploration, it is a downside for some applications of communication. For example, if someone were to try to use this to help victims of a crime create a picture of the perpetrator, they wouldn’t want the system’s suggestions to be influencing the user. That would be a case where the options that come up on the screen would likely bias the user’s memory of the criminal. Even if the generator could produce every face with perfect fidelity, extensive tests would have to be done to make sure the process doesn’t alter a person’s memory. It is clear that the interface affects what the user discovers. For example, by not allowing users to save their best images, they are more conservative in exploration. By not giving them enough diversity, in phenotypes, they feel powerless to guide the process in a direction they are interested in. This means that the optimal interface might make a big difference in ensuring that the user always feels in control and engaged in the process. Going forward there are many directions for this. Different generative models and evolution ideas can be experimented with. New and interesting types of domains can be explored, such as time-based domains. Finally, the way the user interacts, as discussed, can itself go in many directions. 6 Conclusion This paper presented a novel approach to interactive evolution. Users are able to interactively evolve the latent code of a GAN that is pre-trained on a certain class of images (e.g. shoes, faces, chairs). Our approach tries to strike a balance between a system like Picbreeder that allows the evolution of arbitrary images (but is very hard to guide towards a specific target), and systems like GANs, in which images are normally sampled randomly without any user input. We presented initial results that show that users are able to interactively recreate target images through our DeepIE approach, which has been shown difficult with traditional IEC approaches. In the future, it will be interesting to extend the approach to other domains such as video games that can benefit from high-quality and controllable content generation. References 1. Aubry, M., Maturana, D., Efros, A., Russell, B., Sivic, J.: Seeing 3d chairs: exemplar partbased 2d-3d alignment using a large dataset of cad models. In: CVPR (2014) 2. Berthelot, D., Schumm, T., Metz, L.: Began: Boundary equilibrium generative adversarial networks. arXiv preprint arXiv:1703.10717 (2017) 3. Bongard, J.C., Hornby, G.S.: Combining Fitness-based Search and User Modeling in Evolutionary Robotics. Proceeding of the 15th Annual conference on Genetic and Evolutionary Computation Conference - GECCO ’13 pp. 159–166 (2013) 4. Dawkins, R.: The blind watchmaker: Why the evidence of evolution reveals a universe without design. WW Norton & Company (1986) 5. Elgammal, A., Liu, B., Elhoseiny, M., Mazzone, M.: Can: Creative adversarial networks, generating” art” by learning about styles and deviating from style norms. arXiv preprint arXiv:1706.07068 (2017) 6. Goodfellow, I.: Nips 2016 tutorial: Generative adversarial networks. arXiv preprint arXiv:1701.00160 (2016) 7. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Advances in neural information processing systems. pp. 2672–2680 (2014) 8. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.: Improved training of wasserstein gans. arXiv preprint arXiv:1704.00028 (2017) 9. Hastings, E.J., Guha, R.K., Stanley, K.O.: Automatic content generation in the galactic arms race video game. IEEE Transactions on Computational Intelligence and AI in Games 1(4), 245–263 (Dec 2009) 10. Hoover, A.K., Szerlip, P.A., Stanley, K.O.: Interactively evolving harmonies through functional scaffolding. In: Proceedings of the 13th annual conference on Genetic and evolutionary computation. pp. 387–394. ACM (2011) 11. Kamalian, R., Zhang, Y., Takagi, H., Agogino, A.: Reduced human fatigue interactive evolutionary computation for micromachine design. In: Machine Learning and Cybernetics, 2005. Proceedings of 2005 International Conference on. vol. 9, pp. 5666–5671 Vol. 9 (08 2005) 12. Kim, Y., Zhang, K., Rush, A.M., LeCun, Y., et al.: Adversarially regularized autoencoders for generating discrete structures. arXiv preprint arXiv:1706.04223 (2017) 13. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of International Conference on Computer Vision (ICCV) (December 2015) 14. Pallez, D., Collard, P., Baccino, T., Dumercy, L.: Eye-Tracking Evolutionary Algorithm to minimize user fatigue in IEC applied to Interactive One-Max problem. Proceedings of the 2007 GECCO conference companion on Genetic and evolutionary computation (2007) 15. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434 (2015) 16. Risi, S., Lehman, J., D’Ambrosio, D.B., Hall, R., Stanley, K.O.: Petalz: Search-based procedural content generation for the casual gamer. IEEE Transactions on Computational Intelligence and AI in Games 8(3), 244–255 (Sept 2016) 17. Secretan, J., Beato, N., D Ambrosio, D.B., Rodriguez, A., Campbell, A., Stanley, K.O.: Picbreeder: evolving pictures collaboratively online. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. pp. 1759–1768. ACM (2008) 18. Sims, K.: Artificial evolution for computer graphics, vol. 25. ACM (1991) 19. Takagi, H.: Interactive evolutionary computation: Fusion of the capabilities of ec optimization and human evaluation. Proceedings of the IEEE 89(9), 1275–1296 (2001) 20. Todd, S., Latham, W.: Evolutionary art and computers academic press (1992) 21. Woolley, B.G., Stanley, K.O.: On the deleterious effects of a priori objectives on evolution and representation. In: Proceedings of the 13th annual conference on Genetic and evolutionary computation. pp. 957–964. ACM (2011) 22. Yu, A., Grauman, K.: Fine-Grained Visual Comparisons with Local Learning. In: Computer Vision and Pattern Recognition (CVPR) (June 2014) 23. Zhang, J., Taarnby, R., Liapis, A., Risi, S.: Drawcompileevolve: Sparking interactive evolutionary art with human creations. In: International Conference on Evolutionary and Biologically Inspired Music and Art. pp. 261–273. Springer (2015)
9cs.NE
arXiv:1608.04151v2 [math.GR] 29 Dec 2016 The Congruence Subgroup Problem for low rank Free and Free Metabelian groups David El-Chai Ben-Ezra, Alexander Lubotzky January 2, 2017 T o Ef im Zelmanov, a f riend and a leader Abstract The congruence subgroup problem for a finitely generated group Γ \ asks whether Aut (Γ) → Aut(Γ̂) is injective, or more generally, what is its kernel C (Γ)? Here X̂ denotes the profinite completion of X. In this paper we first give two new short proofs of two known results (for Γ = F2 and Φ2 ) and a new result for Γ = Φ3 : (1) C (F2 ) = {e} when F2 is the free group on two generators. (2) C (Φ2 ) = F̂ω when Φn is the free metabelian group on n generators, and F̂ω is the free profinite group on ℵ0 generators. (3) C (Φ3 ) contains F̂ω . Results (2) and (3) should be contrasted with an upcoming result of the first author showing that C (Φn ) is abelian for n ≥ 4. Mathematics Subject Classification (2010): primary: 19B37, 20H05, secondary: 20E18, 20E36, 20E05. Key words and phrases: congruence subgroup problem, profinite groups, automorphism groups, free groups, free metabelian groups. 1 Introduction The classical congruence subgroup problem (CSP) asks for, say, G = SLn (Z) or G = GLn (Z), whether every finite index subgroup of G contains a principal congruence subgroup, i.e. a subgroup of the form G (m) = ker (G → GLn (Z/mZ)) for some 0 6= m ∈ Z. Equivalently, it asks whether the natural map Ĝ → GLn (Ẑ) is injective, where Ĝ and Ẑ are the profinite completions of the group G and the ring Z, respectively. More generally, the CSP asks what is the kernel of this map. It is a classical 19th century result that the answer is negative for n = 2. Moreover (but not so classical, cf. [Me], [Lu1]), the kernel, in this case, 1 is F̂ω - the free profinite group on a countable number of generators. On the other hand, for n ≥ 3, the map is injective and the kernel is therefore trivial. The CSP can be generalized as follows: Let Γ be a group and M a finite index characteristic subgroup of it. Denote: G (M ) = ker (Aut (Γ) → Aut (Γ/M )) . Such a finite index normal subgroup of G = Aut (Γ) will be called a “principal congruence subgroup” and a finite index subgroup of G which contains such a G (M ) for some M will be called a “congruence subgroup”. Now, the CSP for Γ asks whether every finite index subgroup of G is a congruence subgroup. When Γ is finitely generated, Aut(Γ̂) is profinite and the CSP is equivalent to the u u \ question (cf. [BER], 1 and 3): Is the map Ĝ = Aut (Γ) → Aut(Γ̂) injective? More generally, it asks what is the kernel C (Γ) of this map. As GLn (Z) = Aut (Zn ), the classical congruence subgroup results mentioned above can therefore be reformulated as C (A2 ) = F̂ω while C (An ) = {e} for n ≥ 3, when An = Zn is the free abelian group on n generators. Very few results are known when Γ is non-abelian. A very surprising result was proved in [As] by Asada by methods of algebraic geometry: Theorem 1.1. C (F2 ) = {e}, i.e., the free group on two generators has the \ congruence subgroup property, namely Aut (F2 ) → Aut(F̂2 ) is injective. A purely group theoretic proof for this theorem was given by Bux-ErshovRapinchuk [BER]. Our first goal in this paper is to give an easier and more direct proof of Theorem 1.1, which also give a better quantitative estimate: we give an explicitly constructed congruence subgroup G (M ) of Aut (F2 ) which is contained in a given finite index subgroup H of Aut (F2 ) of index n. Our estimates on the index of M in F2 as a function of n are substantially better than those of [BER] - see Theorems 2.7 and 2.9. We then turn to Γ = Φ2 , the free metabelian group on two generators. The initial treatment of Φ2 is similar to F2 , but quite surprisingly, the first named author showed in [Be1] a negative answer, i.e. C (Φ2 ) 6= {e}. We also give a shorter proof of this result, deducing that: Theorem 1.2. C (Φ2 ) = F̂ω . We then go ahead from 2 to 3 and prove: Theorem 1.3. C (Φ3 ) contains a copy of F̂ω . In particular, the congruence subgroup property (strongly) fails for Φ3 . This is also surprising, especially if compared with an upcoming paper of the first author [Be2] showing that C (Φn ) is abelian for n ≥ 4. So, while the dichotomy for the abelian case An = Zn is between n = 2 and n ≥ 3, for the metabelian case, it is between n = 2, 3 and n ≥ 4. A main ingredient of the proof of Theorem 1.3 is showing that Aut (Φ3 ) is large, i.e. it has a finite index subgroup which is mapped onto a non-abelian 2 free group. For this we use the method developed by Grunewald and the second author in [GL] to produce arithmetic quotients of Aut (Fn ). In particular, it is shown there that Aut (F3 ) is large. Our starting point to prove Theorem 1.3 is the observation that the same proof shows also that Aut (Φ3 ) is large. In our proof of Theorem 1.2, the largeness of Aut (Φ2 ) is also playing a crucial role. But, a word of warning is needed here: largeness of Aut (Γ) by itself is not sufficient to deduce negative answer for the CSP for Γ. For example, Aut (F2 ) is large but has an affirmative answer for the CSP. At the same time, as mentioned above, Aut (F3 ) is large and we do not know whether F3 has the congruence subgroup property or not. To prove Theorem 1.3 we use the largeness of Aut (Φ3 ) combined with the fact that every non-abelian finite simple group which is involved in Aut(Φ̂3 ) is already involved in GL3 (R) for some finite commutative ring R, as we will show below. u The paper u is organized as follows: In th2 we give a short proof for Theorem 1.1 and in 3 for Theorem u 1.2. The 4 section is devoted to the proof of Theorem 1.3. We close in 5 with some remarks and open problems, about free nilpotent and solvable groups. Acknowledgements: The first author wishes to offer his deepest thanks to the Rudin foundation trustees for their generous support during the period of the research. This paper is part of his PhD’s thesis at the Hebrew University of Jerusalem. The second author wishes to acknowledge support by the ISF, the NSF, the ERC, Dr. Max Rossler, the Walter Haefner Foundation and the ETH Foundation. 2 The CSP for F2 Before we start, let us quote some general propositions which Bux-ErshovRapinchuk bring throughout their paper. Proposition 2.1. (cf. [BER], Lemma 2.1) Let: β α 1 → G1 → G2 → G3 → 1 be an exact sequence of groups. Assume that G1 is finitely generated and that the center of its profinite completion Ĝ1 is trivial. Then, the sequence of the profinite completions β̂ α̂ 1 → Ĝ1 → Ĝ2 → Ĝ3 → 1 is also exact. Proposition 2.2. (cf. [BER], Corollaries 2.3, 2.4. and 2.7) Let F be the free group on the set X, |X| ≥ 2. then: 1. The center of F̂ , the profinite completion of F , is trivial. 2. If x, y ∈ X, x 6= y, then the centralizer of [y, x] in F̂ is ZF̂ ([y, x]) = h[y, x]i, the closure of the cyclic group generated by [y, x]. We start now with the following lemma whose easy proof is left to the reader: 3 Lemma 2.3. Let H ≤ G = Aut (Γ) be a congruence subgroup. Then: ker(Ĝ → Aut(Γ̂)) = ker(Ĥ → Aut(Γ̂)) In particular, the map Ĝ → Aut(Γ̂) is injective if and only if the map Ĥ → Aut(Γ̂) is injective. Denote now F2 = hx, yi = the free group on x and y. It is a well known theorem of Nielsen (cf. [MKS], 3.5) that the kernel of the natural surjective map:  Aut (F2 ) → Aut (F2 /F2′ ) = Aut Z2 = GL2 (Z) is Inn  (F2 ), the inner group of F2 . It is also well known that the   automorphism  1 2 1 0 ∼ group , = F2 is free on two generators and of finite in0 1 2 1 dex in GL2 (Z) which 2 (Z) → GL2 (Z/4Z)). Now, if we denote  contains  ker (GL 1 2 1 0 the preimage of , under the map Aut (F2 ) → GL2 (Z) 0 1 2 1 by Aut′ (F2 ), then Aut′ (F2 ) is of finite index in Aut (F2 ) and contains the principal congruence subgroup:  . ker Aut (F2 ) → GL2 (Z) → GL2 (Z/4Z) = Aut F2 / F24 F2′ \ ′ (F ) → Aut(F̂ ) is injective. So, by Lemma 2.3 it is enough to prove that Aut 2 2 Now, by the description above, we deduce the exact sequence:     1 2 1 0 1 → Inn (F2 ) → Aut′ (F2 ) → , → 1. 0 1 2 1     1 2 1 0 As , is free , this sequence splits by the map: 0 1 2 1 ( (     x 7→ x x 7→ xy 2 1 2 1 0 7→ α = , 7→ β = 2 0 1 2 1 y 7→ yx y 7→ y and thus: Aut′ (F2 ) = Inn (F2 ) ⋊ hα, βi. By Propositions 2.1 and 2.2, the exact sequence: 1 → Inn (F2 ) → Aut′ (F2 ) → hα, βi → 1 yields the exact sequence: ′ (F ) → hα, \ \ \ 1 → Inn (F2 ) → Aut βi → 1 2 which gives: \ ′ (F ) = Inn \ \ Aut (F2 ) ⋊ hα, βi 2 Thus, all we need to show is that the following map is injective: \ \ Inn (F2 ) ⋊ hα, βi → Aut(F̂2 ). \ We will prove this, in three parts: The first part is that the map Inn (F2 ) → \ ∼ Aut(F̂2 ) is injective, but this is obvious as Inn (F2 ) = F̂2 is mapped isomorphi\ cally to Inn(F̂2 ) ∼ βi → = F̂2 . The second part is to show that the map ρ : hα, 4 Aut(F̂2 ) is injective, and the last part is to show that the intersection of the \ \ images of Inn (F2 ) and hα, βi in Aut(F̂2 ) is trivial, i.e. Inn(F̂2 ) ∩ Imρ = {e}. So it remains to prove the next two lemmas, Lemma 2.4 and Lemma 2.6: \ Lemma 2.4. The map hα, βi → Aut(F̂2 ) is injective. Before proving the lemma, we recall a classical result of Schreier: Theorem 2.5. (cf. [MKS], 2.3 and 2.4) Let F be the free group on the set X where |X| = n, and ∆ a subgroup of F of index m. Let T be a right Schreier transversal of ∆ (i.e. a system of representatives of right cosets containing the identity, such that the initial segment of any element of T is also in T ). Then: 1. ∆ is a freen group on m · (n − 1) + 1 elements. o −1 2. The set tx tx 6= e | t ∈ T, x ∈ X is a free generating set for ∆, where for every g ∈ F we denote by ḡ the unique element in T satisfying ∆g = ∆ḡ. 2 Proof. (of Lemma 2.4) Define ∆ = ker(F2 → (Z/2Z) ). This is a characteristic subgroup of index 4 in F2 , that by the first part of Theorem 2.5, is isomorphic ˆ = ker(F̂2 → (Z/2Z)2 ), and therefore, there is a natural to F5 . We also have: ∆ ˆ ∼ homomorphism: Aut(F̂2 ) → Aut(∆) = Aut(F̂5 ) which induces the composition \ ˆ hα, βi → Aut(F̂2 ) → Aut(∆). Thus, it is enough to show that the composition \ ˆ is injective. map hα, βi → Aut(∆) Now, let X = {x, y} and T = {1, x, y, xy} be a right Schreier transversal of ∆. By applying the second part of Theorem 2.5 for X and T , we get the following set of free generators for ∆: e1 = x2 , e2 = yxy −1 x−1 , e3 = y 2 , e4 = xyxy −1 , e5 = xy 2 x−1 Hence, the automorphisms α and β act on ∆ in the following way:   e1 = x2 7→ x2 = e1    −1 −1 −1 −1   = e2 e2 = yxy x 7→ yxy x 2 2 2 α = e3 = y 7→ yx yx = e2 e4 e3 e1   e4 = xyxy −1 7→ xyxy −1 = e4    e = xy 2 x−1 7→ xyx2 yx = e4 e2 e5 e1 5  2 2 2  = e5 e1 e3  e1 = x 7→ xy xy  −1 −1 −1 −1   = e2 e2 = yxy x 7→ yxy x 2 2 β = e3 = y 7→ y = e3   e4 = xyxy −1 7→ xy 3 xy = e5 e4 e3    e = xy 2 x−1 7→ xy 2 x−1 = e5 5 5 ∼ F2 (yes! these are the same α Let us now define the map π : ∆ → hα, βi = and β) by the following way:   e1 7→ α      e  2 7→ 1 π = e3 7→ β    e4 7→ α−1    e 7→ β −1 5 It is easy to see that N = ker π is the normal subgroup of ∆ generated as a normal subgroup by e2 , e1 e4 and e3 e5 , and that N is invariant under the action of the automorphisms α and β, since:  = e2 ∈ N  α (e2 )  α (e e ) = e e ∈ N 1 4 1 4 α (e3 e5 ) = e2 e4 e3 e1 e4 e2 e5 e1       (e3 e5 ) (e1 e4 ) e−1 e3 ((e1 e4 ) e2 ) e−1 = e4 e−1 4 ∈ N 4 e2 e4 3   = e2 ∈ N β (e2 )   β (e1 e4 ) = e5 e1 e3 e5 e4 e3 = e5 e1 (e3 e5 ) e1−1 (e1 e4 ) (e3 e5 ) e−1 5 ∈ N   β (e3 e5 ) = e3 e5 ∈ N \ ˆ induces a homomorphism: Therefore, the hα, βi → Aut(∆)  homomorphism  \ \ hα, βi → Aut hα, βi , and thus it is enough to show that the last map is injective. Now, under this map, α and β act on hα, βi in the following way: ( α = e1 N 7→ α (e1 N ) = α (e1 ) N = e1 N =α α = β = e3 N 7→ α (e3 N ) = α (e3 ) N = e2 e4 e3 e1 N = α−1 βα ( α = e1 N 7→ β (e1 N ) = β (e1 ) N = e5 e1 e3 N = β −1 αβ β = β = e3 N 7→ β (e3 N ) = β (e3 ) N = e3 N =β \ Namely, α and β act via π on hα, βi by the inner α and β  automorphisms  \ \ and hence hα, βi is mapped isomorphically to Inn hα, βi , yielding that the   \ \ \ map hα, βi → Aut hα, βi is injective and hα, βi → Aut(F̂2 ) is injective as well, as required. \ Lemma 2.6. Inn(F̂2 ) ∩ Imρ = {e}, where ρ : hα, βi → Aut(F̂2 ) is the map defined above. Proof. First we observe that α and β fix e2 = [y, x]. Thus, by the second part of Proposition 2.2, we have: Inn(F̂2 ) ∩ Imρ ⊆ ZInn(F̂2 ) (Inn ([y, x])) = hInn ([y, x])i = hInn (e2 )i. 6 Now, as e2 ∈ ker π, where  π is  as defined in the proof of Lemma 2.4, the \ image of hInn (e2 )i in Inn hα, βi is trivial. Thus, the image of Inn(F̂2 ) ∩ Imρ   \ in Inn hα, βi is trivial , and isomorphic to Inn(F̂2 ) ∩ Imρ as we saw that Imρ   \ is mapped isomorphically to Inn hα, βi . So Inn(F̂2 ) ∩ Imρ is trivial. This finishes the proof of Theorem 1.1. In [BER], the authors give an explicit construction of a congruence subgroup which is contained in a given finite index subgroup of Aut(F̂2 ). They prove the following theorem: Theorem 2.7. (cf. [BER], Theorem 5.1) Let H be a finite index normal subgroup of G = Aut (F2 ) such that Inn (F2 ) ≤ H ≤ Aut′ (F2 ) and let n = [Aut′ (F2 ) : H]. Pick two distinct odd primes p, q ∤ n, and set m = n · pn+1 . Then, there exists an explicitly constructed normal subgroup M ≤ F2 of index 4 dividing 144 · m4 · q 36·m +1 such that G (M ) ≤ H, when for a general normal subgroup M ⊳ F2 we define: G (M ) = {σ ∈ G | σ (M ) = M, σ acts trivially on F2 /M } We end this section with a much simpler explicit construction of a congruence subgroup and with a better bound for the index of M . But before, let us recall the “discrete version” of Proposition 2.2 from [BER]: Proposition 2.8. (cf. [BER], Propositions 2.2. and 2.6.) Let F be the free group on the set X, |X| ≥ 2, and let F/N be a finite quotient of F . Pick a prime p not dividing the order of F/N and set M = N p N ′ . Then: 1. The image of every normal abelian subgroup of F/M through the natural projection F/M → F/N , is trivial. 2. If N ⊆ F2′ F26 , x, y ∈ X, x 6= y, then the image of the centralizer ZF/M ([y, x] · M ) through the natural projection F/M → F/N , is h[y, x] · N i. Theorem 2.9. Let H be a finite index normal subgroup of G = Aut (F2 ) such that Inn (F2 ) ≤ H ≤ Aut′ (F2 ) = Inn (F2 ) ⋊ hα, βi and let n = [Aut′ (F2 ) : H]. Then for every prime p ∤ 6n, there exists an explicitly constructed normal sub4 group M ≤ F2 of index dividing 144 · n4 · p36·n +1 such that G (M ) ≤ H. Proof. Recall the map π : F2 ⊇ ∆ → hα, βi from the proof of Lemma 2.4, and let t1 = 1, t2 = x, t3 = y, t4 = xy be the system of representatives of right cosets of ∆ in F2 . Denote also K = H ∩ hα, βi and define: N = F2′ F26 = F2′ F24 \ g −1 π −1 (K) g = F2′ F26 −1 t−1 (K) ti i π i=1 g∈F2 M 4 \ ′ ∩N N p T4 Then π −1 (K) is a subgroup of index n in ∆ and i=1 ti−1 π −1 (K) ti is a normal subgroup of F2 of index dividing n4 in ∆, and of index dividing 4n4 in F2 . So as F2′ F26 is of index 9 in ∆, N is a normal subgroup of index dividing 36 · n4 in F2 . 7 4 Thus, by the Schreier formula, the index of N ′ N p in F2 divides 36 · n4 · p36·n +1 4 and the index of M in F2 is dividing 4 · 36 · n4 · p36·n +1 . So it remains to show that G (M ) ≤ H. Let σ ∈ G (M ). As M ≤ F2′ F24 we have:  G (M ) ≤ ker G → Aut F2 / F2′ F24 ≤ Aut′ (F2 ) = Inn (F2 ) ⋊ hα, βi and therefore we can write σ = Inn (f )·δ for some f ∈ F2 and δ ∈ hα, βi. By assumption, σ acts trivially on F2 /M and thus δ acts on F2 /M as Inn f −1 . Now, as α and β fix [y, x], we deduce that so does δ. Thus, f · M ∈ ZF2 /M ([y, x] · M ) and by Proposition 2.8, f · N ∈ h[y, x] · N i. Hence, δ acts on the group F2 /M as Inn ([y, x]r · n) for some r ∈ Z and n ∈ N . Therefore, δ acts on ∆/M as Inn (er2 · n) for some r ∈ Z and n ∈ N . So, δ acts on π (∆) /π (M ) = ∆/ (M · ker π) as Inn (π (er2 · n)) for some r ∈ Z and n ∈ N . But e2 ∈ ker π, so δ acts on π (∆) /π (M ) as Inn (π (n)) for some n ∈ N . Now, by the definition of N , π (N ) ⊆ K and also π (M ) ⊆ K ′ K p , so δ acts on π (∆) /K ′ K p as Inn (k) for some k ∈ K. Moreover, by the definition of π we have π (∆) = hα, βi and by the computations we made in the proof of Lemma 2.4, δ acts on hα, βi as Inn (δ). −1 Thus, there exists some k ∈ K such that Inn (δ) · Inn (k) acts trivially on ′ p −1 ′ p hα, βi /K K , i.e. δ · k ∈ Z (hα, βi /K K ). Now, by the first part of Proposition 2.8, as Z (hα, βi /K ′ K p ) is an abelian normal subgroup of hα, βi /K ′ K p it is mapped trivially to hα, βi /K . I.e. δ · k −1 ∈ K, so also δ ∈ K ⊆ H. Thus, σ = Inn (f ) · δ ∈ H, as required. 3 The CSP for Φ2 In this section we will prove Theorem 1.2, and will show that the congruence kernel of the free metabelian group on two generators is the free profinite group on a countable number of generators. Before we start, let us observe that for a group Γ, one can also ask a parallel congruence subgroup problem for G = Out (Γ). I.e. one can ask whether every finite index subgroup of G contains a principal congruence subgroup of the form: G (M ) = ker (G → Out (Γ/M )) for some finite index characteristic subgroup M ≤ Γ. When Γ is finitely generb → Out(Γ̂) ated, this is equivalent to the question whether the congruence map G is injective. Moreover, it is easy to see that Lemma 2.3 has a parallel version for G, namely, if H ≤ G is a congruence subgroup of G, then: b → Out(Γ̂)) = ker(H b → Out(Γ̂)) ker(G We start now with the next proposition which is slightly more general than Lemma 3.1. in [BER]. Nevertheless, it is proven by the same arguments: Proposition 3.1. (cf. [BER], Lemma 3.1.) Let Γ be a finitely generated residually finite group such that Γ̂ has a trivial center. Considering the congruence 8 \ map Out (Γ) → Out(Γ̂), we have: \ \ C (Γ) = ker(Aut (Γ) → Aut(Γ̂)) ∼ (Γ) → Out(Γ̂)) = ker(Out It is well known that Φ2 is a residually finite group (cf. [Be1], Theorem 2.11). It is also proven there that Z(Φ̂2 ) is trivial (proposition 2.10). So by the above proposition: \ \ C (Φ2 ) = ker(Aut (Φ2 ) → Aut(Φ̂2 )) ∼ (Φ2 ) → Out(Φ̂2 )) = ker(Out In addition, it is an old result by Bachmuth [Ba] that the kernel of the surjective map:   ker Aut (Φ2 ) → Aut (Φ2 /Φ′2 ) = Aut Z2 = GL2 (Z) = Inn (Φ2 )     1 2 1 0 i.e., Out (Φ2 ) ∼ , is a = GL2 (Z). Now, the free group 0 1 2 1 congruence subgroup of Out (Φ2 ) as it contains:  ker Out (Φ2 ) → Out Φ2 /Φ′2 Φ42 = ker (GL2 (Z) → GL2 (Z/4Z)) . So by the appropriate version of Lemma 2.3 and by Proposition 3.1, we obtain that: C (Φ2 ) = = = Now, as  C (Φ2 ) 1 0 2 1 \ ker(Out (Φ2 ) → Out(Φ̂2 )) \ ker(GL 2 (Z) → Out(Φ̂2 )) !     1 2\1 0 ker , → Out(Φ̂2 ) 0 1 2 1   1 , 2 = ker  0 1  is a free group, we can also state that: !    1 2\1 0 , → Out(Φ̂2 ) 0 1 2 1 (3.1) \ ∼ βi → Aut(F̂2 ) → Aut(Φ̂2 ) → Out(Φ̂2 )) = ker(hα, where α and β are the automorphisms  of F2that we  defined  in the previous 1 2 1 0 section, which are the preimages of and under the map 0 1 2 1 Aut (F2 ) → GL2 (Z), respectively. So all we need to show is that: \ Lemma 3.2. C (Φ2 ) = ker(hα, βi → Out(Φ̂2 )) = F̂ω . 9     1 2 1 0 Proof. As the free group , is a congruence subgroup of 0 1 2 1   the group Aut Z2 = Out Z2 = GL2 (Z), we have: 2 C Z  = ker  !    1 2\1 0 2 , → Out(Ẑ ) 0 1 2 1 \ = ker(hα, βi → Out(Ẑ2 )) \ = ker(hα, βi → Aut(Φ̂2 ) → Out(Φ̂2 ) → Out(Ẑ2 ) = Aut(Ẑ2 )) \ Thus, if we denote: C = ker(hα,  βi → Aut(Φ̂2 )), then using equation (3.1), we have: C ≤ C (Φ2 ) ≤ C Z2 . Now, if we consider the action of Φ̂2 on Φ′2 = ker(Φ̂2 → Ẑ2 ) by conjugation, then as Φ′2 is abelian, we actually obtain an action on Φ′2 as a Z[Φ̂2 /Φ′2 ] = Z[Ẑ2 ]-module, which is generated by the 2 element [y, x] as a Z[Ẑ2 ]-module, since hx, y | [y, x] = 1i is a presentation of  Z . 2 Moreover, as we observed previously, α and β fix [y, x]. Therefore, C Z acts trivially not only on Φ̂2 /Φ′2 = Ẑ2 but also on Φ′2 . Let us now make the following observation: if σ, τ are two automorphisms of a group Γ which act trivially on Γ/M and on M , where M ⊳ Γ is abelian, then σ and τ commute. Indeed, if g ∈ Γ, then σ (g) = g · m and τ (g) = g · n for some m, n ∈ M , and thus: τ (σ (g)) = τ (g · m) = g · n · m = g · m · n = σ (g · n) = σ (τ (g)) . The conclusion from the above observation and from the previous discussion  is that C Z2 /C is abelian, and thus, C Z2 /C (Φ2 ) is also abelian. Finally,  C Z2 is known to be isomorphic to F̂ω ([Me], [Lu1]). Moreover, by Proposition 1.10 and Corollary 3.9 of [LV] every normal closed subgroup N of F̂ω such that F̂ω /N is abelian, is also isomorphic to F̂ω . Thus, C (Φ2 ) ∼ = F̂ω as well, as required. Remark 3.3. Our proof of Theorem 1.2 is much shorter than the one given in [Be1], but the latter gives more information. We show here that C Z2 /C  (Φ2 ) 2 is abelian, while from [Be1] one can deduce that, infact, C (Φ ) = C Z . See 2 u 5 fore more. 4 The CSP for Φ3 In this section we will prove Theorem 1.3 which claims that C (Φ3 ) contains a copy of F̂ω . Let us start by showing that Aut (Φ3 ) is large: Proposition 4.1. The group Aut (Φ3 ) is large, i.e. it has a finite index subgroup that can be mapped onto a non-abelian free group. 10 Proof. The proof will follow the method developed in [GL] to produce arithmetic quotients of Aut (Fn ). Denote the free group on 3 generators by F3 = hx, y, zi, and the cyclic ( group of order 2 by C2 = {1, g}. Define the map π : F3 → C2 x 7→ g by: π = , and its kernel by R = ker π. Then, using the right y, z 7→ 1 transversal T = {1, x}, we deduce by Theorem 2.5 that R is freely generated by: x2 , y, xyx−1 , z, xzx−1 . Thus, R̄ = R/R′ = Z5 is generated as a free abelian group by the images: v1 = x2 , v2 = y, v3 = xyx−1 , v4 = z, v5 = xzx−1 Now, the action of F3 on R by conjugation induces an {1, g} on R̄ = R/R′ , sending:    v1 = x2 7→ x2 = v1     −2 −1 2 −1   = v3 v2 = y 7→ x (xyx ) x = xyx  −1 g 7→ v3 = xyx 7→ y = v2 =       v4 = z 7→ x−2 (xzx−1 ) x2 = xzx−1 = v5    v = xzx−1 7→ z̄ = v4 5 action of F3 /R = C2 = 1 0 0 0 0 0 0 1 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 1 0    =B   The above matrix has two eigenvalues λ = ±1 and the eigenspaces are: V1 V−1 = = Sp {v1 , v2 + v3 , v4 + v5 } Sp {v2 − v3 , v4 − v5 } Recall, Φ3 = F3 /F3′′ , and as F3 /R is abelian, F3 /R′ is metabelian. Thus, we have a surjective homomorphism: Φ3 ։ F3 /R′ . Denote now: S = R/F3′′ , so we can identify: F3 /R ∼ = S/S ′ = S̄. So as = Φ3 /S ′ and R̄ = R/R′ ∼ = Φ3 /S, F3 /R′ ∼ before, Φ3 /S = C2 acts on S̄ by the matrix B. Denote also G (S) = {σ ∈ Aut (Φ3 ) | σ (S) = S}. It is clear that G (S) is of finite index in Aut (Φ3 ) with  a natural map: G (S) → Aut (S) which induces a map: ρ : G (S) → Aut S̄ = GL5 (Z). We claim now that if σ ∈ G (S) than ρ (σ) commutes with B. First observe that there exists some s ∈ S such that σ (x) = sx (x now plays the role of the image of x under the map F3 → Φ3 ). Now, let t ∈ S, and remember that the action of B on S̄ is induced by the action of x on S by conjugation. So:  σ x−1 tx = σ (x)−1 σ (t) σ (x) = = = and hence: (ρ (σ) · B) (t̄) x−1 s−1 σ (t) sx = −1 −1   x−1 sx x σ (t) x x−1 sx = σ (x−1 tx) = −1 = (x−1 sx) (x−1 σ (t) x) (x−1 sx) = = x−1 σ (t) x = (B · ρ (σ)) (t̄) 11 Therefore, ρ (G (S)) commutes with B. It follows that the eigenspaces of B are invriant under the action of G (S). Inparticular, we deduce that V−1 is invariant under the action of ρ (G (S)). Thus, we obtain a homomorphism ν : G (S) →  Aut V−1 ∩ S̄ = GL2 (Z). Consider now the following automorphisms of Aut (Φ3 ) (x, y, z play the role of the images of x, y, z under F3 → Φ3 ):     x 7→ x x 7→ x α = y 7→ y , β = y 7→ yz     z 7→ zy z 7→ z So α, β ∈ G (S) act on V−1 = Sp {u1 = v2 − v3 , u2 = v4 − v5 } in the following way:   α (u1 ) = α y − xyx−1 = y − xyx−1 = u1   α (u2 ) = α z − xzx−1 = z + ȳ − xzx−1 − xyx−1 = u2 + u1   β (u1 ) = β y − xyx−1 = ȳ + z − xyx−1 − xzx−1 = u1 + u2   β (u2 ) = β z − xzx−1 = z − xzx−1 = u2   1 1 Therefore, under the map ν : G (S) → GL2 (Z) we have: α 7→ and 0 1       1 0 1 2 1 0 β 7→ . Thus, the image of G (S) contains , 1 1 0 1 2 1 which is free and of finite index in GL (Z). Finally, if we denote the preimage 2     1 2 1 0 H = ν −1 , , then H is a finite index subgroup of 0 1 2 1 Aut (Φ3 ) that can be mapped onto a free group, as required. Let us now continue with the following definition: Definition 4.2. We say that a group P is involved in a group Q, if it isomorphic to a quotient group of some subgroup of Q. It is not difficult to see that if a finite group P is involved in a profinite group Q, than it is involved in a finite quotient of Q. Now, we showed that Aut (Φ3 ) has a finite index subgroup H which can be mapped onto F2 . Thus we b ։ F̂2 , but as F̂2 is free, the map splits, and thus H b and hence have a map: H \ \ Aut (Φ3 ), contains a copy of F̂2 . Thus, any finite group is involved in Aut (Φ3 ). On the other hand, we claim: Proposition 4.3. Let P be a non-abelian finite simple group which is involved  in Aut(Φ̂3 ). Then, for some prime p and some d ∈ N, P is involved in SL3 pd , the special linear group over the field of order pd . 12 Proof. Let Fn be the free group on x1 , . . . , xn . Then there is a natural injective homomorphism from Fn into the matrix group: ) (  n X g 0 Z [Fn ] ti | g ∈ Fn , t ∈ t 1 i=1 defined by the map: xi 7→  xi ti 0 1  , 1≤i≤n where ti is a free basis for a right Z [Fn ]-module. This is called the Magnus embedding. Usually, its properties are studied by Fox’s free differential calculus, but we will not need it here explicitly (cf. [Bi], [RS], [Ma]). One can prove, by induction that for a word w ∈ Fn , under  on its length,  w 0 Pn the Magnus embedding, w 7→ where: 1 i=1 wi ti w−1= n X (4.1) (xi − 1) wi . i=1 The identity (4.1) shows that the polynomials wi determine the word w uniquely. Thus, we have an injective map (which is not homomorphism) J : End (Fn ) → Mn (Z [Fn ]) defined by:   α (x1 )1 · · · α (xn )1 J   .. .. α 7→   . . α (x1 )n ··· α (xn )n It is not difficult to check, using the identity (4.1), that the above map satisfies: J (α ◦ β) = J (α) · α (J (β)) where by α (J (β)) we mean that α acts on every enrty of J (β) separately. Now, for m ∈ N, denote: Kn,m = Fnm Fn′ and Zm = Z/mZ. Then, the natural maps Fn → Fn /Kn,m = Znm and Z → Zm induce a map: ) (  n X g 0 Z [Fn ] ti πn,m : Fn → | g ∈ Fn , t ∈ t 1 i=1 ) (  n X g 0 Zm [Znm ] ti . → | g ∈ Znm , t ∈ t 1 i=1 m ′ It is shown in ([Be1], Proposition 2.6) that ker (πn,m ) = Kn,m Kn,m and m ′ hence Φn,m := Im (πn,m ) ∼ = Fn /Kn,m Kn,m . Moreover, it is proven there (Proposition 2.7) that we have the following equality: Φ̂n = ← lim −m Φn,m . 13 Observe now that for every m2 |m1 , ker (Φn,m1 → Φn,m2 ) is characteristic in Φn,m1 , and for every m, ker(Φ̂n → Φn,m ) is characteristic in Φ̂n . Thus: Aut(Φ̂n ) = Aut(lim lim ←−m Φn,m ) = ← −m Aut (Φn,m ) . Now, observe that the identity (4.1) is also valid for the entries of the elements of Φn,m , and thus, every element of Φn,m is determined by its left lower coordinate. Therefore, as every automorphism of Φn,m can be lifted to an endomorphism of Fn , we have an injective map (which is not a homomorphism) Jm : Aut (Φn,m ) → Mn (Zm [Znm ]) which satisfies the identity: Jm (α ◦ β) = Jm (α) · α (Jm (β)) where the action of α on Zm [Znm ] = Zm [Fn /Kn,m] is through the natural prom ′ Kn,m → Fn /Kn,m ∼ jection Φn,m ∼ = Znm . = Fn /Kn,m We denote now KA (Φn,m ) = ker (Aut (Φn,m ) → Aut (Φn,m /Kn,m )). Observe, that as KA (Φn,m ) acts trivially on Φn,m /Kn,m = Znm , the map Jm gives us a homomorphism, which is also injective, as mentioned above: Jm : KA (Φn,m ) → GLn (Zm [Znm ]) Now, if P is a non-abelian simple group which is involved in Aut(Φ̂3 ), then it must be involved in Aut (Φ3,m ) for some m. Thus, it must be involved either in  Aut (Φ3,m /K3,m ) = GL3 (Zm ) or in KA (Φ3,m ) ≤ GL3 Zm Z3m . So it must be involved in GL3 (R) for some finite commutative ring R. As every finite commutative ring is artinian, it can be decomposed as: R = R1 × . . . × Rl for some local finite rings R1 , . . . , Rl , so: GL3 (R) = GL3 (R1 ) × . . . × GL3 (Rl ) and thus P must be involved in GL3 (R) for some local finite commutative ring R. Denote the unique maximal ideal of R by M ⊳ R. As R is a finite local Noetherian ring, it is well known that M r = 0 for some r ∈ N. Note now that if S, T ⊳ R for some commutative ring R, and I + A ∈ ker (GL3 (R) → GL3 (R/S)) I + B ∈ ker (GL3 (R) → GL3 (R/T )) when I denotes the identity element in GL3 (R), then [I + A, I + B] ∈ ker (GL3 (R) → GL3 (R/ST )) . −1 Indeed, if I + C = (I + A) and I + D = (I + B) AD = CB = 0 (mod ST ) we have: [I + A, I + B] = = = −1 then, as AB = CD = (I + A) (I + B) (I + C) (I + D) = I + AC + A + BD + B + C + D (mod ST ) I + (I + A) (I + C) − I + (I + B) (I + D) − I = I (mod ST ) 14 With the above observation we deduce   that for every k ≥ 1, the kernel of the map GL3 R/M k+1 → GL3 R/M k is abelian. So, P must be involved in GL3 (R/M) = GL3 pd for some prime p and d ∈ N. Finally, using the fact  that GL3 pd /SL3 pd is abelian, we obtain that P is involved in SL3 pd , as required. Corollary 4.4. There exists a finite simple group which is not involved in Aut(Φ̂3 ). Proof. By the proposition above, it is enough to show that there is a finite  simple non-abelian group which is not involved in SL3 pd for any prime p and d ∈ N. Now, by a theorem of Jordan, there exists an integer-valued function J (n) such that for every field F , char (F) = 0, any finite subgroup of GLn (F) contains a normal abelian subgroup of index at most J (n). As a corollary of this theorem, Schur proved that the same holds (with the same function) for any finite subgroup Q ≤ GLn (F) with char (F) = p > 0, provided p ∤ |Q| (cf. [We] chapter 9). Clearly, the same holds for any group which is involved in such a finite group Q.  We claim that for n large enough, Alt (n) is not involved in SL3 pd for any p and d. Indeed, fix two different primes q1 and q2 larger than J (3). Then, for n sufficiently large (e.g. n > qi3 ) the qi -sylow subgroup Si of Alt (n) is non-abelian (since Alt (n) contains the non-abelian qi -group of order qi3 ) and every subgroup of Si of index  ≤ J (3) is equal to Si , so also non-abelian. If Alt (n) were involved in SL3 pd then for at least one of the qi , qi 6= p, a contradiction. Corollary 4.5. The congruence kernel C (Φ3 ) contains a copy of F̂ω . Proof. The immediate conclusion of Corollary 4.4 is that Aut(Φ̂3 ) does not contain a copy of F̂2 . Thus, the intersection of C (Φ3 ) and the copy of F̂2 \ in Aut (Φ3 ) is not trivial. Thus, C (Φ3 ) contains a non-trivial normal closed subgroup N of F̂2 . By Theorem 3.10 in [LV] it contains a copy of F̂ω , as required. 5 Remarks and open problems We end this paper with several remarks and open problems. Denote the free solvable group of derived length r on 2 generators by Φ2,r . By combining the results of ([BFM], Theorem 1) and ([KLM], Theorem 1.4) we have:   ker Aut (Φ2,r ) → Aut Z2 = GL2 (Z) = Inn (Φ2,r ) u for every r, i.e. Out (Φ2,r ) = GL2 (Z). So by the same arguments as in 3 we have: \ C (Φ2,r ) = ker(hα, βi → Out(Φ̂2,r )) 15 As Out(Φ̂2,r+1 ) is mapped onto Out(Φ̂2,r ), we obtain the sequence:  = C (Φ2,1 ) ≥ C (Φ2 ) = C (Φ2,2 ) ≥ C (Φ2,3 ) ≥ C Z2 ≥ C (Φ2,4 ) ≥ . . . ≥ C (Φ2,r ) ≥ . . . ≥ C (F2 ) = {e} and a natural question is whether the inequalities are strict or not. An equivalent reformulation of this question is the following: the cosets of the kernels ker(GL2 (Z) = Out (Φ2,r ) → Out (Φ2,r /K)) for characteristic finite index subgroups K ≤ Φ2,r provide a basis for a topology C (r) on GL2 (Z), called the congruence topology with respect to Φ2,r , which is weaker (equal) than the profinite topology F of GL2 (Z), and stronger (equal) than the classical congruence topology of GL2 (Z). The latter is equal to C (1). So, the question above is equivalent to the question whether these topologies are strictly weaker than F , and whether the topology C (r), for a given r, is strictly weaker than C (r + 1). For example, Theorem 1.1 which states that C (F2 ) = {e} is equivalent to the statement that the congruence topology which Out(F̂2 ) induces on Out (F2 ) = GL2 (Z) is equal to the profinite topology of GL2 (Z). Considering Theorem 1.2 we deduce that C (2) $ F , but with the proof we gave here one can not decide whether C (1)  = C (2) or C (1)  $ C (2). Equivalently, we can not decide whether C Z2 = C (Φ2 ) or C Z2 C (Φ2 ). But, in [Be1] it was shown quite surprisingly, that:  Theorem 5.1. C (1) = C (2), or equivalently C Z2 = C (Φ2 ). The proof in [Be1] suggested to conjecture that C (1) = C (2) = C (r) u for every r. But, the explicit construction of a congruence subgroup we gave in 2 gives a counter example:  Proposition 5.2. C (1) $ C (r) for every r ≥ 3. Equivalently C Z2 C (Φ2,r ) for every r ≥ 3.     1 2 1 0 Proof. Denote G = , ≤ GL2 (Z). Then by a theorem 0 1 2 1 of Reiner [Re], for every p 6= 2, G′ Gp is not a congruence subgroup of GL2 (Z) in the classical manner, i.e. G′ Gp ∈ / C (1). On the other hand, applying the explicit construction given in Theorem 2.9, we obtain a finite index normal subgroup M ⊳ F2 such that F2 /M is of solvability length 3 such that1 : ker (Out (F2 ) = GL2 (Z) → Out (F2 /M )) ≤ G′ Gp . This shows that G′ Gp is a congruence subgroup of GL2 (Z) with respect to the congruence topology induced by Out(Φ̂2,3 ). Equivalently, C (1) $ C (3) or  C Z2 C (Φ2,3 ), as required. 1 We by T remark that if one wants M to be characteristic, all we need to do, is to replace M σ (M ), and this procedure does not change the solvability length of F2 /M . σ∈Aut(F2 ) 16 The proposition suggests the following conjecture: Conjecture 5.3. C (Φ2,r ) C (Φ2,r+1 ) for every r ≥ 2, or equivalently C (r) $ C (r + 1). In particular, C (Φ2,r ) 6= {e} = C (F2 ) and C (r) 6= F for every r. We should remark that we do not even know to decide whether C (Φ2,r ) 6= {e} for r ≥ 3, i.e. we do not know if the congruence subgroup property holds for Φ2,r for r ≥ 3 or not. Note that our proofs of Theorems 1.2 and 1.3 claiming that Φ = Φ2 = Φ2,2 and Φ = Φ3 do not satisfy the CSP were based on two facts: \ 1. Aut (Φ) is large, and hence every finite group is involved in Aut (Φ), and 2. not every finite group is involved in Aut(Φ̂). Now, for Φ = Φd,r , the free solvable group on d ≥ 2 generators and solvability length r,upart 2 is valid for 1 ≤ r ≤ 2 and every d (with the same proof as for d = 3 in 4). But, as C (Φd,1 ) = {e} for every d ≥ 3, and C (Φd,2 ) is abelian for every d ≥ 4 (cf. [Be2]), part 1 is not valid in these cases. On the other hand, for Φ = Φ2,r or Φ = Φ3,r , part 1 is still true for every r ≥ 2 but not part 2. Infact, we have: Proposition 5.4. Let Φd,r be the free solvable group on d ≥ 2 generators and solvability length r. Then if r ≥ 3, then every finite group H is involved in Aut(Φ̂d,r ). Proof. By the same arguments of ([Lu2], 5.2), it can be deduced from Gaschutz’s Lemma that for every surjective homomorphism π : Φ̂d,r → Γ where Γ is finite, the homomorphism n o Aut(Φ̂d,r ) ≥ σ ∈ Aut(Φ̂d,r ) | σ (ker π) = ker π → Aut (Γ) is surjective. Thus, for proving our proposition it suffices to show that Φ̂d,r has a finite quotient Γ such that H is involved in Aut (Γ). Now, by Cayley’s Theorem, H is a subgroup of Sym (n − 1) for some n and the later is a subgroup of SLn (p) for every prime p. Thus, the next lemma due to Robert Guralnick, finishes the proof of the proposition. Lemma 5.5. For every n ≥ 2, there exists a prime p and a finite group Γ generated by two elements and of solvability length three, such that SLn (p) is involved in Aut(Γ). Proof. Fix a prime r such that r > n + 1. Using Dirichlet’s Theorem, pick a prime p such that r divides p − 1. Consider now the general affine group    a 0 ∆ = AGL1 (r) = | a ∈ F∗r , b ∈ Fr = Fr ⋊ F∗r . b 1 17 Then ∆ is of order r (r − 1). In addition, as r| (p − 1), Fp contains the unit roots of order r, fix one of them ξ 6= 1, and consider the diagonal matrix   ξ ··· 0  ..  ∈ GL D =  ... . . . r−1 (p) . .  0 ··· ξ r−1 Now, we can embed ∆ in GLr−1 (p) by sending an element b ∈ {0,. . . , r − 1} = Fr to the diagonal matrix Db (giving rise to a subgroup N = Db | b ∈ Fr ) and an element a ∈ F∗r to the permutation matrix which normalizes N , sending Db to Dba . So ∆ has a module V of dimension r − 1 over Fp . Now, every ∆-submodule of V is also N -submodule. The N -submodules are direct sums of different one dimensional N -modules, the eigen-spaces of D1 , on which F∗r acts transitively. We deduce that V is an irreducible module. Denote now W = ⊕r−2 i=1 V and using the obvious action of ∆ on W , define: Γ = W ⋊ ∆. We claim that Γ is generated by two elements. By the description above, it is clear why ∆ is generated by two element, one of them is D ∈ Fr and we denote the other one by S ∈ F∗r . Let us now define   D′ = ((~e1 , . . . , ~er−2 ), D) , S ′ = (~0, . . . , ~0), S ∈ W ⋊ ∆ where {~e1 , . . . , ~er−1 } is the standard basis of V . For a 1 ≤ j ≤ r − 1 denote k+1 6= 0. η = ξ j . Note, that for every 1 ≤ k ≤ r − 2, 1 + η + . . . + η k = 1−η 1−η  It follows that D′k = (α1~e1 , . . . , αr−2~er−2 ), Dk where 0 6= αi ∈ Fp for every 1 ≤ k ≤ r − 2. Now, there is a power S l of S, 1 ≤ l ≤ r − 2, which sends ~er−1 to ~e1 . We have also S l DS −l = Dr−k for some 1 ≤ k ≤ r − 2. Thus, for some 0 6= αi ∈ Fp , we can write: w = = S ′l D′ S ′−l D′k ((~0, . . . , ~0), S l )((~e1 , . . . , ~er−2 ), D)((~0, . . . , ~0), S −l )((α1~e1 , . . . , αr−2~er−2 ), Dk ) ((S l (~e1 ), . . . , S l (~er−2 )), S l DS −l )((α1~e1 , . . . , αr−2~er−2 ), Dk ) = (S l (~e1 ) + Dr−k (α1~e1 ), . . . , S l (~er−2 ) + Dr−k (αr−2~er−2 ), I) ∈ W. = Now, as S l sends ~er−1 to ~e1 , ~e1 does not appear in any entry of w except the first one. Observe now, that the diagonals of D0 , . . . , Dr−2 , considered as column vectors of V = Fr−1 p , form a basis for V as the matrix:      1 1 .. . ξ ξ2 .. . 1 ξ r−1 ··· ··· ··· ξ r−2 ξ 2(r−2) .. . ξ (r−1)(r−2)      is a Vandermonde matrix, and therefore invertible. Thus, there is a linear 18 combination    C = β0 D0 + . . . + βr−2 Dr−2 =   1 0 0 0 .. . 0 0 ··· .. . ··· 0    ..  , βi ∈ Fp . .  0 Now, observe that D′ acts on W by conjugation via the action of D on V . Thus, we obtain an action of C on W via its action on V , in which C (w) has ~0 in every entry except the first one. This shows, as V is irreducible, that the first copy of V in W is inside the group generated by D′ and S ′ . In a similar way, all the r − 2 copies of V are generated by D′ and S ′ , so Γ is generated by two elements. r−2 Now, ∆ × SLr−2 (p) acts on W = ⊕r−2 in a obvious way. i=1 V = V ⊗ Fp Thus Γ = W ⋊ ∆ is normal in W ⋊ (∆ × SLr−2 (p)), so SLr−2 (p) is involved in Aut (Γ). Let us remark that while we do not know the answer to the congruence subgroup problem for free solvable groups on two generators and solvability rank r (unless r = 1 or 2), the situation with free nilpotent groups on two generators is easier: Proposition 5.6. For every free nilpotent group on two generators Γ, the congruence kernel contains a copy of F̂ω - the free profinite group on countable number of generators. Proof. It is known that if Γ̂ is a pro-nilpotent group, then the kernel of the map Aut(Γ̂) → Aut(Γ̂/Γ′ ) is pro-nilpotent (cf. [Lu2], 5.3). Thus, if Γ is a free nilpotent group (of arbitrary class) then by the same arguments we brought before, there exists a finite group which is not involved in Aut(Γ̂). On the other hand, if Γ is free nilpotent group on two generators, then Aut (Γ) is large, as it \ can be mapped onto GL2 (Z) 2 . Thus, F̂2 is a subgroup of Aut (Γ) and C (Γ)∩ F̂2 is non-trivial, hence contains a copy of F̂ω (cf. [LV]). Our last remark is about the CSP for subgroups of automorphism groups. Considering the classical congruence subgroup problem, one can take G to be a subgroup of GLn (R) where R is a commutative ring, and ask whether every finite index subgroup of G contains a subgroup of the form ker (G → GLn (R/I)) for some finite index ideal I ⊳ R. This direction of generalization of the classical CSP has been studied intensively during the second half of the 20th century (cf. [Rag], [Rap]). One can ask for a parallel generalization for automorphism groups or outer atomorphism groups. I.e. let G ≤ Aut (Γ) (resp. G ≤ Out (Γ)), does every finite index subgroup of G contain a principal congruence subgroup of the form ker (G → Aut (Γ/M )) (resp. ker (G → Out (Γ/M ))) for some characteristic finite index subgroup M ≤ Γ? 2 In general, the kernel of the map Aut (Γ) → GL (Z) strictly contains Inn (Γ) (cf. [BG], 2 [An]) 19 Now, let πg,n be the fundamental group of Sg,n , the surface of genus g with n punctures, such that χ (Sg,n ) = 2−2g−n ≤ 0. Then, there is an injective map of P M od (Sg,n ), the pure mapping class group, into Out (πg,n ) (cf. [FM], chapter 8). Thus, one can ask the CSP for P M od (Sg,n ) as a subgroup of Out (πg,n ). Considering the above problem, it is known that: Theorem 5.7. For g = 0, 1, 2 and every n > 0, P M od (Sg,n ) has the CSP. The cases for g = 0 were proved by [DDH] and in [Mc], the cases for g = 1 were proved by [As], and the cases for g = 2 where proved by [Bo]. It can be shown that for every n > 0, πg,n ∼ = F2g+n−1 = the free group on 2g + n − 1 generators. Thus, the above cases give an affirmative answer for various subgroups of the outer aoutomorphism group of finitely generated free groups. Though, the CSP for the full Out (Fd ) where d ≥ 3 is still unsettled. References [An] S. Andreadakis, On the automorphisms of free groups and free nilpotent groups, Proc. London Math. Soc. 15, 1965, 239–268. [As] M. Asada, The faithfulness of the monodromy representations associated with certain families of algebraic curves, J. Pure Appl. Algebra 159, 2001, 123–147. [Ba] S. Bachmuth, Automorphisms of free metabelian groups, Trans. Amer. Math. Soc. 118, 1965, 93-104. [Be1] D. E-C. Ben-Ezra, The congruence subgroup problem for the free metabelian group on two generators, Groups Geom. Dyn. 10, 2016, 583–599. [Be2] D. E-C. Ben-Ezra, The congruence subgroup problem for the free metabelian group on n ≥ 4 generators, in preparation. [Bi] J. S. Birman, Braids, links, and mapping class groups, Princeton University Press, Princeton, NJ, University of Tokyo Press, Toyko, 1975. [Bo] M. Boggi, The congruence subgroup property for the hyperelliptic modular group: the open surface case, Hiroshima Math. J. 39, 2009, 351–362. [BER] K-U. Bux, M. V. Ershov, A. S. Rapinchuk, The congruence subgroup property for Aut (F2 ): a group-theoretic proof of Asada’s theorem, Groups Geom. Dyn. 5, 2011, 327–353. [BFM] S. Bachmuth, E. Formanek, H. Mochizuki, IA-automorphism of certain two-generator torsion-free groups, J. Algebra 40, 1976, 19–30. [BG] R. M. Bryant, C. K. Gupta, Automorphism groups of free nilpotent groups, Arch. Math. (Basel) 52, 1989, 313–320. 20 [DDH] S. Diaz, R. Donagi, D. Harbater, Every curve is a Hurwitz space, Duke Math. J. 59, 1989, 737–746. [FM] B. Farb, D. Margalit, A primer on mapping class groups, Princeton Mathematical Series, 49. Princeton University Press, Princeton, NJ, 2012. [GL] F. Grunewald, A. Lubotzky, Linear representations of the automorphism group of a free group, Geom. Funct. Anal. 18, 2009, 1564–1608. [KLM] P. H. Kropholler, P. A. Linnell, J. A. Moody, Applications of a new K-theoretic theorem to soluble group rings, Proc. Amer. Math. Soc. 104, 1988, 675–684. [Lu1] A. Lubotzky, Free quotients and the congruence kernel of SL2 , J. Algebra 77, 1982, 411–418. [Lu2] A. Lubotzky, Combinatorial group theory for pro-p-groups, J. Pure Appl. Algebra 25, 1982, 311–325. [LV] A. Lubotzky, L. van den Dries, Subgroups of free profinite groups and large subfields of Q̃, Israel J. Math. 39, 1981, 25–45. [Ma] W. Magnus, On a theorem of Marshall Hall, Ann. of Math. 40, 1939, 764–768. [Mc] D. B. McReynolds, The congruence subgroup problem for pure braid groups: Thurston’s proof, New York J. Math. 18, 2012, 925–942. [Me] O. V. Mel´nikov, Congruence kernel of the group SL2 (Z), in Russian, Dokl. Akad. Nauk SSSR 228, 1976, 1034–1036. [MKS] W. Magnus, A. Karrass, D. Solitar, Combinatorial group theory: Presentations of groups in terms of generators and relations, Interscience Publishers, New York-London-Sydney, 1966. [Rag] M. S. Raghunathan, The congruence subgroup problem, Proc. Indian Acad. Sci. Math. Sci. 114, 2004, 299–308. [Rap] A. S. Rapinchuk, The congruence subgroup problem, Algebra, K-theory, groups, and education, New York, 1997, 175–188. [Re] I. Reiner, Normal subgroups of the unimodular group, Illinois J. Math. 2, 1958, 142–144. [RS] V. N. Remeslennikov, V. G. Sokolov, Some properties of a Magnus embedding. 2 Algebra i Logika 9, 1970, 566–578, in Russian. English translation, Algebra and 3 Logic 9, 1970, 342–349. 21 [We] B. A. F. Wehrfritz, Infinite linear groups. An account of the grouptheoretic properties of infinite groups of matrices, Ergebnisse der Matematik und ihrer Grenzgebiete, Band 76. Springer-Verlag, New York-Heidelberg, 1973. Institute of Mathematics The Hebrew University Jerusalem, ISRAEL 91904 [email protected] [email protected] 22
4math.GR
arXiv:1704.04722v2 [cs.RO] 11 May 2017 Chapter 1 Bounded Distributed Flocking Control of Nonholonomic Mobile Robots Thang Nguyen1 , Hung M. La2 , Vahid Azimi3 , and Thanh-Trung Han4 There have been numerous studies on the problem of flocking control for multiagent systems whose simplified models are presented in terms of point-mass elements. Meanwhile, full dynamic models pose some challenging problems in addressing the flocking control problem of mobile robots due to their nonholonomic dynamic properties. Taking practical constraints into consideration, we propose a novel approach to distributed flocking control of nonholonomic mobile robots by bounded feedback. The flocking control objectives consist of velocity consensus, collision avoidance, and cohesion maintenance among mobile robots. A flocking control protocol which is based on the information of neighbor mobile robots is constructed. The theoretical analysis is conducted with the help of a Lyapunov-like function and graph theory. Simulation results are shown to demonstrate the efficacy of the proposed distributed flocking control scheme. 1.1 INTRODUCTION It is well-known that the collective behavior of self-propelled organisms constitutes flocking [1]. The coherent motion of the flock inspires various research on flocking control of multiagent systems. A typical objective is to achieve a desired collective motion which can be produced by a constructive flocking control procedure. For numerous models, which are described from simplest models such as point-mass models to actual physical models, design protocols have been systematically proposed for multiagent systems [2, 3]. Several control strategies were also addressed in noisy environments where the agent’s position is affected by noise [4–6]. 1 Advanced Robotics and Automation (ARA) Lab, Department of Computer Science and Engineering, University of Nevada, Reno, NV89557, USA (e-mail: [email protected]) 2 Advanced Robotics and Automation (ARA) Lab, Department of Computer Science and Engineering, University of Nevada, Reno, NV89557, USA (e-mail:[email protected]) 3 School of Electrical and Computer Engineering, Georgia Institute of Technology, 777 Atlantic Drive NW Atlanta, GA 30332-0250, USA (e-mail: [email protected]) 4 Faculty of Electrical and Electronics Engineering, Ton Duc Thang University, Ho Chi Minh City, Vietnam (e-mail: [email protected]) 2 Swarm Intelligence - From Concepts to Applications With point-mass models, the problem of flocking control of multiple agents has been addressed with typical results reported in [7–11]. For a wide range of engineering applications, extensive studies in flocking control of mobile robots have been done in various scenarios [12–14]. In this chapter, we study the problem of distributed flocking control of mobile robots by bounded feedback, which takes into consideration nonholonomic nature of mobile robots as well as the implementation issue posed by the physical limit of the motor speed. Our flocking control problem employs the full dynamic model of the mobile robot derived in [15]. Similar to [16, 17], due to the nonholonomic property of the dynamics of mobile robots, our proposed design framework constructed to achieve velocity consensus is modular. In other words, the consensuses on the linear speed and orientation angles are obtained separately. In this chapter, we are interested in agents with nonholonomic dynamics and boundedness constraints. Specifically, a coordination function is proposed to ensure that the induced attractive and repulsive forces are bounded, and hence can be incorporated in the bounded velocity control. Using the results of Barbalat’s lemma and graph theory, the theoretical analysis is conducted, which shows that the maximal value of the coordination function determines the basin of attraction for the flocking convergence. In this chapter, graph theory will be employed as in the case of nearest neighbor communication [7, 18]. We will employ the velocity control law reported in [16, 17] in a decentralised sense, which helps to avoid collision and maintain a linear speed consensus. In addition, the orientation consensus will be achieved using a modified approach, which is inspired by the one in [19], where the input constraint is taken into account. The organization of the chapter is as follows. Section1.2 summarises some research work in the literature related to the topic in this chapter. In Section 1.3, the multiple-goal control problem for flocking of nonholonomic mobile robots and preliminaries are introduced. Section 1.4 describes main results where a modular design framework is proposed for bounded velocity control and bounded orientation control and the theoretical analyses are introduced. In Section 1.5, a description of an obstacle avoidance scheme is presented. Section 1.6 shows some simulation results. Section 1.7 concludes the chapter by some conclusions. Notations: R and R+ are the sets of real numbers and nonnegative real numbers, respectively; for q = [q1 , . . . , qn ]T , ∇q = [∂ /∂ q1 , . . . , ∂ /∂ qn ]T is the del operator [20]; for two vectors a and b, a · b is their scalar product; (a1 , . . . , an ) is [aT1 , . . . , aTn ]T ; | · | is the absolute value of scalars; and k · k is the Euclidean norm of vectors. 1.2 RELATED WORK In many applications, the mission carried out by a single complicated robotic system can be equivalently completed by a coordination of a mobile robotic system with much simpler configurations, whose advantages lie in scalability, flexible deployment, cheaper costs, reliability, etc. Therefore, more sophisticated tasks can be Bounded Distributed Flocking Control of Nonholonomic Mobile Robots 3 fulfilled using a group of small mobile robots with lower cost and higher efficiency than a complex unit; see [2, 14, 21–39] and references therein. Flocking control of mobile robots was widely addressed with different control schemes; see [9, 18, 21, 40–42] and references therein. Recently, a new measuretheoretic approach which systematically provides a framework for obtaining flocking protocols for mobile robots was reported in [13]. The common assumption in many papers is the availability of information of all agents or all-to-all communication. Numerous control protocols for mobile robots have been constructed based on this assumption. This centralized communication control architecture yields inflexibility and large computation costs for the controller of each agent, especially when the number of agents is large. Meanwhile, a distributed control protocol can offer an ease of implementation and less computational burden as each element of the system needs only the information of neighbor agents. In this direction, a range of decentralized control schemes for mobile robots have been proposed [8, 10, 18, 43, 44]. For a wide range of engineering applications, cohesion maintenance and collision avoidance (CMCA) properties of a mobile robotic system are of importance. As reported in [21,45], the attractive and repulsive forces cannot be included in the control for CMCA of mobile robots, as it is possible for point-mass agents [7]. In [13], desired attractive and repulsive forces for CMCA of mobile robots was achieved using a new rearrangement strategy. In [7, 10, 18], the graph theory was employed to generate control protocols that maintain CMCA of multiagent systems with double integrator models. In [19], a distributed flocking control approach was proposed but no constraints on the control inputs are imposed. The work in [16, 17] considers the bounded feedback flocking control problem for nonholonomic mobile robots without a flocking desired heading angle. The problem of interest in this chapter is to address bounded control of nonholonomic dynamic mobile vehicles, which also achieves CMCA and obstacle avoidance. We also consider a flocking desired heading angle, which reveals a collective flocking behaviour. 1.3 PROBLEM FORMULATION Similarly to [16, 17], we investigate a collective system of N identical autonomous mobile robots whose respective equations of motion are [15] q̇i = vi e(θi ) θ̇i = wi v̇i = ui ẇi = τi (1.1) where i = 1, ..., N, qi = [xi , yi ]T ∈ R2 , and θi ∈ R are respectively the position and the heading angle of the i-th robot in the inertial frame Oxy; vi ∈ R is the linear speed, and e(θi ) is the unit vector [cos θi , sin θi ]T ; wi ∈ R is the angular speed, and ui , τi ∈ R are control inputs. 4 Swarm Intelligence - From Concepts to Applications Following the same vein as in [16], we define 0 < r0 < R0 . Then, the flocking control problem for (1.1) is to construct the control inputs ui , τi as bounded functions of the collective state (q1 , . . . , qN , θ1 , . . . , θN , v1 , . . . , vN , w1 , . . . , wN ) in a distributed fashion to satisfy the following multiple goals G1) Velocity consensus: lim (q̇i (t) − q̇ j (t)) = 0, ∀i, j = 1, . . . , N t→∞ G2) G3) (1.2) Collision avoidance: ri j (t) = kqi (t) − q j (t)k ≥ r0 , ∀t ≥ 0, ∀i 6= j Cohesion maintenance: ri j (t) ≤ R0 , ∀t ≥ 0, ∀i 6= j. Similarly to [16, 17], we have the following definition. Definition 1. A control ζ˙ = g(ζ , y), u = c(ζ , y), (ζ , y) ∈ Rd × Rm of a system ẋ = f (x, u), y = h(x, u) is said to be bounded if there is a finite constant M > 0 such that kc(ζ , y)k ≤ M, ∀(ζ , y) ∈ Rd × Rm . To achieve the goals G2) and G3), we consider the coordination function U : R+ → R+ which satisfies the following properties: P1) there is a constant UM > 0 such that 0 ≤ U(r) ≤ UM , ∀r ∈ R P2) P3) P4) (1.3) U(r) is continuously differentiable on [r0 , R0 ]; lim U(r) = UM ; and r→r0+ lim U(r) = UM . r→R− 0 For a link between agents i and j of the flock, we aim to maintain ri j (t) ∈ [r0 , R0 ]. Without loss of generality, we assume that U(0) = 0 and hence U(r) is well defined for rii = 0 [16]. We are interested in the function U with the dead zone [a, A] since even distribution of agents may not be achievable by a common coordination function U. Accordingly, we use the zone [a, A] for free alignment. A function U satisfying the above requirements is shown in Figure 1.1. For bounded control, we shall use the linear saturation functions σ1 , and σ2 , which are continuous and nondecreasing functions and satisfy, for given positive constants Li ≤ Mi , i = 1, 2 [16] i) ii) iii) σi (−s) = −σi (s) for all s; σi (s) = s for s ≤ Li ; and |σi (s)| ≤ Mi , ∀s ∈ R. Similarly to other works on distributed for multiagent systems [7, 10, 17, 18], the graph theory will be utilised to address our problem. A digraph associated with (1.1) is called G (t) = (V , E (t)) where V = 1, . . . , N and E ⊆ V × V . The set V is denoted as the node set of G (t) and the set E (t) is defined as the edge set of G (t). In addition, Ni (t) denotes the neighbor set of the node i for i = 1, . . . , N. Bounded Distributed Flocking Control of Nonholonomic Mobile Robots 5 U(r) 0 r0 a R0 A Figure 1.1 Coordination function (extracted from [16]). As in [10], the description of the edge E (t) is presented as follows. Given any R > 0, ε2 ∈ (0, R), and ε1 ∈ (0, R−ε2 ), for any t ≥ 0, E (t) = {(i, j)|i, j ∈ V } is defined such that 1. 2. 3. 4. E (0) = {(i, j)|ε1 < kqi (0) − q j (0)k < (R − ε2 )}; if kqi (0) − q j (0)k ≥ R, then (i, j) ∈ / E (t); for i = 1, . . . , N, j = 1, . . . , N, if (i, j) ∈ / E (t − ) and kqi (t) − q j (t)k < R − ε2 , then (i, j) ∈ E (t); for i = 1, . . . , N, j = 1, . . . , N, if (i, j) ∈ E (t − ) and kqi (t) − q j (t)k < R, then (i, j) ∈ E (t). As in [17], the following results will be employed for the main results. Lemma 1. Let σ : R → R be a function satisfying σ (−s) = −σ (s), ∀s ∈ R. Then, for all ai , bi , it holds true that N 1 N (a − a )σ (b − b ) = i j i j ∑ ∑ ∑ ∑ ai σ (bi − b j ). 2 i=1 i=1 j∈N (t) j∈N (t) i (1.4) i Proof: Since σ (−s) = −σ (s) and G (t) is an undirected graph, we have N ∑ ∑ i=1 j∈Ni (t) N a j σ (bi − b j ) = − ∑ ∑ a j σ (b j − bi ) ∑ ai σ (bi − b j ). i=1 j∈Ni (t) N =−∑ i=1 j∈Ni (t) (1.5) 6 Swarm Intelligence - From Concepts to Applications Hence, N ∑ ∑ (ai − a j )σ (bi − b j ) i=1 j∈Ni (t) N N =∑ ∑ i=1 j∈Ni (t) ai σ (bi − b j ) − ∑ ∑ a j σ (bi − b j ) i=1 j∈Ni (t) N =2∑ ∑ ai σ (bi − b j ) (1.6) i=1 j∈Ni (t) which implies (1.4). Remark 1. Lemma 1 plays an important role in the theoretical analysis of the main results. Here, the lemma is similar to the one in [46]. However, [46] considers all-to-all communication in the multiagent system. Our problem in this chapter is focused on the distributed fashion, which requires the employment of the neighbour set Ni (t) of robot i. Lemma 2. The linear saturation functions σi , i = 1, 2, 3 satisfy (σi (θ1 ) − σi (θ2 ))σi (θ1 − θ2 ) ≥ 0, ∀θ1 , θ2 . (1.7) Proof: Without loss of generality, suppose that θ1 ≥ θ2 . Since σi are nondecreasing functions, this implies that σi (θ1 ) − σi (θ2 ) ≥ 0. (1.8) Furthermore, as σi (0) = 0, θ1 ≥ θ2 and the nondecreasing property of σi imply that σi (θ1 − θ2 ) ≥ 0. (1.9) Multiplying (1.8) and (1.9) side-by-side, we obtain (1.7). 1.4 MAIN RESULTS Our constructive strategy is to design ui to achieve consensus on vi , and τi to achieve consensus on θi . The design for ui is derived from [17], while the construction for τi is built based on the approach in [19]. Note that U(ri j ) = U(kqi − q j k), which is the symmetric function of qi and q j . As a result, we write U(qi , q j ) with the understanding that U(qi , q j ) = U(q j , qi ). The control protocols ui and τi are constructed based on Lyapunov theory. Specifically, a positive definite function V is presented such that the time derivative of V is a negative definite function. Regarding the distribution control problem, the graph theory will be employed to show the connectivity preservation for our multiagent system. Then, we apply the LaSalle’s invariance principle [47] to conclude the desired consensuses. Bounded Distributed Flocking Control of Nonholonomic Mobile Robots 7 Similarly to [17], the initial state of the collective system of agents (1.1) is chosen such that the graph G (0) is connected. The parameters of the graph G (0) are chosen as follows R = R0 , (1.10) r0 ≤ ε1 < a, (1.11) < ε2 ≤ R0 − a. (1.12) 0 1.4.1 Speed consensus and connectivity preservation The derivation of this subsection is essentially similar to the control design for the linear speed in [17]; hence, it is presented here for completeness. Consider the energy function for system (1.1) V1 = 1 N 2 1 N U(q , q ) + i j ∑ ∑ ∑ vi . 2 i=1 2 i=1 j∈N (t) (1.13) i We assume that U(r) is designed such that U(R0 ) = UM > V1max , (1.14) where V1max , 1 N 2 N(N − 1) ∑ v (0) + 2 U(R0 − ε2 ). 2 i=1 (1.15) Let m0 be the number of the links of the initial graph. The simplest connected graph of N agents is a tree whose number of links is n − 1. Hence, m0 ≥ n − 1. Let (N − 1)(N − 2) U(R0 − ε2 ). (1.16) 2 Note that U(qi , q j ) is a symmetric function of qi and q j . We compute the derivative of V1 with respect to (1.1) V1 (0) ≤ V1max − N N V̇1 = ∑ ∑ i=1 j∈Ni (t) ∇qi U(qi , q j ) · q̇i + ∑ vi ui i=1 ! N = ∑ vi i=1 ∑ ∇qi U(qi , q j ) · e(θi ) + ui . (1.17) j∈Ni (t) From (1.17), a control law for the speed consensus protocol is chosen as ui = − ∑ ∇qi U(qi , q j ) · e(θi ) − j∈Ni (t) ∑ σ1 (vi − v j ) (1.18) j∈Ni (t) where σ1 is the linear saturation function introduced in Section 1.3. Substituting (1.18) into (1.17), we obtain N V̇1 = − ∑ ∑ i=1 j∈Ni (t) vi σ1 (vi − v j ). (1.19) 8 Swarm Intelligence - From Concepts to Applications We have the following speed consensus theorem [17]. Theorem 1. Suppose that the collective system (1.1) subject to the protocol (1.18) is initiated such that V1 (0) < V1max . Then, the following properties hold: i) ii) iii) G (t) is connected for all t ≥ 0 and there exists tk such that for t ≥ tk , G (t) = G (tk ) Collision avoidance is guaranteed, i.e. kqi − q j k > r0 for all i, j ∈ N and i 6= j. lim (vi (t) − v j (t)) = 0 t→∞ Proof: Assume that G (t) switches at time tk (k = 1, 2, . . . ). Hence, G(t) = G(0) for all t ∈ [0,t1 ). In other words, G (t) = G(0), t ∈ [0,t1 ) G (t1 ) 6= G(0). (1.20) We prove that G(0) ⊂ G(t1 ). Using control law (1.18), we have N V̇1 = − ∑ vi i=1 ∑ σ1 (vi − v j ). (1.21) j∈Ni (t) According to Lemma 1, V̇1 = − 1 N ∑ ∑ (vi − v j )σ1 (vi − v j ). 2 i=1 j∈N (t) (1.22) i Because σ1 (s) is an odd function, (vi − v j )σ1 (vi − v j ) ≥ 0. Thus, V̇1 ≤ 0, which deduces V1 (t) ≤ V1 (0) < V1max < UM for [0,t1 ). (1.23) From the definition of U(r), U(R0 ) > V1max ≥ V1 (0). Hence for any (i, j) ∈ G (t) for t ∈ [0,t1 ) U(qi , q j ) ≤ V1 (t) < UM = U(r0 ) = U(R0 ). (1.24) By the continuity of U(r), (1.24) shows that r0 < kqi − q j k < R0 . This implies that no existing links are deleted at time t1 and collision avoidance is achieved. As a result, new links must be added to the current graph at the switching time t1 . We assume that there are m1 new links being added to the network at time t1 . Since the number of the current links before switching is m0 ≥ N − 1 and the complete graph edges, m1 ≤ N(N−1) − (N − 1) = (N−1)(N−2) . Hence, we have possesses N(N−1) 2 2 2 V1 (t1 ) = V1 (t1− ) + m1U(R0 − ε2 ). (1.25) Due to (1.16), V1 (t1− ) ≤ V1 (0) < V1max − (N − 1)(N − 2) U(R0 − ε2 ). 2 (1.26) Thus, V1 (t1 ) < V1max . (1.27) Bounded Distributed Flocking Control of Nonholonomic Mobile Robots 9 By induction, for t ∈ [tk−1 ,tk ), V̇1 = − 1 N ∑ ∑ (vi − v j )σ1 (vi − v j ) ≤ 0, 2 i=1 j∈N (t) (1.28) i and therefore V1 (t) ≤ V1 (tk−1 ) ≤ V1max . This shows that no edges are lost at time tk and V1 (tk ) ≤ V1max . As a result, the size of the set of the links of G (t) forms an increasing sequence, bounded above by N(N−1) , which is the number of the links of 2 a complete graph. Thus, there exists a finite integer k > 0 such that G (t) = G (tk ), t ∈ [tk , ∞). (1.29) Hence, for t ≥ tk , we have V̇1 = − 1 N ∑ ∑ (vi − v j )σ1 (vi − v j ) ≤ 0. 2 i=1 j∈N (t ) (1.30) i k Next, we will show that the linear velocities of all agents converge to the same value. Using the fact that U(qi , q j ) ≤ V1 (t) ≤ V1max < UM and the properties of U, we deduce kqi − d j k > r0 . This shows that no collision takes place among agents. Since 0 ≤ V1 (t) ≤ V1max and V̇1 ≤ 0, by Barbalat’s lemma, limt→∞ V̇1 (t) = 0. Because the graph G (t) is connected for all t and sσ1 (s) ≥ 0 for all s, from (1.30), lim (vi − v j ) = 0, for all i, j = 1, 2, . . . , N. t→∞ (1.31) Remark 2. The proof of the theorem follows similar approaches as in [10, 48], where graph theory was employed as a means for proving the connectivity of mobile networks. The potential in this chapter is similar to the one in [10] in the sense that it is bounded. In contrast, the potential function used in [48] goes to infinity at singularities. Note that the mobile robots in this work are nonholonomic while [10, 48] addressed double integrator systems. Remark 3. The first sum in (1.18) consists of the gradients of U(qi , q j ) and the unit vector e(θi ) which are bounded by definition. The second sum in (1.18) is comprised of σ1 (.), which is a linear saturation function defined in Section 1.3. Hence, as a whole, the control law (1.18) for each agent is bounded. This satisfies our objective on the boundedness of the control input ui . Theorem 1 shows that the design (1.18) achieves speed consensus and the goals G2) and G3). In the next subsection, we will design τi for orientation consensus completing the goal G1). 1.4.2 Orientation Consensus Motivated by the orientation consensus design method presented in [19], we shall develop a bounded control approach which employs a saturation function in Section 1.3. 10 Swarm Intelligence - From Concepts to Applications Define the orientation trajectory error for agent i as ei = θi − θr , (1.32) where θr is the desired orientation of the flock. Thus, the angle difference between two agents i and j is θi − θ j = θi − θr − (θ j − θr ) = ei − e j . (1.33) Similarly to [19], the following lemma is employed for our convergence analysis. Lemma 3. Suppose that the flock possesses a graph G (t), then the trajectory error signals of the group have the following property: N 1 N (ei − e j )(ėi − ė j ) = ∑ ∑ ei (ėi − ė j ). ∑ ∑ 2 i=1 j∈N (t) i=1 j∈N (t) i (1.34) i Proof: The proof is similar to the one employed in Lemma 1. We have the following orientation consensus theorem. Theorem 2. Assume that the desired orientation θr and its first and second derivation are bounded, and the collective system (1.1) is subject to the following protocol τi = θ̈r − σ2 (θ̇i − θ̇r ) − kθ [(ni + 1)θi − ∑ θ j − θr ], ni + 1 j∈N (t) (1.35) i where ni is the number of the neighbors of robot i and kθ is a positive parameter. Then, all the mobile robots eventually reach consensus on the heading angles θi in the sense that lim (θi (t) − θ j (t)) = 0, ∀i, j. t→∞ (1.36) Proof: Consider the following Lyapunov function candidate V2 (t) = 1 N kθ 2 1 N 2 1 N kθ ∑ ni + 1 ei + 2 ∑ ėi + 4 ∑ ∑ ni + 1 (θi − θ j )2 . 2 i=1 i=1 i=1 j∈N (t) i (1.37) Bounded Distributed Flocking Control of Nonholonomic Mobile Robots 11 According to Theorem 1, there exists tk such that for t ≥ tk , G (t) = G (tk ). The derivative of V2 (t) with respect to t for t ≥ tk is given as N V̇2 (t) = = N kθ N kθ (θi − θ j )(θ̇i − θ̇ j ) n +1 i i=1 j∈N (t) ∑ ni + 1 ei ėi + ∑ ėi (τi − θ̈r ) + ∑ ∑ i=1 i=1 N N kθ i N kθ ∑ ni + 1 ei ėi − ∑ ni + 1 ėi ei − ∑ ėi σ2 (ėi ) i=1 i=1 i=1 1 N kθ kθ ėi (θi − θ j ) + ∑ ∑ (θi − θ j )(θ̇i − θ̇ j ) n + 1 2 n i=1 j∈N (t) i + 1 i=1 j∈N (t) i N −∑ ∑ i i N N kθ ėi (ei − e j ) n +1 i i=1 j∈N (t) = − ∑ ėi σ2 (ėi ) − ∑ i=1 ∑ i N + 1 kθ ∑ ∑ ni + 1 (ei − e j )(ėi − ė j ). 2 i=1 j∈N (t) (1.38) i Using Lemma 3, we obtain N V̇2 (t) = − ∑ ėi σ2 (ėi ) ≤ 0 (1.39) i=1 since σ2 (.) is a nondecreasing function defined in Section 1.3. Since θi , θr ∈ [−pi, pi] and θ̈r is bounded, the control law (1.35) implies ëi is bounded. By the Barbalat’s lemma, from (1.39), ėi → 0. Also, since θ̈r is bounded, ëi → 0. Therefore, the control law (1.35) implies that θi → θ j , which proves the theorem. Remark 4. The boundedness of the control law (1.35) is guaranteed by the properties of linear saturation function σ2 and the fact that θi , θ j , θr ∈ [−π, π]. This demonstrates that the proposed control scheme meets the requirement on the physical limits of the control inputs. Remark 5. It should be noted that our control law for orientation consensus is similar to the one in [19] but here the boundedness of the control input is taken into account. The scheme in this chapter also shares the same objective as the one in [17] but offers a more simple form and implementation. Combining Theorems 1 and 2, we have the following bounded flocking theorem. Theorem 3. Suppose that the collective system (1.1) is subject to the bounded protocols (1.18) and (1.35). Suppose further that the initial configuration of the collective system (1.1) is such that N (0) is connected. Then, all the multiple flocking goals of velocity consensus, cohesion maintenance, and collision avoidance are achieved. Proof: The proof is straightforward from the results of Theorems 3.1 and 3.2. 12 Swarm Intelligence - From Concepts to Applications 1.5 Avoidance of Obstacles The problem of obstacle avoidance has been extensively studied in the literature [17, 19, 22]. In this section, we employ the idea from [19] to derive our control algorithm in which the agents are able to pass obstacles. It is shown that a convex obstacle can be extrapolated by a round shape or a rectangle [17, 19]. In [19], a convex obstacle is presented by a circle, which is used in our work. Let qr = [xr , yr ]T be the coordinate of the robot, q jobs = [xobs , yobs ]T be the projection point of the robot onto obstacle j = 1, 2, . . . , nobs where nobs is the number of obstacles, and Ōk = [xk , yk ]T be the centre of the obstacle. From [19], r r qr + (1 − )Ōr (1.40) q jobs = kqr − Ōk k kqr − Ōk k where r is the radius of the obstacle. The projection point has the following velocity v jobs = vr sin α kqr − q jobs k (1.41) where v is the velocity of the agent, α is the angle between the heading of the robot and the straight line which connects the robot to the centre of the obstacle. The projection point moves in the direction [19] π θ jobs = − + α + θ , α > 0 (1.42) 2 π θ jobs = + α + θ , otherwise. (1.43) 2 The fact that the projection point possesses a position, velocity, and orientation enables it to be an agent. The above descriptions of the robot and obstacle are demonstrated in Figure 1.2. We have the following orientation consensus theorem. Theorem 4. The following control protocol guarantees that the robot avoids obstacles with arbitrary boundary shapes: ui = − ∇qi U(qi , q j ) · e(θi ) − ∑ j∈Nobs ∑ σ1 (vi − v j ) (1.44) j∈Nobs τi = θ̈r − σ2 (θ̇i − θ̇r ) − kθ [(nobs + 1)θi − ∑ θ j − θr ], nobs + 1 j∈N (t) (1.45) i where Nobs is the set of obstacles, nobs is the number of obstacles and θr = 1 nobs ∑ θ jobs . (1.46) j∈Nobs Proof: The proof can be derived using the same approach as in [19]. Remark 6. It should be noted that the speed control law in (1.44) enjoys the boundedness due to the saturation function σ1 (.), which is different from the one in [19]. Since θi , θr ∈ (−π, π] and the saturation function σ2 (.) is bounded, (1.45) reveals that the heading control law is also bounded. Bounded Distributed Flocking Control of Nonholonomic Mobile Robots r Y 13 Ok Obstacle q jobs   qr Robot X O Figure 1.2 Illustration of a robot and a convex obstacle. 1.6 SIMULATION We conducted simulation for a multi-agent system of 15 mobile robots of the model (1.1). A bump function is used to generate the smooth coordination function U. As the control (1.18) invokes the gradient forces ∇qi U, we designed the coordination function in the form Z r U(r) = (1.47) ϕ(s)ds 0 where ϕ is a compact support function given by    −(s−s0 )2  p exp 1  2 2  0)   ((a−r0 )/2) −(s−s −(s−s1 )2 ϕ(s) = p2 exp ((R0 −A)/2)2 −(s−s1 )2    0 if s ∈ (r0 , a) if s ∈ (A, R0 ) otherwise where s0 = r02+a , s1 = R02+A , and p1 , p2 , a, A, r0 and R0 are design parameters. The parameters of the coordinate function are r0 = 1, a = 3, A = 6, R0 = 8, and UM = 15. The parameter for the control law (1.35) is kθ = 1.5. The initial positions of 15 mobile robots are randomly distributed on three circles. Their coordinates are x(i) = Γ sin(z π (i − 1)/Γ + π) y(i) = Γ cos(z π (i − 1)/Γ + π) 14 Swarm Intelligence - From Concepts to Applications where z = 1, and ( 4 if i < 6 . Γ= 8 otherwise (1.48) The initial values of θi and vi are randomly chosen. The initial value of wi is 0. The desired orientation of the flock is θr = π/2. We obtained the simulation results shown in Figures 1.9–1.7. It is shown in Figures 1.3 and 1.5, the heading angles converge to θr and angular speeds of all agents converge to 0 after t = 40s. The linear speeds are depicted in Figure 1.4, where the convergence of all agents takes place after t = 45s. In Figure 1.5, the angular speeds converge after 40 seconds. Hence, Figure 1.3 and Figure 1.4 demonstrate that consensuses on orientation and linear speed of the mobile robots have been obtained. The speed control and steering control signals are shown to be bounded in Figure 1.6 and Figure 1.7. The minimum distance among agents is described in Figure 1.8, which shows that collision avoidance is guaranteed. The flocking behavior is shown in Figure 1.9, where no collision occurred. 2 1.5 θi 1 0.5 0 -0.5 0 10 20 30 40 50 60 70 80 time (s) Figure 1.3 Orientation consensus Next, we consider a simulation for a multiagent system with an obstacle. The obstacle is a circle whose coordinate is [12, −1] with a radius of 1. The configuration of the multiagent system is the same as in the obstacle-free case above. The desired group heading is chosen as θr = π/4. An agent is chosen as a leader of the group. Bounded Distributed Flocking Control of Nonholonomic Mobile Robots 2 1.5 vi 1 0.5 0 -0.5 0 10 20 30 40 50 60 70 80 70 80 time (s) Figure 1.4 Linear speed consensus 0.45 0.4 0.35 0.3 wi 0.25 0.2 0.15 0.1 0.05 0 -0.05 0 10 20 30 40 50 time (s) Figure 1.5 Angular speed 60 15 Swarm Intelligence - From Concepts to Applications 6 4 ui 2 0 -2 -4 -6 0 10 20 30 40 50 60 70 80 60 70 80 time (s) Figure 1.6 Translational force 0.6 0.5 0.4 0.3 τi 16 0.2 0.1 0 -0.1 -0.2 0 10 20 30 40 50 time (s) Figure 1.7 Steering control Bounded Distributed Flocking Control of Nonholonomic Mobile Robots 17 3.5 3 dmin 2.5 2 1.5 1 0 10 20 30 40 50 60 70 80 time (s) Figure 1.8 The minimum distance among agents When the collision avoidance mechanism of the leader is inactive, a speed control law is designed to drive the group at a constant speed, that is ul = −σ1 (vl − vr ) (1.49) where ul is the speed control of the leader, vl is its linear speed, and vr = 0.2. During the first 50 seconds of their evolution, the mobile robots encounter the obstacle. The control laws (1.44) and (1.45) enable them to avoid potential collisions with the obstacle and with other neibouring agents. The orientations of the robots in Figure 1.10 converge to π/4 after 60 seconds. Similarly, in Figure 1.11, the linear speeds of the agents converge after 50 seconds. In Figure 1.12, the angular speeds converge faster to 0 after 50 seconds. Figs. 1.13 and 1.14 demonstrate that the control inputs both are bounded. Finally, Figure 1.15 shows that no collision takes place during the evolution of the robots. The evolution of the multiagent system is shown in Figure 1.16 in which the robots cooperate to form a flocking and avoid the obstacle. 1.7 CONCLUSIONS This chapter has presented a bounded decentralized control protocol for the flocking problem of mobile robots by a systematic fashion, where the control laws only require the information of neighbor agents. The proposed scheme is modular in designing the speed control and steering control separately. Theoretical and numerical results have shown that using our proposed method, a collective system of mobile Swarm Intelligence - From Concepts to Applications 60 50 40 y 18 30 20 10 0 -10 -10 0 10 x Figure 1.9 Distributed flocking of 15 mobile robots Bounded Distributed Flocking Control of Nonholonomic Mobile Robots 3 2.5 2 θi 1.5 1 0.5 0 -0.5 -1 0 10 20 30 40 50 60 70 80 90 100 time (s) Figure 1.10 Orientation consensus for the obstacle avoidance case 4 3.5 3 2.5 vi 2 1.5 1 0.5 0 -0.5 -1 0 10 20 30 40 50 60 70 80 90 100 time (s) Figure 1.11 Linear speed consensus for the obstacle avoidance case 19 Swarm Intelligence - From Concepts to Applications 1.5 wi 1 0.5 0 -0.5 0 10 20 30 40 50 60 70 80 90 100 time (s) Figure 1.12 Angular speed for the obstacle avoidance case 5 0 ui 20 -5 -10 0 10 20 30 40 50 60 70 80 90 time (s) Figure 1.13 Speed control for the obstacle avoidance case 100 Bounded Distributed Flocking Control of Nonholonomic Mobile Robots 21 3 2 τi 1 0 -1 -2 -3 0 10 20 30 40 50 60 70 80 90 100 time (s) Figure 1.14 Steering control for the obstacle avoidance case 3.5 3 dmin 2.5 2 1.5 1 0 10 20 30 40 50 60 70 80 90 100 time (s) Figure 1.15 The minimum distance among agents for the obstacle avoidance case 22 Swarm Intelligence - From Concepts to Applications 20 15 y 10 5 0 -5 -10 -10 -5 0 5 10 15 20 x Figure 1.16 Distributed flocking of 15 mobile robots for the obstacle avoidance case robots achieves all the multiple objectives of the flocking control: velocity consensus, cohesion maintenance, and collision avoidance. Future work would consider the shape and size of each mobile robot and obstacle avoidance for which a similar context was studied in [49]. Noisy and uncertain environments can affect the performance of the proposed scheme. Robustness analysis and improved methods can be proposed to address this issue. 1.8 Acknowledgement This work was partially supported by the National Science Foundation under grant NSF-NRI #1426828, the National Aeronautics and Space Administration (NASA) under Grant No. NNX15AI02H issued through the Nevada NASA Research Infrastructure Development Seed Grant, and the University of Nevada, Reno. REFERENCES 23 References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] Toner J, Tu Y. Flocks, herds, and schools: A qualitative theory of flocking. Phys Rev E. 1998;58:4828–4858. La HM, Sheng W. Distributed Sensor Fusion for Scalar Field Mapping Using Mobile Sensor Networks. IEEE Trans Cybernetics. 2013 April;43(2):766– 778. Muñoz F, Espinoza Quesada ES, La HM, et al. Adaptive consensus algorithms for real-time operation of multi-agent systems affected by switching network events. International Journal of Robust and Nonlinear Control. 2017;27(9):1566–1588. Rnc.3687. Available from: http://dx.doi.org/ 10.1002/rnc.3687. La HM, Sheng W. Flocking control of multiple agents in noisy environments. In: 2010 IEEE International Conference on Robotics and Automation; 2010. p. 4964–4969. La HM, Sheng W. Multi-Agent Motion Control in Cluttered and Noisy Environments. Journal of Communications. 2013;8(1):32–46. Dang AD, La HM, Horn J. Distributed formation control for autonomous robots following desired shapes in noisy environment. In: 2016 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI); 2016. p. 285–290. Olfati-Saber R. Flocking for multi-agent dynamic systems: algorithms and theory. IEEE Trans Autom Control. 2006 Mar;51(3):401–420. Tanner HG, Jadbabaie A, Pappas GJ. Flocking in fixed and switching networks. IEEE Transactions on Automatic Control. 2007;52(5):863–868. Cucker F, Smale S. Emergent behavior in flocks. IEEE Trans Autom Control. 2007 May;52(5). Dong Y, Huang J. Flocking with connectivity preservation of multiple double integrator systems subject to external disturbances by a distributed control law. Automatica. 2015;55:197 – 203. Ghapani S, Mei J, Ren W, et al. Fully distributed flocking with a moving leader for Lagrange networks with parametric uncertainties. Automatica. 2016;67:67–76. Moshtagh N, Jadbabaie A. Distributed geodesic control laws for flocking of nonholonomic agents. IEEE Transactions on Automatic Control. 2007;52(4):681–686. Han TT, Ge SS. Styled-velocity flocking of autonomous vehicles: a systematic design. IEEE Trans Autom Control. 2015 Aug;60(8):2015–2030. La HM, Sheng W, Chen J. Cooperative and Active Sensing in Mobile Sensor Networks for Scalar Field Mapping. IEEE Trans Systems, Man, and Cybernetics: Systems. 2015 Jan;45(1):1–12. Amar K, Mohamed S. Stabilized feedback control of unicycle mobile robots. Int J Adv Robot Syst. 2013;10(187):1–8. 24 [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] Swarm Intelligence - From Concepts to Applications Han TT, La HM, Dinh BH. Flocking of mobile robots by bounded feedback. In: 2016 IEEE International Conference on Automation Science and Engineering (CASE); 2016. p. 689–694. Nguyen T, Han TT, La HM. Distributed flocking control of mobile robots by bounded feedback. In: 2016 54th Annual Allerton Conference on Communication, Control, and Computing (Allerton); 2016. p. 563–568. Jadbabaie A, Lin J, Morse AS. Coordination of groups of mobile autonomous agents using nearest neighbor rules. IEEE Trans Autom Control. 2003 Jun;48(6):988–1001. Liang Y, Lee HH. Avoidance of Multiple Obstacles for a Mobile Robot With Nonholonomic Constraints. In: ASME proceedings of Dynamic Systems and Control; 2005. p. 1657–1663. Riley KF, Hobson MP, Bence SJ. Mathematical Methods for Physics and Engineering. 3rd ed. Cambridge: Cambridge University Press; 2006. Tanner HG, Jadbabaie A, Pappas GJ. Flocking in teams of nonholonomic agents. In: Kumar V, Leonard N, Morse S, editors. Cooperative Control. vol. 309 of Lecture Notes in Control and Information Sciences. Springer; 2004. p. 229–239. Liang Y, Lee HH. Decentralized formation control and obstacle avoidance for multiple robots with nonholonomic constraints. In: 2006 American Control Conference; 2006. p. 5596–5601. Nguyen T, La HM. Formation Control of Multiple Rectangular Agents with Limited Communication Ranges. In: 10th International Symposium on Advances in Visual Computing. Springer International Publishing; 2014. p. 915– 924. Nguyen T, La HM, Jafari M. On the Formation Control of a Multi Vehicle System. In: ISSAT International Conference on Modeling of Complex Systems and Environments; 2015. . Cruz D, McClintock J, Perteet B, et al. Decentralized cooperative control-a multivehicle platform for research in networked embedded systems. IEEE control systems. 2007;27(3):58–78. Nguyen T, La HM. Distributed Formation Control of Nonholonomic Mobile Robots by Bounded Feedback in the Presence of Obstacles. arXiv preprint arXiv:170404566. 2017;. Pham HX, La HM, Feil-Seifer D, et al. A Distributed Control Framework for a Team of Unmanned Aerial Vehicles for Dynamic Wildfire Tracking. arXiv preprint arXiv:170402630. 2017;. La HM, Sheng W. Flocking control of a mobile sensor network to track and observe a moving target. In: 2009 IEEE International Conference on Robotics and Automation; 2009. p. 3129–3134. La HM, Sheng W. Moving targets tracking and observing in a distributed mobile sensor network. In: 2009 American Control Conference; 2009. p. 3319–3324. REFERENCES [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] 25 La HM, Sheng W. Adaptive flocking control for dynamic target tracking in mobile sensor networks. In: 2009 IEEE/RSJ International Conference on Intelligent Robots and Systems; 2009. p. 4843–4848. Nguyen MT, La HM, Teague KA. Compressive and collaborative mobile sensing for scalar field mapping in robotic networks. In: 2015 53rd Annual Allerton Conference on Communication, Control, and Computing (Allerton); 2015. p. 873–880. La HM, Sheng W, Chen J. Cooperative and active sensing in mobile sensor networks for scalar field mapping. In: 2013 IEEE International Conference on Automation Science and Engineering (CASE); 2013. p. 831–836. La HM, Sheng W. Cooperative sensing in mobile sensor networks based on distributed consensus. In: SPIE Optical Engineering+ Applications. International Society for Optics and Photonics; 2011. p. 81370Y–81370Y. La HM, Nguyen TH, Nguyen CH, et al. Optimal flocking control for a mobile sensor network based a moving target tracking. In: 2009 IEEE International Conference on Systems, Man and Cybernetics; 2009. p. 4801–4806. La HM, Lim RS, Sheng W. Hybrid system of reinforcement learning and flocking control in multi-robot domain. In: Proc. Conf. Theoretical Appl. Comput. Sci.; 2010. p. 7–13. La HM, Lim RS, Sheng W, et al. Cooperative flocking and learning in multirobot systems for predator avoidance. In: 2013 IEEE International Conference on Cyber Technology in Automation, Control and Intelligent Systems; 2013. p. 337–342. Jafari M, Sengupta S, La HM. In: Bebis G, Boyle R, Parvin B, et al., editors. Adaptive Flocking Control of Multiple Unmanned Ground Vehicles by Using a UAV. Cham: Springer International Publishing; 2015. p. 628–637. Available from: http://dx.doi.org/10.1007/978-3-319-27863-6 58. La HM. Multi-Robot Swarm for Cooperative Scalar Field Mapping. Handbook of Research on Design, Control, and Modeling of Swarm Robotics. 2015;p. 383–395. La HM, Sheng W. Flocking control algorithms for multiple agents in cluttered and noisy environments. In: Bio-Inspired Self-Organizing Robotic Systems. Springer Berlin Heidelberg; 2011. p. 53–79. Carrillo JA, Fornasier M, Rosado J, et al. Asymptotic flocking dynamics for the kinetic Cucker-Smale model. SIAM Journal on Mathematical Analysis. 2010;42(1):218–236. La HM, Lim R, Sheng W. Multirobot cooperative learning for predator avoidance. IEEE Trans Control Syst Technol. 2015 Jan;23(1):52–63. Sepulchre R, Paley DA, Leonard NE. Stabilization of planar collective motion: all-to-all communication. IEEE Trans Autom Control. 2007 May;52(5):811–824. Zavlanos MM, Tanner HG, Jadbabaie A, et al. Hybrid control for connectivity preserving flocking. IEEE Transactions on Automatic Control. 2009;54(12):2869–2875. 26 [44] [45] [46] [47] [48] [49] Swarm Intelligence - From Concepts to Applications Dong W. Flocking of multiple mobile robots based on backstepping. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics). 2011;41(2):414–424. Moshtagh N, Michael N, Jadbabaie A, et al. Vision-based, distributed control laws for motion coordination of nonholonomic robots. IEEE Trans Robotics. 2009 Aug;25(4):851–860. Han TT, La HM, Dinh BH. Flocking of mobile robots by bounded feedback. In: 2016 IEEE International Conference on Automation Science and Engineering (CASE); 2016. p. 689–694. Khalil HK. Nonlinear Systems. 3rd ed. Upper Saddle River, NJ: Prentice Hall; 2002. Zavlanos MM, Jadbabaie A, Pappas GJ. Flocking while preserving network connectivity. In: 2007 46th IEEE Conference on Decision and Control; 2007. p. 2919–2924. Nguyen T, La H, Le TD, et al. Formation Control and Obstacle Avoidance of Multiple Rectangular Agents with Limited Communication Ranges. IEEE Transactions on Control of Network Systems. 2016;.
3cs.SY
On Koszulity in homology of moduli spaces of stable n-pointed curves of genus zero arXiv:1304.6343v2 [math.RA] 20 Nov 2017 Natalia Iyudu Abstract We prove Koszulity of the homology of the of moduli spaces of stable n-pointed curves of genus zero M 0,n for n = 5, using its presentation due to Keel and the Priddy criterion of Koszulity. For ! n = 6 we establish that M 0,6 is potential, find the expression for the potential, and based on that prove that it is Koszul. 1 Introduction In the paper [2], due to Keel, dedicated to the intersection theory of the moduli space of n-pointed stable curves of genus zero, it was shown that the canonical map from the Chow ring to the homology is an isomorphism, and the presentation of this ring by generators and relations have been derived. The variables D S in this presentation are parameterized by subsets S ⊂ {1, ⋯, n} consisting of C two or more elements, variables corresponding to complimentary sets are coincide: D S = D S . The relations then looks as follows: (1). For any four distinct elements i, j, k, l ∈ {1, ⋯, n}: DS = ∑ i,j∈S,k,l∉S ∑ DS = i,k∈S,j,l∉S ∑ DS i,l∈S,j,k∉S (2). D S D T unless one of the following holds: S ⊂ T, T ⊂ S, S ⊂ T C , T C ⊂ S. We will use this presentation to examine Koszulity of the ring A = H● (M 0,n ). For n = 5 we suggest the following choice of linearly independent variables: out of all generators {δi,j } = {δ12 , δ13 , δ14 , δ15 , δ23 , δ24 , δ25 , δ34 , δ35 , δ45 } we take those, sitting on the sides of the pentagon: xi = δi,i+1 , i ∈ Z/5Z, then we rewrite all other variables and relations via them. This presentation wakes up some associations with cluster algebras. Using the obtained relations we compute a presentation for the Koszul dual (noncommutative) algebra A! . It turns out that it can be presented by one relation of the form: ∑ xa xb − a≠b,a≠b+1,a≠b−1 2 ∑ xk = 0. k=1,...,5 1 We can find a change of variables, such that the new relations does not have a square term x2 . Thus we have an algebra A! presented by a quadratic Gröbner basis. Hence due to the Priddy criterion [5, 3], we can conclude that A! is Koszul, and hence A itself as well. Then we consider A = H● (M 0,6 ). After change of generators to a smaller set of variables, which span the same space modulo linear relations, we are able to show that the dual algebra A! is potential, and find an expression for the potential. Then we construct a potential complex, which is coincide with the Koszul complex. We show its exactness by checking it is exact in all but one places, then numeric Koszulity implies Koszulity, according to lemma??. Our methods here could be compared to those appeared in a vast literature, where Koszulity of various algebras given by relation, originated from geometric objects, have been proved. To name a few, there are series of papers [6, 7, 8] dedicated to Koszulity of algebras associated to graphs and sell complexes, papers [?] on Koszulity of objects associated to the Coexter group action, etc. 2 Koszulity of H● (M 0,5) We consider first the presentation of A = H● (M 0,n ) due to Keel [2] with generators {δi,j } = {δ12 , δ13 , δ14 , δ15 , δ23 , δ24 , δ25 , δ34 , δ35 , δ45 } and corresponding linear and monomial quadratic relations. Let V = spank {δi,j }. We suggest to choose as a linear basis in V the following set: {xi = δi,i+1 , i ∈ Z/5Z}. Via this basis we are going to express other elements of {δi,j }, then we rewrite quadratic relations in terms of variables xi , i ∈ Z/5Z. This will allow us later on to pass to the Koszul dual algebra A! , which could be shown to be Koszul. The presentation for the A! which we obtain, after suitable change of variables, will have a quadratic Gröbner bases. This due to the Priddy’s PBW criterion [5, 3] will imply Koszulity of A! and hence of A. STEP 1 Expression of {δi,j } via xi = δi,i+1 , i ∈ Z/5Z. Linear relation which we have for quadruple {a, b, c, d} ∈ Z/5Z of pairwise distinct elements from 1 to 5, are δab + δcd = δac + δbd = δad + δbc . In particular, δad = δab + δcd − δbc , and for, say a = i, b = i + 1, c = i + 2, d = i + 3 we get δi,i+3 = δi,i+1 + δi+2,i+3 − δi+1,i+2 , that is δi,i+3 = xi + xi+2 − xi+1 . The set of elements {δi,j }, which are our original generators can be presented ’geometrically’ as a sides and diagonals of the pentagon (analogously to the presentation of cluster variables). Sides correspond to variables δi,i+1 and diagonals to δi,i+3 , for appropriate i. We will call these two types of generators side type and diagonal type generators respectively. Hence, by expressing the diagonal type variables δi,i+3 , via those δi,j with nearby indexes, e.i. via the side type variables, we express all generators via the new basis {xi , i ∈ Z/5Z}. STEP 2 Calculation of quadratic relations in terms of xi = δi,i+1 , i ∈ Z/5Z. 2 Our initial quadratic relations have a shape δi,j δj,k , i ≠ k This means, for those variables which are side type, we will have (1) δi,i+1 δi+1,i+2 = xi xi+1 = 0. If a relation contains both side and diagonal type generators, it have a shape: δi,i+3 xk , k ≠ i + 1, k ∈ Z/5Z. In the pentagon picture it could be interpreted as a fact, that only the side xi+1 does not intersect the diagonal δi,i+3 . So in our new variables these relations look like: (2) (xi + xi+2 − xi+1 )xi = 0(k = i) (xi + xi+2 − xi+1 )xi−1 = 0(k = i − 1) (xi + xi+2 − xi+1 )xi+2 = 0(k = i + 2) (xi + xi+2 − xi+1 )xi+3 = 0(k = i + 3). In the case when relations contain two diagonal variables, we also will have: δi,i+3 δj,j+3 = 0, which means δi,i+3 δi−3,i = 0. That is, (3) (xi + xi+2 − xi+1 )(xi−3 + xi−1 − xi−2 ) = 0. Since it is obviously the same as (xi + xi+2 − xi+1 )(xi+2 + xi−1 − xi+3 ) = 0, we can see that (3) follows from (2) and hence the relations of A on variables xi are formed by (1) and (2). This system of relations is equivalent to the following one: xi xi+1 = xi+1 xi = 0 ′ ′ and for any other pair a, b, a , b ∈ Z/5Z xa xb = xa′ xb′ = −x2a . The Hilbert series here is clearly HA = 1 + 5t + t2 . This also implies that codimension of the second graded component of the ideal I2 is 1. Now using this system of relations, and additional information that the dimension of the space orthogonal to I2 is one, we can write down generating relation for the Koszul dual algebra A! : 3 2 ∑ xk = 0. xa xb − ∑ a≠b,a≠b+1,a≠b−1 k=1,...,5 This generating system of A! is not yet form a Gröbner basis of the ideal. But we can make the change of variables which will make this relation free from square of one of variables (say, x2 ). This will make a presentation of algebra combinatorially free, that is new relations is a Gröbner basis. So the Gröbner basis is quadratic Due to the Priddy PBW criterion [5] this implies that A! is Koszul. So we have proved Theorem 2.1. The algebra A = H● (M 0,5 ) is Koszul. 3 Koszulity in the case of M 0,6 Our goal in this section will be the following. Theorem 3.1. The algebra A = H● (M 0,6 ) is Koszul. The canonical generating set for A from the described above Keel’s presentation we denote {aij = a{i,j} , Sijk = S{i,j,k}} for two-element subsets {i, j} and three-element subsets {i, j, k} of Ω = {1, 2, 3, 4, 5, 6}.. We have 15 generators aij = aji with i ≠ j and 10 generators SR = SΩ∖R for 3-element subsets R of Ω. It will be convenient to use the following notation: ijR ⇐⇒ i, j ∈ R or i, j ∈ Ω ∖ R; iRj ⇐⇒ i ∈ R, j ∈ Ω ∖ R or j ∈ R, i ∈ Ω ∖ R. The defining relations of A are as follows. The linear relations are: 0 = ηi,j,k,m = aij + akm + Sijα + Sijβ − aik − ajm − Sikα − Sikβ , for all i, j, k, m, α, β ∈ Ω such that {i, j, k, m, α, β} = Ω. The quadratic relations are: SR SR′ = 0 if SR ≠ SR′ ; SR aij = 0 if iRj, aij aik = 0 if i, j, k are distinct +commutativity. Remark 3.2. Note that there are automorphisms of A corresponding to elements of the group S6 of permutations of points Ω. Indeed, the natural action of S6 on the set {aij , Sijk } of generators extends to action by automorphisms on A. This follows from the fact that this action on the set of generators preserves the way in which corresponding sets intersect and therefore preserves the set of defining relations. Consider the set of variables vi = ∑ aij . j≠i Our goal is to show that 4 Theorem 3.3. The linear span of all aij and SR and the linear span of all vi and SR coincide modulo L, where L is the linear span of all linear relations ηi,j,k,m. Furthermore, SR and vk are linearly independent modulo L. Finally, the expressions of aij via vi and SR modulo L are: 1 1 3 1 aij = (vi + vj ) − ∑ vk + ∑ SR − ∑ SR . 5 20 k≠i,j 5 R∶iRj 10 R∶ijR First, we need the following lemma. Lemma 3.4. The space L⊥ is spanned by vi for i ∈ Ω and ξR = SR − 12 ∑ apq . Furthermore, the six pqR vectors vi and 10 vectors ξR form a linear basis in L⊥ . Proof. First, note that introduced above new variables vs are orthogonal to each ηi,j,k,m. This gives us six linearly independent elements vi of L⊥ . Next, we search for vectors orthogonal to L of the form ξR = SR + α ∑ apq + β ∑ apq . The scalar products of such a vector with vectors ηi,j,k,m have pRq pqR to be equal to 0, we see that w ∈ L⊥ precisely when 1 + 2β − 2α = 0. Now each ξR belongs to L⊥ . Obviously, all 16 vectors vi and ξR are linearly independent. On the other hand, the Hilbert series of A is known [3]: HA = 1 + 16t + 16t2 + t3 and therefore, the dimension of L is 25-16=9 (25 is the number of generators). Hence, the dimension of L⊥ is 25 − 9 = 16. Hence the 16 vectors vi and ξR form a linear basis in L⊥ . Proof of Theorem 3.3. We shall express aij as a linear combination of vk and SR modulo L. We are going to use the basis in L⊥ from Lemma 3.4. We shall try to find α, β, γ, δ ∈ K such that aij = α(vi + vj ) + β ∑ vk + γ ∑ SR + δ ∑ SR R∶iRj k≠i,j R∶ijR modulo L. This is equivalent to the inclusion ζij = aij − α(ti + tj ) − β ∑ vk − γ ∑ SR − δ ∑ SR ∈ L. R∶iRj k≠i,j R∶ijR This, in turn, is equivalent to ζij being orthogonal to all vk and ξR of Lemma 3.4. Orthogonality of ζij to vi (or vj ) reads as the equation 1 − 6α − 4β = 0, while orthogonality of ζij to vk with k ∉ {i, j} is equivalent to the equation 2α + 8β = 0. Orthogonality of ζij to ξR satisfying ijR reads as the equation −1 + 4α + 8β − δ = 0. Finally, orthogonality of ζij to ξR satisfying iRj is equivalent to the equation 4α + 12β − γ = 0. Solving this system of linear equations, we obtain that ζij ∈ L precisely 1 3 , γ = 15 and δ = − 10 . Hence when α = 15 , β = − 20 1 1 3 1 aij = (ti + tj ) − ∑ vk + ∑ SR − ∑ SR 5 20 k≠i,j 5 R∶iRj 10 R∶ijR modulo L. All other statements of Theorem 3.3 follow via elementary dimension arguments.. The above results immediately yield the following presentation of A. Theorem 3.5. A is given by 16 generators vi , SR and quadratic relations∶ SR SR′ = 0 if SR ≠ SR′ ; SR tij = 0 if iRj, tij tik = 0 if i, j, k are distinct +commutativity, where tij = 4(vi + vj ) − ∑ vk + 4 ∑ SR − 6 ∑ SR . R∶iRj k≠i,j 5 R∶ijR Theorem 3.6. The dual algebra B = A! is potential with symmetric potential which is also invariant under the S6 -action corresponding to permutations of Ω. Furthermore, the corresponding potential, when written in terms of generators vi and SR has the form ⟲ 3 2 P = a1 ∑ S R + a2 ∑ S R vi + a3 ∑ SR vi2 R + a5 R,i + a4 + a8 i,j∶i≠j i SR vi vj i,j,R∶ijR ⟲ vi vj2 + a6 ∑ vi3 + a7 ∑ i,j,R∶iRj ⟲ ∑ R,i ⟲ S R vi vj ∑ ⟲ ∑ vi vj vk i,j,k i,j,k are distinct for some non-zero (a1 , . . . , a8 ) ∈ K8 . Proof. Let V = A1 be the linear span of the generators vi and SR of A. Let R be the subspace of V 2 of quadratic relations of A! . Then R⊥ is the space of quadratic relations of A. Since the Hilbert series of A is HA = 1 + 16t + 16t2 + t3 [3],, dim A3 = 1. Hence the codimension of V R⊥ + R⊥ V in V 3 equals 1. It follows that (V R⊥ + R⊥ V )⊥ = RV ∩ V R is one-dimensional. Let P be a (unique up to a non-zero scalar multiple) element of V 3 spanning RV ∩ V R. The fact that R⊥ contains commutators of generators easily implies that P is symmetric. In particular it is cyclicly symmetric, aka a potential. The fact that P spans RV ∩ V R together with HA = 1 + 16t + 16t2 + t3 now yields that R is spanned by the first derivatives of P . Since R⊥ is invariant under S6 -action, so is R and so is RV ∩ V R. It follows that P is invariant under this action. Finally, since SR SR′ ∈ R⊥ , we have that two distinct SR never feature in the same monomial of P . It follows that P must have the form as stated. Now we shall compute P . Theorem 3.7. The dual algebra B = A! is potential with potential (in generators vi , SR ) given by the formula ⟲ 3 2 P = 2 ∑ SR − 2 ∑ SR vi + 4 R R,i + 5 ∑ vi3 − 11 ∑ S R vi vj i,j,R∶iRj ⟲ vi vj2 −3 i,j∶i≠j i ⟲ ∑ ∑ vi vj vk . i,j,k i,j,k are distinct Proof. By Theorem 3.6, ⟲ 3 2 P = a1 ∑ S R + a2 ∑ SR vi + a3 ∑ SR vi2 R + a5 R,i ⟲ S R vi vj ∑ i,j,R∶iRj ⟲ + a4 R,i SR vi vj i,j,R∶ijR + a6 ∑ vi3 + a7 ∑ i ⟲ ∑ ⟲ vi vj2 + a8 i,j∶i≠j ∑ vi vj vk i,j,k i,j,k are distinct for some non-zero (a1 , . . . , a8 ) ∈ K8 . The derivatives of P , being relations of A! must be orthogonal to quadratic relations of A listed in Theorem 3.5. Note that they are automatically orthogonal to ∂P to SR′ tij , iR′ j with all commutators as well as to products of distinct SR . Orthogonality of ∂S R SR ≠ SR′ is also obvious. ∂P Orthogonality of ∂S to SR tij with iRj is easily seen to be equivalent to 4a1 + 4a2 = 0, which R yields a2 = −a1 . Orthogonality of ∂P ∂SR to tij tik with ijR and ikR reads (taking into account a2 = −a1 ): 84a1 + 11a3 + 57a4 − 42a5 = 0. 6 Orthogonality of a2 = −a1 ): ∂P ∂SR to tij tik with ijR, iRj, or iRj, ikR, or iRj, iRk reads (taking into account −16a1 + 11a3 − 3a4 + 8a5 = 0. Orthogonality of ∂P ∂vi to SR tij with iRj reads (taking into account a2 = −a1 ): −4a1 + 4a3 − 2a4 + 2a5 = 0. Orthogonality of a2 = −a1 ): ∂P ∂vi to SR tjk with jRk and i, j, k pairwise distinct reads (taking into account −4a1 − a3 + 3a4 + 2a5 = 0. The equations we have so far form a closed system of linear equations on a1 , . . . , a5 . The space of its solutions is one-dimensional and consists of (a1 , . . . , a5 ) satisfying a2 = −a1 , a5 = 2a1 , a3 = a4 = 0. We go on. Orthogonality of (3.1)): ∂P ∂vi (3.1) to tij tik with i, j, k pairwise distinct reads (taking into account −60a1 + 16a6 − 5a7 + 5a8 = 0. Orthogonality of ∂P ∂vi to tij tjk with i, j, k pairwise distinct reads (taking into account (3.1)): 180a1 − 4a6 + 35a7 − 15a8 = 0. Finally, orthogonality of (3.1)): ∂P ∂vi to tjk tjm with i, j, k, m pairwise distinct reads (taking into account 20a1 + a6 + 15a8 = 0. Note that by now we have covered all options. The system of last three equations has onedimensional space of solutions consisting of (a1 , a6 , a7 , a8 ) satisfying 11 3 5 a6 = a1 , a7 = − a1 , a8 = − a1 .. 2 2 2 (3.2) It remains to plug in some non-zero number as a1 (we choose a1 = 2 to kill denominators) and use (3.1) and (3.2) to obtain the required formula for P . Now we consider the potential complex for A! . It is easily seen to coincide with the Koszul complex. In order to prove that A! is Koszul (which is equivalent to Koszulity of A), it is enough to show that the potential complex for A! is exact. We do this by calculating the Hilbert series of A! and applying the following lemma. The main part is the calculation of the Hilbert series. We will establish numerical Koszulity for ! A. Theorem 3.8. The Hilbert series of dual algebra A! for A is H A! = 1 . 1 − 16t + 16t2 − t3 First we find minimal series in the variety P (16, 3) of potential algebras with 16 generators and potential of degree 3. We use Proposition 3.9. The potential complex is exact in general position in P(n,m) if in P(n,m) there exists an exact potential complex. 7 This could be shown in the same way as in [1], Theorem 3.2. Then we construct an exact potential(=Koszul) complex. ∂P Example Consider a potential algebra C = K⟨x, y1 , . . . , ys ⟩/ ∂P ∂x , ∂yj , s ⩾ 2 (we need s = 15), with the potential ⟲ ⟲ s ⟲ P = xy1 y2 + xy2 y1 + ∑ xyj2 . j=3 The relations are as follows: ∂P = y1 y2 + y2 y1 + y32 + . . . + ys2 ; ∂x ∂P = xy2 + y2 x; ∂y1 ∂P = xy1 + y1 x; ∂y2 ∂P = xyj + yj x for j ⩾ 3. ∂yj It is easy to see that the only ambiguity xy1 y2 resolves, so the relations form a quadratic Gröbner basis. Thus algebra C is PBW and due to Priddy criterion [5, 3] it is Koszul. Thus the exact potential complex according to proposition and the above Example is in general position in P (n, m). Since generic series is minimal (standard arguments, see for example [4]), and exactness of potential complex defines the series uniquely via the recurrence relation, we know that 1 minimal series in P (16, 3) is the same as in the Example and is equal to 1−16t+16t 2 −t3 . So we have Proposition 3.10. HA! ⩾ 1 . 1−16t+16t2 −t3 Now we need to show that an equality in the opposite direction holds. For this we consider algebra A! over the field of characteristic 7. After coefficients of relations considered as elements of Z7 the pattern of Gröbner basis of relations becomes much simpler, namely the highest words can be made exactly the same as those in Example 1. And it is possible to check that these words form a Gröbner basis, hence the Hilbert series of A! is equal to 1 . Z7 1 − 16t + 16t2 − t3 But again standard arguments (as in [4]) show that after passing from Z to Zp , the series can only become bigger. Hence we have H A! = Proposition 3.11. HA! ⩽ HA! = Z7 1 . 1−16t+16t2 −t3 Combining Proposition3.10 and Proposition3.11 we have that the Hilbert series of A! is 1 . 1 − 16t + 16t2 − t3 To prove Theorem3.1 we combine Theorem3.8, lemma?? and note that the potential complex is exact in all but one terms. H A! = 4 Acknowledgements We are grateful to IHES and MPIM for hospitality, support, and excellent research atmosphere. I would like to thank Yu.I.Manin for asking this question and constant encouragement. This work is funded by the ERC grant 320974, and partially supported by the project PUT9038. 8 References [1] N.Iyudu, A.Smoktunowicz Golod-Shafarevich type theorems and potential algebras, preprint IHES, 2016. [2] S.Keel, Theory of moduli spaces of stable n-pointed curves of genus zero, Trans. AMS, 330 (1992), no.2, 545–574. [3] Yu.I.Manin, Some remarks on Koszul algebras and quantum groups, Ann. Inst. Fourier (Grenoble) 37 (1987), no. 4, 191–205. [4] A. Polishchuk and L. Positselski, Quadratic algebras, University Lecture Series, 37 (2005), American Mathematical Society, Providence, RI. [5] S.Priddy, Koszul resolutions, Trans. Amer. Math. Soc. 152 (1970), 39–60. [6] Robert Laugwitz, Vladimir Retakh, Algebras of Quasi-Plcker Coordinates are Koszul, J.Algebra, 296, N2, p.440-461. [7] Vladimir Retakh, Shirlei Serconek, Robert Lee Wilson, Koszulity of splitting algebras associated with cell complexes, J.Algebra, 323, N4, p.983-999. [8] Vladimir Retakh, Shirlei Serconek, Robert Lee Wilson, On a class of Koszul algebras associated to directed graphs, J.Algebra, 304 (2005), N2, p.1114-1129. [9] GRAAL System of computations in (noncommutative) Graded Algebras, Ulyanovsk brunch of the Moscow State University, A.Kondratiev, supervised by A.Verevkin, 1998. Natalia Iyudu School of Mathematics The University of Edinburgh James Clerk Maxwell Building The King’s Buildings Peter Guthrie Tait Road Edinburgh Scotland EH9 3FD E-mail addresses: [email protected], [email protected] 9
0math.AC
arXiv:1709.05000v2 [cs.DM] 23 Sep 2017 Weighted and locally bounded list-colorings in split graphs, cographs, and partial k-trees Cédric Bentz∗ Abstract For a fixed number of colors, we show that, in node-weighted split graphs, cographs, and graphs of bounded tree-width, one can determine in polynomial time whether a proper list-coloring of the vertices of a graph such that the total weight of vertices of each color equals a given value in each part of a fixed partition of the vertices exists. We also show that this result is tight in some sense, by providing hardness results for the cases where any one of the assumptions does not hold. The edge-coloring variant is also studied, as well as special cases of cographs and split graphs. Keywords: locally bounded list-colorings, dynamic programming, NPcompleteness, maximum flows, tree-width, split graphs, cographs. 1 Introduction A proper coloring of a given graph G is an assignment of colors (integers) to its vertices, such that any two vertices linked by an edge of G take different colors. For any given color, the set of vertices taking this color is called a color class. In [2], a coloring problem with non global constraints on the sizes of the color classes was studied. More precisely, the following problem was considered: LocallyBoundedColoring Instance: A graph G = (V, E), a partition V1 , . . . , Vp of the vertex set V , and a list of pk integral bounds (n11 , . . . , n1k , n21 , . . . , n2k , . . . , np1 , . . . , npk ) such that Pk j=1 nij = |Vi | for each i ∈ {1, . . . , p}. Question: Decide whether there exists a proper k-coloring of G (i.e., a proper coloring of G using k colors) that is such that, for each i ∈ {1, . . . , p} and for each j ∈ {1, . . . , k}, the number of vertices having color j in Vi is nij . In this previous paper, it was shown that this problem can be solved in O(n2pk−1 log n) time when G is a tree, where n = |V |. A link with a scheduling problem consisting in processing a set of unit tasks on a set of processors with various unavailability constraints was also presented. ∗ CNAM & CEDRIC ([email protected]) Laboratory, 292 1 rue Saint-Martin 75003 Paris, France Moreover, it was shown in [10] that the following list-coloring problem (i.e., where each vertex must take a color from a list of possible colors) with global constraints on the size of the color classes is tractable in several classes of graphs: BoundedListColoring Instance: A graph G = (V, E), a list of k integral bounds (n1 , . . . , nk ), and, for each vertex v of G, a list of possible colors L(v) ⊆ {1, . . . , k}. Question: Decide whether there exists a proper k-coloring of G such that, for each i ∈ {1, . . . , k}, the number of vertices having color i is at most equal to the color bound ni , and the color of v belongs to L(v) for each vertex v. More precisely, the authors of [10] described efficient algorithms to solve this problem in graphs of bounded tree-width and in cographs (that can be defined as graphs with no induced path on four vertices [6]). The notion of tree-width will be recalled formally when needed: it can be viewed as a measure of the tree-likeness of a graph, and equals 1 in forests (and 0 in graphs with no edges). Graphs of tree-width at most t for some t ≥ 0 are also called partial t-trees [6]. In this paper, we generalize both results, and show how to solve efficiently the list-coloring problem with non global and weighted constraints on the sizes of the color classes in graphs of bounded tree-width (and also in cographs). More precisely, the problem we shall study can be formally defined as follows: WeightedLocallyBoundedListColoring Instance: A graph G = (V, E), a weight function w : V → N∗ , a partition V1 , . . . , Vp of the vertex set V , a list of pk integral bounds (W11 , . . . , W1k , W21 , . . . , Pk P W2k , . . . , Wp1 , . . . , Wpk ) with j=1 Wij = v:v∈Vi w(v) for each i ∈ {1, . . . , p}, and, for each vertex v of G, a list of possible colors L(v) ⊆ {1, . . . , k}. Question: Decide whether there exists a proper k-coloring of G such that, for each i ∈ {1, . . . , p} and for each j ∈ {1, . . . , k}, the total weight of vertices having color j in Vi is Wij , and the color of v belongs to L(v) for each vertex v. In the scheduling application mentioned above, the weight of each vertex would correspond to the amount of resource needed to process the associated task. Moreover, apart from list-coloring problems, other constrained coloring problems related to ours have been studied in the literature, see for instance [1, 3, 4, 5, 7, 8, 11, 14]. We will not review all the associated known results, as we will not need them all, and will only recall the useful ones when needed. The equitable coloring problem consists in finding a coloring where the sizes of any two color classes differ by at most one, while the bounded coloring problem consists in finding a coloring where the size of any color class does not exceed a given bound (that is common to all color classes). Moreover, without the requirement of being a list-coloring, the problem studied in [10] is defined as the capacitated coloring problem in [5]. Hence, the bounded coloring problem is a special case of the capacitated coloring problem, where all the color bounds (i.e., all the bounds on the sizes of the color classes) are equal. 2 It was observed in [5] that the equitable coloring problem is in fact a special case of the capacitated coloring problem   (and one in which the color bounds   are actually reached), by setting ni = nk for any i ≤ n mod k, and ni = nk for any other i, where n is the number of vertices. It was also observed in [3] that any instance of the bounded coloring problem is equivalent to an instance of the equitable coloring problem with the same number of colors, by adding sufficiently many isolated vertices (which destroys neither the property of having bounded tree-width nor the one of being a cograph or a split graph, i.e., a graph whose vertex set can be partitioned into a clique and an independent set [6]). In fact, all of the above problems (including the k-coloring problem, which is a list-coloring problem where L(v) = {1, . . . , k} for each vertex v) are special cases of BoundedListColoring, and this is true for WeightedLocallyBoundedListColoring as well, in most of the graphs that we shall consider: take any instance of BoundedListColoring, keep the same color bounds, set p and the weight of any vertex to 1, and add isolated vertices of weight 1 (that Pk can take any color), so that the total number of vertices reaches i=1 ni . Numerous applications of the list-coloring problem and of the capacitated coloring problem were described respectively in [14] and in [4, 5]. Many papers also consider a precolored variant of the equitable or capacitated coloring problem, where some of the vertices are already colored [3]. Observe that this can also be achieved, when considering an instance of the list-coloring problem, by defining the list of possible colors for each vertex that is already colored as a singleton, corresponding to the given color. Similarly, one can transform a listcoloring instance into a precolored one, by adding pendant vertices adjacent to each vertex v (these pendant vertices being colored by the colors not in L(v)). Such an operation does not destroy the property of having bounded tree-width, but can destroy the one of being a cograph (or a split graph, depending on whether v lies in the clique or in the independent set). Finally, when considering instances of WeightedLocallyBoundedListColoring where p is not fixed and L(v) = {1, . . . , k} for each vertex v, then precoloring some vertices can be achieved by putting each of these vertices in one set Vi of the partition V1 , . . . , Vp (Vi contains only this vertex), and defining the Wij ’s appropriately. The present paper is organized as follows. We begin by providing, in Section 2, tight hardness results for WeightedLocallyBoundedListColoring. Then, we describe in Section 3 a dynamic programming algorithm for this problem, that runs in polynomial time whenever we are not in the cases covered by Section 2 (i.e., when p, k and the tree-width of the graph are fixed, and the vertex weights are polynomially bounded). In Section 4, we prove that the problem is tractable in cographs under similar assumptions, and that, under weaker assumptions, it can be solved in polynomial time in several subclasses of cographs. Section 5 is devoted to split graphs, and we show that, in such graphs, the problem can be solved in (pseudo)polynomial time even when p is not fixed (provided that k is), and provide additional tractable special cases. Finally, in the last section of the paper, we extend to edge colorings all our results concerning (vertex) colorings in the graphs studied in Sections 3 to 5. 3 2 NP-completeness proofs In this section, we prove that WeightedLocallyBoundedListColoring, which is clearly in NP, is NP-complete, even in very restricted special cases. Before detailing these special cases, we note that, when k = 3, WeightedLocallyBoundedListColoring is NP-complete in general graphs (i.e., when the tree-width is unbounded), even when p = 1 and w(v) = 1 for each vertex v, as deciding whether a graph can be colored using three colors is. Also note that this latter result does not hold when k = 2: in this case, the coloring problem is trivial, as the input graph is bipartite (otherwise, the answer is no). Actually, WeightedLocallyBoundedListColoring itself can be solved in polynomial time when k = 2 and p = O(1), provided that the vertex weights are polynomially bounded (in the next subsection, we deal with the case where they are not). Indeed, if k = 2, then any connected component of the graph is bipartite (otherwise, the answer is no), and hence admits two possible colorings. For each one, one can check whether such a coloring is a valid list-coloring in the associated connected component. This implies that the whole problem can be solved by a dynamic programming algorithm similar to the one given in [10, Theorem 2]: we consider the connected components one after the other, keeping track for each one of them of all the possible weights of vertices of color c in the set Vh , for each c and h, and then combining these weights whenever a new component is considered. The total amount of information that we need to keep track of is thus O((maxh,c Whc )2p ), which is polynomial under our assumptions. 2.1 When vertex weights are arbitrary We first give a short proof that WeightedLocallyBoundedListColoring is NP-complete in general, even when k = 2 (there are two colors) and p = 1 (the vertex partition contains only one set). To do this, we consider the following well-known weakly NP-complete problem [9]: Partition Pn Instance: n + 1 positive integers (a1 , . . . , an , B) such that i=1 aP i = 2B. Question: Decide whether there exists I ′ ⊂ {1, . . . , n} such that i∈I ′ ai = B. Given an instance (a1 , . . . , an , B) of Partition, we define the following instance I of WeightedLocallyBoundedListColoring: the graph G consists of n isolated vertices v1 , . . . , vn . We define p = 1, k = 2 and W11 = W12 = B. We also define, for each i ∈ {1, . . . , n}, L(vi ) = {1, 2} (i.e., any vertex can take any color) and w(vi ) = ai . Then, (a1 , . . . , an , B) has a solution if and only if I admits a solution, i.e., a 2-coloring where the total weight of the vertices of each color is B: a vertex vi will take color 1 if and only if i ∈ I ′ . This implies: Theorem 1. WeightedLocallyBoundedListColoring is NP-complete in the weak sense in graphs consisting of isolated vertices, even if k = 2, p = 1, and each vertex can take any color. 4 Note that this instance can be made connected by setting k = 3 and adding, for each i ∈ {1, . . . , n − 1}, two edges (vi , ui ) and (ui , vi+1 ), where the ui ’s are new vertices of color 3 (e.g., L(ui ) = {3} for each i ∈ {1, . . . , n − 1}). The graph G we obtain is then a tree (in fact, a chain), and so has tree-width 1. Also note that, if k = 2 and G is connected, then solving WeightedLocallyBoundedListColoring is a trivial task, since a connected bipartite graph has only two proper (vertex) colorings using two colors. In [3], Bodlaender and Fomin managed to show that the equitable coloring problem can be solved in polynomial time in graphs of bounded tree-width (i.e., even if the number of colors is not fixed). In the next subsections, we rule out this possibility by proving that, even if all vertex weights are polynomially bounded, WeightedLocallyBoundedListColoring remains NP-complete in graphs of bounded tree-width, if either k or p is not fixed. 2.2 When the number of colors is not fixed First, we look at the case where the number k of colors is not fixed. We consider the following well-known strongly NP-complete problem [9]: 3−Partition Instance: A set of 3n positive integers A = (a1 , . . . , a3n ), and an integer B such P3n that i=1 ai = nB and, for each i ∈ {1, . . . , 3n}, B/4 < ai < B/2. Question: Decide whether A can be partitioned into P n disjoint sets of three elements A1 , . . . , An such that, for each j ∈ {1, . . . , n}, ai ∈Aj ai = B. Since this problem is strongly NP-complete, we can assume w.l.o.g. that the ai ’s are polynomially bounded in n. We construct an instance of WeightedLocallyBoundedListColoring in a graph G consisting of 3n isolated vertices vi , by defining p = 1, k = n, W1c = B for each c ∈ {1, . . . , n}, and, for each i ∈ {1, . . . , 3n}, w(vi ) = ai and L(vi ) = {1, . . . , k}. It is easy to see that the equivalence between the solutions of the two instances is given by: for each h ∈ {1, . . . , n}, a vertex vi will take color h in G if and only if ai ∈ Ah . Hence: Theorem 2. WeightedLocallyBoundedListColoring is NP-complete in the strong sense in graphs consisting of isolated vertices, even if p = 1, all the vertex weights are polynomially bounded, and each vertex can take any color. As in the case of Theorem 1, we can make this instance connected (for instance, obtaining a chain, or a star) by defining k = n + 1 and adding new vertices (and associated edges) that can only take color n + 1. Note, however, that in this reduction we need non uniform vertex weights. Actually, this is unavoidable, from the following fact: the problem becomes tractable in graphs with no edges (and hence also in stars, by a simple reduction) when w(v) = 1 for each v ∈ V . For the sake of completeness, we give a proof of this fact. Proposition 1. If all vertex weights are 1, WeightedLocallyBoundedListColoring is polynomial-time solvable in graphs consisting of isolated vertices. 5 Proof. Given an instance in such a graph G, we can construct a bipartite graph H as follows: there is a vertex vi in H for each vertex vi in G, and, for each h ∈ {1, . . . , p} and c ∈ {1, . . . , k}, there are Whc vertices ujh,c in H. There is an edge between vi and ujh,c in H if and only if vi ∈ Vh and c ∈ L(vi ): since Pk P v:v∈Vh w(v) = |Vh | for each h ∈ {1, . . . , p}, this implies that the c=1 Whc = number of vertices is the same in each side of the bipartition of H. Then, there exists a feasible coloring for the initial WeightedLocallyBoundedListColoring instance in G if and only if H admits a perfect matching. (H being bipartite, such a perfect matching can be computed as a maximum flow.) Indeed, we define the following equivalence: vertex vi ∈ Vh has color c if and only if the edge vi ujh,c belongs to this perfect matching, for some j. On the one hand, each vertex vi will be incident to one edge of the matching, and hence vi will take one color (by construction, this is a color in L(vi )); on the other hand, each vertex ujh,c will be incident to one edge of the matching, and hence in each Vh there will be exactly Whc vertices with color c. We can nevertheless prove the strong NP-completeness of WeightedLocallyBoundedListColoring in graphs of bounded tree-width when w(v) = 1 for each v ∈ V , p = 1 and k is arbitrary, by using a more complex reduction from 3−Partition. Actually, such a reduction was already given in [3, Theorem 8] for the precolored variant of the equitable coloring problem. In this reduction, each instance is a set of trees of height 3 (together with additional isolated vertices). The leaves are in fact precolored vertices, which are used to restrain the set of possible colors for other vertices. Hence, if we consider list-colorings, we can ignore (i.e., remove) these leaves, and simply define suitable lists of possible colors for the other vertices. This yields instances consisting of trees of height at most 2, and leaves as open the case where each connected component is a tree of height at most 1 (which, unlike trees of height 2, are also cographs). We now describe an alternative reduction, that needs more colors but only makes use of stars (i.e., trees of height 1), and that will prove useful in Sections 3 to 5. Assume we are given an instance I = (a1 , . . . , a3n , B) of 3−Partition with n ≥ 2. We construct an instance I ′ of WeightedLocallyBoundedListColoring as follows: we define p = 1, k = 3n2 + 4n, w(v) = 1 for each vertex v, and the graph, that contains 3n2 (nB +1)+3n vertices, consists of 3n2 vertexdisjoint stars and 3n isolated vertices u1 , . . . , u3n . We denote each star by Sij , for each i ∈ {1, . . . , 3n} and each j ∈ {1, . . . , n}, and its central vertex by vij . For each i ∈ {1, . . . , 3n} and each j ∈ {1, . . . , n}, vij can take any of the n + 1 colors in {n + i, ni + 3n + 1, ni + 3n + 2, . . . , ni + 4n}, and Sij has 3nai leaves, which can only take colors j and n + i. Moreover, for each i ∈ {1, . . . , 3n}, vertex ui can take any of the n colors in {ni + 3n + 1, . . . , ni + 4n}. We also define W1j = 3nB for each j ∈ {1, . . . , n}, W1(n+i) = 3nai (n − 1) + 1 for each i ∈ {1, . . . , 3n}, and W1(ni+3n+h) = 1 for each i ∈ {1, . . . , 3n} and each h ∈ {1, . . . , n}. Then, the following holds: Lemma 1. I has a solution if and only if I ′ has a solution. 6 Proof. We shall prove that the following equivalence holds: ai ∈ Aj if and only if vij takes color n + i, for each i ∈ {1, . . . , 3n} and each j ∈ {1, . . . , n}. Assume that we are given a solution to the 3−Partition instance I. For each i ∈ {1, . . . , 3n}, if ai ∈ Aj for some j, then vij takes color n + i, and the ′ 3nai leaves in Sij take color j, while, for each j ′ 6= j, vij takes color ni + 3n + j ′, ′ and all the leaves in Sij take color n + i. Moreover, for each i ∈ {1, . . . , 3n}, ui takes color ni + 3n + j. This implies, in I ′ , that there are 3nai (n − 1) + 1 vertices of color n+ i for each i ∈ {1, . . . , 3n}, and there are 3nai1 + 3nai2 + 3nai3 = 3nB (assuming Aj = {ai1 , ai2 , ai3 }) vertices of color j for each j ∈ {1, . . . , n}. Conversely, assume that we are given a solution to the WeightedLocallyBoundedListColoring instance I ′ , and let ai ∈ Aj if vij takes color n + i for each i ∈ {1, . . . , 3n} and each j ∈ {1, . . . , n}. Observe that, for each i ∈ {1, . . . , 3n}, exactly one of the vij ’s takes color n + i (from the fact that W1(ni+3n+h) = 1 for each h, and from the possible colors that can be taken by ui and the vij ’s). This means that the leaves of such a Sij must take color j. For each i ∈ {1, . . . , 3n}, the leaves of all the other Sij ’s must take color n + i, since W1(n+i) = 3nai (n − 1) + 1. Hence, for each j ∈ {1, . . . , n}, there are exactly three i’s in {1, . . . , 3n} such that vij has color n + i: if there are at most two such i, then ai < B/2 for each i implies that at most 3n(B − 1) vertices have color j, a contradiction; if there are at least four such vertices, then ai > B/4 for each i implies that at least 3n(B + 1) vertices have color j, a contradiction. The sum of these three ai ’s is then W1j /3n = B, which concludes the proof. Hence, we have proved: Theorem 3. WeightedLocallyBoundedListColoring is NP-complete in the strong sense in star forests, even if p = 1 and w(v) = 1 for each v ∈ V . Corollary 1. WeightedLocallyBoundedListColoring is strongly NPcomplete in cographs of tree-width 1, even if p = 1 and w(v) = 1 ∀v ∈ V . Again, we can make the instance I ′ connected by adding a new color, as well as one or more vertices that can only take this new color. After doing this, we can obtain a tree of height 2 (or a caterpillar), while recall that, from a direct consequence of Proposition 1, WeightedLocallyBoundedListColoring is tractable in stars (i.e., in trees of height 1) when w(v) = 1 for each v ∈ V . Moreover, one of the features of the above reduction is that all the ui ’s and vij ’s must take different colors, which will be useful in Section 5. However, one can use a simpler reduction, in which |L(v)| = 2 for each vertex v. Too see this, remove all the ui ’s, and set k = 4n and L(v) = {j, n + i} for each vertex v in each Sij . The color bounds are then W1(n+i) = 3nai (n − 1) + 1 for each i ∈ {1, . . . , 3n} and W1j = 3nB + 3(n − 1) for each j ∈ {1, . . . , n}. As in the proof of Theorem 3, and for similar reasons, we have ai ∈ Aj if and only if vij takes color n + i for each i ∈ {1, . . . , 3n} and each j ∈ {1, . . . , n}. Note that the above reductions leave as open the case of chains (which are not cographs, except when they are also stars) where the vertex weights are 1, p 7 is fixed and k is not. The reduction given in [7, Theorem 3] closes this gap, and shows that the problem is strongly NP-complete in this case, even when p = 1 and |L(v)| = 2 for each v ∈ V . However, this reduction is rather complicated, in particular since all the chains do not play the same role. We now give a different kind of reduction to prove this result, in which we have p = 1 and |L(v)| = 2 for each v ∈ V , as well as an additional restriction on the lists of possible colors: Theorem 4. WeightedLocallyBoundedListColoring is NP-complete in the strong sense in linear forests, even if p = 1, w(v) = 1 and |L(v)| = 2 for each v ∈ V , and all the vertices of each chain share one common possible color. Proof. Given a 3−Partition instance, instead of defining 3n2 stars as in Theorem 3, we define 3n2 disjoint chains µji , for each i ∈ {1, . . . , 3n} and each j ∈ {1, . . . , n}. We also define p = 1, k = 7n + 3n2 , and w(v) = 1 for each v ∈ V . For each i ∈ {1, . . . , 3n} and each j ∈ {1, . . . , n}, µji has 2ai + 2 vertices, and (i) the first and the last vertices in µji can only take colors (i − 1)n + j + 7n and n+ i, (ii) the vertices with an odd index different from 1 in µji can only take colors j and n + i, (iii) the vertices with an even index different from 2ai + 2 in µji can only take colors n + i and 4n + i. Hence, for each vertex v in µji , we have n + i ∈ L(v). To end the reduction, we define W1(n+i) = n(ai + 1) and W1(4n+i) = (n − 1)ai for each i ∈ {1, . . . , 3n}, W1(7n+j+(i−1)n) = 1 for each i ∈ {1, . . . , 3n} and each j ∈ {1, . . . , n}, and W1j = B for each j ∈ {1, . . . , n}. Note that, for each i ∈ {1, . . . , 3n} and each j ∈ {1, . . . , n}, any feasible coloring has exactly one vertex (either the first one or the last one) in µji of color 7n + j + (i − 1)n, while there can be at most ai + 1 vertices of color n + i in this chain. Since each W1(n+i) is equal to n(ai + 1), this implies that in any feasible coloring there are exactly ai + 1 vertices of color n + i in µji , for each i ∈ {1, . . . , 3n} and each j ∈ {1, . . . , n}: the vertices of color n + i in µji are either all the vertices with an odd index, or all the vertices with an even index. Hence, in each chain, the ai remaining vertices either all have color j, or all have color 4n + i. Moreover, since each W1(4n+i) is equal to (n − 1)ai , this implies that in any feasible coloring there are exactly n − 1 integers j in {1, . . . , n} such that there are ai vertices of color 4n + i in µji , for each i ∈ {1, . . . , 3n}. We can thus define the following equivalence between the two instances: for each i and each j, ai ∈ Aj if and only if there are ai vertices of color j in µji (for each j, there will be exactly three such i’s, for reasons similar to the ones mentioned in the proof of Theorem 3), which concludes the proof. We conclude by pointing out that we can turn the previous instance into a chain by adding a new color and new vertices that can only take this new color. 2.3 When p is not fixed Now, we prove the NP-completeness of the case where k is fixed but p is not, by considering the following well-known strongly NP-complete problem [9]: 8 MonotoneOneInThreeSAT Instance: A set X of ν boolean variables x1 , . . . , xν , and a set of µ clauses, each one containing exactly three (non negated) boolean variables from X. Question: Decide whether there exists a truth assignment for the variables in X such that in every clause there is exactly one variable equal to true. Theorem 5. WeightedLocallyBoundedListColoring is NP-complete in the strong sense in star forests, even if k = 2, and each vertex has weight 1 and can take any color. Proof. Given an instance of MonotoneOneInThreeSAT, we construct an instance of WeightedLocallyBoundedListColoring as follows: the graph G consists of ν vertex-disjoint stars, one star per variable xi . We denote by vi occ(i) its leaves, where occ(i) the central vertex of the ith star, and by u0i , . . . , ui is the number of occurrences of xi in the set of clauses. Then, we define k = 2, p = ν + µ, and, for each vertex v, w(v) = 1 and L(v) = {1, 2}. The partition of the vertices of G and their target weights are given by: Vh = {vh , u0h } and Wh1 = Wh2 = 1 for each h ∈ {1, . . . , ν}, and Vh = {uai , ubj , uck }, Wh1 = 1 and Wh2 = 2 for each h ∈ {ν + 1, . . . , ν + µ}, where we consider that the (h − ν)th clause consists of the ath occurrence of the variable xi , of the bth occurrence of the variable xj , and of the cth occurrence of the variable xk . It is easy to check that we have the following equivalence between the initial MonotoneOneInThreeSAT instance and this WeightedLocallyBoundedListColoring instance: for each i ∈ {1, . . . , ν}, variable xi is equal to true if and only if vi has color 2, which concludes the proof. Corollary 2. WeightedLocallyBoundedListColoring is strongly NPcomplete in cographs of tree-width 1, even if k = 2, and each vertex has weight 1 and can take any color. As in the case of Theorem 3, we can make the previous instance connected (obtaining a tree of height 2, or a caterpillar) by defining k = 3 and adding new vertices (and associated edges, to link the stars together) that can only take color 3 (for instance, we can define p = ν + µ + 1 and Wp3 = |Vp |, and include all these new vertices of weight 1 in the new Vp , keeping the other Vh ’s unchanged). Again, this reduction, that will prove useful both in Sections 3 and 4 (thanks to Corollary 2), leaves as open the case of chains where w(v) = 1 for each v ∈ V , k is fixed and p is not. The next theorem closes this gap. Theorem 6. WeightedLocallyBoundedListColoring is NP-complete in the strong sense in linear forests, even if k = 2, and each vertex has weight 1 and can take any color. Proof. Given a MonotoneOneInThreeSAT instance, we define a WeightedLocallyBoundedListColoring instance as follows: the graph G consists of ν vertex-disjoint chains (one per variable xi ) and of isolated vertices. The verocc(i) occ(i) (in this order), where occ(i) is , ui tices of the ith chain are vi1 , u1i , . . . , vi 9 defined as in the proof of Theorem 5. For each chain, we also add occ(i) isolated Pν occ(i) 1 . Then, we define k = 2, p = µ + i=1 occ(i), and, for vertices v ′ i , . . . , v ′ i each vertex v, w(v) = 1 and L(v) = {1, 2}. It remains to define the partition of the vertices of G and their target weights. We have Vh = {uai , ubj , uck }, Wh1 = 1 and Wh2 = 2 for each h ∈ {1, . . . , µ}, where we consider that the hth clause consists of the ath occurrence of the variable xi , of the bth occurrence of the variable xj , and of the cth occurrence of the variable xk . For each i ∈ {1, . . . , ν} l and each l ∈ {1, . . . , occ(i)}, we define Vh = {vil , v ′ i } and Wh1 = Wh2 = 1, Pi−1 where h = l + µ + j=1 occ(j). As in the proof of Theorem 5, for each i ∈ {1, . . . , ν}, variable xi is equal to true if and only if vi1 has color 2. Indeed, for each i, all the vij ’s must have the same color, and all the uji ’s must have the same color (the v ′ ji ’s are only useful to ensure that vi1 can freely take color 1 or 2). Theorem 6 follows. As in the case of Theorem 4, we can make the previous instance connected (obtaining a chain) by defining k = 3 and adding new vertices (and associated edges) that can only take color 3. Also note that, in the reductions used to prove Theorems 5 and 6, we only need colorings (not list-colorings), although we did need to make use of list-colorings in the proofs of Theorems 3 and 4. 3 An algorithm for graphs of bounded tree-width This section deals with WeightedLocallyBoundedListColoring in partial k-trees.  Given a graph G = (V, E), a tree decomposition of G is a pair {Xi |i ∈ I}, T where Xi ⊆ V, ∀i ∈ I, is a bag of vertices of G, T = (I, F ) is a tree, and: S (1) i∈I Xi = V , (2) For every edge uv ∈ E, there is an i ∈ I such that u, v ∈ Xi , (3) For all i, j, l ∈ I, if j lies in the path between i and l, then Xi ∩ Xl ⊆ Xj . The width of a given tree decomposition of a graph G is equal to maxi∈I |Xi |− 1. The tree-width of a graph G, denoted by tw(G), is the minimum width of a tree decomposition of G, taken over all tree decompositions of G. Note that trees (and hence chains and stars) have tree-width 1. Without loss of generality, we can also assume that the tree decomposition is nice [15], i.e.: • T is rooted at some node r, • T is binary and has O(|V |) nodes, • If a node i has two children j and k then Xi = Xj = Xk (join node) • If a node i has one child j, then either (a) |Xi | = |Xj | − 1 and Xi ⊂ Xj (forget node) (b) |Xi | = |Xj | + 1 and Xj ⊂ Xi (introduce node) 10 Given two vertices i, j of T , we will use the notation j  i to denote the fact that j is either i or a descendant of i with respect to r. Given a node i ∈ I, let Yi = E ∩ (Xi × XiS ), i.e., Yi is the subset of E induced by the vertices in Xi . Moreover, let Ti = ji Xj and let G[Ti ] be the subgraph of G induced by Ti . In order to design a standard dynamic programming algorithm that solves WeightedLocallyBoundedListColoring, we use the following function f : • f (i, ci , ω11 , . . . , ω1k , . . . , ωp1 , . . . , ωpk ) = true if there exists a list-coloring of G[Ti ] where each vertex u ∈ Xi has color ci (u) and where the total weight of vertices of G[Ti ] having color c in Vh is ωhc , and false otherwise. To describe our algorithm, we now simply need to write down the induction equations defining the values of f (·), for each type of nodes of T . Then, these values will be computed in a bottom-up fashion, starting from the leaves of T . Note that, by the definition of tree decompositions, any two vertices linked by an edge in G must both belong to at least one common bag of the tree decomposition (Condition (2)). This implies that a (list-)coloring that is proper in each bag is also proper in the whole graph G, provided that each vertex has the same color in each bag it belongs to. Moreover, Condition (3) ensures that the subgraph of T induced by the bags any given vertex belongs to is connected, and hence in all these bags this vertex will have the same color if we do not change its color whenever we move (in T ) from one bag to an adjacent one. If i is a forget node. Let j denote its child such that Xj = Xi ∪ {v}: _ f (j, cj , ω11 , . . . , ωpk ) f (i, ci , ω11 , . . . , ωpk ) = cj :(cj (u)=ci (u) ∀u∈Xi )∧(cj (v)∈L(v)) If i is an introduce node. Let j denote its child such that Xj = Xi \ {v}, and assume that v ∈ Vh : ^ f (i, ci , ω11 , . . . , ωpk ) = E f (j, cj , ω11 , . . . , ωhci (v) − w(v), . . . , ωpk ) V V where V E = (ci (v) ∈ L(v)) (ci (v) 6= ci (u) ∀u : uv ∈ Yi ) (cj (u) = ci (u) ∀u ∈ Xj ) (ωhci (v) ≥ w(v)). i If i is a join node. Let j and l denote its two children, and let whc be the total weight of vertices v ∈ Xi ∩ Vh such that ci (v) = c, for each h and c. _  i i f (j, ci , ω11 +w11 −q11 , . . . , ωpk +wpk −qpk )∧f (l, ci , q11 , . . . , qpk ) f (i, ci , ω11 , . . . , ωpk ) = (q11 ,...,qpk )∈Qi i where Qi = {(q11 , . . . , qpk ) : whc ≤ qhc ≤ ωhc ∀h ∀c}. 11 If i is a leaf node. In this case, we just have to check that the coloring function ci provides a valid locally bounded list-coloring of G[Xi ]. ^ f (i, ci , ω11 , . . . , ωpk ) = E (w({v ∈ Xi ∩ Vh : ci (v) = c}) = ωhc ∀h ∀c) V where E = (ci (v) ∈ L(v) ∀v ∈ Xi ) (ci (v) 6= ci (u) ∀u, v ∈ Xi : uv ∈ Yi ). Root value. The answer true or false for the initial instance is obtained at the root r by computing the following value: _ f (r, cr , W11 , . . . , Wpk ) cr :cr (u)∈L(u) ∀u∈Xr Lemma 2. The values of f (i, ci , ω11 , . . . , ω1k , . . . , ωp1 , . . . , ωpk ) computed by the above algorithm are correct. Proof. In order to show that a vertex list-coloring is proper in Ti , first notice that, from our preliminary remark, it suffices to show that (i) it is a proper listcoloring in Xi and (ii) the color of any vertex remains the same when moving from one bag of Ti to an adjacent one. We now show the correctness of the above equations by considering each possible node type for i. Assume i is a forget node. Then, since Ti = Tj , f (i, ci , ω11 , . . . , ωpk ) is true if and only if f (j, cj , ω11 , . . . , ωpk ) is true for some coloring cj , such that cj (u) = ci (u) for each u ∈ Xi (each vertex must keep its color when moving from Xj to Xi ) and vertex v ∈ Xj takes some color in L(v). Assume i is an introduce node. Then, f (i, ci , ω11 , . . . , ωpk ) is true if and only if f (j, cj , ω11 , . . . , ωhci (v) − w(v), . . . , ωpk ) is true (v ∈ Vh has weight w(v) and color ci (v), so the total weight of vertices of color ci (v) in Vh is ωhci (v) −w(v) ≥ 0 in Tj ), each vertex keeps its color when moving from Xj to Xi , and the color of v defines a valid list-coloring in the subgraph of G induced by Xi (i.e., ci (v) ∈ L(v) and ci (v) 6= ci (u) for each u such that uv ∈ Yi ). Assume i is a join node. Then, f (i, ci , ω11 , . . . , ωpk ) is true if and only if ′ ′ ) and f (l, ci , q11 , . . . , qpk ) are true for some q11 , . . . , qpk , . . . , qpk both f (j, ci , q11 ′ i ′ ′ for each h and c (with obviously = ωhc + whc and q11 , . . . , qpk such that qhc + qhc i i ′ whc ≤ qhc ≤ ωhc and whc ≤ qhc ≤ ωhc for each h and c), since the weights of the vertices in Xi are counted twice, i.e., both in Tj and in Tl (and any other vertex weight in Tj ∪ Tl is counted only once). Assume i is a leaf node. Then, f (i, ci , ω11 , . . . , ωpk ) is true if and only if the coloring function ci provides a valid locally bounded list-coloring of G[Xi ]. Finally, the root value is obtained by requiring that cr is a list-coloring in the subgraph of G induced by Xr , which concludes the proof. Running time. Let Wmax = maxh,c Whc (with Wmax ≤ n maxv∈V w(v)). The running time for a given node of T , that depends on its type, is given by: 12 node type forget introduce join leaf running time O(tw(G) + k) O(tw(G) + k) pk O(Wmax ) O(tw(G)(tw(G) + pk)) There are O(n) nodes in T , O(k tw(G)+1 ) possible colorings of any given bag, pk and O(Wmax ) possible pk-tuples ω11 , . . . , ωpk , so when running the algorithm we pk have O(nWmax k tw(G)+1 ) values f (·) to compute. Since computing the optimal pk value only takes O(k tw(G)+1 ) time, the overall running time is O(nWmax k tw(G)+1 pk (Wmax + tw(G)(tw(G) + pk))). Together with Lemma 2, this implies: Theorem 7. In graphs of bounded tree-width, WeightedLocallyBoundedListColoring can be solved: • in pseudopolynomial time when p and k are fixed, • in polynomial time when (i) p and k are fixed, and (ii) all vertex weights are polynomially bounded. Observe that these results are best possible, in the sense that, from Theorems 1 to 6, dropping any of the assumptions (on p, k, and the vertex weights) leads to NP-completeness. They also generalize the results in [2, 10, 14]. We close this section by mentioning that this approach can be adapted to solve an optimization version of WeightedLocallyBoundedListColoring. More precisely, one can associate a profit function π : V × {1, . . . , k} → Z to the vertices of G; the profit of a vertex then depends on its color. By slightly modifying the above dynamic programming algorithm, one can compute a valid weighted and locally bounded list-coloring of maximum (or minimum) profit (if any). In this case, the value of f (·) is no longer equal to true or false, i.e.: • f (i, ci , ω11 , . . . , ω1k , . . . , ωp1 , . . . , ωpk ) = the maximum total profit of a listcoloring of G[Ti ] where each vertex u ∈ Xi has color ci (u) and where the total weight of vertices of G[Ti ] having color c in Vh is ωhc (if any). In order to compute the value of this “new” function f (·), we must make some changes in the equations. We provide them without proofs, as the arguments are quite similar to the ones used in the proof of Lemma 2 (note that, by convention, infeasible solutions will have a value of −∞, as we maximize f (·)): W • (in forget nodes and root value) “ ” becomes “max”. • (in introduce nodes) f (i, ci , ω11 , . . . , ω1k , . . . , ωpk ) is equal to −∞ if E is false, and to f (j, cj , ω11 , . . . , ωhci (v) −w(v), . . . , ωpk )+π(v, ci (v)) otherwise. W • (in join nodes) “ ” becomes “max”, “∧” becomes “+”, and we add the P value “− v∈Xi π(v, ci (v))” at the end of the line. P • (in leaf nodes) f (i, ci , ω11 , . . . , ω1k , . . . , ωpk ) is equal to v∈Xi π(v, ci (v)) if both E and the condition “w({v ∈ Xi ∩ Vh : ci (v) = c}) = ωhc ∀h ∀c” are true, and to −∞ otherwise. 13 4 Locally bounded list-colorings in cographs In this section, we study the tractability of WeightedLocallyBoundedListColoring in cographs. Cographs, as defined in Section 1, can be characterized in several ways. For instance, a graph G = (V, E) is a cograph if and only if it can be associated with a cotree T : the leaves of T are the vertices of G, and the internal nodes of T are either union nodes or join nodes. A subtree of T having a union node as a root corresponds to the disjoint union of the subgraphs of G associated with the children of this node, and a subtree of T having a join node as a root corresponds to the complete union of the subgraphs of G associated with the children of this node (i.e., we add an edge between every pair of vertices with one vertex in each subgraph). Moreover, this cotree can easily be transformed in linear time into a binary cotree with O(|V |) nodes [6]. First note that WeightedLocallyBoundedListColoring is still NPcomplete in cographs, even when k is arbitrary, p = 1, and each vertex has weight 1. Indeed, on the one hand, it was proved in [14] that the list-coloring problem is NP-complete in complete bipartite graphs (which are cographs), when k is not fixed. On the other hand, the bounded coloring problem was proved to be NP-complete in cographs in [4], by a reduction from bin packing (which, by [13], also shows its W[1]-hardness with respect to k in cographs). However, the instances used in [4, 14] have a large tree-width (as otherwise the list-coloring and bounded coloring problems are tractable [3, 14]): since star forests and isolated vertices are cographs, Corollary 1 shows that this remains true even when the tree-width is 1 (the case of tree-width 0 and polynomially bounded vertex weights being covered by Theorem 2). Corollary 2 shows that, under the same assumption of being a cograph of tree-width 1 (the case of treewidth 0 will be discussed later in this section), this is also true as soon as k = 2 (provided that p is arbitrary), even when each vertex has weight 1 and can take any color. Finally, as in the case of partial k-trees, Theorem 1 shows that allowing arbitrary vertex weights leads to weak NP-completeness in cographs of tree-width 0, even if k = 2, p = 1, and any vertex can take any color. Moreover, the instances from the previous reductions can be made connected by adding a new vertex, adjacent to all the other vertices, that must take a new color (this increases the tree-width by 1). In particular, when the graph in the reduction consisted of isolated vertices, it then becomes a star. However, when both p and k are fixed, we can design an efficient dynamic programming algorithm, based on standard techniques, to solve WeightedLocallyBoundedListColoring in cographs, by using the associated (binary) cotrees. In order to describe this algorithm, we define the following function: • f ′ (i, ω11 , . . . , ω1k , . . . , ωp1 , . . . , ωpk ) = true if there exists a list-coloring of the subgraph of G induced by the leaves of the subgraph of T rooted at node i, where the total weight of vertices of this induced subgraph of G having color c in Vh is ωhc , and false otherwise. The value of each f ′ (·) is then computed in a bottom-up fashion, as follows. 14 If i is a join node. Let j and l denote its two children. _  f ′ (j, ω11 −q11 , . . . , ωpk −qpk )∧f ′ (l, q11 , . . . , qpk ) f ′ (i, ω11 , . . . , ωpk ) = (q11 ,...,qpk )∈Qijoin where Qijoin = {(q11 , . . . , qpk ) : 0 ≤ qhc ≤ ωhc ∀h ∀c} ∩ {(q11 , . . . , qpk ) : P P ( ph=1 qhc ) × ( ph=1 (ωhc − qhc )) = 0 ∀c}. If i is a union node. Let j and l denote its two children. _  f ′ (j, ω11 −q11 , . . . , ωpk −qpk )∧f ′ (l, q11 , . . . , qpk ) f ′ (i, ω11 , . . . , ωpk ) = (q11 ,...,qpk )∈Qiunion where Qiunion = {(q11 , . . . , qpk ) : 0 ≤ qhc ≤ ωhc ∀h ∀c}. If i is a leaf node. Let vi ∈ Vh be the vertex of G corresponding to this leaf. _  ωhj = w(vi ) ∧ (ωh′ j ′ = 0 ∀(h′ , j ′ ) 6= (h, j)) f ′ (i, ω11 , . . . , ωpk ) = j∈L(vi ) Lemma 3. The values of f ′ (i, ω11 , . . . , ω1k , . . . , ωp1 , . . . , ωpk ) computed by the above algorithm are correct. Proof. In order to show that a vertex list-coloring is proper in the whole graph, it is sufficient to prove that it is a valid list-coloring at each step (i.e., during the computation of the value of f ′ (·) at each node of the cotree). Assume i is a join node. Then, the graph induced by the leaves of the subgraph of T rooted at node i is the complete union of two graphs. For each h and j, the sum of the weights of the vertices of color j in Vh in these two graphs must be equal to ωhc . However, when taking the complete union of two subgraphs of G, vertices not belonging to the same subgraph cannot have the same color (otherwise, the coloring would not be proper). Assume i is a union node. Then, the graph induced by the leaves of the subgraph of T rooted at node i is the disjoint union of two graphs, and thus, for each h and j, the sum of the weights of the vertices of color j in Vh in these two graphs must be equal to ωhc . Assume i is a leaf node. Then, in any valid list-coloring, the vertex vi of G associated with i takes only one color, which belongs to L(vi ). pk pk The above algorithm runs in O(nWmax (Wmax + pk)) time, where Wmax = maxh,c Whc . Together with Lemma 3, this yields the following result: Theorem 8. WeightedLocallyBoundedListColoring can be solved: • in pseudopolynomial time in cographs when p and k are fixed, • in polynomial time in cographs when (i) p and k are fixed, and (ii) all vertex weights are polynomially bounded. 15 Note that this result generalizes the ones in [4, 10, 14]. Again, one can associate a profit function π to the vertices of W G, and modify slightly this dynamic programming algorithm (by replacing “ ” by “max” and “∧” by “+” in union and join nodes, and by returning maxj∈L(vi ) π(vi , j) in leaf nodes if the above conditions are satisfied, and −∞ otherwise) in order to compute a feasible weighted and locally bounded list-coloring of maximum profit (if any). Note that we return −∞ if Qiunion = ∅ or Qijoin = ∅ for some i. Theorem 8 can also be used to show the following proposition: Proposition 2. If k = O(1), WeightedLocallyBoundedListColoring can be solved in (pseudo)polynomial time in graphs consisting of isolated vertices. Proof. Since there are no edges in this case, any coloring will be proper, and hence any set Vi of the partition can be considered independently from the p − 1 other Vj ’s. Hence, solving such an instance of WeightedLocallyBoundedListColoring is then equivalent to solving p independent instances where the partition contains only one vertex set. Since k is fixed as well, any such instance can be solved in (pseudo)polynomial time thanks to Theorem 8. To close this section, we study WeightedLocallyBoundedListColoring in particular cographs. We begin by studying complete bipartite graphs Kn1 ,n2 (which are cographs represented by binary cotrees containing n1 + n2 − 2 union nodes and only one join node, at the root), where this problem is trivial when k = 2 (so we shall assume k ≥ 3). Since the complexity of WeightedLocallyBoundedListColoring is the same in graphs consisting of isolated vertices and in stars (which are the graphs K1,n2 ), Theorems 1 and 2 imply that WeightedLocallyBoundedListColoring is NP-complete in complete bipartite graphs if k is not fixed or if the vertex weights are not polynomially bounded, even if any vertex can take any color. Moreover, it was proved in [14] that the list-coloring problem (without constraints on the sizes of the color classes) is strongly NP-complete in complete bipartite graphs when the number k of colors is arbitrary. However, adding isolated vertices cannot yield complete bipartite graphs, so this result does not directly apply to our problem. We now show that a more complex reduction, partly inspired by the one given in [14], can be obtained for WeightedLocallyBoundedListColoring in this case: Theorem 9. WeightedLocallyBoundedListColoring is strongly NPcomplete in complete bipartite graphs, even if p and all vertex weights are 1. Proof. Given a MonotoneOneInThreeSAT instance I with µ clauses and ν variables x1 , . . . , xν , we construct a WeightedLocallyBoundedListColoring instance I ′ in a complete bipartite graph G as follows. On the left side of G, there are µ vertices u1 , . . . , uµ associated with the occ(i) µ clauses, and, for each i ∈ {1, . . . , ν}, there are occ(i) vertices vi1 , . . . , vi associated with each variable xi , where occ(i) is the number of occurrences of xi in the set of clauses. On the right side of G, there are occ(i) vertices 16 occ(i) associated with each variable xi . We set p = 1, k = 2ν+1, w(v) = wi1 , . . . , wi 1 for each vertex v, and, for each i ∈ {1, . . . , ν} and each j ∈ {1, . . . , occ(i)}, we define L(vij ) = {i, 2ν + 1} and L(wij ) = {i, ν + i}. Furthermore, for each h ∈ {1, . . . , µ}, we set L(uh ) = {ν + i1 , ν + i2 , ν + i3 }, assuming that the hth clause is xi1 ∨ xi2 ∨ xi3 . Finally, we set W1i = W1(ν+i) = occ(i) for each i ∈ {1, . . . , ν}, and W1(2ν+1) = µ. Then, we have the following equivalence between the solutions of I and I ′ : for each i ∈ {1, . . . , ν}, all the wij ’s take color i if and only if variable xi takes value true. Let us justify such an equivalence. First assume that we know a solution to I. For each h ∈ {1, . . . , µ}, if we denote by xi1 ∨ xi2 ∨ xi3 the hth clause, then vertex uh takes color ν + il , where il ∈ {i1 , i2 , i3 } is such that xil is the only variable of value true in the hth clause. Moreover, for each i ∈ {1, . . . , ν}, all the vij ’s take color 2ν + 1 if xi has value true, and color i otherwise (meaning that the number of vertices of color 2ν + 1 is exactly the number of occurrences of variables of value true). It can be easily checked that this yields a feasible solution to I ′ , since there are exactly µ vertices of color 2ν + 1 (as each of the µ clauses has exactly one true literal). Assume now that there is a solution to I ′ . Consider any i ∈ {1, . . . , ν}. If there is some wij that takes color ν + i, then, by construction, no vertex uh can take color ν + i. Hence, the only vertices that can take color ν + i are the wij ’s. So, the only way to have W1(ν+i) = occ(i) vertices of color ν + i is that each vertex wij takes color ν + i. This implies, in turn, that the only way to have W1i = occ(i) vertices of color i is that each vertex vij takes color i. Now, if there is some wij that takes color i, then, by construction, each vertex vij must take color 2ν + 1. So, the only way to have W1i = occ(i) vertices of color i is that each vertex wij takes color i. This implies, in turn, that the only way to have W1(ν+i) = occ(i) vertices of color ν + i is that occ(i) vertices uh take color ν + i. The only such vertices that can take color ν + i are the occ(i) ones associated with the clauses where xi appears, so all these uh ’s must take color ν + i. In short, for each i, either all the wij ’s take color i (in which case the occ(i) vertices vij take color 2ν + 1, and the occ(i) vertices uh corresponding to clauses where xi appears take color ν + i), or all the wij ’s take color ν + i (in which case no vertex uh takes color ν + i, and the occ(i) vertices vij take color i). Hence, whenever a vertex uh (associated with the clause xi1 ∨ xi2 ∨ xi3 ) takes color ν + il for some il ∈ {i1 , i2 , i3 }, then this means that all the wijl ’s take color il , and that all the wijl′ ’s take color ν + il′ for each il′ ∈ {i1 , i2 , i3 } \ {il }. In other words, there is exactly one literal of value true in each clause of I. On the positive side, we show that WeightedLocallyBoundedListColoring can be solved in (pseudo)polynomial time in complete bipartite graphs if k is fixed but p is not. Note that, in such a graph, all the vertices of a given color belong to the same side of the bipartition of G. So, if k is fixed, we can “guess” which colors are on each side by enumerating all the possibilities (there are 2k such possibilities). For each configuration (i.e., for each such assignment of colors to both sides) we enumerate, we consider each side of the bipartition independently, and for each one we check whether this configuration is feasible 17 (note that, for each side, we can consider each Vi independently, since they do not interact). We are left with a set of independent instances where p = 1, k is fixed and all the vertices are isolated vertices, so we can solve each one of them in (pseudo)polynomial time thanks to Proposition 2. This yields: Theorem 10. If k = O(1), WeightedLocallyBoundedListColoring can be solved in (pseudo)polynomial time in complete bipartite graphs. Now, we turn to the case of complete graphs G = Kn (which are cographs represented by binary cotrees containing n − 1 join nodes and no union node). Notice that, in a complete graph, all the nodes must take different colors. So, we must have k ≥ n. If k > n, then in any feasible Pp coloring one of the colors will not be used: this is possible only if we have h=1 Whc = 0 for this color c, and thus we can remove it. Hence, we know that we necessarily have k = n, and that each color will appear exactly once. In particular, since the Vi ’s are disjoint, this implies that, for each color c, we have Whc c > 0 for some hc and Wh′ c = 0 for each h′ 6= hc . We solve WeightedLocallyBoundedListColoring in this case (for arbitrary values of p and for arbitrary vertex weights) by reducing it to a matching instance in a bipartite graph H: there is one vertex in H for each vertex vi in G = Kn , and one vertex for each color c ∈ {1, . . . , n}. There is an edge between vertex vi and color c when (i) c ∈ L(vi ), (ii) w(vi ) = Whc c and (iii) vi ∈ Vhc . Due to the rule we used to define the edges of H (by linking vertices of G with “compatible” colors only), it is easy to see that we have a solution to the WeightedLocallyBoundedListColoring instance if and only if we have a perfect matching in H (a vertex vi is linked to color j in this matching if and only if it has color j in G): therefore, WeightedLocallyBoundedListColoring can be solved in polynomial time in this case (using for instance an efficient algorithm computing a maximum flow, since H is bipartite). This yields: Theorem 11. WeightedLocallyBoundedListColoring can be solved in polynomial time in complete graphs. Again, one can associate a profit function π : V × {1, . . . , k} → Z to the vertices of G. Then, we can obtain a feasible solution to WeightedLocallyBoundedListColoring with maximum profit (if any) by finding a perfect matching of maximum weight in H (or, equivalently, by solving a linear assignment instance optimally), where, for each i and each j, the weight of the edge between a vertex vi and a color j in H is π(vi , j). 5 Locally bounded list-colorings in split graphs In this section, we study the tractability of WeightedLocallyBoundedListColoring in split graphs. Recall that a split graph G is a graph whose set of vertices can be partitioned into a vertex set inducing a clique K, and a vertex set inducing a stable (or independent) set S. For the sake of simplicity, we will denote by |K| (resp. |S|) the number of vertices of K (resp. of S). 18 From Section 2, WeightedLocallyBoundedListColoring remains NPcomplete in split graphs: from Theorems 1 and 2, it is weakly NP-complete when the vertex weights are arbitrary, even if p = 1, k = 2, and each vertex can take any color, and strongly NP-complete when k is arbitrary, even if the vertex weights are polynomially bounded, p = 1, and each vertex can take any color. However, when k is fixed, it can be solved in (pseudo)polynomial time: Theorem 12. WeightedLocallyBoundedListColoring can be solved in (pseudo)polynomial time in split graphs when k = O(1). Proof. Since we are looking for a proper vertex coloring, any vertex in K cannot have more than k − 1 neighbors. Hence, K cannot contain more than k vertices. When k = O(1), one can then enumerate in constant time all the possible colorings of K (for each one, we must check that it is proper and that the color of each vertex v ∈ K is in L(v)). Then, for each such coloring, we remove K, update the value of each of the Whc ’s and the list L(v) for each vertex v ∈ S accordingly. This way, we obtain an instance in a graph consisting of isolated vertices, and where k = O(1): from Proposition 2, such an instance can be solved in (pseudo)polynomial time, which concludes the proof. The list-coloring problem in complete bipartite graphs (and thus in cographs) is strongly NP-complete from [14]. Actually, in the reduction that is used, one side of the bipartite instances can be replaced by a clique: hence, the proof also holds in split graphs (while the one of Theorem 9 does not, despite the fact that the former one partly inspired the latter one). Adding, for each initial vertex, (k − 1) new isolated vertices that can take any color, yields (NP-complete) instances of the equitable list-coloring problem. However, a straightforward adaptation of the proof of Theorem 3 yields a stronger result: Theorem 13. WeightedLocallyBoundedListColoring is strongly NPcomplete in split graphs, even when p = 1, w(v) = 1 for each vertex v, and the degree of each vertex in the part S inducing an independent set is one. Proof. Start from the reduction described in the proof of Theorem 3, and add a clique K on 3n(n+1) vertices, namely all the ui ’s and all the vij ’s. As any two of these vertices do receive different colors in the original reduction, we obtain an equivalent instance. Then, the independent set S contains all the leaves of the Sij ’s, and hence any vertex of S has degree one, which concludes the proof. Note, however, that the list-coloring problem is easy in split graphs if every vertex in S has degree at most 1: remove any v ∈ S with only one possible color, update the possible colors for each u ∈ K accordingly, use an optimal flow to color the vertices of K appropriately (as in Theorem 11), and then give to any other vertex a color not used by its unique neighbor (if such a neighbor exists). Moreover, the lists of possible colors play a major role in the above reduction, so one may wonder what happens when each vertex can take any color. When p and k are arbitrary, and w(v) = 1 for each vertex v, the following result holds: 19 Theorem 14. WeightedLocallyBoundedListColoring is strongly NPcomplete in split graphs, even when Whc ∈ {0, 1} for each h and c, and each vertex has weight 1 and can take any color. Proof. Take an instance of 3SAT [9], that consists of µ clauses of size 3 defined on ν boolean variables x1 , . . . , xν . To each variable xi , we associate two vertices yi and zi . To the jth clause, we associate three vertices uj , vj , wj , for each j. Then, we add edges in the following way, in order to obtain a split graph: the vertices yi and zi over all i induce a clique on 2ν vertices, and the vertices uj , vj , wj over all j induce an independent set on 3µ vertices. For each j ∈ {1, . . . , µ}, if we let the jth clause be x′i1 ∨ x′i2 ∨ x′i3 (where x′il ∈ {xil , x̄il } for each l ∈ {1, 2, 3}), we also add an edge between wj and yil , for each l ∈ {1, 2, 3}. Finally, we let k = 2ν and p = ν + µ, and then define Vi = {yi , zi } and Wii = Wi(ν+i) = 1 for each i ∈ {1, . . . , ν}, as well as Vν+j = {uj , vj , wj } and W(ν+j)c̄i1 = W(ν+j)c̄i2 = W(ν+j)c̄i3 = 1 for each j ∈ {1, . . . , µ}, where, for each l ∈ {1, 2, 3}, c̄il = ν + il if x′il = xil , and c̄il = il otherwise. Note that any other Whc is equal to 0, and that, for each vertex v, w(v) = 1 and L(v) ∈ {1, . . . , 2ν}. We can define the following equivalence between the solutions of these two instances: for each i, xi = true if and only if the color of yi is i. Note that, for each j ∈ {1, . . . , µ}, the vertex wj must take a color c̄il for some l ∈ {1, 2, 3} such that yil takes color cil (and hence the literal x′il is true), where cil = il if c̄il = ν + il , and cil = ν + il otherwise. This concludes the proof. Actually, we can even prove a stronger result, and get rid of the assumption that p must be arbitrary, while allowing each vertex to take any color. Theorem 15. WeightedLocallyBoundedListColoring is strongly NPcomplete in split graphs, even when p = 1, W1c ∈ {1, 4} for each color c, and each vertex has weight 1 and can take any color. Proof. We reduce from the NP-complete problem 3-DimensionalMatching, in which we are given three sets X, Y, Z of elements such that |X| = |Y | = |Z|, and a list of triples T ⊆ X × Y × Z. One wants to decide whether there exists T ′ ⊆ T covering all the elements of X, Y , and Z, and such that |T ′ | = |X|. Given an instance I of 3-DimensionalMatching, we define the following instance I ′ of WeightedLocallyBoundedListColoring. There are 3|X| + |T | vertices, that can take any color, and have weight 1: one vertex th for each triple in T , one vertex xi for each element in X, one vertex yj for each element in Y , and one vertex zl for each element in Z. Moreover, there is an edge between each th and all the other vertices, except xih , yjh and zlh , provided that the hth triple of T is composed of the ih th element of X, of the jh th element of Y , and of the lh th element of Z. In particular, we have |K| = |T | and |S| = 3|X|, and we define k = |T | and p = 1. Finally, we define the color bounds as follows: W1c = 4 for each c ∈ {1, . . . , |X|}, and W1c = 1 for each c ∈ {|X| + 1, . . . , k}. Then, it is easy to see that the following equivalence holds: for each h, the hth triple in I is in T ′ if and only if th takes a color c with W1c = 4 in I ′ . Indeed, in I ′ , each color must appear exactly once in K, as |K| = k. Moreover, 20 each vertex th ∈ K is non adjacent to exactly three vertices: hence, if th takes some color c with W1c = 4, then the three vertices non adjacent to it must take color c. The |X| such colors must cover all the vertices of S, and |X| vertices of K. Each of the k − |X| other vertices of K takes a color c′ with W1c′ = 1. Corollary 3. The capacitated coloring problem is NP-complete in split graphs. Note that, on the contrary, the bounded coloring problem is easy in split graphs [4]. We end this section by providing three last tractable special cases. The first one is obtained by assuming that |S| = O(1). In this case, we enumerate the O(k |S| ) potential colorings of S (for each one, we must check that it is proper, and that the color of each vertex v ∈ S is in L(v)). Then, for each such coloring, we remove S, update the values Whc and the lists L(v) for all the vertices v ∈ K accordingly, and we are left with an instance in a complete graph, that we can solve in polynomial time thanks to Theorem 11. The second one is obtained by assuming that the tree-width is O(1) (i.e., that |K| = O(1)) and w(v) = 1 for each vertex v (without the latter assumption, the problem is NP-complete from Theorems 1 and 2). In this case, we enumerate the O(k |K| ) potential colorings of K (for each one, we must check that it is proper, and that the color of each u ∈ K is in L(u)). Then, for each such coloring, we remove K, update the values Whc and the list L(v) of each v ∈ S accordingly, and we are left with an instance in a set of isolated vertices and where each vertex has weight 1, that we can solve efficiently thanks to Proposition 1. The third one complements Corollary 3, and is obtained by assuming that each vertex has weight 1 and can take any color, and that, for all colors c except a constant number k ′ of them, Whc = B for each h, for some common constant B (the assumption k ′ = O(1) being less restrictive than k = O(1)). Recall that it was proved in [4] that the bounded coloring problem can be solved in polynomial time in split graphs, by means of a reduction to an intermediate problem, which is then solved using a maximum flow algorithm. The construction we will use to solve this third special case (in which the number k of colors is arbitrary) also solves the bounded coloring problem, with a more direct reduction to a maximum flow problem. Namely, we prove: Theorem 16. WeightedLocallyBoundedListColoring can be solved in polynomial time in split graphs when (i) each vertex has weight 1 and can take any color, and (ii) for all colors c except a constant number k ′ of them (called the singular colors), Whc = B for each h, for some constant B. Proof. Let us consider such an instance in a split graph G, and let u1 , . . . , u|K| be the vertices of K, and v1 , . . . , v|S| the vertices of S. We start by “guessing” the singular colors that will appear in K, and  color which (unique)  for each such ′ vertex of K uses it. This can be done in O (|K| + 1)k time by a brute-force enumeration. Any other vertex of K will take an arbitrary (but different for each of them) color c such that Whc = B for each h, and some singular colors ′ may appear only in S. Also, for each h and each c, we set Whc = Whc − w(ui ) = ′ Whc − 1 if some vertex ui ∈ K ∩ Vh takes color c, and Whc = Whc otherwise. 21 Since k ≥ |K| (otherwise, there is no solution), we always end up with a proper coloring of K. Then, we define for each vj ∈ S a list of possible colors L(vj ) = {1, . . . , k} \ {c : there exists ui ∈ K adjacent to vj that takes color c}. Now, we can remove the vertices of K, and we are left with an instance of WeightedLocallyBoundedListColoring in a graph consisting of isolated vertices of weight 1 (i.e., in the subgraph of G induced by the vertices of S). The ′ color bounds are the Whc ’s, and we can solve this instance in polynomial time (as a maximum matching problem in a bipartite graph, and hence as a maximum flow problem) thanks to Proposition 1, which concludes the proof. Again, note that this generalizes the polynomial-time solvability of the case where p = 1 and k ′ = 0, i.e., of the bounded coloring problem in split graphs. Moreover, without the assumption that each vertex can take any color, the problem becomes hard, even when p = 1. Indeed, take the reduction from Theorem 15, and, for each color c such that W1c = 1, add three isolated vertices that can only take color c. This way, we obtain (NP-complete) instances where W1c = 4 for each color c, and hence where k ′ = 0. Similarly, adapting the proof of Theorem 14 shows that, without the above assumption, the problem is NP-complete when p is arbitrary and Whc = 1 for each h and each c. However, if each vertex in the clique part K of G has both a list of possible colors and an arbitrary weight, then the proof of Theorem 16 can easily be adapted to show that the problem remains tractable, provided that each vertex of S has weight 1 and is allowed to take any color. Indeed, after the “guess” part, we can give appropriate non singular colors to the remaining uncolored vertices of K by solving another matching instance (and not in a greedy way anymore), in a way similar to the one used in the proof of Theorem 11. 6 Extension to edge colorings In this last section, we extend to edge colorings all our previous results concerning (vertex) k-colorings in cographs, split graphs, and graphs of bounded tree-width. An edge coloring of a graph is an assignment of colors (integers) to its edges, in such a way that any two edges sharing a vertex take different colors. The problem we consider in this section is thus the following: WeightedLocallyBoundedListEdgeColoring Instance: A graph G = (V, E), a weight function w : E → N∗ , a partition E1 , . . . , Ep of the edge set E, a list of pk integral bounds W11 , . . . , W1k , . . . , Wp1 , Pk .P . . , Wpk such that j=1 Wij = w(Ei ) for each i ∈ {1, . . . , p} (where w(Ei ) = e:e∈Ei w(e)), and, for each edge e, a list of possible colors L(e) ⊆ {1, . . . , k}. Question: Decide whether there exists an edge coloring of G such that, for each i ∈ {1, . . . , p} and for each j ∈ {1, . . . , k}, the total weight of edges having color j in Ei is Wij , and the color of e belongs to L(e) for each edge e. 22 6.1 NP-completeness proofs for edge colorings We can prove that this problem is NP-complete by using a reduction from Partition, similar to the one in Theorem 1 (so we have p = 1, k = 2 and W11 = W12 = B). Each isolated vertex vi becomes an isolated edge ei , and we define w(ei ) = ai and L(ei ) = {1, 2} for each i ∈ {1, . . . , n}. Again, we can make this instance connected, and obtain a chain, by setting k = 3 and adding new edges of color 3 between the ei ’s. However, the problem is trivial when k = 2 and G is connected, since G has maximum degree 2 in this case, and hence G is either a path or an (even) cycle, and has only two possible edge-colorings. The main goal of this section is to prove that WeightedLocallyBoundedListEdgeColoring is tractable in the graphs considered in the previous sections, under the same assumptions as its vertex coloring counterpart. However, we shall first justify that, as in the case of vertex colorings, dropping any of these assumptions leads to NP-completeness, without looking further into other possible special cases, for the sake of brevity. The previous paragraph settles the case of arbitrary edge weights, so let us look at the other cases. When the input graph is neither a cograph nor a split graph, and has an unbounded tree-width, WeightedLocallyBoundedListEdgeColoring is NP-complete even when k = 3, p = 1, and we have both L(e) = {1, 2, 3} and w(e) = 1 for each edge e. Indeed, deciding whether the edges of a graph can be properly colored using three colors is known to be NP-complete [12]. When p is arbitrary, we need an adaptation of the proof of Theorem 5: Theorem 17. WeightedLocallyBoundedListEdgeColoring is strongly NP-complete in graphs where each connected component is either a cycle of length four or a single edge, even when k = 2, w(e) = 1 for each edge e, and each edge can take any color. Proof. Given an instance of MonotoneOneInThreeSAT, we construct such a graph. More precisely, for each variable xi , there are occ(i) isolated edges (denoted by eji ) and occ(i) vertex-disjoint cycles of length four, where occ(i) is the number of occurrences of xi in the set of clauses. For each i ∈ {1, . . . , ν} and each j ∈ {1, . . . , occ(i)}, let us denote by aji , bji , cji , dji the four edges of the jth of these cycles, and assume for instance that aji and cji are vertex-disjoint, and that so are bji and dji . Moreover, each edge has weight 1 and can take any of the two colors. Intuitively, in the jth cycle associated with variable xi , the edge bji will be used to interact with the j + 1th one, the edge cji will be used to interact with the j − 1th one, the edge dji will only be useful to obtain a cograph, and the color of the edge aji will reflect the value of xi (true or false). Let us define formally the partition that we use, and the associated color bounds. For each i ∈ {1, . . . , ν} and each j ∈ {1, . . . , occ(i)}, we set Wh1 = occ(i) 1 , ci } otherwise, } if j < occ(i) and Vh = {bi Wh2 = 1, as well as Vh = {bji , cj+1 i Pi−1 where h = 2 q=1 occ(q) + j. Moreover, for each i ∈ {1, . . . , ν} and each j ∈ {1, . . . , occ(i)}, we set Vh′ = {dji , eji } and Wh′ 1 = Wh′ 2 = 1, where h′ = Pi−1 2 q=1 occ(q)+ occ(i)+ j. Note that k = 2, and hence, in the jth cycle of length 23 four associated with variable xi , edges aji and cji must take the same color, and edges bji and dji must take the same color. For each i, by the way we defined the Vh ’s and Vh′ ’s so far, all the aji ’s must take the same color (1 or 2). We end the reduction by defining one set of the partition for each clause: if the lth clause consists of the j1 th occurrence of variable xi1 , of the j2 th occurrence of then we define Vh′′ = variable xi2 , and of the j3 th occurrence of variable xi3 , P ν {aji11 , aji22 , aji33 }, Wh′′ 1 = 1 and Wh′′ 2 = 2, where h′′ = 2 q=1 occ(q) + l. This Pν implies that p = 2 q=1 occ(q) + µ, and this yields the following equivalence between the solutions of the two instances: for each i, xi = true if and only if aji has color 1 for each j, which concludes the proof. When the number of colors is arbitrary, WeightedLocallyBoundedListEdgeColoring is strongly NP-complete, as shown by an easy reduction from 3−Partition, similar to the one in Theorem 2 (so we have p = 1, k = n, and W1c = B for each c ∈ {1, . . . , n}). Each isolated vertex vi becomes an isolated edge ei , and we define w(ei ) = ai and L(ei ) = {1, . . . , k} for each i ∈ {1, . . . , 3n}. Note that the graphs used in the reductions described above consist of vertexdisjoint cycles of length four (i.e., complete bipartite graphs with two vertices on each side) and/or of isolated edges. In other words, they are both cographs and graphs of tree-width at most 2. Moreover, we can easily turn the instances used in the case where the number k of colors is arbitrary into split graphs. To see it, start from the above reduction, but define k = n + 3n(3n−1) 2 instead of k = n. Then, choose one endpoint of each ei , i ∈ {1, . . . , 3n}, and edges (i.e., a clique) between these 3n vertices (this yields a split add 3n(3n−1) 2 graph). These edges will be denoted by f1 , f2 , . . . , and can take any color. We define w(fj ) = 1 and W1(n+j) = 1 for each j ∈ {1, . . . , 3n(3n−1) }. As we can 2 assume without loss of generality that ai > 1 for each i (simply multiply B and each ai by two in the original instance of 3−Partition, in order to obtain such an equivalent instance), this implies that each color c ≥ n+1 will be used by one and only one edge fj , which shows the equivalence between the two reductions. Hence, from now on, we shall assume that k is fixed, as otherwise WeightedLocallyBoundedListEdgeColoring is NP-complete. In the following subsection, we give efficient algorithms to solve this problem when k = O(1). 6.2 Efficient algorithms for edge colorings when k = O(1) We start by considering graphs of bounded tree-width. We assume that such a graph G is given, together with a tree-decomposition T = (I, F ) of G having minimum width, and that p is fixed (as otherwise WeightedLocallyBoundedListEdgeColoring is NP-complete from Theorem 17). We then define the following function, before showing how to compute it for each node type: • g(i, ci , ω11 , . . . , ω1k , . . . , ωp1 , . . . , ωpk ) = true if there is a list-edge-coloring of G[Ti ] where each edge e ∈ Yi has color ci (e) and where the total weight of edges of G[Ti ] having color c in Eh is ωhc , and false otherwise. 24 If i is a forget node. Let j denote its child such that Xj = Xi ∪ {v}: _ g(j, cj , ω11 , . . . , ωpk ) g(i, ci , ω11 , . . . , ωpk ) = cj :(cj (e)=ci (e) ∀e∈Yi )∧(cj (uv)∈L(uv) ∀u∈Xi :uv∈Yj ) If i is an introduce node. Let j denote its child such that Xj = Xi \ {v}. We denote by e1 , . . . , ed the edges incident to v in Yi (i.e., the edges in Yi \ Yj ), and assume that eh ∈ Eφ(h) for each h ∈ {1, . . . , d}. ^ g(i, ci , ω11 , . . . , ωpk ) = E g(j, cj , ω11 , . . . , ωφ(1)ci (e1 ) −w(e1 ), . . . , ωφ(d)ci (ed ) −w(ed ), . . . , ωpk ) V V where E = (cj (e) = ci (e) ∀e ∈ Yj ) (ci (e) ∈ L(e) ∀e ∈ {e1 , . . . , ed }) (ci (e) 6= ci (f ) ∀e ∈ {e1 , . . . , ed }, ∀f ∈ Yi : e 6= f and both edges share a common vertex) V (ωφ(h)ci (eh ) ≥ w(eh ) ∀h ∈ {1, . . . , d}). i If i is a join node. Let j and l denote its two children, and let whc be the total weight of edges e ∈ Yi ∩ Eh such that ci (e) = c, for each h and c. _  i i g(i, ci , ω11 , . . . , ωpk ) = −qpk )∧g(l, ci , q11 , . . . , qpk ) g(j, ci , ω11 +w11 −q11 , . . . , ωpk +wpk (q11 ,...,qpk )∈Qi i where Qi = {(q11 , . . . , qpk ) : whc ≤ qhc ≤ ωhc ∀h ∀c}. If i is a leaf node. In this case, we just have to check that the coloring function ci provides a valid locally bounded list-edge-coloring of G[Xi ]. ^ g(i, ci , ω11 , . . . , ωpk ) = E (w({e ∈ Yi ∩ Eh : ci (e) = c}) = ωhc ∀h ∀c) where E = (ci (e) ∈ L(e) ∀e ∈ Yi ) Yi , u2 v ∈ Yi , u1 6= u2 ). V (ci (u1 v) 6= ci (u2 v) ∀u1 , u2 , v ∈ Xi : u1 v ∈ Root value. Again, the answer true or false for the initial instance is obtained at the root r by computing the following value: _ g(r, cr , W11 , . . . , Wpk ) cr :cr (e)∈L(e) ∀e∈Yr The proof of correctness is similar to the one given in Lemma 2, and the analysis of the running time is similar to the one given just before stating Theorem 7. Therefore, they are both omitted. This yields the following result: Theorem 18. In graphs of bounded tree-width, WeightedLocallyBoundedListEdgeColoring can be solved: • in pseudopolynomial time when p and k are fixed, • in polynomial time when (i) p and k are fixed, and (ii) all edge weights are polynomially bounded. 25 Again, one can associate a profit function to the edges of G, and then modify slightly this dynamic programming algorithm, in order to compute a feasible weighted and locally bounded list-edge-coloring of maximum profit (if any). We then turn our attention towards cographs, and assume that p is fixed (as otherwise an NP-completeness result holds, from Theorem 17). Since we are looking for a proper edge coloring, the maximum degree of such a graph G is bounded by the number k of colors. As a cograph contains no induced path on four vertices, the diameter of any of its connected components is at most two. In other words, any vertex v has at most k neighbors, each of its neighbors has at most k − 1 other neighbors, and there can be no other vertex in the connected component v belongs to. This implies that any connected component contains at most 1 + k + k(k − 1) = O(k 2 ) vertices, and hence O(k 4 ) edges. Therefore, when k = O(1), each connected component has O(1) possible edge colorings. For each one of them, one can easily check whether it defines a proper listcoloring of the component. Moreover, as p and k are fixed, these colorings can be combined one by one, by using an efficient dynamic programming algorithm similar to the one outlined at the beginning of Section 2, that needs to store a (pseudo)polynomial amount of information, namely O((maxh,c Whc )kp ). Finally, let us study the case of split graphs. Again, since we are looking for a proper edge coloring, the maximum degree of such a graph G is bounded by the number k of colors, and isolated vertices play no role in this case (they can be removed). Hence, each vertex in K, the clique part of G, cannot have more than k neighbors, which implies that K cannot contain more than k + 1 vertices. As any edge of G is incident to one of these vertices, and as each of these vertices cannot have more than k incident edges, the graph G contains O(k 2 ) edges. Therefore, when k = O(1) (and hence O(k 2 ) = O(1)), WeightedLocallyBoundedListEdgeColoring is trivial in split graphs. 7 Conclusion and open problems We conclude by giving an overview of all our results concerning WeightedLocallyBoundedListColoring in cographs, split graphs, and graphs of bounded tree-width. The three following tables summarize these results, each one being devoted to one of these classes of graphs (here, NPC means NPcomplete, P means solvable in polynomial time, ”?” means arbitrary, and tw(G) denotes the tree-width of the input graph G). 26 tw(G) tw(G) = 0 tw(G) = 0 tw(G) = 0 tw(G) = 0 tw(G) = 1 tw(G) = 1 tw(G) = O(1) w(v) ? w(v) = poly(n) ∀v ∈ V w(v) = 1 ∀v ∈ V w(v) = poly(n) ∀v ∈ V w(v) = 1 ∀v ∈ V w(v) = 1 ∀v ∈ V w(v) = poly(n) ∀v ∈ V w(v) ? w(v) = poly(n) ∀v ∈ V w(v) = 1 ∀v ∈ V w(v) = 1 ∀v ∈ V w(v) = poly(n) ∀v ∈ V w(v) = 1 ∀v ∈ V w(v) = poly(n) ∀v ∈ V ? w(v) ? w(v) = poly(n) ∀v ∈ V w(v) = 1 ∀v ∈ V w(v) = 1 ∀v ∈ V w(v) = 1 ∀v ∈ V w(v) = 1 ∀v ∈ V w(v) = poly(n) ∀v ∈ V k k ? ? k k ? k ? =2 =2 = O(1) = O(1) k k=2 ? ? ? ? ? k = O(1) GRAPHS OF BOUNDED TREE-WIDTH k p List-coloring Result k=2 p=1 NO NPC (in the weak sense) from Theorem 1 ? p=1 NO NPC from Theorem 2 ? ? YES P from Proposition 1 k = O(1) ? YES P from Proposition 2 ? p=1 YES NPC from [3, 7] and Theorems 3 and 4 k=2 ? NO NPC from Theorems 5 and 6 k = O(1) p = O(1) YES P from Theorem 7 p p=1 p=1 p=1 ? p = O(1) p=1 ? ? p p=1 p=1 p=1 ? p=1 ? ? COGRAPHS List-coloring Result NO NPC (in the weak sense) from Theorem 1 NO NPC from Theorem 2 YES NPC from Corollary 1 and [4] (without list-coloring) NO NPC from Corollary 2 YES P from Theorem 8 YES NPC in complete bipartite graphs from Theorem 9 YES P in complete bipartite graphs from Theorem 10 YES P in complete graphs from Theorem 11 SPLIT List-coloring NO NO YES NO NO NO YES GRAPHS Result NPC (in the weak sense) from Theorem 1 NPC from Theorem 2 NPC from [14] and Theorem 13 NPC from Theorem 14 (even if Whc ∈ {0, 1} for each h and c) NPC from Theorem 15 P if O(1) singular colors from Theorem 16 P from Theorem 12 The following table summarizes results concerning arbitrary graphs: w(v) w(v) = poly(n) ∀v ∈ V w(v) = 1 ∀v ∈ V k k=2 k=3 p p = O(1) p=1 List-coloring YES NO Result P from Section 2 NPC (since 3-coloring is) Again, we highlight that, in Section 6, similar results for edge colorings (including some specific hardness proofs, such as the one of Theorem 17) are provided in cographs, split graphs, and graphs of bounded tree-width. 27 One may actually wonder how WeightedLocallyBoundedListColoring behaves in bipartite graphs, which we did not mention yet. The instances used in the proofs of Theorems 1 to 6 being such graphs, the only relevant question would be: is this problem polynomial-time solvable in bipartite graphs when k = O(1), p = O(1), and all the vertex weights are polynomially bounded? Unfortunately, the answer is no. Indeed, Bodlaender and Jansen proved in [4, Theorem 5.1] that the equitable coloring problem is strongly NP-complete in bipartite graphs (this theorem was stated only for the bounded coloring problem, but the proof actually holds in this case as well, without any modification), even with only k = 3 colors, and so is WeightedLocallyBoundedListColoring when p = 1, k = 3, and any vertex has weight 1 and can take any color. One may also be interested in FPT algorithms for WeightedLocallyBoundedListColoring with polynomially bounded vertex weights. Unfortunately, the results in Theorems 7, 8 and 12 are best possible, as this problem is W[1]-hard with respect to k in graphs of tree-width 0 (which are also cographs and split graphs) if p = 1, even with no list-coloring. To see it, take any instance of unary bin packing with l bins, which is W[1]-hard with respect to l [13], and build an equivalent instance of the target problem: each of the n items becomes a vertex (of the same weight wP i ), the k = l color bounds equal the common bin n capacity B, and we add kB − i=1 wi more isolated vertices of weight 1. Finally, the main open question we would like to mention as worth studying is a direct consequence of our results in Sections 2 and 3. Indeed, Theorems 1 to 7 leave as open the case where p = 1, the tree-width is O(1), and each vertex has weight 1 and can take any color. In other words, what is the complexity of the capacitated coloring problem (as defined in [5]) in graphs of bounded tree-width when k is arbitrary? Bodlaender and Fomin managed to prove in [3] that the bounded coloring problem is tractable in this case, a question that was open for a long time before they finally settled it. The capacitated coloring problem is more general, and should probably be harder (even in this case), but no one has been able to prove it so far. Actually, there is a reduction from the capacitated coloring problem to the bounded (or equitable) coloring problem: to see it, take any instance of the former problem with color bounds n1 ≥ n2 P ≥ . . . ≥ nk , k add isolated vertices so that the total number of vertices reaches i=1 ni , and then add k sets of vertices inducing independent sets, in such a way that the ith of these sets contains n1 − ni + 1 isolated vertices and any two vertices in any two different sets are linked by an edge. We obtain an instance of the latter problem by setting the common color bound to n1 + 1, as it can be checked that all the vertices in the ith set must take color i. However, in general, this reduction preserves neither the property of having bounded tree-width nor the one of being a split graph (although it does preserve the one of being a cograph). Also recall that the capacitated coloring problem is NP-complete in cographs and split graphs when k is arbitrary, from [4] and Corollary 3 respectively. 28 References [1] B. Baker and E. Coffman Jr. Mutual exclusion scheduling. Theoretical Computer Science 162 (1996) 225–243. [2] C. Bentz and C. Picouleau. Locally bounded k-colorings of trees. RAIRORO 43 (2009) 27–34. [3] H.L. Bodlaender and F.V. Fomin. Equitable colorings of bounded treewidth graphs. Theoretical Computer Science 349 (2005) 22–30. [4] H.L. Bodlaender and K. Jansen. Restrictions of graph partition problems: Part I. Theoretical Computer Science 148 (1995) 93–109. [5] F. Bonomo, S. Mattia and G. Oriolo. Bounded coloring of co-comparability graphs and the pickup and delivery tour combination problem. Theoretical Computer Science 412 (2011) 6261–6268. [6] A. Brandstädt, V.B. Le and J.P. Spinrad. Graph Classes: A Survey. SIAM Monographs on Discrete Mathematics and Applications 2, Society for Industrial and Applied Mathematics, Philadelphia (1999). [7] M. Dror, G. Finke, S. Gravier and W. Kubiak. On the complexity of a restricted list-coloring problem. Discrete Mathematics 195 (1999) 103–109. [8] J. Fiala, P.A. Golovach and J. Kratochvil. Parameterized complexity of coloring problems: Treewidth versus vertex cover. Theoretical Computer Science 412 (2011) 2513–2523. [9] M.R. Garey and D. S. Johnson. Computers and Intractability, a Guide to the Theory of NP-Completeness. Ed. Freeman, New York (1979). [10] S. Gravier, D. Kobler and W. Kubiak. Complexity of list coloring problems with a fixed total number of colors. Discrete Applied Mathematics 117 (2002) 65–79. [11] P. Hansen, A. Hertz and J. Kuplinsky. Bounded vertex colorings of graphs. Discrete Mathematics 111 (1993) 305–312. [12] I. Holyer. The NP-Completeness of Edge-Colouring. SIAM Journal on Computing 10 (1981) 718–720. [13] K. Jansen, S. Kratsch, D. Marx and I. Schlotter. Bin packing with fixed number of bins revisited. Journal of Computer and System Sciences 79 (2013) 39–49. [14] K. Jansen and P. Scheffler. Generalized coloring for tree-like graphs. Discrete Applied Mathematics 75 (1997) 135–155. [15] T. Kloks. Treewidth: Computations and Approximations. Lecture Notes in Computer Science 842 (1994). 29
8cs.DS
Quantifying topological invariants of neuronal morphologies Lida Kanari,1 Pawel Dlotko,2 Martina Scolamiero,3 Ran Levi,4 Julian Shillcock,1 Kathryn Hess,3 and Henry Markram1 1 Blue Brain Project, EPFL∗ DataShape, INRIA Saclay, Ile-de-France 3 Laboratory for Topology and Neuroscience at the Brain Mind Institute, EPFL 4 Institute of Mathematics, University of Aberdeen (Dated: March 29, 2016) arXiv:1603.08432v1 [q-bio.NC] 28 Mar 2016 2 Nervous systems are characterized by neurons displaying a diversity of morphological shapes. Traditionally, different shapes have been qualitatively described based on visual inspection and quantitatively described based on morphometric parameters. Neither process provides a solid foundation for categorizing the various morphologies, a problem that is important in many fields. We propose a stable topological measure as a standardized descriptor for any tree-like morphology, which encodes its skeletal branching anatomy. More specifically it is a barcode of the branching tree as determined by a spherical filtration centered at the root or neuronal soma. This Topological Morphology Descriptor (TMD) allows for the discrimination of groups of random and neuronal trees at linear computational cost. Keywords: topological analysis, neuronal, structures, trees, morphologies The analysis of complex tree structures, such as neurons, branched polymers [1], viscous fingering [2] and fractal trees [3, 4], is important for understanding many physical and biological processes. Yet an efficient method for quantitatively analysing the morphology of such trees has been difficult to find. Biological systems provide many examples of complex trees. The nervous system is one of the most complex biological systems known, whose fundamental units, neurons, are sophisticated information processing cells possessing highly ramified arborizations [5]. The structure and size of neuronal trees determines the input sources to a neuron and the range of target outputs and is thought to reflect their involvement in different computational tasks [6–8]. In order to understand brain functionality, it is fundamental to understand how neuronal shape determines neuronal function [9]. As a result, much effort has been devoted to grouping neurons into distinguishable morphological classes [10–13], a categorization process that is important in many fields [14]. Neurons come in a variety of shapes with different branching patterns (e.g., frequency of branching, branching angles, branching length, overall extent of the branches, etc). Classifying these different morphologies has traditionally focused on visually distinguishing the shapes as observed under a microscope [15]. This method is inadequate as it is subject to large variations between experts studying the morphologies [10] and is made even more difficult by the presence of an enormous variety of morphological types [11]. An objective method of discriminating between neuronal morphologies could advance progress in generating a parts-list of neurons in the nervous system. For this reason, experts now generate a digital version of the cells structure - a 3D reconstructed model of the neuron [16] that can be studied computationally. The reconstructed morphology is encoded as a set of points in R3 along each branch and edges connect- ing pairs of points. The reconstruction forms a mathematical tree representing the neurons morphology. Figure 1: Illustration of the separation of similar tree structures into distinct groups, using topological analysis. The colored pie segments show six distinct tree types: three neuronal types (upper half) and three artificial ones (lower half). The thick blue lines show that our topological analysis can reliably separate similar-looking trees into groups. It is accurate both for artificially-generated trees and neuronal morphologies. The dashed green lines show that classification using an improper set of user-selected features cannot distinguish the correct groups. In general, the properties of geometric trees, described by a set of points, have been rigorously studied in two extreme cases: in the limit of the full complexity of the trees [17], where the entire set of points is used, and in the opposite limit of a feature space [10, 18], where a (typically small) number of selected morphometrics (i.e., statistical features of the branching pattern) [19, 20] are 2 extracted from the digital morphology and form the input for standard classification algorithms. Topological data analysis (TDA) has been shown to reliably identify geometric objects that are built out of well-understood pieces, such as spheres, cylinders and tori, based on a point cloud sampled from the object [17]. It suffers, however, from the deficiency that reliable classification of complex geometric tree structures by standard TDA methods, such as Rips complexes [21], requires thousands of sampled points, which is expensive in terms of both computational complexity and memory requirements. Thus, feature extraction is the only currently feasible solution to establishing a more quantitative approach to morphologies[13, 22, 23]. While this approach has been efficiently used in image recognition [24], the wide diversity of neuronal shapes, even for cells identified as of the same type, has made it difficult to isolate an optimal set of features that reliably describe all neuronal shapes. Indeed, alternative sets of morphometrics result in different classifications [10], as illustrated in Fig 1, because the statistical features commonly overlap even across markedly different morphological types. This in turn is because traditional feature extraction results in significant loss of information as the dimensionality of the data is reduced. As a result, neither of these methods is suitable for the study of complex tree structures that describe physical or biological objects. The algorithm we propose here overcomes both of these limitations by constructing a Topological Morphology Descriptor (TMD) of a tree embedded in R3 . A distance inspired by persistent homology [17] is defined to distinguish trees, because alternative methods for computing distances between trees, such as the edit distance [25], the blastneuron distance [20] and the functional distortion distance [26], are in general computationally expensive (see SI, Distances between trees). Our algorithm preserves the overall shape of the tree while reducing the computational cost by discarding the local fluctuations. It takes as input the set of key functional points in a tree [7]: the branch points, i.e., the nodes with more than one child, and the leaves, i.e., the nodes with no children and transforms them into a multi-set of intervals on the real line known as a persistence barcode [17], Fig 2b. Each interval encodes the lifetime of each connected component in the underlying structure, delimited by the points at which a branch is first detected (birth) and when it connects to a larger subtree (death). This information can be equivalently represented in a persistence diagram[17], Fig 2c in which the birth-death times are points in the real plane with the advantage that this structure simplifies the mathematical analysis of the data. Our method is applicable to any tree-like structure and we demonstrate its generality by first applying it to a collection of artificial random trees, (see SI, Random Trees), and to groups of neuronal trees (see SI, Data provenance). Our results show that the TMD of tree shapes is highly effective for reliably and efficiently distinguishing random and neuronal trees (Fig 1). Figure 2: Application of topological analysis to a neuronal tree (A) showing the largest persistent component (red). The persistence barcode (B) represents each component as a horizontal line whose endpoints mark its birth and death in units that depend on the choice of the function f used for the ordering of the nodes of the tree. In our case, it is radial distance of the nodes from the soma, so the units are microns. The largest component is again shown in red together with its birth (I) and death (II). This barcode can be equivalently represented as points in a persistence diagram (C) where the birth (I) and death (II) of a component are the X and Y coordinates of a point respectively (in red). The diagonal line is a guide to the eye and marks points whose birth and death are simultaneous. Note that the persistence barcode (B) has an arbitrary ordering of the components along the Y axis, which prevents the calculation of an average barcode. The corresponding persistence diagram (C), however, has physical coordinates on both axes, allowing an average persistence diagram to be defined. Methods The operation F to extract the barcode D(T ) of an embedded tree T is described in Algorithm 1. Let T be a rooted, and therefore oriented, tree [27], embedded in R3 . Note that the operation described here is in fact generalizable to trees embedded in any metric space. We denote by N := B ∪ L the set of nodes of T , which is the union of the set of branch points B and the set of leaves L. In the case of a neuron, the root R is the point representing the soma. Each node n ∈ N has references to its parent, i.e., the first node on the path towards the root and to its children. Nodes with the same parent are called siblings. Let f be a real-valued function defined on the set of nodes N . For the purpose of this study f is the radial distance from the root R. Another possible choice would be the path distance from R. Based on the function f , we construct a real-valued function v, defined on the set of nodes N as the largest value of the function f on the leaves of the subtree starting at n. An ordering of the siblings can be defined based on v: if n1 , n2 ∈ N , are siblings and v(n1 ) < v(n2 ), then n1 is younger than n2 . 3 The algorithm is initialized by setting the value of v(l), l ∈ L equal to the value of f (l). Following the path of each leaf l ∈ L towards the root R, all but the oldest (with respect to v) siblings are killed at each branch point. If siblings are of the same age it is equivalent to terminate either one of them. For each killed component one interval (birth-death) is added to the persistence barcode (Fig 2). The oldest sibling survives until it merges with a larger component at a subsequent branch point. This operation is applied iteratively to all the nodes until the root is reached. At this point only one component, the oldest one, is still alive. Algorithm 1 F : extracting persistence barcode from tree Require: T with R, B, L, f : R ∪ L → R and v 1: D(T ) : empty list to contain pairs of real numbers 2: A ← L . A : set of active nodes 3: for every l ∈ L do 4: v(l) = f (l) 5: while R 6∈ A do 6: for l in A do 7: p : parent of l 8: C : children of p 9: if ∀n ∈ C, n ∈ A then 10: cm : randomly choose one of {c | v(c) = maxc0 (v(c0 )) for c0 ∈ C} 11: Add p to A 12: for ci in C do 13: Remove ci from A 14: if ci 6= cm then 15: Add (v(ci ), f (p)) to D(T ) 16: v(p) ← v(cm ) 17: Add (v(R), f (R)) to D(T ) 18: Return D(T ) R and as such it is invariant under rotations about the root and rigid translations of the tree in R3 . The most common topological metric that is used for the comparison of persistence diagrams is the bottleneck distance [21], denoted dB . The bottleneck distance is the infimum of the infinity norm L∞ (maximum distance) between matching points over all possible matchings of two persistence diagrams [21]. We prove that the operation F : T 7→ D(T ) is stable with respect to standard topological distances, such as the Bottleneck distance (see SI, Bottleneck and Wasserstein stability of the barcode construction). For  modifications in the tree T with respect to Hausdorff distance (see SI, Distances between metric spaces), the persistence diagram D(T ) is not modified more than O(L), as long as function f is L-Lipschitz. Note that this requirement is always satisfied (with L = 1) by the radial distance. As a result, the method is robust with respect to small perturbations in the positions of the nodes and the addition/ deletion of small branches. 0 0 250 500 Distance between tree pairs When all the branches are outgoing, i.e., the radial distance of the origin is smaller than the radial distance of the terminal point of the branch, the operation F is equivalent to a filtration of concentric spheres of decreasing radii, centered at R (Fig 2). In this case, the birth time of a component is the supremum of the radii of the spheres that do not contain the entire component. The death time is the infimum of the radii of the spheres that contain the branch point at which the component merges with a longer one. The computational complexity of this operation is linear in the number of nodes. Note that the if statement in line 9 of the algorithm is critical for the linear complexity. The number of currently active children is saved at each parent node to avoid quadratic complexity. This process results in a set of intervals on the real line, each of which represents the lifetime of one component of the tree. The operation F that associates a persistence barcode D(T ) to a tree T is invariant under rotations and translations, as long as the function f is also. For the purposes of this paper, f is the radial distance from 1000 2000 3000 4000 750 1000 1250 1500 1750 2000 2250 Hierarchical clustering based on dendrogram Figure 3: Application of topological analysis to a set of artificial trees generated using a well-defined stochastic process. Three sets of trees are shown in which each set differs from the others only in the tree depth. Each group contains 20 individuals of which only 4 are shown for clarity. Each individual is generated using the same tree parameters but a different random number seed. The TMD separates each group accurately using the distance metric between barcodes of the trees described in the main text. This is seen in the separation of colors in the colormap, and is supported by the clear distinction into three groups resulting from a simple hierarchical clustering algorithm [34]. However, none of the standard topological distances between persistence diagrams is appropriate for the grouping of neuronal trees. The bottleneck distance as well as distances based on the bottleneck, such as the persistence distortion distance [28] (see SI, Distances between trees) cannot distinguish diagrams that differ on the short components, which are nevertheless important for the distinction of neuronal morphologies. We therefore define an alternative distance dBar in the space of the barcodes that we use to classify branching morphologies. For each barcode we generate a density profile as follows: ∀x ∈ R the value of the histogram is 4 AI B C 120 D 1500 1500 100 80 1000 1000 500 500 60 40 II y 100um x 20 0 0 200 400 600 800 1000 1200 1400 y IV 100um x 300 250 200 200 150 150 100 100 100 50 50 0 0 0 50 100 150 200 250 300 150 200 250 300 350 400 300 20 200 200 10 100 0 100 200 300 400 500 40 0 100 150 200 250 300 350 150 100 50 300 400 500 250 200 150 150 100 100 0 200 400 600 800 100 150 200 250 300 350 0 1000 1000 800 800 600 600 400 400 0 100 200 300 400 500 50 0 50 50 100 150 200 250 300 350 200 200 0 0 200 0 50 200 250 0 0 100 300 50 100 150 200 250 300 350 400 100 0 300 20 50 400 300 60 y 100 um x 100 400 80 V 0 50 30 0 200 400 600 800 1000 1200 1400 1600 50 0 40 100 y 100um x 0 800 1000 1200 1400 1600 250 150 III 600 300 300 x 400 350 200 100 um 0 200 350 250 y 0 0 200 400 600 800 1000 0 200 400 600 800 1000 Figure 4: Topological comparison of neurons from different animal species. Each row corresponds to a species: (I) cat, (II) dragonfly, (III) drosophila, (IV) mouse and (IV) rat. Trees from several exemplar cells for each species are shown in the first column (A). Representative persistence barcodes for the cells in A are shown in the second column (B). The structural differences of the trees are clearly evident in these barcodes. II, III and V have clusters of short components, clearly distinct from the largest component, while I and IV have bars of a quasi-continuous distribution of decreasing lengths. Also, barcodes III, and V show empty regions between dense regions of bars, indicating the existence of two clusters in the morphologies, while barcodes I and IV are dense overall. The persistent image for each representative barcode in B and its superimposed persistence diagram are shown in the third column (C). By combining the persistence diagrams in (C) for several tress we can define an average persistent image (D) in order to study the qualitative differences between sets of morphologies. The trees in the first row (cat) are more tightly grouped than those in the second row (dragonfly), and two clusters are visible in the dragonfly trees. Considering rows 1 and 4, the extension of the elliptical peak perpendicular to the diagonal line reflects the variance in the length-scale mentioned earlier for a single cell’s barcode. Note that the trees, barcodes, and persistent images are not shown to the same scale: see the scale bar in each case. the number of intervals that contain x, i.e., the number of components alive at that point. The distance between two barcodes D(T1 ) and D(T2 ) is defined as the sum of the differences between the density profiles of the barcodes. This distance is not stable with respect to Hausdorff distance, but it is the only distance we are aware of that succeeds in capturing the differences between distinct neuronal persistence barcodes. Finally, the persistence diagram can be converted into a persistent image, as introduced in [29]. This method is based on the discretization of a sum of Gaussian kernels [30], centered at the points of the persistence diagram. This discretization can generate a matrix of pixel values, encoding the persistence diagram in a vector, called the persistent image. Machine learning tools, such as decision trees and support vector machines can then be applied to this vectorization for the classification of the persistence diagrams. Results We demonstrate the discriminative power of the TMD by applying it to three examples of increasing complexity. First, we applied it to artificial random trees that provide a well-defined test case to explore the method’s performance because they are generated by a stochastic algorithm (described in Random Trees) and therefore have properties that can be precisely modified. Next, we analyzed two datasets of more biological relevance: neurons from different species, downloaded from [31], and distinct types of trees obtained from several morphological types of rat cortical pyramidal cells (see SI, Data provenance). This last example is interesting because, although there is biological support for their separation into distinct groups, no consistent objective classification has previously been performed. The persistence diagrams of trees from all these example datasets can be classified correctly with high accuracy, indicating good performance of the algorithm across non-trivial examples. 5 UTPC A STPC TTPC1 TTPC2 B C 1.0 Expert Randomized 0.8 Accuracy (fraction) Mathematical random trees are defined by a set of parameters that constrain their shape: the tree depth Td , the branch length Bl , the bifurcation angle Ba and the degree of randomness Dr (see SI, Generation of random trees). We defined a control group as a set of trees generated with predefined parameters (Td = 5, Bl = 10, Ba = π/4, Dr = 10%) but independent random seeds. Each parameter was varied individually to generate groups of trees that differed from the control group in only one property. For all trees we extracted the persistence barcode using Algorithm 1. A comparison of the distances dBar between each tree’s barcode and the barcodes of the trees in every groups constitutes one trial. The trial is successful if the minimum barcode distance occurs for a tree and the group generated with the same parameters. The TMD method accurately separates groups of random trees that have different tree depth as seen in Fig 3, with an accuracy of 99% ± 2% (see SI, Random trees results). The distance matrix in Fig 3 shows three distinct groups and a simple hierarchical clustering algorithm [32] correctly assigns each tree to its group. The TMDs effectiveness is demonstrated by its accuracy in assigning individual trees to their originating group. The TMD distinguishes random trees generated by varying the parameters Td , Ba , Bl , Dr with an accuracy of 99%, 94%, 99% and 77% respectively. Further results are given in the Random trees results. Next, the algorithm is used to quantify differences between neuronal morphologies. Neurons that serve distinct functional purposes exhibit unique branching patterns. The morphologies of neuronal trees from different species, as seen in Fig 4, can be classified into different groups. For the purposes of this study, we used cat, drosophila, dragonfly, mouse and rat neurons. The qualitative differences between the neuronal tree types are evident from the individual geometrical tree shapes (Fig 4A) as well as the extracted barcodes (Fig 4B). From the superposition of persistence diagrams for a set of trees, we constructed the persistent images of the group (Fig 4D). Note that the persistence barcode has an arbitrary ordering of the trees components along the Y-axis, which prevents the calculation of an average barcode. The corresponding persistence diagram, and the associated persistent image, have physical coordinates on both axes, allowing an average persistence diagram to be defined. This representation is useful for quantifying the differences between tree types. It is trivial to identify in Fig 4D areas of high branching density at different radial distances from the soma by the position of the clusters. Since density is usually correlated with connection probability, we can identify the anatomical parts of the trees that are important for the functionality and the connectivity patterns of different cell types. Finally, we applied the TMD to a case for which it is difficult for a non-expert to distinguish the morpholo- 0.6 0.4 0.2 0.0 UTPC - ΤTPC2 UTPC - ΤTPC1 STPC - ΤTPC2 UTPC - STPC STPC - ΤTPC1 ΤTPC1 - ΤTPC2 Figure 5: Comparison of the topological analysis of apical dendrite trees extracted from several types of rat neuron. Four cell types are shown in (A): UTPC, STPC, TTPC1, TTPC2 (left to right). The morphological differences between these cell types are subtle, but the persistent images (B) clearly reveal them, particularly the presence of two clusters in the TTPC1 and TTPC2 cell types. From these persistent images we train a decision tree classifier on the expert-assigned groups of cells. Then we perform binary classification on trees taken from each pair of groups (C). The blue bars show that the performance of the classifier using the expert-assigned groups in order of decreasing accuracy (98%, 93%, 82%, 77%, 74%, 64%). The red bars are obtained from a random assignment of trees to groups and are significantly worse than the expert assignment. This shows that the persistent images objectively support the expert’s classification. gies. While pyramidal cell (PC) morphologies (Fig 5A) of the rat appear superficially similar, the persistent images (5B) reveal fundamental morphological differences between the four neuronal types, related to the existence and the shape of the apical tuft. The apical tuft of PCs is known to play a key role in the integration of neuronal inputs through their synapses in higher cortical layers, and is therefore a key indicator for the functional role of the cell. Therefore, the TMD can successfully discriminate between these functionally distinct neuronal types. Discussion The morphological diversity of neurons supports the complex information processing capabilities of the nervous system. A major challenge in neuroscience has therefore been to objectively classify the neurons by shape. Long-standing techniques, such as feature extraction, have so far been unable to reliably distinguish cell types, without expert manual interaction, in a reasonable computational time. We have introduced here a Topological Morphology Descriptor derived from principles of persistent topology that retains enough topological information about trees to be able to distinguish them in linear computational time. Its ability to efficiently separate neurons into distinct groups with high accuracy provides objective support for expert-assigned neuronal classes, and may lead to a more comprehensive understanding of the relation between neuronal morphology and function. 6 This technique can be applied to any rooted tree equipped with a Lipschitz function defined on its nodes. Further biological examples include botanic trees and the roots of plants. The method is not restricted to trees in R3 , however, but can be generalized to any subset T of a metric space M , with a base-point R. A persistence barcode can then be extracted using a filtration by concentric spheres in M centered at R, enabling us to efficiently study the shape of complex multidimensional objects. Another generalization is to use a different function f , such as the path distance from the root, which should serve to reveal shape characteristics that are independent of the radial distance and thus not captured by the current approach. While the static neuronal structures presented in this paper are biologically interesting themselves, our method can also be used to track the morphological evolution of trees. The topological study of the growth of an embedded tree could be addressed through Multidimensional Persistence [33], as the spherical filtration identifying relevant topological features of the tree could be enriched with a second filtration representing temporal evolution. This application could be useful in agriculture to study growing roots [34] and trees, and would also extend the method to neurons in the developing brain. Acknowledgments LK and JCS were supported by Blue Brain Project. P.D. and R.L. were supported in part by the Blue Brain Project and by the start-up grant of KH. Partial support for P.D. has been provided by the Advanced Grant of the European Research Council GUDHI (Geometric Understanding in Higher Dimensions). MS was supported by the SNF NCCR ”Synapsy”. Author contributions LK and KH conceived the study. LK developed the topological algorithm with the contribution of PD. PD, KH contributed topological ideas used in the analysis. HM suggested the biological datasets analyzed. RL, MS and JCS gave conceptual advice. LK, PD, MS, JCS, KH and HM wrote the main paper. All authors discussed the results and commented on the manuscript at all stages. The authors declare no competing financial interests. 7 Supplementary information Demonstration of the algorithm g h d e R f 1 2 3 45 6 a b c Figure 6: Illustration of the Algorithm 1. The idea of Algorithm 1 is presented in Figure 6. The root, denoted by R, is shown in red, while the other nodes of the tree are labelled a − h. During the initialization of the algorithm, the leaves (a, c, e, g, h) are inserted into the list of alive nodes, A. The algorithm then iterates over the members of A, as follows. If the node a is the first element of the list A, then the first step of the algorithm begins by finding its parent, b. Since both a and c are in A, the algorithm then computes cm = c and therefore v(b) = 3. The interval [1, 2] is added to the list of intervals D(T ) representing the lifetime of the node a. Nodes a and c are removed from A, and b is added to A. If the next vertex in the list A is e, the algorithm begins the next step by finding its parent, d. This node is not processed further at this stage, however, since its sibling, f , is not in A. The next node to be processed is therefore g. Both children g and h of its parent f are in A. The oldest component is cm = h and therefore v(f ) = 6. The interval [5, 4] is added in D(T ) to represent the lifetime of node g. The node f is added to A, and both g and h are removed from A. The list of alive components then consists of b, e, and f . The node b cannot be processed since its sibling d is not in A. The next node to be processed is therefore e, whose parent d has all of its children in A. In this case, cm = f , and v(d) = v(f ) = 6. The interval [4, 3] is added to D(T ), while nodes e and f are removed from A, and d is added to A. The next node in A is b, whose parent is R. Since both children of R are now in A, we compute cm = d, v(R) = v(d) = 6, and add the interval [3, 2] to D(T ). The root R is added to A, and b and d are removed from the list. Since now only the root R is alive, the while loop in the algorithm terminates. The last step adds the interval [6, 0] to D(T ). 8 Definition of Distances Distances between metric spaces In order to analyze the stability of Algorithm 1, we need to define a notion of distance between trees embedded in a metric space, as well as of distance between persistence diagrams. We will use the standard Hausdorff distance, defined as follows, to measure the distance between trees. Given T1 , T2 , trees in embedded in a metric space with metric d, the Hausdorff distance between T1 and T2 is given by the formula: dH (T1 , T2 ) = max(supx∈T1 infy∈T2 d(x, y)), supx∈T2 infy∈T1 d(x, y)). Distances between persistence diagrams Below we recall various representations of persistence diagrams and the possible notions of distance between them. We also provide a reference to software that computes the distances considered if such exist. All of the metrics summarized below can be applied directly to the output of Algorithm 1. The most classical distances used in topological data analysis are the bottleneck and Wasserstein distances. Given two persistence diagrams D1 and D2 , we may assume without loss of generality that they both contain the points from the diagonal with infinite multiplicity. We construct a matching (i.e., a bijection) φ : D1 → D2 and define two numbers Bφ = sup d(x, φ(x)) x∈D1 and Wφp = ( X 1 d(x, φ(x))p ) p , x∈D1 where d is the standard euclidean distance in R2 . Note that Bφ is simply the longest distance that φ shifts a point in D1 , while (Wφp )p is a sum of p-th powers of lengths of the line segments joining x and φ(x), for all x. The infimum of Bφ over all possible matchings is the bottleneck distance between D1 and D2 . The infimum of Wφp over all possible matchings is the p−Wasserstein distance between D1 and D2 . Given two persistence diagrams D1 and D2 , their bottleneck distance will be denoted by dB (D1 , D2 ) and their p−Wasserstein distance by Wp (D1 , D2 ). One implementation of these distances can be found in [35] and a faster approximation in [36]. Please follow the link in the paper for the implementation. A persistence diagram can also be represented by a persistence landscape, i.e., a piecewise linear function f : R × N → R. Given two persistence landscapes, we can compute the distance between them in Lp space [37]. The implementation is described in [38]. One can also represent persistence diagrams by persistence images, which were introduced in [29]. The idea is to place a Gaussian kernel at every point of the diagram and to discretize the distribution obtained in a pixel-based image. It is then straightforward to compute a distance between two persistence images, using common image-recognition techniques. This representation is used in the classification of morphological types of neurons in the experimental section of this paper. We are not aware of a publicly available implementation of this approach. An implementation is provided with the software of this paper. Distances between trees A classic metric to compare trees, the edit distance [25], is based on the transformation of one tree T1 into another T2 by a sequence of operations (deletion and insertion of vertices), each of which has a non-negative cost. The edit distance [25] between T1 and T2 is defined to be the supremum of the total cost of all possible transformations from 9 T1 to T2 . However, the edit distance is not relevant to our problem, since it does not involve geometric information about the tree structure and is known to be NP-complete [39]. A metric that is useful for distinguishing neuronal trees was proposed in [20]. A set of morphological features is extracted from the trees, and the initial estimation of the distance between them is defined by the distance between the extracted features. An alignment algorithm is then applied to pairs of trees in order to identify local similarities. The local alignment requires the comparison of all pairs of branches, making the computation very expensive. This method is designed for the efficient matching of trees with highly similar structures, but the high variability within the groups of rat cortical neurons does not allow similar trees to be grouped together by local alignment, since local structures are often altered. Another important class of distances between trees are the distances between merge trees introduced in [40] and [41], which are applied to merge trees of sublevel sets of functions. For a function f : X → R, where X is a metric space, the sublevel set at level a ∈ R is {x ∈ X|f (x) ≤ a}. The differences captured by merge trees are considerably more subtle than the differences captured by the persistent homology of the function’s sublevel sets. The authors of [40] and [41] provide examples of pairs of simple merge trees T and T 0 that have the same persistence diagrams, but that are a nonzero distance apart. It is clear that in this particular case, Algorithm 1 would provide the persistent homology of the sublevel sets of the function. Therefore, by rescaling T and T 0 , the difference between the distances used in those papers and the distances used in the current paper can get arbitrarily large. Another relevant metric, the persistence distortion distance [28] between two graphs G1 and G2 , is expressed in terms of the shortest paths from every v1 ∈ G1 and from every v2 ∈ G2 . Given this shortest path metric, the persistence distortion is defined as the minimal bottleneck distance between the zero-dimensional persistent homology modules of the superlevel sets of the distance functions P (G1 , v1 ) and P (G2 , v2 ). The persistence diagrams obtained in the process of computing the persistence distortion distance are conceptually very close to the diagrams we get from Algorithm 1. In our case, we obtain a significant computational advantage from working with rooted trees, since there is always a unique path between every pair of vertices. Moreover the logical choice of initial vertices v1 and v2 from which to compute shortest paths is to take the root of the trees considered. In this case, the persistence diagram arising when computing the persistence distortion distance is the one we would get from Algorithm 1 when the function v is the path distance from the root. The computational cost of the distortion distance is considerable in the general case, but linear in our case. However, since it is based on the bottleneck distance, it suffers from that metric’s limitations, i.e., the shortest components, which are important for the neuronal morphologies, are not taken into account. The code to compute persistence distortion distance is available here [42]. Functional distortion distance was first introduced in [26] as a distance to compare Reeb graphs. Since this distance, more generally, compares graphs equipped with a function defined on edges, we will consider it for neural tree comparison. Let T be a tree, viewed as a CW-complex, with a real-valued function g : T → R that is monotone on edges. Following [26] we give T a metric space structure. For any two points u, v in T (not necessarily vertices), let λ be the unique path between u and v. The height of λ is defined as height(λ) := maxx∈λ g(x) − minx∈λ g(x). In particular, if e = (v0 , v1 ) is an edge in T , since g is monotone on edges, height(e) = |g(v0 ) − g(v1 )|. The distance between two points u and v in T is defined as df (u, v) := height(λ). The function df defines a metric on T whenever there is no path of constant function value between two points u 6= v in G. Otherwise, df is a pseudo-metric. Let T and T 0 be two trees equipped with real-valued functions f : T → R and f 0 : T 0 → R. For every pair of continuous functions φ : T → T 0 and ψ : T 0 → T , let G(φ, ψ) := {(x, φ(x))|x ∈ T } ∪ {(ψ(y), y)|y ∈ T 0 } and 1 D(φ, ψ) := sup(x,y),(x̃,ỹ)∈G(φ,ψ) |df (x, x̃) − df 0 (y, ỹ)|. 2 The functional distortion distance between T and T 0 is defined by: dF D (T, T 0 ) = infφ,ψ max{D(φ, ψ), ||f − f 0 ◦ φ||∞ , ||f ◦ ψ − f 0 ||∞ }. In Section 3.2 we will use the properties of functional distortion distance to prove that Algorithm 1 is robust with respect to specific types of perturbation in the input data. 10 Proof of Stability Bottleneck and Wasserstein stability of the barcode construction We prove now that the operation F to transform a tree T into a persistence diagram D(T ) is robust under  small perturbations. In the space of metric trees, we use the Hausdorff distance defined in Section 8. Since our input consists only of leaves and branch points, and their respective connectivity, when perturbing the tree, we perturb only those points. In the space of persistence diagrams we consider the bottleneck and Wasserstein distances defined in Section Distances between persistence diagrams. As a result, the method is robust with respect to noisy data. Given a tree T , with a real-valued function f defined on the set of nodes N , we construct a real-valued function v, defined on the set of nodes N as the largest value of the function f on the leaves of the subtree starting at n (see Methods). In order to prove the robustness of the method, we require the function v to be Lipschitz with a constant L. For the purpose of this study f is the radial distance from the root R and as a result v is 1−Lipschitz. The proof of stability is carried out in three steps. First we prove stability in the case when the branch point’s position can modified by . We then consider the removal or addition of short branches. At the end, we unite these results to prove the main theorem. h' g e h' ϵ h ϵ e d h g d R R f f 1 2 3 45 1 2 6 3 45 6 a a b b c c Figure 7: Perturbation of node position: h’ has been moved  from h. Figure 8: Addition of a small branch: h’ has been added  close to h. We denote the perturbed tree as T 0 , with the same function v defined on its nodes. We consider the following types of perturbations, for which dH (T, T 0 ) ≤ . (a) Perturbation of branch points’ positions The location of a branch point is modified by at most , as illustrated in Fig 7 for node h. The branching structure of the tree is preserved. In this case, the value of function v of the node h of tree T will be modified by at most L in the node h0 of tree T 0 , since v is L-Lipschitz. Since the connectivity has not been modified, the transformation of T 0 to the persistence diagram D(T 0 ), using Algorithm [1], will result in the same number of points. The difference between D(T ) and D(T 0 ) will be in the position of the point(s) that depend on the modification of the node h0 and on v(h0 ). Assuming that the correspondence between the nodes of T and T 0 is known, all the points in D(T 0 ) that have not been modified are matched with the corresponding points in D(T ). Since |v(h) − v(h0 )| < L the distance between the corresponding point(s) that have been modified in the persistence diagram will be at most L. In this case the bottleneck distance between D(T ) and D(T 0 ) is bounded by L and the Wasserstein distance in bounded by cL, where c is the number of points influenced by the modification of the node h0 . We can generalize this concept to a perturbation of any number of nodes of the tree T . The maximum difference on the persistence diagrams can occur if both nodes of a component are moved by  in opposite directions. In 11 this case the distance between the two points is at most 2L. Therefore, the bottleneck distance between D(T ) 1 and D(T 0 ) is bounded by 2L. The p−th Wasserstein distance between D(T ) and D(T 0 ) is bounded by 2n p L, where n is the number of leaves in T , which is equal to the number of points in D(T ). Note that if the function v depends on the location of the root, for example in our case v is based on the radial distance v from R, we need to double the Lipschitz constant in the analysis above. When a node of the tree is moved by at most , the value of the function v from the root may change by 2L (respectively 2 in our case) since the position of both the root and any other node of a tree will be perturbed by at most . Therefore the distance between the root and the node will change by no more than 2 and the value of the function v on the node will change by no more than 2L. (b) Small branches are added to the tree. In this case, branches of at most length  are added to the tree. The locations of the other nodes of the tree T 0 remain fixed. In the simplest case, a single branch h0 is added to the component that contains node h, as presented in Fig 8. This new branch will result in one more point in the corresponding persistence diagram. We create a matching between the points of D(T ) and D(T 0 ). Assuming that the correspondence of the nodes of T and T 0 is known, all the points in D(T 0 ) that have not been modified are matched with the corresponding points in D(T ). The new point will represent the termination of either the h0 or the h component, depending on which node the value of function v is smaller. In both cases, the new point will be matched with the diagonal at a distance less than L and therefore the bottleneck and the Wasserstein distances between D(T ) and D(T 0 ) are bounded by L. We can generalize this concept to the addition of m  small branches to the tree T . In this case, m new points will be added to the persistence diagram of T 0 . The rest of the points in D(T 0 ) will be matched with the points of the corresponding branches in D(T ). Using the previous argument, the bottleneck distance between T and 1 T 0 is bounded by L. The p−th Wasserstein distance between T and T 0 is bounded by 2m p L, where m is the number of new at most  long branches in T 0 that are not in T . (c) Small branches are removed from the tree. Another perturbation is to remove small branches of length at most  from the tree. This case is the inverse of the previous case. As a result, stability in this case is implied by the stability of the previous case. (d) Generalization of previous results. Finally, we discuss the most general case, where we can have perturbation of branch points’ positions as well as addition and deletion of small branches. From the previous cases, using the triangle inequality, it is easy to show that the bottleneck distance between T and T 0 is bounded by 3L and the p−th Wasserstein distance 1 between T and T 0 is bounded by 3(m + n) p L, where m is the number of branches of length  that are added or removed, and n is the number of branch points in T . Note that the persistence intervals introduced by Algorithm 1 also give rise to a decomposition of a tree into a distinct collection of branches. As an example, let us consider a tree in Figure 9.a. The point R denotes the root, a and b are leaves while c is the only branch point in the tree. Function v depends on f which is the radial distance from the root R. In Figure 9.a we assume that b is slightly further away from R than a. In this case, when Algorithm 1 processes the node c, the branch originating at a will terminate at c and the branch originating at b will continue. This gives rise to a branch decomposition illustrated on the right corner. However, if the leaves a and b are slightly perturbed, as in Figure 9.b, the resulting branch decomposition in the lower right corner will be very different from the previous case even though the persistence intervals will be very close for those two cases. This example illustrates that persistence diagrams are stable, but the corresponding branch decompositions are not. Since Algorithm 1 extracts the persistence diagram of a tree, the instability of the branch decomposition do not concern us for the current paper. Stability proof based on functional distortion distance In this section we will use properties of the functional distortion distance dF D to prove that sampling errors in the input neural tree will not result in a significant difference in the associated persistence diagram. The space of persistence diagrams is equipped with the bottleneck distance dB . 12 R R F c a F c c a b R R b a b (a) c a b (b) Figure 9: An illustration of a region of a tree corresponding to persistence intervals computed by Algorithm 1. Recall that a neural tree is represented by a tree T in R3 with a selected vertex R called the root and a function f : T → R associating to each vertex in T its radial distance to the root. In Algorithm 1 the tree structure is represented by the set of branch points, the set of leaves and their pairwise connectivity. In this section instead, in order to use the results from [26], the tree is viewed as a regular one-dimensional CW-complex. Intuitively we have glued an homeomorphic image of the interval [0, 1] between any two adjacent vertices in the data structure for Algorithm 1. In accordance with [26] we denote by Dg0 (Tf ) the 0-th persistence diagram of the sublevel set filtration of f : T → R. The persistence diagram of the superset filtration of f is equivalent to Dg0 (T−f ). Let T and T 0 be trees, equipped with functions f : T → R and f 0 : T 0 → R. Theorem 4.2 of [26] states that: dB (Dg0 (Tf ), Dg0 (Tf0 0 )) ≤ dF D (T, T 0 ). The persistence diagram D(T ) obtained from Algorithm 1 is equivalent to considering the disjoint union of the persistence diagrams Dg0 (Tf ) and Dg0 (T−f ). In particular Dg0 (Tf ) corresponds to points above the diagonal in D(T ) and identifies leaves that are closer to the root than their corresponding branch point. Points in Dg0 (T−f ) are instead under the diagonal in D(f ) and represent leaves that are farther from the root than their corresponding branch point. The diagonal of a persistence diagram is considered with infinite multiplicity, which implies that a matching between D(T ) and D(T 0 ) corresponds to a pair: a matching of Dg0 (Tf ) with Dg0 (Tf0 0 ) and a matching of 0 Dg0 (T−f ) with Dg0 (T−f 0 ). It is therefore sufficient to consider the optimal matchings on the upper diagonal and lower diagonal part of the diagram given by Theorem 4.2 to conclude that: dB (D(T ), D(T 0 )) ≤ dF D (T, T 0 ). We will now use the property above to prove that the bottleneck distance between the persistence diagrams of trees in R3 is stable with respect to adding branches of length at most  and perturbing the position of a branching point by . Lemma .1. Let T be a tree in R3 and T 0 a tree obtained by adding an edge of length at most  to T , i.e. T 0 = T ∪v e. If f and f 0 are functions on T and T 0 that associate to each point of a rooted tree its euclidean distance to the root, then dF D (T, T 0 ) < . Proof. Let φ : T ,→ T 0 be the natural inclusion of T in T 0 , and let ψ : T 0 → T be defined by: ( x if x ∈ /e ψ(x) = v if x ∈ e. By definition f = f ◦ φ. By the triangle inequality, ||f ◦ ψ − f ||∞ ≤ . Furthermore since the length of the edge e is at most , D(φ, ψ) ≤ /2. Therefore dF D (T, T 0 ) ≤ . The proof above can immediately be generalised to the case in which finitely many edges, each of length at most , are added to the tree. 13 Lemma .2. Let h : R3 → R3 be a homeomorphism such that the euclidean distance between x and h(x) is at most  for any x in R3 . If f and f 0 are functions on T and h(T ) that associate to each point of a rooted tree its euclidean distance to the root, then dF D (T, h(T )) < . Proof. Since the euclidean distance between x and h(x) is at most , the triangle inequality implies that ||f −f 0 ◦h||∞ ≤ . Let g : R3 → R3 be the inverse to h. As before since the distance between x and g(x) is at most , the triangle inequality implies that ||f ◦ g − f 0 ||∞ ≤ . By the definition of h and g, it is also true that D(h, g) < . It follows that  dF D T, h(T ) ≤ . We have now proved that adding branches of length at most  and moving the points of the tree T in R3 by  results in a tree which is -close to T in functional distortion distance. In this section we also showed that the bottleneck distance between persistence diagrams of trees is stable with respect to functional distortion distance. Combining these two facts guarantees that under the mentioned -perturbations, the bottleneck distance between the persistence diagrams is at most . Random Trees bifurcation angle degree of randomness th branch leng tree depth Generation of random trees Figure 10: Definition of growth parameters for artificial random trees. The random trees that were used for testing the algorithm were generated with software developed within the Blue Brain Project (BBP). Each tree consists of branches, i.e., paths between two branch points, which are generated based on a random walk in R3 . The position of the walk at each step is given as a weighted sum of a predefined direction dn and a simple random walk Xn : Xn+1 = ws · ((1 − Dr ) · dn + Dr · Xn ) , where ws is the step size, and Dr defines the randomness of each step. For Dr = 0 the branch is a straight line, while for Dr = 1 the branch is a simple random walk (SRW). The number of steps is given by the preselected branch length Bl . Once the number of steps is reached, the tree bifurcates, i.e., two new branches are created. The angle between the initial points of the branches is defined by the bifurcation angle Ba . The tree generated this way is binary, i.e., every leaf has the same depth, since new branches are added at every branch point until the preselected tree depth Td (i.e., the number of nodes from a leaf to the root node) is reached. The total number of branches in the tree is then 2Td − 1. For example, the tree in Fig 10 has Td = 4 and consists of 24 − 1 = 15 branches. This set of parameters {Td , Bl , Ba , Dr } defines the global properties of the tree. Random trees that are generated with the same set of parameters share common morphometric properties, but have unique spatial structures, due to the stochastic component of the growth. This allowed us to check the effectiveness of the algorithm at identifying sets of trees that have been generated with the same input parameters {Td , Bl , Ba , Dr } and that differ only in the random seed. 14 Random trees results We defined a control group as a set of trees generated with predefined parameters (Td = 5, Bl = 10, Ba = π/4, Dr = 10%) but independent random seeds. Then, we varied each parameter individually to generate groups of trees that differed from the control group in only one property. For all trees we extracted the persistence barcode using Algorithm 1. A comparison of the distances dBar between each tree’s barcode and the barcodes of the trees in every groups constitutes one trial. The trial is successful if the minimum barcode distance occurs for a tree and the group generated with the same parameters. The overall accuracy of the TMD to separate groups of trees generated with different values for each of the described parameters is given in the following table: Td : (4, 6, 8) Ba : ( π4 , π2 , π) Bl : (5, 10, 30) Dr : (0.01, 0.10, 0.90) 99 ± 2% 94 ± 2% 99 ± 1% 77 ± 9% Software availability The software used for the extraction of persistence barcodes will be made available upon publication, under the BSD-3 license, more details in https://opensource.org/licenses/BSD-3-Clause. The code will be accessible in https://github.com/BlueBrain/Topaz. Data provenance The artificial random trees used in Fig 1 and Fig 3 were generated by software developed in BBP and can be made available upon request. The biological morphologies used in Fig 1, Fig 2 and Fig 5 are provided from LNMC, EPFL [43]. The biological morphologies used in Fig 4 were downloaded from http://neuromorpho.org/ . In particular, cat neurons were provided by [44], dragonfly neurons by [45], drosophila neurons by [46], mouse neurons by [47] and rat neurons by [43]. ∗ [email protected] [1] Z. Alexandrowicz. Growth and shape of branched polymers, aggregates and percolating clusters. Physics Letters A, 109(4):169 – 173, 1985. [2] O. Agam, E. Bettelheim, P. Wiegmann, and A. Zabrodin. Viscous fingering and the shape of an electronic droplet in the quantum hall regime. May 2002. [3] B Mandelbrot, WH Freeman, and San Francisco co. The fractal geometry of nature. Earth Surface Processes and Landforms, 8(4):406, 1983. [4] J.F. Garst. Alternate realities - mathematical models of nature and man. Journal of Computational Chemistry, 11(2):266– 267, 1990. [5] Y.N. Jan and L.Y. Jan. Branching out: mechanisms of dendritic arborization. Nature reviews. Neuroscience, 11:316–28, May 2010. [6] H. Cuntz, A. Borst, and I. Segev. Optimization principles of dendritic structure. Theor Biol Med Model, 4(21), 2007. [7] M Ferrante, M Migliore, and GA Ascoli. Functional impact of dendritic branch point morphology. J Neurosci, 33(5):2156– 65, Jan 2013. [8] R.A. Silver. Neuronal arithmetic. Nat Rev Neurosci, 11(7):474–489, Jul 2010. [9] C.J. Honey, J.P. Thivierge, and O. Sporns. Can structure predict function in the human brain? NeuroImage, 52(3):766 – 776, 2010. Computational Models of the Brain. [10] J. DeFelipe et al. New insights into the classification and nomenclature of cortical gabaergic interneurons. Nature, 2013. [11] H Markram, M Toledo-Rodriguez, Y Wang, A Gupta, G Silberberg, and C Wu. Interneurons of the neocortical inhibitory system. Nat Rev Neurosci, 5(10):793–807, Oct 2004. [12] M Halavi, KA Hamilton, R Parekh, and GA Ascoli. Digital reconstructions of neuronal morphology: Three decades of research trends. Frontiers in Neuroscience, 6(49), 2012. [13] PING The Petilla Interneuron Nomenclature Group. Petilla terminology: nomenclature of features of gabaergic interneurons of the cerebral cortex. Nat Rev Neurosci, 9(7), Jul 2008. [14] M. J. Lyons, J. Budynek, and S. Akamatsu. Automatic classification of single facial images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 21(12):1357–1362, Dec 1999. [15] M Masseroli, A Bollea, and G Forloni. Quantitative morphology and shape classification of neurons by computerized image analysis. Computer Methods and Programs in Biomedicine, 41(2):89 – 99, 1993. [16] J Dieter. Accurate Reconstruction of Neuronal Morphology. In Computational Neuroscience, Frontiers in Neuroscience. CRC Press, Nov 2000. [17] G. Carlsson. Topology and data. Bull. Amer. Math. Soc., 46:255–308, 2009. 15 [18] P Gomez-Gil, M Ramirez-Cortes, J Gonzalez-Bernal, AG Pedrero, CI Prieto-Castro, D Valencia, R Lobato, and JE Alonso. A feature extraction method based on morphological operators for automatic classification of leukocytes. In Artificial Intelligence, 2008. MICAI ’08. Seventh Mexican International Conference on, pages 227–232, Oct 2008. [19] A.V. Blackman, S. Grabuschnig, R. Legenstein, and P.J. Sjstrm. A comparison of manual neuronal reconstruction from biocytin histology or 2-photon imaging: morphometry and computer modeling. Frontiers in Neuroanatomy, 8(65), 2014. [20] Y. Wan, F. Long, .L Qu, H. Xiao, M. Hawrylycz, E.W. Myers, and H/ Peng. Blastneuron for automated comparison, retrieval and clustering of 3d neuron morphologies. Neuroinformatics, 13(4):487–499, 2015. [21] H Edelsbrunner and JL Harer. Persistent homologya survey, volume 453, pages 257–282. American Mathematical Society, Providence, Rhode Island, 2008. [22] R Scorcioni, S Polavaram, and GA Ascoli. L-measure: a web-accessible tool for the analysis, comparison, and search of digital reconstructions of neuronal morphologies. Nat Protoc, 3:866–76, 2008. [23] C. Ling, M.L.Hendrickson, and R.E. Kalil. Morphology, classification, and distribution of the projection neurons in the dorsal lateral geniculate nucleus of the rat. PLoS ONE, 7:e49161, 11 2012. [24] T. Schurer. An experimental comparison of different feature extraction and classification methods for telephone speech. In Interactive Voice Technology for Telecommunications Applications, 1994., Second IEEE Workshop on, pages 93–96, Sep 1994. [25] P. Bille. A survey on tree edit distance and related problems. Theoretical Computer Science, 2005. [26] U. Bauer, X. Ge, and .Y Wang. Measuring distance between reeb graphs. In Proceedings of the Thirtieth Annual Symposium on Computational Geometry, SOCG’14. ACM, 2014. [27] D.E. Knuth. The art of computer programming, volume 2: Seminumerical algorithms, addison-wesley. Reading, Massachusetts, 1998. [28] T.K. Dey, D. Shi, and Y. Wang. Comparing graphs via persistence distortion. SOCG, 2015. [29] S. Chepushtanova, T. Emerson, E.M. Hanson, M. Kirby, F.C. Motta, R. Neville, C. Peterson, P.D. Shipman, and L. Ziegelmeier. Persistence images: An alternative persistent homology representation. CoRR, abs/1507.06217, 2015. [30] DW Scott. Kernel Density Estimators, pages 125–193. John Wiley and Sons, Inc., 2008. [31] G.A. Ascoli, D.E. Donohue, and M. Halavi. Neuromorpho.org: a central resource for neuronal morphologies. J Neurosci., 27(35):9247–51, 2007. [32] J.H. Ward Jr. Hierarchical grouping to optimize an objective function. Amer. Statist. Assoc., 58(301):236–244, 1963. [33] G Carlsson and A Zomorodian. The theory of multidimensional persistence. Discrete & Computational Geometry, 42(1):71– 93, 2009. [34] H Wang, J Siopongco, LJ Wade, and A Yamauchi. Fractal analysis on root systems of rice plants in response to drought stress. Environmental and Experimental Botany, 65(23):338 – 344, 2009. [35] D. Morozov. Dionysus, http://mrzv.org/software/dionysus/, 2016. [36] M Kerber, D Morozov, and A Nigmetov. Geometry Helps to Compare Persistence Diagrams, chapter 8, pages 103–112. [37] P Bubenik. Statistical topological data analysis using persistence landscapes. J. Mach. Learn. Res., 16(1):77–102, 1 2015. [38] P. Bubenik and P. Dlotko. A persistence landscapes toolbox for topological statistics. Journal of Symbolic Computations, abs/1501.00179, 2015. [39] A. Feragen. Structural, Syntactic, and Statistical Pattern Recognition: Joint IAPR International Workshop, SSPR&SPR 2012, Hiroshima, Japan, November 7-9, 2012. Proceedings, chapter Complexity of Computing Distances between Geometric Trees, pages 89–97. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012. [40] K. Beketayev, D. Yeliussizov, D. Morozov, G.H. Weber, and B. Hamann. Topological Methods in Data Analysis and Visualization III: Theory, Algorithms, and Applications, chapter Measuring the Distance Between Merge Trees, pages 151–165. Springer International Publishing, Cham, 2014. [41] D Morozov, K Beketayev, and GH Weber. Interleaving distance between merge trees. Discrete and Computational Geometry, 49:22–45, 2013. [42] T.K. Dey and D. Shi. ”PersistenceDistortion” software, 2016. [43] S Romand, Y Wang, M Toledo-Rodriguez, and H Markram. Morphological development of thick-tufted layer v pyramidal cells in the rat somatosensory cortex. Frontiers in neuroanatomy, 5:5, 2011. [44] PK Rose, T Jones, R Nirula, and T Corneil. Innervation of motoneurons based on dendritic orientation. Journal of Neurophysiology, 73(3):1319–1322, 1995. [45] PT Gonzalez-Bellido, H Peng, J Yang, AP Georgopoulos, and RM Olberg. Eight pairs of descending visual neurons in the dragonfly give wing motor centers accurate population vector of prey direction. Proceedings of the National Academy of Sciences, 110(2):696–701. [46] AS Chiang, CY Lin, CC Chuang, HM Chang, CH Hsieh, CW Yeh, CT Shih, JJg Wu, GT Wang, YC Chen, CC Wu, GY Chen, YT Ching, PC Lee, CY Lin, HH Lin, CC Wu, HW Hsu, YA Huang, JY Chen, HJ Chiang, CF Lu, RF Ni, CY Yeh, and JK Hwang. Three-dimensional reconstruction of brain-wide wiring networks in drosophila at single-cell resolution. Current Biology, 21(1):1 – 11, 2011. [47] TC Badea and J Nathans. Morphologies of mouse retinal ganglion cells expressing transcription factors brn3a, brn3b, and brn3c: Analysis of wild type and mutant cells using genetically-directed sparse labeling. Vision Research, 51(2):269 – 279, 2011.
8cs.DS
On completeness of logical relations for monadic types ⋆ Sławomir Lasota1 ⋆⋆ David Nowak2 Yu Zhang3 ⋆ ⋆ ⋆ arXiv:cs/0612106v1 [cs.LO] 21 Dec 2006 1 Institute of Informatics, Warsaw University, Warszawa, Poland 2 Research Center for Information Security, National Institute of Advanced Industrial Science and Technology, Tokyo, Japan 3 Project Everest, INRIA Sophia-Antipolis, France Abstract. Software security can be ensured by specifying and verifying security properties of software using formal methods with strong theoretical bases. In particular, programs can be modeled in the framework of lambda-calculi, and interesting properties can be expressed formally by contextual equivalence (a.k.a. observational equivalence). Furthermore, imperative features, which exist in most real-life software, can be nicely expressed in the so-called computational lambdacalculus. Contextual equivalence is difficult to prove directly, but we can often use logical relations as a tool to establish it in lambda-calculi. We have already defined logical relations for the computational lambda-calculus in previous work. We devote this paper to the study of their completeness w.r.t. contextual equivalence in the computational lambda-calculus. 1 Introduction Contextual equivalence. Two programs are contextually equivalent (a.k.a. observationally equivalent) if they have the same observable behavior, i.e. an outsider cannot distinguish them. Interesting properties of programs can be expressed using the notion of contextual equivalence. For example, to prove that a program does not leak a secret, such as the secret key used by an ATM to communicate with the bank, it is sufficient to prove that if we change the secret, the observable behavior will not change [18,3,19]: whatever experiment a customer makes with the ATM, he or she cannot guess information about the secret key by observing the reaction of the ATM. Another example is to specify functional properties by contextual equivalence. For example, if sorted is a function which checks that a list is sorted and sort is a function which sorts a list, then, for all list l, you want the expression sorted(sort(l)) to be contextually equivalent to the expression true. Finally, in the context of parameterized verification, contextual equivalence allows the verification for all instantiations of the parameter to be reduced to the ⋆ Partially supported by the RNTL project Prouvé, the ACI Sécurité Informatique Rossignol, the ACI jeunes chercheurs “Sécurité informatique, protocoles cryptographiques et détection d’intrusions”, and the ACI Cryptologie “PSI-Robuste”. ⋆⋆ Partially supported by the Polish K BN grant No. 4 T11C 042 25 and by the European Community Research Training Network Games. This work was performed in part during the author’s stay at LSV. ⋆⋆⋆ This work was mainly done when the author was a PhD student under an MENRT grant on ACI Cryptologie funding, École Doctorale Sciences Pratiques (Cachan). verification for a finite number of instantiations (See e.g. [6] where logical relations are one of the essential ingredients). Logical relations. While contextual equivalence is difficult to prove directly because of the universal quantification over contexts, logical relations [15,8] are powerful tools that allow us to deduce contextual equivalence in typed λ-calculi. With the aid of the so-called Basic Lemma, one can easily prove that logical relations are sound w.r.t. contextual equivalence. However, completeness of logical relations is much more difficult to achieve: usually we can only show the completeness of logical relations for types up to first order. On the other hand, the computational λ-calculus [10] has proved useful to define various notions of computations on top of the λ-calculus: partial computations, exceptions, state transformers, continuations and non-determinism in particular. Moggi’s insight is based on categorical semantics: while categorical models of the standard λcalculus are cartesian closed categories (CCCs), the computational λ-calculus requires CCCs with a strong monad. Logical relations for monadic types, which are particularly introduced in Moggi’s language, can be derived by the construction defined in [2] where soundness of logical relations is guaranteed. However, monadic types introduce new difficulties. In particular, contextual equivalence becomes subtler due to the different semantics of different monads: equivalent programs in one monad are not necessarily equivalent in another! This accordingly makes completeness of logical relations more difficult to achieve in the computational λ-calculus. In particular the usual proofs of completeness up to first order do not go through. Contributions. We propose in this paper a notion of contextual equivalence for the computational λ-calculus. Logical relations for this language are defined according to the general derivation in [2]. We then explore the completeness and we prove that for the partial computation monad, the exception monad and the state transformer monad, logical relations are still complete up to first-order types. In the case of the non-determinism monad, we need to restrict ourselves to a subset of first-order types. As a corollary, we prove that strong bisimulation is complete w.r.t. contextual equivalence in a λ-calculus with monadic non-determinism. Not like previous work on using logical relations to study contextual equivalence in models with computational effects [16,13,11], most of which focus on computations with local states, our work in this paper is based on a more general framework for describing computations, namely the computational λ-calculus. In particular, very different forms of computations like continuations and non-determinism are studied, not just those for local states. Plan. The rest of this paper is structured as follows: we devote Section 2 to preliminaries, by introducing basic knowledge of logical relations in a simple version of typed λ-calculus; then from Section 3 on, we move to the computational λ-calculus and we rest on a set-theoretical model. In particular, Section 3.4 sketches out the proof scheme of completeness of logical relations for monadic types and shows the difficulty of getting a general proof; we then switch to case studies and we explore, in Section 4, the completeness in the computational λ-calculus for a list of common monads: partial computations, exceptions, state transformers, continuations and the non-determinism; the last section consists of a discussion on related work and perspectives. 2 Logical relations for the simply typed λ-calculus 2.1 The simply typed λ-calculus λ→ Let λ→ be a simple version of typed λ-calculus: Types: Terms: τ, τ ′ , ... ::= b | τ → τ ′ t, t′ , ... ::= x | c | λx · t | tt′ where b ranges over a set of base types (booleans, integers, etc.), c over a set of constants and x over a set of variables. We write t[u/x] the result of substituting the term u for free occurrences of the variable x in the term t. Typing judgments are of the form Γ ⊢ t : τ where Γ is a typing context, i.e. a finite mapping from variables to types. We say that x : τ is in Γ whenever Γ (x) = τ . We write Γ, x : τ for the typing context which agrees with Γ except that it maps x to τ . Typing rules are as standard. We consider the set theoretical semantics of λ→ . The semantics of any type τ is given by a set Jτ K. Those sets are such that Jτ → τ ′ K is the set of all functions from Jτ K to Jτ ′ K, for all types τ and τ ′ . A Γ -environment ρ is a map such that, for every x : τ in Γ , ρ(x) is an element of Jτ K. We write ρ[x := a] for the environment which agrees with ρ except that it maps x to a. We write [x := a] for the environment just mapping x to a. Let t be a term such that Γ ⊢ t : τ is derivable. The denotation of t, w.r.t. a Γ -environment ρ, is given as usual by an element JtKρ of Jτ K. We write JtK instead of JtKρ when ρ is irrelevant, e.g., when t is a closed term. When given a value a ∈ Jτ K, we say that it is definable if and only if there exists a closed term t such that ⊢ t : τ is derivable and a = JtK. Let Obs be a subset of base types, called observation types, such as booleans, integers, etc. A context C is a term such that x : τ ⊢ C : o is derivable, where o is an observation type. We spell the standard notion of contextual equivalence in a denotational setting: two elements a1 and a2 of Jτ K, are contextually equivalent (written as a1 ≈τ a2 ), if and only if for any context C such that x : τ ⊢ C : o (o ∈ Obs) is derivable, JCK[x := a1 ] = JCK[x := a2 ]. We say that two closed terms t1 and t2 of the same type τ are contextually equivalent whenever Jt1 K ≈τ Jt2 K. Without making confusion, we shall use the same notation ≈τ to denote the contextual equivalence between terms. We also define a relation ∼τ : for every pair of values a1 , a2 ∈ Jτ K, a1 ∼τ a2 if and only if a1 , a2 are definable and a1 ≈τ a2 . 2.2 Logical relations Essentially, a (binary) logical relation [8] is a family (Rτ )τ type of relations, one for each type τ , on Jτ K such that related functions map related arguments to related results. More formally, it is a family (Rτ )τ type of relations such that for every f1 , f2 ∈ Jτ → τ ′ K, f1 Rτ →τ ′ f2 ⇐⇒ ∀a1 , a2 ∈ Jτ K . a1 Rτ a2 =⇒ f1 (a1 ) Rτ ′ f2 (a2 ) There is no constraint on relations at base types. In λ→ , once the relations at base types are fixed, the above condition forces (Rτ )τ type to be uniquely determined by induction on types. We might have other complex types, e.g., products in variations of λ→ , and in general, relations of these complex types should be also uniquely determined by relations of their type components. For instance, pairs are related when their elements are pairwise related. A unary logical relation is also called a logical predicate. A so-called Basic Lemma comes along with logical relations since Plotkin’s work [15]. It states that if Γ ⊢ t : τ is derivable, ρ1 , ρ2 are two related Γ -environments, and every constant is related to itself, then JtKρ1 Rτ JtKρ2 . Here two Γ -environments ρ1 , ρ2 are related by the logical relation, if and only if ρ1 (x) Rτ ρ2 (x) for every x : τ in Γ . Basic Lemma is crucial for proving various properties using logical relations [8]. In the case of establishing contextual equivalence, it implies that, for every context C such that x : τ ⊢ C : o is derivable (o ∈ Obs), JCK[x := a1 ] Ro JCK[x := a2 ] for every pair of related values a1 , a2 in Jτ K. If Ro is the equality, then JCK[x := a1 ] = JCK[x := a2 ], i.e., a1 ≈τ a2 . Briefly, for every logical relation (Rτ )τ type such that Ro is the equality for every observation type o, logically related values are necessarily contextually equivalent, i.e., Rτ ⊆ ≈τ for any type τ . Completeness states the inverse: a logical relation (Rτ )τ type is complete if every contextually equivalent values are related by this logical relation, i.e., ≈τ ⊆ Rτ for every type τ . Completeness for logical relations is hard to achieve, even in a simple version of λ-calculus like λ→ . Usually we are only able to prove completeness for types up to first order (the order of types is defined inductively: ord(b) = 0 for any base type b; ord(τ → τ ′ ) = max(ord(τ ) + 1, ord(τ ′ )) for function types). The following proposition states the completeness of logical relations in λ→ , for types up to first order: Proposition 1. There exists a logical relation (Rτ )τ type for λ→ , with partial equality on observation types, such that if ⊢ t1 : τ and ⊢ t2 : τ are derivable, for any type τ up to first order, t1 ≈τ t2 =⇒ Jt1 K Rτ Jt2 K. Proof. Let (Rτ )τ type be the logical relation induced by Rb = ∼b at every base type b and we show that it is complete for types up to first order. The proof is by induction over τ . Case τ = b is obvious. Let τ = b → τ ′ . Take two terms t1 , t2 of type b → τ ′ such that Jt1 K and Jt2 K are related by ≈b→τ ′ . Let f1 = Jt1 K and f2 = Jt2 K. Assume that a1 , a2 ∈ JbK are related by Rb , therefore a1 ∼b a2 since Rb = ∼b . Clearly, a1 and a2 are thus definable, say by terms u1 and u2 , respectively. Then, for any context C such that x : τ ′ ⊢ C : o (o ∈ Obs) is derivable, JCK[x := f1 (a1 )] = JC[xu1 /x]K[x := f1 ] = JC[xu1 /x]K[x := f2 ] = JCK[x := f2 (a1 )] = JC[t2 x/x]K[x := a1 ] = JC[t2 x/x]K[x := a2 ] = JCK[x := f2 (a2 )]. (since a1 = Ju1 K) (since f1 ≈b→τ ′ f2 ) (since f2 = Jt2 K) (since a1 ≈b a2 ) Hence f1 (a1 ) ≈τ ′ f2 (a2 ). Moreover, f1 (a1 ) and f2 (a2 ) are therefore definable by t1 u1 and t2 u2 respectively. By induction hypothesis, f1 (a1 ) Rτ ′ f2 (a2 ). Because a1 and a2 are arbitrary, we conclude that f1 Rb→τ ′ f2 . ⊓ ⊔ Note that an equivalent way to state completeness of logical relations is to say that there exists a logical relation (Rτ )τ type which is partial equality on observation types and such that, for all first-order types τ , ∼τ ⊆ Rτ . 3 Logical relations for the computational λ-calculus 3.1 The computational λ-calculus λComp From the section on, our discussion is based on another language — Moggi’s computational λ-calculus. Moggi defines this language so that one can express various forms of side effects (exceptions, non-determinism, etc.) in this general framework [10]. The computational λ-calculus, denoted by λComp , extends λ→ : Types: Terms: τ, τ ′ , ... ::= b | τ → τ ′ | Tτ t, t′ , ... ::= x | c | λx · t | tt′ | val(t) | let x ⇐ t in t′ An extra unary type constructor T is introduced in the computational λ-calculus: intuitively, a type Tτ is the type of computations of type τ . We call Tτ a monadic type in the sequel. The two extra constructs val(t) and let x ⇐ t in t′ represent respectively the trivial computation and the sequential computation, with the typing rules: Γ ⊢t:τ Γ ⊢ val(t) : Tτ Γ ⊢ t : Tτ Γ, x : τ ⊢ t′ : Tτ ′ Γ ⊢ let x ⇐ t in t′ : Tτ ′ Note that the let construct here should not be confused with that in PCF: in λComp , we bind the result of the term t to the variable x, but they are not of the same type — t must be a computation. Moggi also builds a categorical model for the computational λ-calculus, using the notion of monads [10]. Whereas categorical models of simply typed λ-calculi such as λ→ are usually cartesian closed categories (CCCs), a model for λComp requires additionally a strong monad (T, η, µ, t) be defined over the CCC. Consequently, a monadic type is interpreted using the monad T : JTτ K = T Jτ K, and each term in λComp has a unique interpretation as a morphism in a CCC with the strong monad [10]. Semantics of the two additional constructs can be given in full generality in a categorical setting [10]: the denotations of val construct and let construct are defined by the follwoing composites respectively: JΓ ⊢ val(t) : Tτ K : JΓ ⊢t:τ K ηJτ K JΓ K −−−−−→ Jτ K −−−→ T Jτ K, hidJΓ K ,JΓ ⊢t1 :Tτ Ki tJΓ K,Jτ K JΓ ⊢ let x ⇐ t1 in t2 : Tτ ′ K : JΓ K −−−−−−−−−−−→ JΓ K × T Jτ K −−−−−→ T JΓ K × Jτ K µJτ ′ K T JΓ,x:τ ⊢t2 :Tτ ′ K −−−−−−−−−−→ T T Jτ ′ K −−−→ T Jτ ′ K. In particular, the interpretation of terms in the computational λ-calculus must satisfy the following equations: Jlet x ⇐ val(t1 ) in t2 Kρ = Jt2 [t1 /x]Kρ, (1) Jlet x2 ⇐ (let x1 ⇐ t1 in t2 ) in t3 Kρ = Jlet x1 ⇐ t1 in let x2 ⇐ t2 in t3 Kρ,(2) Jlet x ⇐ t in val(x)Kρ = JtKρ. (3) We shall focus on Moggi’s monads defined over the category Set of sets and functions. Figure 1 lists the definitions of some concrete monads: partial computations, exceptions, state transformers, continuations and non-determinism. We shall write λPESCN Comp to refer to λComp where the monad is restricted to be one of these five monads. Partial computation: JTτ K = Jτ K ∪ {⊥} Jval(t)Kρ = JtKρ Jlet x ⇐ t1 in t2 Kρ = Exception: Continuation: Non-determinism: Jt2 Kρ[x := Jt1 Kρ], if Jt1 Kρ 6= ⊥ ⊥, if Jt1 Kρ = ⊥  Jt2 Kρ[x := Jt1 Kρ], if Jt1 Kρ 6∈ E Jt1 Kρ, if Jt1 Kρ ∈ E JTτ K = Jτ K ∪ E Jval(t)Kρ = JtKρ Jlet x ⇐ t1 in t2 Kρ = State transformer:  JTτ K = (Jτ K × St )St Jval(t)Kρ = λ s · (JtKρ, s) Jlet x ⇐ t1 in t2 Kρ = λ s · (Jt2 Kρ[x := a1 ])s1 , where a1 = π1 ((Jt1 Kρ)s), s1 = π2 ((Jt2 Kρ)s) Jτ K JTτ K = RR Jval(t)Kρ = λ kJτ K→R · k(JtKρ) Jlet x ⇐ t1 in t2 Kρ = λ kJτ2 K→R · (Jt1 Kρ)k′ where k′ is a function: λ v Jτ1 K · (Jt2 Kρ[x := v])k JTτ K = Pfin (Jτ K) Jval(t)Kρ = {JtKρ} S Jlet x ⇐ t1 in t2 Kρ = Jt2 Kρ[x := a] a∈Jt1 Kρ Fig. 1. Concrete monads defined in Set The computational λ-calculus is strongly normalizing [1]. The reduction rules in λComp are called βc-reduction rules in [1], which, apart from standard β-reduction in the λ-calculus, contains especially the following two rules for computations: let x ⇐ val(t1 ) in t2 →βc t2 [t1 /x], (4) let x2 ⇐ (let x1 ⇐ t1 in t2 ) in t →βc let x1 ⇐ t1 in (let x2 ⇐ t2 in t).(5) With respect to the βc rules, every term can be reduced to a term in the βc-normal form. Considering also the following η-equality rule for monadic types [1]: let x ⇐ t in t′ [val(x)/x′ ] =η t′ [t/x′ ], (6) we can write every term of a monadic type in the following βc-normal η-long form let x1 ⇐ d1 u11 · · · u1k1 in · · · let xn ⇐ dn un1 · · · unkn in val(u), where n = 0, 1, 2, . . ., every di (1 ≤ i ≤ n) is either a constant or a variable, u and uij (1 ≤ i ≤ n, 1 ≤ j ≤ kj ) are all βc-normal terms or βc-normal-η-long terms (of monadic types). In fact, the rules (4-6) just identify the equations (1-3) respectively. Lemma 1. For every term t of type Tτ in λComp , there exists a βc-normal-η-long term t′ such that Jt′ Kρ = JtKρ, for every valid interpretation J_Kρ (i.e., interpretations satisfying the equations (1-3)). Proof. Because the computational λ-calculus is strongly normalizing, we consider the βc-normal form of term t and prove it by the structural induction on t. – If t is either a variable, a constant or an application, according to the equation (3): JtKρ = Jlet x ⇐ t in val(x)Kρ. In particular, if t is an application t1 t1 , then t1 must be either a variable or a constant since t is βc-normal. Therefore, the term let x ⇐ t in val(x) is in the βc-normalη-long form. – If t is a trivial computation val(t′ ), by induction there is a βc-normal-η-long term t′′ such that Jt′ Kρ = Jt′′ Kρ, for every valid ρ, then Jval(t′ )Kρ = Jval(t′′ )Kρ as well. – If t is a sequential computation let x ⇐ t1 in t2 , since it is βc-normal, t1 should not be any val or let term — t1 must be of the form du1 · · · un (n = 0, 1, 2, . . .) with d either a variable or a constant. By induction, there is a βc-normal-η-long term t′2 such that Jt2 Kρ = Jt2 Kρ, for every valid ρ, then JtKρ = Jlet x ⇐ t′1 in t′2 Kρ and the latter is in the βc-normal-η-long form. ⊓ ⊔ 3.2 Contextual equivalence for λComp As argued in [3], the standard notion of contextual equivalence does not fit in the setting of the computational λ-calculus. In order to define contextual equivalence for λComp , we have to consider contexts C of type To (o is an observation type), not of type o. Indeed, contexts should be allowed to do some computations: if they were of type o, they could only return values. In particular, a context C such that x : Tτ ⊢ C : o is derivable, meant to observe computations of type τ , cannot observe anything, because the typing rule for the let construct only allows us to use computations to build other computations, never values. Taking this into account, we get the following definition: Definition 1 (Contextual equivalence for λComp ). In λComp , two values a1 , a2 ∈ Jτ K are contextually equivalent, written as a1 ≈τ a2 , if and only if, for all observable types o ∈ Obs and contexts C such that x : τ ⊢ C : To is derivable, JCK[x := a1 ] = JCK[x := a2 ]. Two closed terms t1 and t2 of type τ are contextually equivalent if and only if Jt1 K ≈τ Jt2 K. We use the same notation ≈τ to denote the contextual equivalence for terms. 3.3 Logical relations for λComp A uniform framework for defining logical relations relies on the categorical notion of subscones [9], and a natural extension of logical relations able to deal with monadic types was introduced in [2]. The construction consists in lifting the CCC structure and the strong monad from the categorical model to the subscone. We reformulate this construction in the category Set. The subscone is the category whose objects are binary relations (A, B, R ⊆ A × B) where A and B are sets; and a morphism between two objects (A, B, R ⊆ A × B) and (A′ , B ′ , R′ ⊆ A′ × B ′ ) is a pair of functions (f : A → A′ , g : B → B ′ ) preserving relations, i.e. a R b ⇒ f (a) R′ g(b). The lifting of the CCC structure gives rise to the standard logical relations given in Section 2.2 and the lifting of the strong monad will give rise to relations for monadic types. We write T̃ for the lifting of the strong monad T . Given a relation R ⊆ A × B and two computations a ∈ T A and b ∈ T B, (a, b) ∈ T̃ (R) if and only if there exists a computation c ∈ T (R) (i.e. c computes pairs in R) such that a = T π1 (c) and b = T π2 (c). The standard definition of logical relation for the simply typed λ-calculus is then extended with: (c1 , c2 ) ∈ RTτ ⇐⇒ (c1 , c2 ) ∈ T̃ (Rτ ). (7) This construction guarantees that Basic Lemma always holds provided that every constant is related to itself [2]. A list of instantiations of the above definition in concrete monads is also given in [2]. Figure 2 cites the relations for those monads defined in Figure 1. Partial computation: Exception: State transformer: Continuation: Non-determinism: c1 RTτ c2 ⇔ c1 Rτ c2 or c1 = c2 = ⊥ c1 RTτ c2 ⇔ c1 Rτ c2 or c1 = c2 ∈ E where E is the set of exceptions c1 RTτ c2 ⇔ ∀s ∈ St . π1 (c1 s) Rτ π1 (c2 s) & π2 (c1 s) = π2 (c2 s) where St is the set of states c1 RTτ c2 ⇔ c1 (k1 ) = c2 (k2 ) for every k1 , k2 such that ∀a1 , a2 . a1 Rτ a2 =⇒ k1 (a1 ) = k2 (a2 ) c1 RTτ c2 ⇔ (∀a1 ∈ c1 . ∃a2 ∈ c2 . a1 Rτ a2 ) & (∀a2 ∈ c2 . ∃a1 ∈ c1 . a1 Rτ a2 ) Fig. 2. Logical relations for concrete monads We restrict our attention to logical relations (Rτ )τ type such that, for any observation type o ∈ Obs, RTo is a partial equality. Such relations are called observational in the rest of the paper. Note that we require partial identity on To, not on o. But if we assume that denotation of val(_), i.e., the unit operation η, is injective, then that RTo is a partial equality implies that Ro is a partial equality as well. Indeed, let a1 Ro a2 , and by Basic Lemma, Jval(x)K[x := a1 ] RTo Jval(x)K[x := a2 ], that is to say ηJoK (a1 ) = ηJoK (a2 ). By injectivity of η, a1 = a2 . Theorem 1 (Soundness of logical relations in λComp ). If (Rτ )τ tional logical relation, then Rτ ⊆ ≈τ for every type τ . type is an observa- It is straightforward from the Basic Lemma. 3.4 Toward a proof on completeness of logical relations for λComp Completeness of logical relations for λComp is much subtler than in λ→ due to the introduction of monadic types. We were expecting to find a general proof following the general construction defined in [2]. However, this turns out extremely difficult although it might not be impossible with certain restrictions, on types for example. The difficulty arises mainly from the different semantics for different forms of computations, which actually do not ensure that equivalent programs in one monad are necessarily equivalent in another. For instance, consider the following two programs in λComp : let x ⇐ t1 in let y ⇐ t2 in val(x), let y ⇐ t2 in let x ⇐ t1 in val(x), where both t1 and t2 are closed term. We can conclude that they are equivalent in the non-determinism monad — they return the same set of possible results of t1 , no matter what results t2 produces, but this is not the case in, e.g., the exception monad when t1 and t2 throw different exceptions. Being with such an obstacle, we shall switch our effort to case studies in Section 4 and we explore the completeness of logical relations for a list of common monads, precisely, all the monads listed in Figure 1. But, let us sketch out here a general structure for proving completeness of logical relations in λComp . In particular, our study is still restricted to first-order types, which, in λComp , are defined by the following grammar: τ 1 ::= b | Tτ 1 | b → τ 1 , where b ranges over the set of base types. Similarly as in Proposition 1 in Section 2.2, we investigate completeness in a strong sense: we aim at finding an observational logical relation (Rτ )τ type such that if ⊢ t1 : τ and ⊢ t2 : τ are derivable and t1 ≈τ t2 , for any type τ up to first order, then Jt1 K Rτ Jt2 K. Or briefly, ∼τ ⊆ Rτ , where ∼τ is the relation defined in Section 2. As in the proof of Proposition 1, the logical relation (Rτ )τ type will be induced by Rb = ∼b , for any base type b. Then how to prove the completeness for an arbitrary monad T ? Note that we should also check that the logical relation (Rτ )τ type , induced by Rb = ∼b , is observational, i.e., a partial equality on To, for any observable type o. Consider any pair (a, b) ∈ RTo = T̃ (Ro ). By definition of the lifted monad T̃ , there exists a computation c ∈ T Ro such that a = T π1 (c) and b = T π2 (c). But Ro = ∼o ⊆ idJoK , hence the two projections π1 , π2 : Ro → JoK are the same function, π1 = π2 , and consequently a = T π1 (c) = T π2 (c) = b. This proves that RTo is a partial equality. As usual, the proof of completeness would go by induction over τ , to show ∼τ ⊆ Rτ for each first-order type τ . Cases τ = b and τ = b → τ ′ go identically as in λ→ . The only difficult case is τ = Tτ ′ , i.e., the induction step: ∼τ ⊆ Rτ =⇒ ∼Tτ ⊆ RTτ (8) We did not find any general way to show (8) for an arbitrary monad. Instead, in the next section we prove it by cases, for all the monads in Figure 1 except the non-determinism monad. The non-determinism monad is an exceptional case where we do not have completeness for all first-order types but a subset of them. This will be studied separately in Section 4.3. At the heart of the difficulty of showing (8), we find an issue of definability at monadic types in the set-theoretical model. We write def τ for the subset of definable elements in Jτ K, and we eventually show that the relation between def Tτ and def τ can be shortly spelled-out: def Tτ ⊆ T def τ (9) for all the monads we consider in this paper. This is a crucial argument for proving completeness of logical relations for monadic types, but to show (9), we need different proofs for different monads. This is detailed in Section 4.1. 4 Completeness of logical relations for monadic types 4.1 Definability in the set-theoretical model of λPESCN Comp As we have seen in λ→ , definability is involved largely in the proof of completeness of logical relations (for first-order types). This is also the case in λComp and it apparently needs more concern due to the introduction of monadic types. Despite we did not find a general proof for (9), it does hold for all the concrete monads in λPESCN Comp . To state it formally, let us first define a predicate Pτ on elements of Jτ K, by induction on types: – Pb = def b , for every base type b; – PTτ = T (def τ ∩ Pτ ); – Pτ →τ ′ = {f ∈ Pτ →τ ′ | ∀a ∈ def τ , f (a) ∈ Pτ ′ }. We say that a constant c (of type τ ) is logical if and only if τ is a base type or JcK ∈ Pτ . ESCN We then require that λP contains only logical constants. Note that this restriction Comp is valid because the predicates PT τ and Pτ →τ ′ depend only on definability at type τ . Some typical logical constants for monads in λPESCN are as follows: Comp – Partial computation: a constant Ωτ of type Tτ , for every τ . Ωτ denotes the nontermination, so JΩτ K = ⊥. – Exception: a constant raiseeτ of type Tτ for every type τ and every exception e ∈ E. raiseeτ does nothing but raises the exception e, so Jraiseeτ K = e. – State transformer: a constant updates of type Tunit for every state s ∈ St, where unit is the base type which contains only a dummy value ∗. updates simply changes the current state to s, so for any s′ ∈ St, Jupdates K(s′ ) = (∗, s). – Continuation: a constant callkτ of type τ → T bool for every τ and every continuation k ∈ RJτ K . callkτ calls directly the continuation k — it behaves somehow likey “goto” command, so for any a ∈ Jτ K and any continuation k ′ ∈ Rbool, q callkτ (a)(k ′ ) = k(a). – Non-determinism: a constant +τ of type τ → τ → Tτ for every non-monadic type τ . +τ takes two arguments and returns randomly one of them — it introduces the non-determinism, so for any a1 , a2 ∈ Jτ K, J+τ K(a1 , a2 ) = {a1 , a2 }. 1 We assume in the rest of this paper that the above constants are present in λPESCN Comp . Note that Pτ being a predicate on elements of Jτ K is equivalent to say that Pτ can be seen as subset of Jτ K, but in the case of monadic types, PTτ (i.e., T (def τ ∩ Pτ )) is not necessary a subset of JTτ K (i.e., T Jτ K). Fortunately, we prove that all the monads in preserves inclusions, which ensures that the predicate P is well-defined: λPESCN Comp Proposition 2. All the monads in λPESCN preserve inclusions: A ⊆ B ⇒ T A ⊆ T B. Comp Proof. We check it for every monad in λPESCN Comp : – Partial computation: according to the monad definition, if A ⊆ B, then for every c ∈ T A: c ∈ T A ⇐⇒ c ∈ A or c = ⊥ =⇒ c ∈ B or c = ⊥ ⇐⇒ c ∈ T B. – Exception: for every element c ∈ T A: c ∈ T A ⇐⇒ c ∈ A or c ∈ E =⇒ c ∈ B or c ∈ E ⇐⇒ c ∈ T B. – State transformer: for every a ∈ T A: c ∈ T A ⇐⇒ ∀s ∈ St . π1 (cs) ∈ A =⇒ ∀s ∈ St . π1 (cs) ∈ B ⇐⇒ c ∈ T B. A – Continuation: this is a special case because apparently T A = RR is not a subset of B T B = RR , since they contain functions that are defined on different domains, but we shall consider here the functions coinciding on the smaller set A as equivalent. We say that two functions f1 and f2 defined on a domain B coincide on A (A ⊆ B), written as f1 |A = f2 |A , if and only if for every x ∈ A, f1 (x) = f2 (x). Then for every c ∈ T A: ∀k1 , k2 ∈ RB . k1 = k2 =⇒ k1 |A = k2 |A =⇒ c(k1 ) = c(k2 ), so c is also function from RB to R, i.e., c ∈ T B. – Non-determinism: for every c ∈ T A: c ∈ T A ⇐⇒ ∀a ∈ c . a ∈ A =⇒ ∀a ∈ c . a ∈ B ⇐⇒ c ∈ T B. ⊓ ⊔ Introducing such a constraint on constants is mainly for proving (9). Let us figure out the proof. Take an arbitrary element c in def Tτ . By definition, there exists a closed term t of type Tτ such that JtK = c. While it is not evident that c ∈ T def τ , we are expecting to show that JtK ∈ T def τ , by considering the βc-normal-η-long form of t, since 1 It is easy to check that each of these constants is related to itself, except callkτ for continuations. However, we still assume the presence of callkτ for the sake of proving completeness, while we are not able to prove the soundness with it. Note that Theorem 1 and Theorem 2 still hold, but they are not speaking of the same language. λComp is strongly normalizing, Take the partial computation monad as an example, where T def τ = def τ ∪ {⊥}. Consider the βc-normal-η-long form of t: let x1 ⇐ d1 u11 · · · u1k1 in · · · let xn ⇐ dn un1 · · · unkn in val(u). We shall make the induction on n. It is clear that JtK ∈ T def τ when n = 0. For the induction step, we hope that the closed term d1 u11 · · · u1k1 (of type Tτ1 ) would produce either ⊥ (the non-termination), or a definable result (of type τ1 ) so that we can substitute x1 in the rest of the normal term with the result of d1 u11 · · · u1k1 and make use of induction hypothesis. The constraint on constants helps here: to ensure that after the substitution, the resulted term is still in the proper form so that the induction would go through. The following lemma shows that for every computation term t, JtK ∈ T def τ if t is in a particular form, which is a more general form of βc-normal-η-long form. Lemma 2. In λPESCN Comp , JtK ∈ T def τ , for every closed computation term t (of type Tτ ) of the following form: t ≡ let x1 ⇐ t1 w11 · · · w1k1 in · · · let xn ⇐ tn wn1 · · · wnkn in val(w), where n = 0, 1, 2, . . . and ti (1 ≤ i ≤ n) is either a variable or a closed term such that P(Jti K) holds, and w, wij (1 ≤ i ≤ n, 1 ≤ j ≤ ki ) are valid λPESCN terms. Comp Proof. We prove it by induction on n, for every monad: – Partial computation (T def τ = def τ ∪ {⊥}): if n = 0, it is clear that JtK ∈ T def τ . When n > 0, because P(Jt1 K) holds (t1 must be closed), Jt1 w11 · · · w1k1 K ∈ T (def τ1 ∩ Pτ1 ). If Jt1 w11 · · · w1k1 K = ⊥, then JtK = ⊥ ∈ T def τ ; otherwise, assume that Jt′1 K = Jt1 w11 · · · w1k1 K where t′1 is a closed term of type τ1 (assuming that t1 w11 · · · w1k1 is of type Tτ1 ). According to the definition of P, P(Jt′1 K) holds. Let t′ be another closed term: ′ ′ ′ ′ in val(w[t′1 /x1 ]), in · · · let xn ⇐ t′n wn1 · · · wnk t′ ≡ let x2 ⇐ t′2 w21 · · · w2k n 2 ′ where t′i (2 ≤ i ≤ n) is either t′1 or ti , wij ≡ wij [t′1 /x1 ] (2 ≤ i ≤ n, 1 ≤ j ≤ ki ). ′ By induction, Jt K ∈ T def τ holds. Furthermore, Jt′ K = Jlet x2 ⇐ t2 w21 · · · w2k2 in · · · let xn ⇐ tn wn1 · · · wnkn in val(w)K[x1 := Jt′1 K] = Jlet x1 ⇐ t1 w11 · · · w1k1 in · · · let xn ⇐ tn wn1 · · · wnkn in val(w)K = JtK, hence JtK ∈ T def τ . – Exception (T def τ = def τ ∪ E): if n = 0, clearly JtK ∈ T def τ . When n > 0, because P(Jt1 K) holds, Jt1 w11 · · · w1k1 K ∈ T (def τ1 ∩ Pτ1 ). If Jt1 w11 · · · w1k1 K ∈ E, then JtK ∈ E ⊆ T def τ ; otherwise, exactly as in the case of partial computation, build a term t′ . Similarly, we prove that JtK = Jt′ K ∈ T def τ by induction. – State transformer (T def τ = (def τ × St)St ): when n = 0, for every s ∈ St, π 1 (JtKs) = JwK ∈ def τ hence JtK ∈ T def τ . When n > 0, for every s ∈ St, assume that Jts1 K = π 1 (Jt1 w11 · · · w1k1 Ks) where t′1 is a closed term of type τ1 (assuming that t1 w11 · · · w1k1 is of type Tτ1 ). According to the definition of P, P(Jts1 K) holds. Let ts be another closed term: s s s s in val(w[ts1 /x1 ]), in · · · let xn ⇐ tsn wn1 · · · wnk ts ≡ let x2 ⇐ ts2 w21 · · · w2k n 2 s where tsi (2 ≤ i ≤ n) is either ts1 or ti , wij ≡ wij [ts1 /x1 ] (2 ≤ i ≤ n, 1 ≤ j ≤ ki ). s By induction, Jt K ∈ T def τ holds. Furthermore, for every s ∈ St, JtKs = Jlet x1 ⇐ t1 w11 · · · w1k1 in · · · let xn ⇐ tn wn1 · · · wnkn in val(w)Ks = (Jlet x2 ⇐ t2 w21 · · · w2k2 in · · · let xn ⇐ tn wn1 · · · wnkn in val(w)K[x1 := Jts1 K])s′ s ′ = Jt Ks , where s′ = π2 (Jt1 w11 · · · w1k1 Ks). Since Jts K ∈ T def τ for every s ∈ St, π1 (JtKs) = π1 (Jts Ks′ ) ∈ def τ , hence JtK ∈ T def τ . Jτ K def τ – Continuation (T def τ = RR ): we say that an element c ∈ JTτ K = RR is in Jτ K T def τ if and only if for every pair of continuations k1 , k2 ∈ R , k1 |def τ = k2 |def τ =⇒ c(k1 ) = c(k2 ). If n = 0, JtK = λ k.k(JwK) ∈ T def τ . When n > 0, according to the definition of the continuation monad: JtK = λ k · Jt1 w11 · · · wnkn K(k ′ ), where k ′ = λ a·(Jlet x2 ⇐ t2 w21 · · · w2k2 in · · · let xn ⇐ tn wn1 · · · wnkn in val(w)K[x1 := a])k. For every continuations k1 , k2 ∈ RJτ K such that k1 |def τ = k2 |def τ let ki′ = λ a·(Jlet x2 ⇐ t2 w21 · · · w2k2 in · · · let xn ⇐ tn wn1 · · · wnkn in val(w)K[x1 := a])ki , i = 1, 2. Because Jt1 w11 · · · w1k1 K ∈ T (Pτ1 ∩def τ1 ), if we can prove k1′ |Pτ1 ∩def τ1 = k2′ |Pτ1 ∩def τ1 , which implies JtK(k1 ) = JtK(k2 ), we can conclude JtK ∈ T def τ . For every a ∈ Pτ1 ∩def τ1 , let Jta1 K = a where ta1 is a closed term. Define another closed term ta : a a a a in val(w[ta1 /x1 ]), in · · · let xn ⇐ tan wn1 · · · wnk ta ≡ let x2 ⇐ ta2 w21 · · · w2k n 2 a where tai (2 ≤ i ≤ n) is either ta1 or ti , wij ≡ wij [ta1 /x1 ] (2 ≤ i ≤ n, 1 ≤ a ′ j ≤ ki ). By induction, Jt K ∈ T def τ , so k1 (a) = Jta Kk1 = Jta Kk2 = k2′ (a), i.e., k1′ |Pτ1 ∩def τ1 = k2′ |Pτ1 ∩def τ1 . – Non-determinism (T def τ = Pfin (def τ )): when n = 0, JtK = {JwK} ∈ T def τ . When n > 0, for every a ∈ Jt1 w11 · · · w1k1 K, assume that Jta1 K = a where t′1 is a closed term of type τ1 (assuming that t1 w11 · · · w1k1 is of type Tτ1 ). According to the definition of P, P(Jta1 K) holds. Let ta be another closed term: a a a a in val(w[ta1 /x1 ]), in · · · let xn ⇐ tan wn1 · · · wnk ta ≡ let x2 ⇐ ta2 w21 · · · w2k n 2 a where tai (2 ≤ i ≤ n) is either ta1 or ti , wij ≡ wij [ta1 /x1 ] (2 ≤ i ≤ n, 1 ≤ j ≤ ki ). a By induction, Jt K ∈ T def τ holds. Furthermore, JtK = Jlet S x1 ⇐ t1 w11 · · · w1k1 in · · · let xn ⇐ tn wn1 · · · wnkn in val(w)K Jlet x2 ⇐ t2 w21 · · · w2k2 in · · · = a∈Jt1 K xn ⇐ tn wn1 · · · wnkn in val(w)K[x1 := a] S let = Jta K. a∈Jt1 K Because Jta K ∈ T def τ holds for every a ∈ Jt1 w11 · · · w1k1 K, JtK ∈ T def τ . ⊓ ⊔ From the above lemma, we conclude immediately that for every closed βc-normalη-long computation term t in λPESCN with logical constants, JtK ⊆ T def τ . Comp Proposition 3. def Tτ ⊆ T def τ holds in the set-theoretical model of λPESCN with Comp logical constants. Proof. It follows from Lemma 2 by considering the βc-normal-η-long terms that define elements in JTτ K since λComp is strongly normalizing. ⊓ ⊔ 4.2 Completeness of logical relations in λPESC Comp for first-order types We prove (8) in this section for the partial computation monad, the exception monad, the state monad and the continuation monad. We write λPESC Comp for λComp where the monad is restricted to one of these four monads. Proofs depend typically on the particular semantics of every form of computation, but a common technique is used frequently: given two definable but non-related elements of JTτ K, one can find a context to distinguish the programs (of type Tτ ) that define the two given elements, and such a context is usually built based on another context that can distinguish programs of type τ . Lemma 3. Let (Rτ )τ type be a logical relation in λPESC Comp with only logical constants. ∼τ ⊆ Rτ =⇒ ∼Tτ ⊆ RTτ holds for every type τ . Proof. Take two arbitrary elements c1 , c2 ∈ JTτ K such that (c1 , c2 ) 6∈ RTτ , we prove that c1 6∼Tτ c2 for every monad in λPESC Comp : – Partial computation: the fact (c1 , c2 ) 6∈ RTτ amounts to the following two cases: • c1 , c2 ∈ Jτ K but (c1 , c2 ) 6∈ Rτ , then c1 6∼τ c2 . If one of these two values is not definable at type τ , by Proposition 3, it is not definable at type Tτ either. If both values are definable at type τ but they are not contextually equivalent, then there is a context x : τ ⊢ C : To such that JCK[x := c1 ] 6= JCK[x := c2 ]. Thus, the context y : Tτ ⊢ let x ⇐ y in C : To can distinguish c1 and c2 (as two values of type Tτ ). • c1 ∈ Jτ K and c2 = ⊥ (or symmetrically, c1 = ⊥ and c2 ∈ Jτ K), then the context let x ⇐ y in val(true) can be used to distinguish them. c1 6∼Tτ c2 in both cases. – Exception: the fact (c1 , c2 ) 6∈ RTτ amounts to three cases: • c1 , c2 ∈ Jτ K but (c1 , c2 ) 6∈ Rτ , then c1 6∼τ c2 . Suppose both values are definable at type τ , otherwise by Proposition 3, they must not be definable at type Tτ . Similar as in the case of partial computation we can build a context that distinguishes c1 and c2 as values of type Tτ , from the context that distinguishes c1 and c2 as values of type τ . • c1 ∈ Jτ K, c2 ∈ E. Consider the following context: y : Tτ ⊢ let x ⇐ y in val(true) : Tbool. When y is substituted by c1 and c2 , the context evaluates to different values, namely, a boolean and an exception. • c1 , c2 ∈ E but c1 6= c2 . Try the same context as in the second case, which will evaluate to two different exceptions that can be distinguished. c1 6∼Tτ c2 in all the three cases. – State transformer: because (c1 , c2 ) 6∈ RTτ , there exists some s0 ∈ St such that • either (π1 (c1 s0 ), π1 (c2 s0 )) 6∈ Rτ . Then by induction π1 (c1 s0 ) 6∼τ π1 (c2 s0 ). If π1 (ci s0 ) (i = 1, 2) is not definable, then by Proposition 3, ci is not definable either. If both π1 (c1 s0 ) and π1 (c2 s0 ) are definable, but π1 (c1 s0 ) 6≈τ π1 (c2 s0 ), then there is a context x : τ ⊢ C : To such that JCK[x := π1 (c1 s0 )] 6= JCK[x := π1 (c2 s0 )], i.e., for some state s′0 ∈ St , JCK[x := π1 (c1 s0 )](s′0 ) 6= JCK[x := π1 (c1 s0 )](s′0 ). Now we can use the following context: y : Tτ ⊢ let x ⇐ y in let z ⇐ updates′0 in C : T o, Let fi = s ∈ St, r z let x ⇐ y in let z ⇐ updates′0 in C [y := ci ], then for every r z fi (s) = let z ⇐ updates′0 in C [x := π1 (ci s)](π2 (ci s)) = JCK[x := π1 (ci s)](s′0 ), (i = 1, 2). f1 6= f2 , because when applied to the state s0 , they will return two different pairs, so the above context can distinguish the two values c1 and c2 ; • or π2 (c1 s0 ) 6= π2 (c2 s0 ). we use the context y : Tτ ⊢ let x ⇐ y in val(true) : Tbool, then Jlet x ⇐ y in val(true)K[y := ci ] = λs.(true, π2 (ci s)) (i = 1, 2). These two functions are not equal since they return different results when applied to the state s0 . In both cases, c1 6∼Tτ c2 . – Continuation: first say that two continuations k1 , k2 ∈ RJτ K are R-related, if and only if for every a1 , a2 ∈ Jτ K, a1 Rτ a2 =⇒ k1 (a1 ) = k2 (a2 ). The fact (c1 , c2 ) 6∈ RTτ means that there are two R-related continuations k1 , k2 such that c1 (k1 ) 6= c2 (k2 ). Because ∼τ ⊆ Rτ , for every definable value a ∈ def τ , clearly, a ∼τ a =⇒ a1 R a2 =⇒ k1 (a1 ) = k2 (a2 ), so k1 and k2 coincide over def τ . Suppose that both c1 and c2 are definable, then by Proposition 3, c1 (k1 ) = c1 (k2 ) and c2 (k1 ) = c2 (k2 ), hence c1 (k1 ) 6= c2 (k1 ). Consider the context y : Tτ ⊢ let x ⇐ y in callkτ 1 (x) : T bool. For every k ∈ RJboolK , q y k1 let x ⇐ q y inkcallyτ (x) [y := ci ](k) = ci (λ a · ( callτ 1 (x) [x := a])k) = ci (λ a · k1 (a)) = ci (k1 ). (i = 1, 2), Since c1 (k1 ) 6= c2 (k1 ), this context distinguishes the two computations, hence c1 6∼Tτ c2 . ⊓ ⊔ Theorem 2. In λPESC Comp , if all constants are logical and in particular, if the following constants are present – updates for the state transformer monad; – callkτ for the continuation monad, then logical relations are complete up to first-order types, in the strong sense that there exists an observational logical relation (Rτ )τ type such that for any closed terms t1 , t2 of any type τ 1 up to first order, if t1 ≈τ 1 t2 , then Jt1 K Rτ 1 Jt2 K. Proof. Take the logical relation (Rτ )τ type induced by Rb =∼b , for any base type b. We prove by induction on types that ∼τ ⊆ Rτ for any first-order type τ . In particular, the induction step ∼τ ⊆ Rτ =⇒∼Tτ ; ⊆ RTτ is shown by Lemma 3. ⊓ ⊔ 4.3 Completeness of logical relations for the non-determinism monad The non-determinism monad is an interesting case: the completeness of logical relations for this monad does not hold for all first-order types! To state it, consider the following two programs of a first-order type that break the completeness of logical relations: ⊢ val(λx.(true +bool false)) : T(bool → Tbool), ⊢ λx.val(true) +bool→Tbool λx.(true +bool false) : T(bool → Tbool). Recall the logical constant +τ of type τ → τ → Tτ : J+τ K(a1 , a2 ) = {a1 , a2 } for every a1 , a2 ∈ Jτ K. The two programs are contextually equivalent: what contexts can do is to apply the functions to some arguments and observe the results. But no matter how many time we apply these two functions, we always get the same set of possible values ({true, false}), so there is no way to distinguish them with a context. Recall the logical relation for non-determinism monad in Figure 2: c1 RTτ c2 ⇔ (∀a1 ∈ c1 . ∃a2 ∈ c2 . a1 Rτ a2 ) & (∀a2 ∈ c2 . ∃a1 ∈ c1 . a1 Rτ a2 ). Clearly the denotations of the above two programs are not related by that relation because the function Jλx.val(true)K from the second program is not related to the function in the first. However, if we assume that for every non-observable base type b, there is an equality test constant testb : b → b → bool (clearly, P(testb ) holds), logical relations for the non-determinism monad are then complete for a set of weak first-order types: τw1 ::= b | Tb | b → τw1 . Compared to all types up to first order, weak first-order types do not contain monadic types of functions, so it immediately excludes the two programs in the above counterexample. Theorem 3. Logical relations for the non-determinism monad are complete up to weak first-order types. in the strong sense that there exists an observational logical relation (Rτ )τ type such that for any closed terms t1 , t2 of a weak first-order type τw1 , if t1 ≈τw1 t2 , then Jt1 K Rτw1 Jt2 K. Proof. Take the logical relation R induced by Rb =∼b , for any base type b. We prove by induction on types that ∼τw1 ⊆ Rτ 1 for any weak first-order type τw1 . w Cases b and b → τw1 go identically as in standard typed lambda-calculi. For monadic types Tb, suppose that (c1 , c2 ) 6∈ RTb , which means either there is a value in c1 such that no value of c2 is related to it, or there is such a value in c2 . We assume that every value in c1 and c2 is definable (otherwise it is obvious that c1 6∼Tb c2 because at least one of them is not definable, according to Proposition 3). Suppose there is a value a ∈ c1 such that no value in c2 is related to it, and a can be defined by a closed term t of type b. Then the following context can distinguish c1 and c2 : x : Tτ ⊢ let y ⇐ x in testb (y, t) : Tbool since every value in c2 is not contextually equivalent to a, hence not equal to a.  Now let state and label be base types such that label is an observation type, whereas state is not. Using non-determinism monad, we can define labeled transition systems as elements of Jstate → label → TstateK, with states in JstateK and labels in JlabelK, as functions mapping states a and labels l to the set of states b such that l / b . The logical relation at type state → label → Tstate is given by [2]: a (f1 , f2 ) ∈ Rstate→label→Tstate ⇐⇒ ∀a1 , a2 , l1 , l2 · (a1 , a2 ) ∈ Rstate & (l1 , l2 ) ∈ Rlabel =⇒ (∀b1 ∈ f1 (a1 , l1 ) · ∃b2 ∈ f2 (a2 , l2 ) · (b1 , b2 ) ∈ Rstate ) & (∀b2 ∈ f2 (a2 , l2 ) · ∃b1 ∈ f1 (a1 , l1 ) · (b1 , b2 ) ∈ Rstate ) In case Rlabel is equality, f1 and f2 are logically related if and only if Rstate is a strong bisimulation between the labeled transition systems f1 and f2 . Sometimes we explicitly specify an initial state for certain labeled transition system. In this case, the encoding of the labeled transition system in the nondeterminism monad is a pair (q, f ) of Jstate × (state → label → Tstate)K, where q is the initial state and f is the transition relation as defined above. Then (q1 , f1 ) and (q2 , f2 ) are logically related if and only if they are strongly bisimular, i.e., Rstate is a strong bisimulation between the two labeled transition systems and q1 Rstate q2 . Corollary 1 (Soundness of strong bisimulation). Let f1 and f2 be transition systems. If there exists a strong bisimulation between f1 and f2 , then f1 and f2 are contextually equivalent. Proof. There exists a strong bisimulation between f1 and f2 , therefore f1 and f2 are logically related. By Theorem 1, f1 and f2 are thus contextually equivalent.  In order to prove completeness, we need to assume that label has no junk, in the sense that every value of JlabelK is definable. Corollary 2 (Completeness of strong bisimulation). Let f1 and f2 be transition systems which are definable. If f1 and f2 are contextually equivalent and label has no junk, then there exists a strong bisimulation between f1 and f2 . Proof. Let R be the logical relation given by Theorem 3. f1 and f2 are definable and contextually equivalent, so f1 Rstate→label→Tstate f2 . Moreover, because label has no junk, Rlabel is equality. Rstate is thus a strong bisimulation between f1 and f2 . ⊓ ⊔ 5 Conclusion The work presented in this paper is a natural continuation of the authors’ previous work [2,3]. In [2], we extend [9] and derive logical relations for monadic types which are sound in the sense that the Basic Lemma still holds. In [3], we study contextual equivalence in a specific version of the computational λ-calculus with cryptographic primitives and we show that lax logical relations (the categorical generalization of logical relations [14]) derived using the same construction is complete. Then in this paper, we explore the completeness of logical relations for the computational λ-calculus and we show that they are complete at first-order types, for a list of common monads: partial computations, exceptions, state transformers and continuations, while in the case of continuation, the completeness depends on a natural constant call, with which we cannot show the soundness. Pitts and Stark have defined operationally based logical relations to characterize the contextual equivalence in a language with local store [13]. This work can be traced back to their early work on the nu-calculus [12] which can be translated in a special version of the computational λ-calculus and be modeled using the dynamic name creation monad [17]. Logical relations for this monad are derived in [19] using the construction from [2]. It is also shown in [19] that such derived logical relations are equivalent to Pitts and Stark’s operational logical relations up to second-order types. An exceptional case of our completeness result is the non-determinism monad, where logical relations are not complete for all first-order types, but a subset of them. We effectively show this by providing a counter-example that breaks the completeness at first-order types. This is indeed an interesting case. A more comprehensive study on this monad can be found in [4], where Jeffrey defines a denotational model for the computational λ-calculus specialized in non-determinism and proves that this model is fully abstract for may-testing. The relation between our notion of contextual equivalence and the may-testing equivalence remains to be clarified. Recently, Lindley and Stark introduce the syntactic ⊤⊤-lifting for the computational λ-calculus and prove the strong normalization [7]. Katsumata then instantiates their liftings in Set [5]. The ⊤⊤-lifting of strong monads is an essentially different approach from that in [2]. It would be interesting to establish a formal relationship between these two approaches, and to look for a general proof of completeness using the ⊤⊤-lifting. References 1. P. N. Benton, G. M. Bierman, and V. C. V. de Paiva. Computational types from a logical perspective. J. Functional Programming, 8(2):177–193, 1998. 2. J. Goubault-Larrecq, S. Lasota, and D. Nowak. Logical relations for monadic types. In Proceedings of CSL’2002, volume 2471 of LNCS, pages 553–568. Springer, 2002. 3. J. Goubault-Larrecq, S. Lasota, D. Nowak, and Y. Zhang. Complete lax logical relations for cryptographic lambda-calculi. In Proceedings of CSL’2004, volume 3210 of LNCS, pages 400–414. Springer, 2004. 4. A. Jeffrey. A fully abstract semantics for a higher-order functional language with nondeterministic computation. Theoretical Computer Science, 228(1-2):105–150, 1999. 5. S. Katsumata. A semantic formulation of ⊤⊤-lifting and logical predicates for computational metalanguage. In Proceedings of CSL’2005, volume 3634 of LNCS, pages 87–102. Springer, 2005. 6. R. Lazić and D. Nowak. A unifying approach to data-independence. In Proceedings of CONCUR’2000, volume 1877 of LNCS, pages 581–595. Springer, 2000. 7. S. Lindley and I. Stark. Reducibility and ⊤⊤-lifting for computation types. In Proceedings of TLCA’2005, number 3461 in LNCS, pages 262–277. Springer, 2005. 8. J. C. Mitchell. Foundations of Programming Languages. MIT Press, 1996. 9. J. C. Mitchell and A. Scedrov. Notes on sconing and relators. In Proceedings of CSL’1992, volume 702 of LNCS, pages 352–378. Springer, 1993. 10. E. Moggi. Notions of computation and monads. Information and Computation, 93(1):55–92, 1991. 11. P. W. O’Hearn and R. D. Tennent. Parametricity and local variables. J. ACM, 42(3):658–709, 1995. 12. A. Pitts and I. Stark. Observable properties of higher order functions that dynamically create local names, or: What’s new? In Proceedings of MFCS’1993, number 711 in LNCS, pages 122–141. Springer, 1993. 13. A. Pitts and I. Stark. Operational reasoning for functions with local state. In Higher Order Operational Techniques in Semantics, pages 227–273. Cambridge University Press, 1998. 14. G. Plotkin, J. Power, D. Sannella, and R. Tennent. Lax logical relations. In Proceedings of ICALP’2000, volume 1853 of LNCS, pages 85–102. Springer, 2000. 15. G. D. Plotkin. Lambda-definability in the full type hierarchy. In To H. B. Curry: Essays on Combinatory Logic, Lambda Calculus and Formalism, pages 363–373. Academic Press, 1980. 16. K. Sieber. Full abstraction for the second order subset of an algol-like language. Theoretical Computer Science, 168(1):155–212, 1996. 17. I. Stark. Categorical models for local names. Lisp and Symbolic Computation, 9(1):77–107, 1996. 18. E. Sumii and B. C. Pierce. Logical relations for encryption. J. Computer Security, 11(4):521– 554, 2003. 19. Y. Zhang. Cryptographic logical relations. Ph. d. dissertation, ENS Cachan, France, 2005.
6cs.PL
Robust Fusion Methods for Structured Big Data Catherine Aarona , Alejandro Cholaquidisb , Ricardo Fraimanb and Badih Ghattasc a arXiv:1804.01858v1 [math.ST] 5 Apr 2018 Université Clermont Auvergne, Campus Universitaire des Cézeaux, France. b Universidad de la República, Facultad de Ciencias, Uruguay. c Aix Marseille Université, CNRS, Centrale Marseille, I2M UMR 7373, 13453, Marseille, France. Abstract We address one of the important problems in Big Data, namely how to combine estimators from different subsamples by robust fusion procedures, when we are unable to deal with the whole sample. We propose a general framework based on the classic idea of ‘divide and conquer’. In particular we address in some detail the case of a multivariate location and scatter matrix, the covariance operator for functional data, and clustering problems. 1 Introduction Big Data has arisen in recent years to deal with problems in several domains, such as social networks, biochemistry, health care systems, politics, and retail, among many others. New developments are necessary to address most of the problems in the area. Typically, classical statistical approaches that perform reasonably well for small data sets fail when dealing with huge data sets. To handle these challenges, new mathematical and computational methods are needed. The challenges posed by Big Data cover a wide range of various problems, and have been recently considered in a huge literature (see, for instance, Wang et al. (2016), Yu (2014), Ahmed (2017) and the references therein). We address one of these problems, namely, how to combine, using robust techniques, estimators obtained from different subsamples in the case where we are computationally unable to deal with the whole sample. In what follows, we will refer to such approaches as robust fusion methods (RFM). A general algorithm is proposed, which is, in spirit, related with the well known idea of divide-and-combine. We consider the case where the data belong to finite and infinite dimensional spaces (functional data). Functional Data Analysis (FDA) has become a central area of statistics in recent years, having gained much momentum from the work of Ramsay in the early 2000s. Since then, both the quantity and the quality of its results have enjoyed a marked growth, while addressing a great diversity of problems. FDA faces several specific challenges, most of them associated with the infinite-dimensional nature of the data. Some recent important and unavoidable references for FDA are Hovárt and Kokoszka (2012), Ferraty and Vieu (2006), Aneiros et al (2017), as well as the recent surveys Cuevas (2014) and Vardi and Zhand (2000). Divide-and-combine (see for instance Aho et al. (1974)) is a well known technique for dealing with hugh data-sets. In the FDA setting have, in Tang et al. (2016), also been considered recently for the linear regression problem involving Lasso, a problem that is not addressed in the present paper, where we focus on a general robust procedure for different problems. 1 The consistency and robustness of our method is studied in the general setting of FDA, and we apply the proposed algorithm to some statistical problems in finite and infinite dimensional settings, namely, the location and scatter matrix, clustering, and impartial trimmed k-means. Also, a new robust estimator of the covariance operator is proposed. We start by describing one of the simplest problems in this area as a toy example. Suppose we are interested in the median of a huge set of iid random variables {X1 , . . . , Xn } with common density fX , and we split the sample into m subsamples of size l, so that n = ml. We calculate the median of each subsample and obtain m random variables Y1 , . . . ,Ym . Then we take the median of the set Y1 , . . . ,Ym , i.e. we consider the well known median of medians, which, in this case, will be our RFM estimator. It is clear that it does not coincide with the median of the whole original sample {X1 , . . . , Xn }, but it will be close. What else can we say about this estimator regarding its efficiency and robustness? In this particular case, the RFM estimator is nothing but the median of m iid random variables, but now with a different distribution, given by the distribution of the median of l random variables with density fX . Suppose for simplicity that l = 2k + 1. Then, the density of the random variables Yi is given by gY (t) = (2k + 1)! FX (t)k (1 − FX (t))k fX (t). (k!)2 (1) On the one hand, if fX (FX−1 (0.5)) 6= 0, the empirical median θ̂ = med(X1 , . . . , Xn ) behaves, asymptotically, like the normal distribution centred at the true median θ with variance V(θ̂ ) = 1/(4n fX (θ )2 ). On the other hand, θ̃ RFM , the median of medians, behaves asymptotically like the normal distribution centred at θ with variance V(θ̃ RFM ) = 1/(4mgY (θ )2 ), where gY (θ ) = (1/2)2k (2k + 1)!/(k!2 ) fX (θ ) ∼ p 2k/π. So we can explicitly calculate the asymptotic relative loss of efficiency, i.e. limn→∞ V(θ̂ )/V(θ̃ RFM ) = 2/π. In Section 2 we generalize this RFM idea and study its consistency, robustness, breakdown point, and efficiency. Section 3 shows how the RFM may be applied to multivariate location and scatter matrix estimation, covariance operator estimation for functional data, and robust clustering. The last section provides some simulation results for these problems. 2 A general setup for RFM. We start by introducing a general framework for RFM. The idea is quite simple: given a sample {X1 , . . . , Xn } of iid random elements in a metric space E (for instance E = Rd ) and a statistical problem, (such as multivariate location, covariance operators, linear regression, or principal components, among many others), we split the sample into m subsamples of equal size. For each subsample we compute a robust solution for the statistical problem considered. The solution given by RFM corresponds to the deepest point among the m solutions (in terms of the appropriate norm associated to the problem) obtained from the subsamples. In order to introduce the notion of depth, we will use throughout this paper the following notation. Let X be a random variable taking values in some Banach space (E, k · k), with probability distribution PX , and let x ∈ E. The depth of x with respect to PX is defined as follows: 2  D(x, PX ) = 1 − EPX X −x kX − xk  . (2) It was introduced by Chaudhuri (1996), formulated (in a different way) by Vardi and Zhand (2000), and extended to a very general setup by Chakraborty and Chaudhuri (2014). Given a sample {X1 , . . . , Xn }, let us write Pn for the empirical measure. The empirical version of (2) is  D(x, Pn ) = 1 − EPn X −x kX − xk  = 1− 1 n n Xi − x ∑ kXi − xk . (3) i=1 Although we suggest using the depth function, for some statistical problems this is unsuitable, for instance in clustering. In such cases, the deepest point may be replaced by other robust estimators, as we will show in Section 3.3. We summarize our approach in Table 1 for a general framework of parameter estimation. This may be easily applied to any situation where robust estimators exist or can be designed. {X1 , . . . , Xn } iid random elements in a Banach space E. θ0 a parameter to estimate a) split the sample into m subsamples with n = ml {X1 , . . . , Xl }, {Xl+1 , . . . , X2l }, . . . , {X(m−1)l+1 , . . . , Xlm }. b) Compute a robust estimate of θ0 on each subsample, obtaining θ̂1 , . . . , θ̂m . c) Compute the final estimate θ̃ RFM by RFM combining θ̂1 , . . . , θ̂m by a robust approach. For instance, θ̃ RFM can be the deepest point, or the average of 40% of the deepest points among the θ̂1 , . . . , θ̂m . Table 1: Parameter estimation using RFM We will address the consistency, efficiency, robustness, and computational time of the RFM proposals. 3 2.1 Consistency, robustness and breakdown point of the RFM We start by proving that, given a sample {X1 , . . . , Xn } of a random element X, its deepest point (i.e. the value that maximizes (3)) converges almost surely to the value that maximizes (2). Although similar results has already been obtained (see for instance Chakraborty and Chaudhuri (2014)), we will need it when Pn is not necessarily the empirical measure associated to a sample, but any measure converging weakly to a probability distribution P. We will need the following assumption. H1 A probability measure P defined on a separable Hilbert space H fulfils H1 if P(∂ B(y, r)) = 0 for all r > 0 and y ∈ H , where ∂ A stands for the boundary of a set A⊂H . Observe that H1 is fulfilled if the random variables kX − yk are absolutely continuous, for all y ∈ H , where X is a random variable with distribution P. Theorem 1. Let {Xn }n be a sequence of random elements with common distribution Pn , defined in a separable Hilbert space (H , k · k). Let P be a probability distribution fulfilling H1. Assume that Pn → P weakly, and kEP ((X − x)/kX − xk)k has a unique minimum. Then arg max D(x, Pn ) → arg max D(x, P) a.s., as n → ∞. x (4) x In order to prove (4) we will use the following fundamental result proved in Billingsley and Topsøe (1967) (which still holds when H is a separable Banach space), see theorem 1 and example 3. Theorem (Billingsley and Topsøe). Suppose S ⊂ H and let B(S, H ) be the class of all bounded measurable functions mapping S into H . Suppose F ⊂ B(S, H ) is a subclass of functions. Then Z sup f dPn − Z f dP → 0, (5) f ∈F for every sequence Pn that converges weakly to P if, and only if, sup{k f (z) − f (t)k : f ∈ F , z,t ∈ S} ≤ ∞, and for all ε > 0, lim sup P({x : ω f (B(x, δ )) ≥ ε}) = 0, δ →0 f ∈F (6) where ω f (A) = sup{| f (x) − f (y)| : x, y ∈ A} and B(x, δ ) is the open ball of radii δ > 0. Proof of Theorem 1. Consider S = H and F the subclass of functions { fy }y∈H where fy (z) = (z − √ y)/kz − yk. Then, sup{k fy (z) − fy (t)k : y, x, z ∈ H } ≤ 2. Let 2 δ < ε. Then, for all y, √ {x : ω fy (B(x, δ )) > ε} = {x ∈ B(y, δ ) : ω fy (B(x, δ )) > ε}∪ √ {x ∈ / B(y, δ ) : ω fy (B(x, δ )) > ε}. √ / B(y, δ ), then Observe that ω fy (B(x, δ )) = 2δ /kx − yk if kx − yk > δ , and so if x ∈ √ √ ω fy (B(x, δ )) ≤ 2 δ < ε, and so {x ∈ / B(y, δ ) : ω fy (B(x, δ )) > ε} = 0. / Lastly we get 4 that for all y, √ √ {x : ω fy (B(x, δ )) > ε} = {x ∈ B(y, δ ) : ω fy (B(x, δ )) > ε} ⊂ B(y, δ ). √ Now, since P(∂ B(y, δ )) = 0 we have that IB(y ,√δ ) (x) → IB(y,√δ ) (x) a.s. w.r.t. k P, whenever every y, and the dominated convergence theorem implies that √ √ yk → y for √ P(B(yk , δ )) → P(B(y, δ )). This entails that P(B(y, δ )) is a continuous function of y, so its maximum in a compact set, is attained. Let ε > 0 and Kε be a compact set such that P((Kε B(0, 1))c ) < √ ε where Kε B(0, 1) = {z ∈ Kε : d(z, Kεc ) > 1}. Denote by √ yε,δ = arg maxy∈Kε P(B(y, δ )), let us prove that for all fixed ε > 0, P(B(yε,δ , δ )) → 0 as δ → √ 0. If this is not the case there exists η > 0, yn ∈ Kε and δn → 0 such that P(B(yn , δn )) > η for all n. Since Kε is compact we can assume that yn → y for some y ∈ Kε (by considering a subsequence). From P(∂ B(x, r)) = 0 for all x, it follows that P({y}) =p0 (indeed, consider x and r > 0 such that y ∈ ∂ B(x, r)). Let us define ρn = max j≥n ( δ j +ky−y j k) and Bn = B(y, ρn ), then P(Bn ) ≥ η and B1 ⊇ B2 ⊇ . . . ⊇ Bn . . . . √ Finally, 0 = P({y}) = lim P(Bn ), which contradict that P(B(yn , δn )) > η. Now for all δ < 1, n √  √ o sup P(B(y, δ )) ≤ max sup P B(y, δ ) , P (Kε B(0, 1))c , y y∈Kε √ √ therefore supy P(B(y, δ )) ≤ max{P(B(yε,δ , δ )), ε} < ε for δ small enough, showing that (6) holds. Lastly (4) is a consequence of the uniform convergence of D(x, Pn ) to D(x, P) and the argmax argument. The following corollary states the consistency of the RFM explained in Table 1 when the sample X1 , . . . , Xn is distributed as a random variable X with a distribution P0 fulfilling H1. Corollary 1. Assume that P0 fulfils H1 and there exists a unique θ0 such that, for all l, ! θ̂1 − θ0 EP0 = 0. kθ̂1 − θ0 k Then, under P0 , θ̃ RFM → θ0 a.s., as m → ∞. Recall that a sequence of estimators {θ̂n }n is qualitatively robust at a probability distribution P if for all ε > 0 there exists δ > 0, for all probability distribution Q, Π(P, Q) < δ ⇒ Π(LP (θ̂n ), LQ (θ̂n )) < ε (see Hampel (1971)), where Π denotes the Prokhorov distance and LF (θ̂n ) denotes the probability distribution of θ̂n under F. As Π metrizes the weak convergence we have the following corollary. Corollary 2. Robustness of RFM estimators. Under the hypotheses of Corollary 1, θ̃ RFM is qualitatively robust. Remark 1. Qualitative robustness ensures the good behaviour of the estimator in a neighbourhood of P0 . However, there are some estimators that still converge to θ0 even if P is far from P0 . For instance “the shorth”, defined as the average of the observations lying on the shortest interval containing half of the data, has this property. Indeed, consider the case where P0 = U(−1, 1), P1 = U(3, 4), and P = (1 − α)P0 + αP1 , for any α < 0.5. This is also the case for the impartial trimmed estimators, the minimum volume ellipsoid, and the redescendent (with compact support) M-estimators (see subsection 2.2). If the estimators for each subsample have this property, the RFM estimator will inherit it. 5 2.2 Efficiency of the fusion of M-estimators In this section we obtain the asymptotic variance of the RFM method, for the special case of M-estimators. Recall that an M-estimator T can be defined (see section 3.2 in R Huber and Ronchetti (2009)) by the implicit functional equation ψ(x; T (F))F(dx) = 0, where ψ(x; θ ) = (∂ /∂ θ )ρ(x; θ ) and F stands for the true underlying common distribution of the observations. For instance, the Maximum Likelihood estimator is obtained with ρ(x; θ ) = − log( f (x, θ )). The estimator Tn is given √by the empirical version of T , based on a sample {X1 , . . . , Xn }. It is well known that n(Tn − T (F)) is asymptotically normal with mean 0 and variance A(F, T ) given by the integral of the square R of the influence curve, i.e. A(F, T ) = IC(x; F, T )2 F(dx), where the influence curve, IC, is ψ(x; T (F)) IC(x; F, T ) = R . − (∂ /∂ θ )ψ(x; T (F))F(dx) R For the location problem (i.e. ψ(x − µ0 )F(dx) = 0), we get IC(x, F, T ) = −ψ(x − R 2 /A(F, T ), µ0 )/ ψ 0 (x− µ0 )F(dx). The asymptotic efficiency of Tn is defined as Eff(Tn ) = σML 2 where σML is the asymptotic variance of the maximum likelihood estimator. Then the asymptotic variance of an M-estimator built from a sample Tn1 , . . . , Tnm of m Mestimators of T can be calculated easily. The strong consistency of the M-estimators under the model (see Huber (1967)) entails that θ̃ RFM built from M-estimators is consistent (see Corollary 1) whenever the empirical version of the implicit functional equation has an unique solution. The choice of m and l has an impact on the robustness of the estimator and on the computation time. Indeed, if the computation time of each θ̂i = O(l a ) and the computation time of the fusion step is O(mb ), then the optimal choice (if b > 1) is l = O(n(b−1)/(a+b−1) ). 2.3 Breakdown point for the RFM Following Donoho (1982) we consider the finite-sample breakdown point, introduced by Donoho. Intuitively the breakdown point corresponds to the maximum percentage of outliers (located at the worst possible positions) we can have in a sample before the estimate breaks in the sense that it can be arbitrarily large (or close to the boundary of the parameter space). Definition 1. Let x = {x1 , . . . , xn } be a data-set, θ an unknown parameter lying in a metric space Θ, and θ̂n = θ̂n (x) an estimate based on x. Let X p be the set of all data-sets y of size n having n − p elements in common with x: X p = {y : card(y) = n, card(x ∩ y) = n − p}. Then the breakdown point of θ̂n at x is εn∗ (θ̂n , x) = p∗ /n, where p∗ = max{p ≥ 0; ∀y ∈ X p , θ̂n (y) is bounded and also bounded away from the boundary ∂ Θ, if ∂ Θ 6= 0}. / To analyse the breakdown point of the RFM, we consider the case where the breakdown point of the robust estimators is 0.5 (high breakdown point estimators). For each observation Xi from the sample, let Bi = 1 if Xi is an outlier and 0 otherwise. Assume that the variables Bi are iid following a Bernoulli distribution with parameter p and let S j = ∑ls=1 B( j−1)l+s be the number of outliers in the subsample j, 6 for j = 1, . . . , m. The RFM estimator will breakdown if and only if there are more than m/2 cases where S j is greater than k (recall that l = 2k + 1). To take a glance of the behaviour of the breakdown point, we performed 5000 replications where we generated n = 30000 binomial random variables with parameter p. We split each of the samples of size 30000 randomly into m subsamples. Next we calculated the number of its subsamples which contained more than k 1’s (outliers). In Table 2 we report the average number of times (over the 5000 replications) that this number was greater than m/2, for different values of p and m. The best result is obtained for m = 5. m 5 10 30 50 100 150 p = 0.45 0 0 0 0 0 0 p = 0.49 0.0020 0.0088 0.0052 0.0080 0.0192 0.0278 p = 0.495 0.0820 0.1564 0.1426 0.1598 0.2162 0.2728 p = 0.499 0.3892 0.5352 0.5186 0.5412 0.6084 0.6780 Table 2: Average (over 5000 replications) of estimator breakdowns for different values of m and p and fixed n = 30000; p is the proportion of outliers. 3 Some applications of RFM In this section we will show how RFM may be used to tackle three classic statistical problems for large samples: estimating the multivariate location and scatter matrix, estimating the covariance operator for functional data, and clustering. For each problem we show how to apply our approach, given in Table 1. Solutions for many other problems may be derived from these cases (Principal Components, for example, both for non-functional and functional data). 3.1 Robust fusion for location and scatter matrix in finite dimensional spaces Given an iid random sample {X1 , . . . , Xn } in Rd , we consider the location and scatter matrix estimation problem. To perform RFM we only need to make explicit the estimators used for each of the m subsamples, and the depth function in the fusion stage. For the location parameters, we propose to use simple robust estimates, denoted by θ̂1 , . . . , θ̂m (see for instance Maronna, Martin and Yohai (2006)). For the depth function we propose to use the empirical version of (2), replacing PX by the empirical distribution Pm of {θ̂1 , . . . , θ̂m }, D(θ , Pm ) = 1 − 1 m m θ̂ j − θ , j −θk ∑ kθ̂ j=1 (7) where θ ∈ Rd , and k · k is the Euclidean distance. Equivalently, for the scatter matrix we use the depth function D(Σ, Pm ) = 1 − 1 m 7 m Σ̂ j − Σ ∑ kΣ̂ j − Σk j=1 , (8) where Σ̂1 , . . . , Σ̂m are robust estimators of the scatter matrix, the norm is kΣk = max1≤i≤d ∑dj=1 |Σi j |. Pm denotes the empirical distribution of {Σ̂1 , . . . , Σ̂m }. A simulation study is presented in Section 4. 3.2 Robust fusion for the covariance operator The estimation of the covariance operator of a stochastic process is a very important topic in FDA, which helps to understand the fluctuations of a random element, as well as to derive the principal functional components from its spectrum. Several robust and non-robust estimators have been proposed, see for instance Chakraborty and Chaudhuri (2014) and the references therein. In order to perform RFM, we introduce a new robust estimator to use for each of the m subsamples, which can be implemented using parallel computing. It is based on the notion of impartial trimming in the Hilbert–Schmidt space where the covariance operators are defined. It was introduced in Gordaliza (1991) and has been shown to be a very successful tool in robust estimation. Next, the RFM estimator is defined as the deepest point among the m estimators (‘impartial trimmed means’) corresponding to each subsample. To better understand the construction of our new estimator, we will first recall the general framework used for the estimation of covariance operators. 3.2.1 A general framework for the estimation of covariance operators Let E = L2 (I), where I is a finite interval in R, and X, X1 , . . . Xn , . . . be iid random eleR R ments taking values in E. Assume that E(X(t)2 ) < ∞ for all t ∈ I, and I I ρ 2 (s,t)dsdt < ∞, so that the covariance function, given by ρ(s,t) = E((X(t)− µ(t))(X(s)− µ(s))) where E(X(t)) = µ(t), is well defined. For notational simplicity we assume that µ(t) = 0, ∀t ∈ I. Under these conditions, the covariance operator, given by Γ0 ( f )(t) = E(hX, f iX(t)) = Z ρ(s,t) f (s)ds, f ∈ E, (9) I is diagonalizable, with non-negative eigenvalues λi such that ∑i λi2 < ∞. Moreover Γ0 belongs to the Hilbert–Schmidt space HS(E) of linear operators with norm and inner product given by kΓk2HS = ∞ ∑ kΓ(ek )k2 < ∞, hΓ1 , Γ2 iHS = ∞ ∑ hΓ1 (ek ), Γ2 (ek )i, k=1 k=1 (10) respectively, where {ek : k ≥ 1} is any orthonormal basis of E, and Γ, Γ1 , Γ2 ∈ HS(E). 2 In particular, kΓ0 k2HS = ∑∞ i=1 λi . Given an iid sample {X1 , . . . , Xn }, we define the Hilbert–Schmidt operators of rank-one, Wi : E → E, as Wi ( f ) = hXi , f iXi (.), i = 1, . . . n. Let φi = Xi /kXi k, then Wi (φi ) = kXi k2 φi =: ηi φi . The standard estimator of Γ0 is just the average of these operators, i.e. Γ̂n = 1 n n ∑i=1 Wi , which is a consistent estimator of Γ0 by the Law of Large Numbers in the space HS(E). We replace this average by a trimmed version in the space HS(E). 8 3.2.2 A new robust estimator for the covariance operator Our proposal is to consider an impartial trimmed estimator as a resistant estimator. The notion of impartial trimming was introduced in Gordaliza (1991), and the functional data setting was considered in Cuesta-Albertos and Fraiman (2006), from where one can can obtain the asymptotic theory for our setting. The construction of our estimator needs an explicit expression of the distances kWi −W j k, 1 ≤ i ≤ j ≤ n, which we will derive using the following lemma. Lemma 2. We have that di2j := kWi −W j k2HS = kXi k4 + kX j k4 − 2hXi , X j i2 for 1 ≤ i ≤ j ≤ n. (11) Proof. Let us write hWi −W j ,Wi −W j iHS = hWi ,Wi iHS + hW j ,W j iHS − 2hWi ,W j iHS ∞ = ηi2 + η 2j − 2 ∑ hWi (ek ),W j (ek )i k=1 ∞ = ηi2 + η 2j − 2 ∑ hXi , ek iXi , hX j , ek iX j k=1 ∞ = ηi2 + η 2j − 2hXi , X j i ∑ hXi , ek ihXk , ek i. k=1 Now Eq. (11) follows from the identity ∞ ∑ hXi , ek ihXk , ek i = hXi , X j i. k=1 Given the sample, which we have assumed with mean zero for notational simplicity, and 0 < α < 1, we provide a simple algorithm to calculate an approximate impartial trimmed mean estimator of the covariance operator which is strongly consistent. STEP 1: Calculate di j = kWi −W j kHS , 1 ≤ i ≤ j ≤ n, using Lemma 1. STEP 2: Let r = b(1 − α)nc + 1. For each i = 1, . . . n, consider the set of indices Ii ⊂ {1, . . . , n} corresponding to the r nearest neighbours of Wi among {W1 , . . . ,Wn }, (1) (n) and the order statistic of the vector (di1 , . . . , din ), di ≤ . . . ≤ di . (r) (r) STEP 3: Let γ = argmin{d1 , . . . , dn }. STEP 4: The impartial trimmed mean estimator of Γ0 is given by the average of the r nearest neighbours of Wγ among {W1 , . . . ,Wn }, i.e the average of the rank-one operators Wi such that i ∈ Iγ . The covariance function is then estimated by ρ̂(s,t) = 1 r ∑ j∈Iγ X j (s)X j (t). Observe that Steps 1 and 2 of the algorithm can be performed using parallel computing. The final estimator given by the RFM may be obtained by taking the deepest point (or the average of the 40% deepest points) among the m estimators obtained from the algorithm above. The norm used for the depth function in this case is the functional analogue of (8). 9 3.3 Robust fusion for cluster analysis In this section we describe a robust fusion method for clustering. Our approach is based on the use of impartial trimmed k–means (ITkM, see Cuesta-Albertos, Gordaliza and Matrán (1997)) in two steps. In the first one we apply ITkM with a given trimming level α1 to each of the m subsamples, and obtain m sets of k centres M̂1 , . . . , Mˆm . In the second step we apply ITkM with a trimming level α2 to the set ∪m i=1 M̂i , as suggested in Cuesta-Albertos, Gordaliza and Matrán (1997) (Section 5.1). We start by describing briefly ITkM. 3.3.1 Impartial trimmed k-means Given a sample {X1 , . . . , Xn } ⊂ Rd , a trimming level 0 < α < 1, and the number of clusters k, ITkM looks for a set {m̂1 , . . . , m̂k } ⊂ Rd and a partition of the space C0 ,C1 , . . . ,Ck that minimizes the loss function 1 n − [nα] k ∑ ∑ kXi − m̂ j k2 . j=1 Xi ∈C j Here, C0 is the set of trimmed data (with cardinality bnαc). Let X ∈ Rd be a random vector with distribution PX , the number of clusters k, and a trimming proportion 0 < α < 1. • For every k-set M = {m1 , . . . , mk }, with m j ∈ Rd for all j = 1, . . . , k, and x ∈ Rd , we define  d(x, M ) := min kx − m1 k, . . . , kx − mk k . • The set of trimming functions for PX at level α is defined by Z n o τα (PX ) = τ : Rd → [0, 1], measurable, fulfilling τ(x)dPX (x) ≥ 1 − α . The functions in τα (PX ) are a natural generalization of the indicator functions 1A with PX (A) = 1 − α. • For each pair (τ, M ) such that τ ∈ τα (PX ) and M ⊂ Rd with #M = k, let us consider the function V (τ, M , PX ) = R 1 τ(x)dPX (x) Z τ(x)d 2 (x, M )dPX (x). Lastly, we define V (PX ) = inf inf V (τ, M , PX ). τ∈τα (PX ) M ⊂Rd #M =k (12) Corollary 3.2 in Cuesta-Albertos, Gordaliza and Matrán (1997) proves that there exists a pair (not necessarily unique) (τ ∗ , M ∗ ) attaining the value V (PX ). Moreover, if PX is absolutely continuous w.r.t. Lebesgue measure, τ ∗ = 1B(M ∗ ,r∗ ) with r∗ = r(α, M ∗ ) = inf{r ≥ 0 : PX (B(M ∗ , r)) ≥ 1 − α} and B(M ∗ , r) = {x ∈ Rd : d(x, M ∗ ) ≤ r}. Let us denote by Pn the empirical distribution based on the sample. Theorem 3.6 in Cuesta-Albertos, Gordaliza and Matrán (1997) proves that if PX is absolutely continuous w.r.t. Lebesgue measure and there exists a unique pair (τ ∗ , M ∗ ) solving (12), then 10 V (Pn ) → V (PX ) a.s. Moreover, if Mˆ is any sequence of empirical trimmed k-means, then dH (Mˆ, M ∗ ) → 0 a.s., where dH denotes the Hausdorff distance. It is clear that in this case τ̂n = 1B(Mˆ,r̂) → τ ∗ PX a.s., where r̂ = inf{r ≥ 0 : Pn (B(Mˆ, r)) ≥ 1 − α}. M ∗ and Mˆ induce partitions of B(M ∗ , r∗ ) and B(Mˆ, r̂∗ ) respectively, into k– clusters, by defining, for i = 1, . . . , k,  Cluster Ci := x ∈ B(M ∗ , r∗ ) : kx − m∗i k ≤ min kx − m∗j k , (13) j6=i  Cluster Ĉi := x ∈ B(Mˆ, r̂) : kx − m̂i k ≤ min kx − m̂ j k . j6=i (14) The points at a boundary between clusters can be assigned arbitrarily. A functional version of ITkM can be found in Cuesta-Albertos and Fraiman (2006). With this in hand, the fusion step of the RFM is done by applying ITkM to the set of the k × m centres. The whole algorithm is summarized in Table 3. 1) Split the sample into m subsamples (recall that n = ml). 2) To each subsample, apply the empirical version of α-ITkM with α = α1 and obtain Mˆ1 , . . . , Mˆm , each one with k points in Rd . ˆ 3) Apply the empirical version of α-ITkM with α = α2 to the set ∪m i=1 Mi . ˆ 4) Obtain the output of the algorithm (MRFM , r̂RMF ). 5) Build the clusters by applying (14). Table 3: RFM algorithm for clustering 4 Simulation results We now describe the simulations done with the RFM for the three applications described in the previous sections. As the design of each simulation is specific to its application, we describe them separately. All the simulations were done using an 8-core PC, Intel core i7-3770 CPU, 8GB of RAM, 64 bit processor, with the R software package v. 3.3.0 running under Ubuntu. 4.1 Location and scatter matrix for finite dimensional spaces We use the same simulations to analyse both the location of the parameters and their scatter matrix. For the robust estimator we have applied the function CovMest in the R-package rrcov with the parameters given by default. We draw samples from a centred 5-dimensional Gaussian distribution with a covariance matrix with all its off-diagonal elements equal to 0.2. For the outliers we use a 5-dimensional Cauchy distribution with independent coordinates centred at 50. We test two contamination levels, p = 0.2 and p = 0.4. We vary the sample size n within the set {0.1E6, 5E6, 10E6} and the number of subsamples m ∈ {100, 500, 1000, 10000}. We replicate each simulation case K = 5 times and report the average. The estimators obtained by the RFM are the values which maximize the depth functions given in Eqs (7) and (8) for the location and the scatter matrix respectively. In each case, the maximization is done over the set of the m estimates obtained from the subsamples. 11 The mean squared error (averaged over 5 replicates) for the location problem are given in Table 4. The estimators considered are the following: the average of the whole sample (MLE), the average of the robust location estimators (avROB), the average of the 40% deepest robust estimators (RFM1), and the deepest robust estimator (RFM). Table 4: Location Estimators for p = 0.2 and p = 0.4. MLE n 0.1 0.1 0.1 1.0 1.0 1.0 1.0 5.0 5.0 5.0 5.0 10.0 10.0 10.0 10.0 m 100 500 1000 100 500 1000 10000 100 500 1000 10000 100 500 1000 10000 31.3 31.3 31.3 21.4 21.4 21.4 21.4 22.0 22.0 22.0 22.0 23.5 23.5 23.5 23.5 avROB RFM1 p = 0.2 0.0098 0.0124 0.0097 0.0112 0.0097 0.0109 0.0021 0.0029 0.0021 0.0037 0.0021 0.0030 0.0022 0.0035 0.0009 0.0014 0.0009 0.0010 0.0009 0.0014 0.0009 0.0015 0.0009 0.0013 0.0009 0.0012 0.0009 0.0012 0.0009 0.0012 RFM MLE 0.0297 0.0426 0.0477 0.0074 0.0110 0.0159 0.0261 0.0032 0.0056 0.0071 0.0147 0.0026 0.0038 0.0047 0.0090 44.2 44.2 44.2 44.1 44.1 44.1 44.1 45.9 45.9 45.9 45.9 47.0 47.0 47.0 47.0 avROB RFM1 p = 0.4 0.0042 0.0076 0.1070 0.0081 1.3400 0.0231 0.0038 0.0045 0.0038 0.0038 0.0038 0.0053 1.3900 0.0186 0.0007 0.0014 0.0007 0.0014 0.0007 0.0014 0.0013 0.0011 0.0005 0.0010 0.0005 0.0009 0.0005 0.0008 0.0005 0.0009 RFM 0.0288 0.0427 0.0416 0.0087 0.0164 0.0186 0.0375 0.0044 0.0073 0.0097 0.0159 0.0033 0.0052 0.0056 0.0102 We can see that the estimator obtained by the RFM behaves very well. Depending on the structure of the outliers, the mean of the robust estimates may behave well or not. Even if only one of the subsamples contains a high proportion of outliers causing the robust estimator to break down, the average of the robust estimators will break down. On the other hand, the deepest M-estimator always behaves well. The performances of both estimators decrease in general with m. The estimation errors for the covariance are given in Table 5 (p = 0.2) and Table 6 (p = 0.4). We compare the MLE estimator (MLE), a robust estimator based on the whole sample (ROB), the average of the robust scatter matrix estimators (avROB), the average of the 40% deepest robust estimators (RFM1), and the deepest robust estimator (RFM). We also report the average time in seconds necessary for both the global estimator (T0, over the whole sample), and T1, the estimator obtained by fusion (including computing the estimators over subsamples and aggregating them by fusion). Since the second step of the algorithm (see point b) in Table 1) can be parallelized, in practice the computational time T1 can be divided almost by m. The results of RFM are very good for the covariance matrix as well. 4.2 Covariance operator To generate the data, we have used a simplified version of the simulation model used in Kraus and Panaretos (2012): √ 10 √ 10 X(t) = µ(t) + 2 ∑ λk ak sin(2πkt) + 2 ∑ νk bk cos(2πkt), k=1 k=1 k where νk = 13 , λk = k−3 , and ak and bk are random standard Gaussian independent observations. The central observations were generated using µ(t) = 0 whereas for the 12 Table 5: Covariance estimators. Using MLE and robust estimates over the entire sample, and aggregating by average, trimmed average or fusion of m subsamples estimators, p = 0.2. n 0.1 0.1 0.1 1.0 1.0 1.0 1.0 5.0 5.0 5.0 5.0 10.0 10.0 10.0 10.0 m 100 500 1000 100 500 1000 10000 100 500 1000 10000 100 500 1000 10000 T0 0.460 0.460 0.460 3.444 3.444 3.444 3.444 20.528 20.528 20.528 20.528 42.174 42.174 42.174 42.174 T1 4.205 14.527 23.992 6.524 24.028 45.307 945.350 15.984 33.289 68.342 1312.800 29.168 49.992 73.701 1291.000 MLE 23688000 23688000 23688000 1617200 1617200 1617200 1617200 1981900 1981900 1981900 1981900 28135000 28135000 28135000 28135000 ROB 0.2594 0.2594 0.2594 0.2342 0.2342 0.2342 0.2342 0.2317 0.2317 0.2317 0.2317 0.2307 0.2307 0.2307 0.2307 avROB 0.2597 0.2675 0.2748 0.2345 0.2353 0.2360 0.2464 0.2316 0.2316 0.2318 0.2342 0.2306 0.2307 0.2308 0.2323 RFM1 0.2598 0.2498 0.2418 0.2368 0.2371 0.2381 0.2075 0.2340 0.2331 0.2336 0.2267 0.2322 0.2322 0.2315 0.2290 RFM 0.3722 0.4810 0.6130 0.2656 0.3189 0.3295 0.4982 0.2495 0.2687 0.2842 0.3810 0.2445 0.2567 0.2660 0.3439 Table 6: Covariance estimators. Using MLE and robust estimates over the entire sample, and aggregating by average, trimmed average or fusion the m subsamples estimators, p = 0.4. n 0.1 0.1 0.1 1.0 1.0 1.0 1.0 5.0 5.0 5.0 5.0 10.0 10.0 10.0 10.0 m 100 500 1000 100 500 1000 10000 100 500 1000 10000 100 500 1000 10000 T0 0.581 0.581 0.581 2.631 2.631 2.631 2.631 16.651 16.651 16.651 16.651 33.922 33.922 33.922 33.922 T1 3.416 13.614 22.827 5.622 21.131 39.752 833.530 14.126 30.762 60.389 1239.300 24.757 43.999 68.787 1486.100 MLE 448210 448210 448210 6030100 6030100 6030100 6030100 29809000 29809000 29809000 29809000 93071000 93071000 93071000 93071000 ROB 0.8247 0.8247 0.8247 0.8081 0.8081 0.8081 0.8081 0.8010 0.8010 0.8010 0.8010 0.7988 0.7988 0.7988 0.7988 avROB 0.8348 16.3120 205.3000 0.8094 0.8143 0.8201 203.9100 0.8012 0.8021 0.8032 0.9024 0.7989 0.7993 0.8001 0.8117 RFM1 0.8378 0.8057 0.7772 0.8114 0.8103 0.8059 0.7706 0.8035 0.8025 0.8020 0.7877 0.8013 0.8007 0.8001 0.7939 RFM 1.0111 1.2548 1.5159 0.8790 0.9462 0.9690 1.2760 0.8299 0.8571 0.8740 1.0311 0.8185 0.8420 0.8555 0.9403 outliers we took µ(t) = 2 − 8 sin(πt). For t we used an equally spaced grid of T = 20 points in [0, 1]. 10 The covariance operator of this process, given by √ Cov(s,t) = ∑k=1 Ak (s)Ak (t)+Bk (s)Bk (t), √ where Ak (t) = 2λk sin(2πkt) and Ak (t) = 2νk cos(2πkt), was computed for the comparisons. We varied the sample size n within the set {0.1E6, 1E6, 5E6, 10E6} and the number of subsamples m ∈ {100, 500, 1000, 10000}. The proportion of outliers was fixed to p = 0.15 and p = 0.20. We replicated each simulation case K = 5 times and report the average performance over the replicates. We report also the average time in seconds necessary for both a global estimate T0, over the whole sample, and T1, the estimate obtained by fusion (including computing 13 Figure 1: Simulated functions and outliers the estimates over subsamples and aggregating them by fusion). We compare the classical estimator (MLE), the global robust estimate (ROB), the average of the robust estimates from the subsamples (avROB) and the robust fusion estimate (RFM). The results are shown in Tables 7 and 8 for two proportions of outliers, p = 0.15 and p = 0.2 respectively. Table 7: Covariance operator estimator. Using the classical and robust estimators over the entire sample, and aggregating by average or fusion of m subsamples estimators. p = 0.15, T = 20. n 0.05 0.05 0.05 0.05 0.10 0.10 0.10 0.10 m 20 50 100 1000 20 50 100 1000 T0 553 543 528 459 2300 2300 2290 1850 T1 18.20 7.81 4.79 19.40 69.00 28.10 15.20 21.60 MLE 24.3 24.3 24.3 24.3 24.2 24.2 24.2 24.3 ROB 5.16 5.20 5.20 5.13 5.14 5.04 5.06 5.21 avROB 5.21 5.24 5.17 5.54 5.22 5.09 5.15 5.35 RFM 5.52 5.60 5.58 6.58 5.43 5.13 5.43 6.13 Table 8: Covariance operator estimator. Using classical and robust estimators over the entire sample, and aggregating by average or fusion of m subsamples estimators. p = 0.2, T = 20. n 0.05 0.05 0.05 0.05 0.10 0.10 0.10 0.10 m 20 50 100 1000 20 50 100 1000 T0 572 649 633 478 1970 2030 2020 1840 T1 17.90 7.88 4.61 19.50 69.10 28.10 15.10 21.60 MLE 30.5 30.5 30.5 30.5 30.4 30.4 30.4 30.4 14 cvRob 0.879 0.876 0.839 0.864 0.914 0.921 0.840 0.961 avROB 3.96 7.34 8.86 13.10 3.83 4.32 8.44 12.10 RFM 1.45 2.10 2.43 7.08 1.36 1.55 2.35 5.20 If the proportion of outliers is moderate, p = 0.15, the average of the robust estimators still behaves well, better than RFM, but if we increase the proportion of outliers to p = 0.2, RFM clearly outperforms all the other estimators. 4.3 Clustering We performed a simulation study for large sample sizes, using a model with three clusters with outliers, introduced in Cuesta-Albertos, Gordaliza and Matrán (1997). The data were generated using bivariate Gaussian distributions with the following parameters for the clusters and the outliers respectively: µ1 = (0, 0), µ2 = (0, 10), µ3 = (6, 0), µ4 = (2, 10/3), Σ1 = Σ2 = Σ3 = 1.5 × Id, Σ4 = 20 × Id where Id is the two dimensional identity matrix. The outliers were generated with µ4 , Σ4 and n4 . The sizes of the clusters were fixed at the following values: n1 = 15, n2 = 30, n3 = 30, n4 = 40. As in Cuesta-Albertos, Gordaliza and Matrán (1997), the outliers lying in the 75% level confidence ellipsoids of the clusters were replaced by others not belonging to that area. The outliers represent almost 35% of the whole sample. We used this base simulation and varied the whole sample size, multiplying each ni by a factor “fac” taking the values in {10, 100, 1000, 10000}. So for the smallest sample, we have n = 1150, and the largest, n = 1150000. 20 RFM output 20 TkMeans output 20 True Clusters ● ● ● ● ● ● ● ● −5 ● ● ● −5 0 10 5 ● ● ● ● ● −15 ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ●●●● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ●● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ● 0 0 5 10 ● −10 −10 −5 0 5 10 ● ● ● ● ● ● ●●●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●●●● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ●● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● 15 ● ● −5 ● ● ● ● ●●●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ●● ● ●●●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ●● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● −10 15 ● ● ● ● ● ● ● ● ● 15 ● ● ● ● ● ● ● ● 5 10 15 −15 −5 0 ● 5 10 15 −15 −5 0 5 10 15 Figure 2: Left panel: the true clusters. Middle panel: Results obtained by ITkM over the whole sample. Right panel: The output obtained by RFM using m = 100 subsamples. The outliers are the blue points and n = 11500. For each value of n we varied the number of subsamples m within the values {10, 50, 100, 1000, 10000} with the restriction m < fac. Lastly, when applying the trimmed k-means to the samples, we have tested three values for the trimming level, α1 = 0.2, 0.35, 0.45, whereas for the fusion we fixed α2 = 0.1. The left hand panel of Figure 2 shows an example of the simulated data-set for n = 11500, the middle panel shows the results obtained by ITkM applied to the whole sample, and the right hand panel shows the output of the algorithm. 15 The partitions obtained by each approach are compared to the true clusters using the matching error defined by ME = min s∈S (k+1) 1 n ∑ 1{yi 6=s(ŷi )} n i=1 (15) where S (k + 1) is the set of permutations of {0, 1, . . . , k + 1}, yi is the true cluster of observation i and ŷi is the cluster assigned by the algorithm. The results of the simulation are given in Table 9, where we compare the RFM method, with the ITkM calculated with the whole sample. Columns ME1 and ME2 give the matching errors for ITkM applied to the whole sample and for RFM respectively. We also report the average time in seconds necessary for both the global estimator (T0, over the whole sample), and T1, the estimator obtained by fusion (including computing the estimators over subsamples and aggregating them by fusion). Finally T2 is the time using parallel computing. As expected, the RFM matching errors are often higher than those of ITkM applied to the whole sample. But the loss of performance is very small in general and increases with m. For the smallest values of m with large samples (n > 10000), RFM has almost the same performance for all values of α. On the other hand, increasing the value of m reduces considerably the computation time of RFM. 4.3.1 A real data example As an example we have chosen the MNIST data-set of handwritten digits (see https://www.kaggle.com/c/digitrecognizer/data) to compare the performance of the RFM clustering algorithm with the same clustering procedure without splitting the sample (impartial trimmed k-means). The digits have been size-normalized and centred in a fixed-size image of 28 × 28 pixels. The data-set consist of a training sample {(X1 ,Y1 ), . . . , (Xn ,Yn )} of n = 42000 data, and a test sample of 10000 data. As it is explained in the aforementioned link: “this classic dataset of handwritten images has served as the basis for benchmarking classification algorithms”. However, as we are interested in clustering we will use only the sample {X1 , . . . , Xn }, searching for k = 10 groups. This is a very difficult task: if the labels are chosen at random the probability to get at least half of the 42000 data well identified is extremely close to zero. We cluster the 42000 data using both methods. The design is the same as for the previous simulations. On the one hand we cluster the whole sample using the impartial trimmed k-mean algorithm for k = 10. On the other hand we use the RFM clustering method given in Table 3 for m = 10, 100, 500 and 1000, with α1 = 0.05 and α1 = 0.1. The labels Y1 , . . . ,Yn are only used to calculate the misclassification error rates ME1 and ME2 defined in (15). The results are given for α1 = 0.05 and α2 = 0.1 in Table 10 left, and for α1 = 0.1 = α2 in Table 10 right . They show that: (a) this clustering problem is very difficult (b) the relative efficiencies of the RFM clustering procedures for α1 = 0.05 are 5%, 2%, 9% and 6% while the computational times fall down drastically to 17%, 5%, 3% and 3%, for m = 10, 100, 500 and 1000 respectively. For α1 = 0.1 the efficiencies are 8%, 7%, 9% and 8%, the computational times fall down to 16%, 7%, 5% and 4% for m = 10, 100, 500 and 1000 respectively. 16 Table 9: RFM for Clustering using different values of the trimming parameter α1 . n m T0 1150 11500 11500 115000 115000 115000 1150000 1150000 1150000 1150000 10 10 100 10 100 1000 10 100 1000 10000 2.89 21.20 21.20 274.90 274.90 274.90 3452.00 3452.00 3452.00 3452.00 1150 11500 11500 115000 115000 115000 1150000 1150000 1150000 1150000 10 10 100 10 100 1000 10 100 1000 10000 3.45 37.87 37.87 427.70 427.70 427.70 4925.00 4925.00 4925.00 4925.00 1150 11500 11500 115000 115000 115000 1150000 1150000 1150000 1150000 10 10 100 10 100 1000 10 100 1000 10000 2.72 55.58 55.58 698.90 698.90 698.90 7190.00 7190.00 7190.00 7190.00 T1 α1 = 0.2 1.34 21.69 14.65 263.80 218.10 141.50 3149.00 2609.00 2158.00 1434.00 α1 = 0.35 1.43 33.38 15.30 391.10 307.20 137.70 4284.00 3660.00 3052.00 1397.00 α1 = 0.45 1.27 34.12 13.11 586.60 323.90 122.50 7087.00 5654.00 3287.00 1258.00 T2 ME1 ME2 0.55 6.83 4.24 75.11 56.44 37.51 873.40 680.70 546.90 374.70 0.1539 0.1594 0.1594 0.1585 0.1585 0.1585 0.1582 0.1582 0.1582 0.1582 0.1678 0.1603 0.1693 0.1585 0.1591 0.1693 0.1582 0.1583 0.1590 0.1689 0.54 9.89 4.29 109.60 85.70 38.36 1166.00 928.20 792.90 372.20 0.1287 0.1037 0.1037 0.1049 0.1049 0.1049 0.1052 0.1052 0.1052 0.1052 0.1310 0.1071 0.1343 0.1050 0.1071 0.1331 0.1053 0.1055 0.1074 0.1336 0.52 9.80 3.65 170.40 86.35 33.53 2115.00 1508.00 829.60 328.10 0.1330 0.1370 0.1370 0.1325 0.1325 0.1325 0.1327 0.1327 0.1327 0.1327 0.1567 0.1403 0.1723 0.1330 0.1355 0.1729 0.1328 0.1330 0.1360 0.1726 Table 10: Robust clustering, α2 = 0.1, α1 = 0.05 (left) and α1 = 0.1 (right) . m 10 100 500 1000 5 T0 8560 8560 8560 8560 T1 1540 503 244 253 ME1 0.477 0.477 0.477 0.477 ME2 0.503 0.486 0.520 0.508 m 10 100 500 1000 T0 9570 9570 9570 9570 T1 1500 705 445 417 ME1 0.492 0.492 0.492 0.492 ME2 0.530 0.525 0.536 0.532 Concluding remarks We have addressed some fundamental statistical problems in the context of Big Data, namely large samples, in the presence of outliers; location and covariance estimation, covariance operator estimation, and clustering. We have proposed a general robust approach, called the robust fusion method (RFM), and shown how it may be applied to these problems. The simulations gave very good results mainly for the last two problems. Different statistical challenges go through these problems. Our approach may be 17 adapted to any other task as soon as a robust efficient estimate is available for the corresponding problem. • We have addressed one of the important problems in Big Data, namely when the size of the data-set is too large and one needs to split it into pieces. • In this setup we think that robustness is mandatory. • We have provided a general procedure, a robust fusion method, to deal with these problems. The method is very general and can be applied to different statistical problems for high dimensional and functional data. • Robust methods should be reasonably simple, in order to work with very large samples. • We have provided a new robust method (RFM) to estimate the covariance operator in the functional data setting. • As particular cases we considered the multivariate location problem, the scatter matrix, the covariance operator, and clustering methods. We have illustrated through simulated examples the behaviour of RFM for all these problems for different (large) sample sizes. Acknowledgement. To the constructive comments and criticisms from an associated editor and two anonymous referees. For the last author this work has been partially supported by the ECOS project: No. U14E02. References Aho, A., Hopcroft, J.E., and Ullman, J.D. (1974) The design and analysis of computer algorithms. Addison-Wesley Pub. Co. Ahmed, S. Ejaz (Eds) (2017) Big and Complex Data Analysis. Methodologies and Applications. Springer-Verlag, Berlin. Aneiros, G., Bongiorno, E.G., Cao, R., and Vieu, P. (Eds) Functional Statistics and Related Fields. Springer-Verlag, Berlin. Billingsley, P., and Topsøe, F. (1967). Uniformity in weak convergence. Z. Wahrs. und Verw. Gebiete 7 1–16. Chakraborty, A., and Chaudhuri, P. (2014) The spatial distribution in infinite dimensional spaces and related quantiles and depths. Annals of Statistics 42(3) 1203–1231. Chaudhuri, P. (1996) On a geometric notion of quantiles for multivariate data. Journal of the American Statistical Association 91(343) 862–872. Cuestas-Albertos, J. A., and Fraiman, R. (2006) Impartial means for functional data. In: R. Liu, R. Serfling, and D. Souvaine, eds, Data Depth: Robust Multivariate Statistical Analysis, Computational Geometry &amp;amp; Applications. Vol. 72 in the DIMACS Series of the American Mathematical Society, pp. 121–145. Cuesta-Albertos, J.A., Gordaliza, A., and Matrán C. (1997) Trimmed k-means: An attempt to robustify quantizers. Annals of Statistics 25 553–576. 18 Cuevas, A. (2014) A partial overview of the theory of statistics with functional data. Journal of Statistical Planning and Inference 147 1–23. Donoho, D.L. (1982) Breakdown properties of multivariate location estimators. Ph.D. qualifying papers, Dept. of Statistics, Harvard University. Ferraty, F., and Vieu, P. (2006) Nonparametric Functional Data Analysis. SpringerVerlag, Berlin. Goia, A., and Vieu, P. (2016) Special Issue on Statistical Models and Methods for High or infinite Dimensional Spaces. Journal of Multivariate Analysis. 146, 1–352. Gordaliza, A. (1991) Best approximations to random variables based on trimming procedures. J. Approx. Theory 64(2) 162180. Hovárt, L., and Kokoszka, P. (2012) Inference for Functional Data with Applications. Springer-Verlag, Berlin. Huber, P. J., and Ronchetti, E. M. (2009) Robust Statistics. Wiley, Hoboken, NJ. Hampel, F.R. (1971) A general qualitative definition of robustness. The Annals of Mathematical Statistics. Vol. 42 (6), 1887–1896. Huber, P. (1967) The behavior of maximum likelihood estimates under nonstandard conditions. Proc. Fifth Berkeley Symp. on Math. Statist. and Prob. Vol. 1. Univ. of Calif. Press, Berkeley, CA, pp. 221–233 Kraus, D., and Panareto, V.M. (2012) Dispersion operators and resistant second-order functional data analysis. Biometrika 101(1), 141–154. Maronna, R., Martin, R., and Yohai, V. (2006) Robust Statistics: Theory and Methods. Wiley, Hoboken, NJ. Tang, L., Zhou, L. and Song, P. X.-K. (2016) Method of divide-and-combine in regularised generalised linear models for big data. https://arxiv.org/abs/1611.06208. Vardi, Y., and Zhang, C. (2000) The multivariate L1-median and associated data depth. Proc. Nat. Acad. Sci. USA 97(4) 1423–1426. Wang, C., Chen, M.-H., Schifano, E., Wu, J., and Yan, J. (2016). Statistical methods and computing for big data. Statistics and Its Interface 9(4), 399414. Yu, B. (2014). Let Us Own Data Science. IMS Bulletin Online 43(7). 19
10math.ST
Multi-Erasure Locally Recoverable Codes Over Small Fields arXiv:1709.09770v1 [cs.IT] 28 Sep 2017 ∗ Electrical Pengfei Huang∗ , Eitan Yaakobi† , and Paul H. Siegel∗ and Computer Engineering Dept., University of California, San Diego, La Jolla, CA 92093 U.S.A † Computer Science Dept., Technion – Israel Institute of Technology, Haifa 32000, Israel {pehuang,psiegel}@ucsd.edu, [email protected] Abstract—Erasure codes play an important role in storage systems to prevent data loss. In this work, we study a class of erasure codes called Multi-Erasure Locally Recoverable Codes (ME-LRCs) for storage arrays. Compared to previous related works, we focus on the construction of ME-LRCs over small fields. We first develop upper and lower bounds on the minimum distance of ME-LRCs. Our main contribution is to propose a general construction of ME-LRCs based on generalized tensor product codes, and study their erasure-correcting properties. A decoding algorithm tailored for erasure recovery is given, and correctable erasure patterns are identified. We then prove that our construction yields optimal ME-LRCs with a wide range of code parameters, and present some explicit ME-LRCs over small fields. Finally, we show that generalized integrated interleaving (GII) codes can be treated as a subclass of generalized tensor product codes, thus defining the exact relation between these codes. I. I NTRODUCTION Recently, erasure codes with both local and global erasurecorrecting properties have received considerable attention [3], [9], [17]–[19], [21], thanks to their promising application in storage systems. The idea behind them is that when only a few erasures occur, these erasures can be corrected fast using only local parities. If the number of erasures exceeds the local erasure-correcting capability, then the global parities are invoked. In this paper, we consider this kind of erasure codes with both local and global erasure-correcting capabilities for a ρ × n0 storage array [3], where each row contains some local parities, and additional global parities are distributed in the array. The array structure is suitable for many storage applications. For example, consider a redundant array of independent disks (RAID) type of architecture for solid-state drives (SSDs) [3], [8]. In this scenario, a ρ × n0 storage array can represent a total of ρ SSDs, each of which contains n0 flash memory chips. Within each SSD, an erasure code is applied to these n0 chips for local protection. In addition, erasure coding is also done across all the SSDs for global protection of all the chips. More specifically, let us give the formal definition of this class of erasure codes as follows. Definition 1. Consider a code C over a finite field Fq consisting of ρ × n0 arrays such that: 1) Each row in each array in C belongs to a linear local code C0 with length n0 and minimum distance d0 over Fq . 2) Reading the symbols of C row-wise, C is a linear code with length ρn0 , dimension k, and minimum distance d over Fq . Then, we say that C is a (ρ, n0 , k; d0 , d)q Multi-Erasure Lo cally Recoverable Code (ME-LRC). Thus, a (ρ, n0 , k; d0 , d)q ME-LRC can locally correct d0 − 1 erasures in each row, and is guaranteed to correct a total of d − 1 erasures anywhere in the array. Our work is motivated by a recent work by Blaum and Hetzler [3]. In their work, the authors studied ME-LRCs where each row is a maximum distance separable (MDS) code, and gave code constructions with field size q > max{ρ, n0 } using generalized integrated interleaving (GII) codes [11], [22], [24]. Our Definition 1 generalizes the definition of the codes in [3] by not requiring each row to be an MDS code. There exist other related works. The ME-LRCs in Definition 1 can be seen as (r, δ ) LRCs with disjoint repair sets. A code C is called an (r, δ ) LRC [19], if for every coordinate, there exists a punctured code (i.e., a repair set) of C with support containing this coordinate, whose length is at most r + δ − 1, and whose minimum distance is at least δ. Although the existing constructions [19], [21] for (r, δ ) LRCs with disjoint repair sets can generate ME-LRCs as in Definition 1, they use MDS codes as local codes and require a field size that is at least as large as the code length. A recent work [1] gives explicit constructions of (r, δ ) LRCs with disjoint repair sets over field Fq from algebraic curves, whose repair √ √ sets have size r + δ − 1 = q or r + δ − 1 = q + 1. Partial MDS (PMDS) codes [2] are also related to but different from ME-LRCs in Definition 1. In general, an ME-LRC is not a PMDS code which needs to satisfy more strict requirements. A ρ × n0 array code is called an (r; s) PMDS code if each row is an [n0 , n0 − r, r + 1]q MDS code and whenever any r locations in each row are punctured, the resulting code is also an MDS code with minimum distance s + 1. The construction ρn of (r, s) PMDS codes for all r and s with field size O(n0 0 ) was known [6]. More recently, a family of PMDS codes with field size O(max{ρ, nr0+s }s ) was constructed [7]. To the best of our knowledge, however, the construction of optimal ME-LRCs over any small field (e.g., the field size less than the length of the local code, or even the binary field) has not been fully explored and solved. The goal of this paper is to study ME-LRCs over small fields. We propose a general construction based on generalized tensor product codes [15], [23], which were first utilized in [12] to construct binary single-erasure LRCs [9], [10], [13], [14], [18], [21]. The contributions of this paper are: 1) We extend our previous construction in [12] to the scenario of multi-erasure LRCs over any field. As a result, the construction in [12] can be seen as a special case of the construction proposed in this paper. 2) In contrast to [3], our construction does not require field size q > max{ρ, n0 }, and it can even generate binary ME-LRCs. We derive upper and lower bounds on the minimum distance of ME-LRCs. For 2d0 > d, we show that our construction can produce optimal ME-LRCs with respect to (w.r.t.) the new upper bound on the minimum distance. 3) We present an erasure decoding algorithm and its corresponding correctable erasure patterns, which include the pattern of any d − 1 erasures. We show that the ME-LRCs from our construction based on Reed-Solomon (RS) codes are optimal w.r.t. certain correctable erasure patterns. 4) So far the exact relation between GII codes [3], [22], [24] and generalized tensor product codes has not been fully investigated. We prove that GII codes are a subclass of generalized tensor product codes. As a result, the parameters of a GII code can be obtained by using the known properties of generalized tensor product codes. The remainder of the paper is organized as follows. In Section II, we study field size dependent upper and lower bounds for ME-LRCs. In Section III, we propose a general construction of ME-LRCs. The erasure-correcting properties of these codes are studied and an erasure decoding algorithm is presented. In Section IV, we study optimal code construction and give several explicit optimal ME-LRCs over different fields. In Section V, we prove that GII codes are a subclass of generalized tensor product codes. Section VI concludes the paper. Throughout the paper, we use the notation [n] to denote the set {1, . . . , n}. For a length-n vector v over Fq and a set I ⊆ [n], the vector vI denotes the restriction of the vector v to coordinates in the set I , and wq (v) represents the Hamming weight of the vector v over Fq . The transpose of a matrix H is written as H T . For a set S , |S| represents the cardinality of the set. A linear code C over Fq of length n, dimension k, and minimum distance d will be denoted by C = [n, k, d]q or [n, k, d]q for simplicity. For a code with only one codeword, the minimum distance is defined as ∞. II. U PPER AND L OWER B OUNDS FOR ME-LRC S In this section, we derive field size dependent upper and lower bounds on the minimum distance of ME-LRCs. The upper bound obtained here will be used to prove the optimality of our construction for ME-LRCs in the following sections. Now, we give an upper bound on the minimum distance of a (ρ, n0 , k; d0 , d)q ME-LRC, by extending the shortening bound for LRCs in [5]. (q) Let dopt [n, k] denote the largest possible minimum distance of a linear code of length n and dimension k over Fq , and (q) kopt [n, d] let denote the largest possible dimension of a linear code of length n and minimum distance d over Fq . Lemma 2. For any (ρ, n0 , k; d0 , d)q ME-LRC C , the minimum distance d satisfies { } (q) dopt [ρn0 − xn0 , k − xk∗ ] , (1) d6 min k 06x6⌈ k∗ ⌉−1, x∈Z and the dimension satisfies { } (q) k6 min xk∗ + kopt [ρn0 − xn0 , d] , 06x6⌈ kk∗ ⌉−1, x∈Z (2) (q) where k∗ = kopt [n0 , d0 ]. Proof: See Appendix A. An asymptotic lower bound for ME-LRCs with local MDS codes was given in [1]. Here, by simply adapting the GilbertVarshamov (GV) bound [20], we have the following GV-like lower bound on ME-LRCs of finite length without specifying local codes. Lemma 3. A (ρ, n0 , k; > d0 , > d)q ME-LRC C exists, if ( ) d−2 (ρ (n − ⌈log ∑d0 −2 (n0 −1)( q − 1 ) j ⌉) − 1 ) 0 q j=0 j ( q − 1 )i ∑ i i =0 ( d −2 n −1 ) j 0 0 ρ(n −⌈logq ∑ j= 0 ( j )(q−1 ) ⌉)−k <q 0 . (3) Proof: See Appendix B. III. ME-LRC S FROM G ENERALIZED T ENSOR P RODUCT C ODES : C ONSTRUCTION AND D ECODING Tensor product codes, first proposed by Wolf in [23], are a family of binary error-correcting codes defined by a paritycheck matrix that is the tensor product of the parity-check matrices of two constituent codes. Later, they were generalized in [15]. In this section, we first introduce generalized tensor product codes over Fq . Then, we give a general construction of ME-LRCs from generalized tensor product codes. The minimum distance of the constructed ME-LRCs is determined, a decoding algorithm tailored for erasure correction is proposed, and corresponding correctable erasure patterns are studied. A. Generalized Tensor Product Codes over Fq We start by presenting the tensor product operation of two ′ ′′ ′ matrices H and H . Let H be the parity-check matrix of a code with length n′ and dimension n′ − v over Fq . The ma′ trix H can be considered as a v (row) by n′ (column) matrix over Fq or as a 1 (row) by n′ (column) matrix of elements ′ ′ ′ ′ ′ from Fqv . Let H be the vector H = [h1 h2 · · · hn′ ], where ′ ′′ h j , 1 6 j 6 n′ , are elements of Fqv . Let H be the paritycheck matrix of a code of length ℓ and dimension ℓ − λ over ′′ Fqv . We denote H by  ′′  ′′ h11 · · · h1ℓ ′′  ..  , .. H =  ... . .  ′′ hλ 1 ′′ ′′ hλ ℓ ··· where hi j , 1 6 i 6 λ and 1 6 j 6 ℓ, are elements of Fqv . ′′ as ′ The tensor product of the matrices H and H is defined HTP  ′′ ′ ′′ ′ h11 H ′′ ⊗ ′  .. =H H = . hλ 1 H ··· .. . ··· ′′ ′  h1ℓ H  .. , . ′′ hλ ℓ H ′ ′′ ′ ′′ ′ ′′ ′ ′′ ′ where hi j H = [hi j h1 hi j h2 · · · hi j hn′ ], 1 6 i 6 λ and 1 6 j 6 ℓ, and the products of elements are calculated according to the rules of multiplication for elements over Fqv . The matrix HTP will be considered as a vλ × n′ ℓ matrix of elements from Fq , thus defining a tensor product code over Fq . We provide an example to illustrate these operations. ′′ Example 1. (cf. [23]) Let H be the following parity-check matrix over F4 for a [5, 3, 3]4 code where α is a primitive element of F4 , ] [ 0 0 α0 α0 ′′ α α 0 H = . 0 α0 α0 α1 α2 ′ Let H be the following parity-check matrix over F2 for a [3, 1, 3]2 Hamming code, [ ] ′ H = 1 0 1 . 0 1 1 [ ] 1 Representing the elements of F4 as α 0 = , α1 = 0 [ ] [ ] [ ] 0 , α 2 = 1 , and 0 = 0 , we have 1 0 1 ′′ ⊗ ′ HTP = H H [ 0 1 2 ] α α α 0 0 0 α0 α1 α2 α0 α1 α2 α0 α1 α2 . = 0 0 0 α0 α1 α2 α0 α1 α2 α1 α2 α0 α2 α0 α1 Using the same symbol-to-binary vector mapping, we represent HTP over F2 as   101 000 101 101 101 HTP =  0 1 1 0 0 0 0 1 1 0 1 1 0 1 1  , 000 101 101 011 110 000 011 011 110 101 which defines a binary [15, 11, 3]2 code.  Our construction of ME-LRCs is based on generalized ten′ ′′ sor product codes [15]. Define the matrices Hi and Hi for ′ i = 1, 2, . . . , µ as follows. The matrix Hi is a vi × n′ matrix over Fq such that the (v1 + v2 + · · · + vi ) × n′ matrix  ′  H1  ′   H  Bi =  . 2   ..  ′ Hi is a parity-check matrix of an [n′ , n′ − v1 − v2 − · · · − ′ vi , di′ ]q code Ci , where d′1 6 d′2 6 · · · 6 di′ . The matrix ′′ Hi is a λi × ℓ matrix over Fqvi , which is a parity-check ′′ matrix of an [ℓ, ℓ − λi , δi ]qvi code Ci . We define a µ -level generalized tensor product code over Fq as a linear code having a parity-check matrix over Fq in the form of the following µ -level tensor product structure  ′′ ⊗ ′  H1 H1  ′′ ⊗ ′   H H  H =  2 . 2 . (4)   .. ′ ′′ ⊗ Hµ Hµ As the matrix HTP , each level in the matrix H is obtained by operations over Fq and its extension field. We denote this µ code by CGTP . Its length is nt = n′ ℓ and the dimension is kt = nt − ∑µi=1 vi λi . By adapting Theorem 2 in [15] from the field F2 to Fq , we directly have the following theorem on the minimum distance µ of CGTP over Fq . Theorem 4. The minimum distance dt of a generalized tensor µ over Fq satisfies product code CGTP dt > min{δ1 , δ2 d′1 , δ3 d′2 , . . . , δµ dµ′ −1 , dµ′ }. Proof: See Appendix C. B. Construction of ME-LRCs Now, we present a general construction of ME-LRCs based on generalized tensor product codes. Construction A ′ Step 1: Choose vi × n′ matrices Hi over Fq and λi × ℓ ma′′ trices Hi over Fqvi , for i = 1, 2, . . . , µ , which satisfy the following two properties: ′′ 1) The parity-check matrix H1 = Iℓ×ℓ , i.e., an ℓ × ℓ identity matrix. ′ ′′ 2) The matrices Hi (or Bi ), 1 6 i 6 µ , and H j , 2 6 j 6 µ , are chosen such that dµ′ 6 δ j d′j−1 , for j = 2, 3, . . . , µ . Step 2: Generate a parity-check matrix H over Fq according ′ ′′ to (4) with the matrices Hi and Hi , for i = 1, 2, . . . , µ . The constructed code corresponding to the parity-check matrix H is referred to as C A .  Theorem 5. The code C A is a (ρ, n0 , k; d0 , d)q ME-LRC with parameters ρ = ℓ, n0 = n′ , k = n′ ℓ − ∑µi=1 vi λi , d0 = d′1 , and d = dµ′ . Proof: According to Construction A, the code parameters ρ, n0 , k, and d0 can be easily determined. In the following, we prove that d = dµ′ . ′′ Since δ1 = ∞ (H1 is the identity matrix) and dµ′ 6 δi di′−1 for all i = 2, 3, . . . , µ , from Theorem 4, d > dµ′ . Now, we show that d 6 dµ′ . For i = 1, 2, . . . , µ , ′ ′ ′ let Hi = [h1 (i), . . . , hn′ (i)] over Fqvi , and let ′′ ′′ [h11 (i), . . . , hλi 1 (i)]T over Fqvi be the first column of ′′ Hi . Since the code with parity-check matrix Bµ has minimum distance dµ′ , there exist dµ′ columns of Bµ , say in the set of positions J = {b1 , b2 , . . . , bdµ′ }, which ′ are linearly dependent; that is, ∑ j∈ J α j h j (i ) = 0, for we have some α j ∈ Fq , for all i = (1, 2, . . . , µ . Thus, ) ′′ ′ ′′ ′ = 0, for ∑ j∈ J α j h p1 (i )h j (i ) = h p1 (i ) ∑ j∈ J α j h j (i ) p = 1, 2, . . . , λi and i = 1, 2, . . . , µ . That is, the columns in positions b1 , b2 , . . . , bdµ′ of H are linearly dependent. C. Erasure Decoding and Correctable Erasure Patterns We present a decoding algorithm for the ME-LRC C A from Construction A, tailored for erasure correction. The decoding algorithm for error correction for generalized tensor product codes can be found in [15]. Let the symbol ? represent an erasure and “e” denote a de′ coding failure. The erasure decoder D A : (Fq ∪ {?})n ℓ → C A ∪ {“e”} for an ME-LRC C A consists of two kinds of ′ ′′ component decoders Di and Di for i = 1, 2, . . . , µ described below. ′ a) First, the decoder for a coset of the code Ci with paritycheck matrix Bi , i = 1, 2, . . . , µ , is denoted by i Di :(Fq ∪ {?})n × (Fq ∪ {?})∑ j=1 ′ ′ vj → (Fq ∪ {?})n ′ which uses the following decoding rule: for a length-n′ input vector y′ , and a length-∑ij=1 v j syndrome vector s′ without ′ erasures, if y′ agrees with exactly one codeword c′ ∈ Ci + e on the entries with values in Fq , where the vector e is a coset ′ leader determined by both the code Ci and the syndrome ′ vector s′ , i.e., s′ = eBiT , then Di ( y′ , s′ ) = c′ ; otherwise, ′ Di ( y′ , s′ ) = y′ . Therefore, if the length-n′ input vector y′ ′ is a codeword in Ci + e with di′ − 1 or less erasures and the ′ syndrome vector s′ is not erased, then the decoder Di can return the correct codeword. ′′ b) Second, the decoder for the code Ci with parity-check ′′ matrix Hi , i = 1, 2, . . . , µ , is denoted by ′′ Di : (Fqvi ∪ {?})ℓ → (Fqvi ∪ {?})ℓ γ ∈ Fq ∪ {?}\{0}, γ ×? =?, and 0×? = 0. 3) If a length-n vector x, x ∈ (Fq ∪ {?})n , contains an entry ?, then x is considered as the symbol ? in the set Fqn ∪ {?}. Similarly, the symbol ? in the set Fqn ∪ {?} is treated as a length-n vector whose entries are all ?. To describe correctable erasure patterns, we use the following notation. Let we (v) denote the number of erasures ? in the vector v. For a received word y = ( y1 , y2 , . . . , yℓ ), let Nτ = |{ ym : we ( ym ) > dτ′ , 1 6 m 6 ℓ}| for 1 6 τ 6 µ . Theorem 6. The decoder D A for a (ρ, n0 , k; d0 , d)q ME-LRC C A can correct any received word y that satisfies the following condition: Nτ 6 δτ +1 − 1, ∀ 1 6 τ 6 µ , (5) where δµ +1 is defined to be 1. Proof: See Appendix D. The following corollary follows from Theorem 6. Corollary 7. The decoder D A for a (ρ, n0 , k; d0 , d)q ME-LRC C A can correct any received word y with less than d erasures. Proof: See Appendix E. which uses the following decoding rule: for a length-ℓ input ′′ vector y′′ , if y′′ agrees with exactly one codeword c′′ ∈ Ci ′′ on the entries with values in Fqvi , then Di ( y′′ ) = c′′ ; other′′ wise, Di ( y′′ ) = y′′ . Therefore, if the length-ℓ input vector ′′ y′′ is a codeword in Ci with δi − 1 or less erasures, then the ′′ decoder Di can successfully return the correct codeword. The erasure decoder D A for the code C A is summarized in Algorithm 1 below. Let the input word of length n′ ℓ for the decoder D A be y = ( y1 , y2 , . . . , yℓ ), where each component ′ yi ∈ (Fq ∪ {?})n , i = 1, . . . , ℓ. The vector y is an erased version of a codeword c = (c1 , c2 , . . . , cℓ ) ∈ C A . IV. O PTIMAL C ONSTRUCTION AND E XPLICIT ME-LRC S OVER S MALL F IELDS In this section, we study the optimality of Construction A, and also present several explicit ME-LRCs that are optimal over different fields. Algorithm 1: Decoding Procedure of Decoder D A which satisfies kopt [n′ , d′1 ] = n′ − v1 . 2) Bµ is the parity-check matrix of an [n′ , n′ − ∑µi=1 vi , dµ′ ]q Input: received word y = ( y1 , y2 , . . . , yℓ ). Output: codeword c ∈ C A or a decoding failure “e”. 1. Let s1j = 0, for j = 1, 2, . . . , ℓ. ( ′ ) ′ 2. ĉ = (ĉ1 , . . . , ĉℓ ) = D1 ( y1 , s11 ), . . . , D1 ( yℓ , s1ℓ ) . 3. Let F = { j ∈ [ℓ] : ĉ j contains ?}. 4. For i = 2, . . . , µ • If F ̸ = ∅, do the following steps; otherwise go to step 5. ( ) ′′ ′T ′T • (si1 , . . . , siℓ ) = Di ĉ1 Hi , . . . , ĉℓ Hi . ( ) ′ • ĉ j = Di ĉ j , (s1j , . . . , sij ) for j ∈ F ; ĉ j remains the same for j ∈ [ℓ]\F . • Update F = { j ∈ [ℓ] : ĉ j contains ? }. end 5. If F = ∅, let c = ĉ and output c; otherwise return “e”. In Algorithm 1, we use the following rules for operations which involve the symbol ?: 1) Addition +: for any element γ ∈ Fq ∪ {?}, γ +? =?. 2) Multiplication ×: for any element A. Optimal Construction We show how to construct ME-LRCs which are optimal w.r.t. the bound (1) by adding more constraints to Construction A. To this end, we specify the choice of the matrices in Construction A. This specification, referred to as Design I, is as follows. ′ 1) H1 is the parity-check matrix of an [n′ , n′ − v1 , d′1 ]q code (q) (q) code with dopt [n′ , n′ − ∑µi=1 vi ] = dµ′ . 3) The minimum distances satisfy dµ′ 6 2d′1 . ′′ 4) Hi is an all-one vector of length ℓ over Fqvi , i.e., the parity-check matrix of a parity code with minimum distance δi = 2, for all i = 2, . . . , µ .  Theorem 8. The code C A from Construction A with Design I is a (ρ = ℓ, n0 = n′ , k = n′ ℓ − v1 ℓ − ∑µi=2 vi ; d0 = d′1 , d = dµ′ )q ME-LRC, which is optimal with respect to the bound (1). Proof: From Theorem 5, the code parameters ρ, n0 , k, (q) d0 , and d can be determined. We have k∗ = kopt [n′ , d′1 ] = n′ − v1 . Setting x = ℓ − 1, we get { } (q) d6 min dopt [ρn0 − xn0 , k − xk∗ ] 06x6⌈ kk∗ ⌉−1 (q) 6dopt [ℓn′ − (ℓ − 1)n′ , k − (ℓ − 1)k∗ ] (q) =dopt [n′ , n′ − µ ∑ vi ] = dµ′ . i =1 This proves that C A achieves the bound (1). B. Explicit ME-LRCs from Construction A Our construction is very flexible and can generate many ME-LRCs over different fields. In the following, we present several examples. 1) ME-LRCs with local extended BCH codes over F2 From the structure of BCH codes [20], there exists a chain of nested binary extended BCH codes: C3 = [2m , 2m − 1 − 3m, 8]2 ⊂ C2 = [2m , 2m − 1 − 2m, 6]2 ⊂ C1 = [2m , 2m − 1 − m, 4]2 . Let the matrices B1 , B2 , and B3 be the parity-check matrices of C1 , C2 , and C3 , respectively. Example 2. For µ = 3, in Construction A, we use the above ′′ ′′ matrices B1 , B2 , and B3 . We also choose H2 and H3 to be the all-one vector of length ℓ over F2m . From Theorem 5, the corresponding (ρ, n0 , k; d0 , d)2 MELRC C A has parameters ρ = ℓ, n0 = 2m , k = 2m ℓ − (m + 1)ℓ − 2m, d0 = 4, and d = 8. This code satisfies the requirements of Design I. Thus, from Theorem 8, it is optimal w.r.t. the bound (1).  2) ME-LRCs with local algebraic geometry codes over F4 We use a class of algebraic geometry codes called Hermitian codes [25] to construct ME-LRCs. From the construction of Hermitian codes [25], there exists a chain of nested 4-ary Hermitian codes: C H (1) = [8, 1, 8]4 ⊂ C H (2) = [8, 2, 6]4 ⊂ C H (3) = [8, 3, 5]4 ⊂ C H (4) = [8, 4, 4]4 ⊂ C H (5) = [8, 5, 3]4 ⊂ C H (6) = [8, 6, 2]4 ⊂ C H (7) = [8, 7, 2]4 . Now, let the matrices B1 , B2 , B3 , and B4 be the paritycheck matrices of C H (4), C H (3), C H (2), and C H (1), respec′′ tively. Let Hi , i = 2, 3, 4, be the all-one vector of length ℓ over F4 . Example 3. For µ = 2, in Construction A, we use the above ′′ matrices B1 , B2 , and H2 . From Theorem 5, the corresponding (ρ, n0 , k; d0 , d)4 ME-LRC C A has parameters ρ = ℓ, n0 = 8, k = 4ℓ − 1, d0 = 4, and d = 5. For µ = 3, in Construction A, we use the above matrices ′′ ′′ B1 , B2 , B3 , H2 , and H3 . From Theorem 5, the corresponding (ρ, n0 , k; d0 , d)4 ME-LRC C A has parameters ρ = ℓ, n0 = 8, k = 4ℓ − 2, d0 = 4, and d = 6. For µ = 4, in Construction A, we use the above matrices ′′ Bi , i = 1, . . . , 4, and H j , j = 2, 3, 4. From Theorem 5, the corresponding (ρ, n0 , k; d0 , d)4 ME-LRC C A has parameters ρ = ℓ, n0 = 8, k = 4ℓ − 3, d0 = 4, and d = 8. All of the above three families of ME-LRCs over F4 are optimal w.r.t. the bound (1).  3) ME-LRCs with local singly-extended Reed-Solomon codes over Fq Let n′ 6 q and α be a primitive element of Fq . We choose ′ H1 to be the parity-check matrix of an [n′ , n′ − d′1 + 1, d′1 ]q singly-extended RS  1  ′  1 H1 =  ..  . code, namely 1 α .. . ′ 1 α d1 −2 ··· ··· .. . ··· 1 ′ α n −2 .. . ′ ′ α (n −2)(d1 −2) Hi′ For i = 2, 3, . . . , µ , we choose to be  ′ ′ ′ 1 α di−1 −1 · · · α (n −2)(di−1 −1) ′ .. ..  . .. Hi =  .. . . . 1 ′ α di − 2 ··· ′ ′ α (n −2)(di −2) where d′1 < d′2 < · · · < dµ′ . We also require that δi = ⌈ 1 0 .. . 0    .   0 ..  . , 0 dµ′ dµ′ dµ′ ⌉=⌈ ′ ⌉=···=⌈ ′ ⌉, ∀i = 2, . . . , µ ′ di −1 di −1 + 1 di − 1 and δ2 > δ3 > · · · > δµ . ′′ For i = 2, 3, . . . , µ , let Hi be the parity-check matrix of d′ an [ℓ, ℓ − δi + 1, δi = ⌈ d′ µ ⌉]qvi MDS code, which exists i −1 whenever ℓ 6 qvi , where vi = di′ − di′−1 . Note that for an MDS code with minimum distance 2, the code length can be arbitrarily long. Example 4. We use the above chosen matrices Hi′ and Hi′′ for Construction A. The corresponding (ρ, n0 , k; d0 , d)q MELRC C A has parameters ρ = ℓ, n0 = n′ , k = (n′ − d′ d′1 + 1)ℓ − ∑µi=2 (⌈ d′ µ ⌉ − 1)(di′ − di′−1 ), d0 = d′1 , and i −1 d = dµ′ ; the field size q satisfies q > max{q′ , n′ }, where ⌈ ′ 1′ ⌉ q′ = maxi=2,...,µ { ℓ di −di−1 }. When µ = 2 and d′1 < d′2 6 2d′1 , the corresponding (ρ, n0 , k; d0 , d)q ME-LRC C A has parameters ρ = ℓ, n0 = n′ , k = (n′ − d′1 + 1)ℓ − (d′2 − d′1 ), d0 = d′1 , and d = d′2 ; the field size q needs to satisfy q > n′ . Since C A satisfies the requirements of Design I, from Theorem 8, it is optimal w.r.t. the bound (1).  The following theorem shows that the µ -level ME-LRC C A constructed in Example 4 is optimal in the sense of possessing the largest possible dimension among all codes with the same erasure-correcting capability. Theorem 9. Let C be a code of length ℓn′ and dimension k over Fq . Each codeword in C consists of ℓ sub-blocks, each of length n′ . Assume that C corrects all erasure patterns satisfyd′ ing the condition in ( 5), where δτ = ⌈ d′ µ ⌉ for 2 6 τ 6 τ −1 µ. Then, we must have dimension k 6 (n′ − d′1 + 1)ℓ − d′ µ ∑i=2 (⌈ d′ µ ⌉ − 1)(di′ − di′−1 ). i −1 Proof: The proof is based on contradiction. Let each codeword in C correspond to an ℓ × n′ array. We index the coordinates of the array row by row from number 1 to ℓn′ . Let I1 be the set of coordinates defined by I1 = {(i − 1)n′ + j : δ2 − 1 < i 6 ℓ, 1 6 j 6 d′1 − 1}. For 2 6 τ 6 µ , let Iτ be the set of coordinates given by Iτ = {(i − 1)n′ + j : δτ +1 − 1 < i 6 δτ − 1, 1 6 j 6 dτ′ − 1}, where δµ +1 is defined to be 1. Let I be the set of all the coordinates of the array. By calculation, we have |I\(I1 ∪ I2 ∪ · · · ∪ Iµ )| = (n′ − d′ and d2 > d1 . An II code C I I is defined as follows: { C I I = c = (c0 , c1 , . . . , cm−1 ) : ci ∈ C1 , 0 6 i < m, d′1 + 1)ℓ − ∑µi=2 (⌈ d′ µ ⌉ − 1)(di′ − di′−1 ). Now, assume that (n′ d′1 i −1 d′ µ ∑i=2 (⌈ d′ µ ⌉ i −1 − 1)(di′ − di′−1 ). codewords c′ and c′′ in Then, − + 1)ℓ − there exist at least two distinct C that agree on the coordinates in the set {i : i ∈ I\(I1 ∪ I2 ∪ · · · ∪ Iµ )}. We erase the values on the coordinates in the set {i : i ∈ I1 ∪ I2 ∪ · · · ∪ Iµ } of both c′ and c′′ . This erasure pattern satisfies the condition in (5). Since c′ and c′′ are distinct, this erasure pattern is uncorrectable. Thus, our assumpd′ tion that k > (n′ − d′1 + 1)ℓ − ∑µi=2 (⌈ d′ µ ⌉ − 1)(di′ − di′−1 ) i −1 is violated. k > Remark 1. The construction by Blaum and Hetzler [3] based on GII codes cannot generate ME-LRCs constructed in Examples 2 and 3. For the ME-LRC in Example 4, since the local code is the singly-extended RS code, the construction in [3] can also be used to produce an ME-LRC that has the same code parameters ρ, n0 , k, d0 and d as those of the MELRC C A from our construction. However, the construction in [3] requires the field size q to satisfy q > max{ℓ, n′ }, which in general is larger than that in our construction. V. R ELATION TO G ENERALIZED I NTEGRATED I NTERLEAVING C ODES Integrated interleaving (II) codes were first introduced in [11] as a two-level error-correcting scheme for data storage applications, and were then extended in [22] and more recently in [24] as generalized integrated interleaving (GII) codes for multi-level data protection. The main difference between GII codes and generalized tensor product codes is that a generalized tensor product code over Fq is defined by operations over the base field Fq and also its extension field, as shown in (4); in contrast, a GII code over Fq is defined over the same field Fq . As a result, generalized tensor product codes are more flexible than GII codes, and generally GII codes cannot be used to construct ME-LRCs over very small fields, e.g., the binary field. The goal of this section is to study the exact relation between generalized tensor product codes and GII codes. We will show that GII codes are in fact a subclass of generalized tensor product codes. The idea is to reformulate the paritycheck matrix of a GII code into the form of a parity-check matrix of a generalized tensor product code. Establishing this relation allows some code properties of GII codes to be obtained directly from known results about generalized tensor product codes. We start by considering the II codes, a twolevel case of GII codes, to illustrate our idea. A. Integrated Interleaving Codes We follow the definition of II codes in [11]. Let Ci , i = 1, 2, be [n, ki , di ]q linear codes over Fq such that C2 ⊂ C1 and m−1 ∑α i =0 bi } (6) ci ∈ C2 , b = 0, 1, . . . , γ − 1 , where α is a primitive element of Fq and γ < m 6 q − 1. According to the above definition, it is known that the parity-check matrix of C I I is [ ] ⊗ H1 , HI I = I (7) ⊗ Γ2 H2 ⊗ where denotes the Kronecker product. The matrices H1 [ ] H1 and over Fq are the parity-check matrices of C1 and H2 C2 , respectively, the matrix I over Fq is an m × m identity matrix, and Γ2 over Fq is the parity-check matrix of an [m, m − γ , γ + 1]q code in the following form   1 1 ··· 1  1  α ··· α m−1   2 2 ( m − 1 )   1 α α ··· Γ2 =  (8) . .. ..   .. ..  .  . . . 1 α (γ −1) · · · α (γ −1)(m−1) Remark 2. The parity-check matrix H I I over Fq in (7) of C I I is obtained by operations over the same field Fq . In contrast, the parity-check matrix H over Fq in (4) of a generalized tensor product code is obtained by operations over both the base field Fq and its extension field. Remark 3. In general, the codes C1 and C2 in (6) are chosen to be RS codes [11]. If C1 and C2 are chosen to be binary codes, then m can only be m = 1. To see the relation between II codes and generalized tensor product codes, we reformulate H I I in (7) into the following form, by splitting the rows of H2 , ⊗   I H1 ⊗  Γ2  H2 (1) ⊗   , Γ2 H2 (2) HI I =  (9)  .  .. ..  ..  . . ⊗ Γ2 H2 (k1 − k2 ) where the matrix H1 over Fq is the parity-check matrix of C1 , and is treated as a vector over the extension field Fqn−k1 here; correspondingly, the matrix I is treated as an m × m identity matrix over Fqn−k1 . For 1 6 i 6 k1 − k2 , H2 (i ) over Fq represents the ith row of H2 , and Γ2 over Fq is the matrix in (8). Now, referring to the matrix in (4), the matrix in (9) can be interpreted as a parity-check matrix of a (1 + k1 − k2 )-level generalized tensor product code over Fq . Thus, we conclude that an II code is a generalized tensor product code. Using the properties of generalized tensor product codes, we can directly obtain the following result, which was proved in [11] in an alternative way. Lemma 10. The code C I I is a linear code over Fq of length N = nm, dimension K = (m − γ )k1 + γ k2 , and minimum distance D > min{(γ + 1)d1 , d2 }. Proof: For 1 6 i 6 k1 − k2 , let the following paritycheck matrix   H1  H2 (1)   .   ..  H2 (i ) define an [n, k1 − i, d2,i ]q code. It is clear that d1 6 d2,1 6 d2,2 6 · · · 6 d2,k1 −k2 = d2 . From the properties of generalized tensor product codes, the redundancy is N − K = nm − K = (n − k1 )m + γ (k1 − k2 ); that is, the dimension is K = k1 (m − γ ) + k{2 γ . Using Theorem 4, the minimum distance is D > min d1 (γ + 1), d2,1 (γ + 1), . . . , d2,k1 −k2 −1 (γ + } { } 1), d2,k1 −k2 = min (γ + 1)d1 , d2 . B. Generalized Integrated Interleaving Codes With the similar idea used in the previous subsection, we continue our proof for GII codes. We use the definition of GII codes from [24] for consistency. Let Ci , i = 0, 1, . . . , γ , be [n, ki , di ]q codes over Fq such that Ci s = · · · = Ci s−1 + 1 ⊂ Ci s−1 = · · · = Ci s−2 + 1 ⊂ · · · ⊂ Ci1 = · · · = C 1 ⊂ C 0 , and ∑ α bi ci ∈ Cγ−b , i =0 Γ(0, is − is−1 − 1; α ) } b = 0, 1, . . . , γ − 1 , (11) where α is a primitive element of Fq and γ < m 6 q − 1. Let us first define some matrices which will be used below. Let the matrix I over Fq be an m × m identity matrix. Let H0 over Fq be [ the parity-check matrix of C0 . For 1 6 j 6 s, ] H0 let the matrix over Fq represent the parity-check Hi j matrix of Ci j , where   Hi1 \i0  Hi \i    2 1 Hi j =  . ..   . Hi j \i j−1 For any i 6 j, let matrix Γ(i, j; α ) over Fq be the paritycheck matrix of an [m, m − ( j − i + 1), j − i + 2]q code in the following form   1 αi ··· α i (m−1)  1 α i+1 · · · α (i+1)(m−1)    (12) Γ(i, j; α ) =  .. . .. .. ..  .  . . . 1 αj ··· α j(m−1) ⊗ is−1 \is−2 His \is−1 To make a connection between GII codes and generalized tensor product codes, we further reformulate the matrix HGI I in (14) as follows, (10) where i0 = 0 and is = γ . The minimum distances satisfy d0 6 d1 6 · · · 6 dγ . A GII code CGI I is defined as: { CGI I = c = (c0 , c1 , . . . , cm−1 ) : ci ∈ C0 , 0 6 i < m, m−1 Now, according to the definition in (11), using the matrices introduced above, the parity-check matrix of CGI I is   ⊗ I H0 ⊗  Γ(0, is − is−1 − 1; α ) His    ⊗  Γ(is − is−1 , is − is−2 − 1; α ) His−1   , .. .. HGI I =  ..  . .  .  ⊗  Γ(is − i2 , is − i1 − 1; α )  H i 2 ⊗ Hi1 Γ(is − i1 , is − i0 − 1; α ) (13) which can be transformed into the form of ⊗   I H0 ⊗ Hi1 \i0  Γ(0, is − i0 − 1; α )  ⊗    Γ(0, is − i1 − 1; α )  Hi2 \i1   . (14) .. .. HGI I =  ..  . .  .   Γ(0, is − is−2 − 1; α ) ⊗ H  HGI I  I  Γ(0, is − i0 − 1; α )  .  .  .  Γ(0, i − i − 1; α )  s 0   Γ(0, is − i1 − 1; α )  .  .  .   Γ(0, is − i1 − 1; α )   . . =  .  ..  .   Γ(0, is − is−2 − 1; α )   .  ..   Γ(0, is − is−2 − 1; α )   Γ(0, i − i s s−1 − 1; α )   .  .. Γ(0, is − is−1 − 1; α ) ⊗ ⊗ .. . ⊗ ⊗ .. . ⊗ .. . .. . ⊗ .. . ⊗ ⊗ .. . ⊗ H0 Hi1 \i0 (1) .. .       Hi1 \i0 (ki0 − ki1 )   Hi2 \i1 (1)   ..   .  Hi2 \i1 (ki1 − ki2 )    .. , .   ..  .   His−1 \is−2 (1)   ..  .  His−1 \is−2 (kis−2 − kis−1 )    His \is−1 (1)   ..  . His \is−1 (kis−1 − kis ) (15) where, in the first level, the matrix H0 over Fq is treated as a vector over the extension field Fqn−ko , and correspondingly the matrix I is treated as an m × m identity matrix over Fqn−ko . For 1 6 x 6 s and 1 6 y 6 kix−1 − kix , Hix \ix−1 ( y) over Fq represents the yth row of the matrix Hix \ix−1 . Now, referring to the matrix in (4), the matrix in (15) can be seen as a parity-check matrix of a (1 + k0 − kis )-level generalized tensor product code over Fq . As a result, we can directly obtain the following lemma, which was also proved in [24] in a different way. Lemma 11. The code CGI I is a linear code over Fq of length N = nm, dimension K = ∑γx=1 k x + (m − γ )k0 = ∑sj=1 (i j − i j−1 )ki j + (m − γ )k0 , and minimum distance { D > min } (γ + 1)d0 , (γ − i1 + 1)di1 , . . . , (γ − is−1 + 1 ) dis−1 , dis . Proof: For 1 6 x 6 s and 1 6 y following parity-check matrix  H0  H  . i1 \i0 ( 1 )  .  .   Hi1 \i0 (ki0 − ki1 )   ..  .   H  i x \i x−1 ( 1 )  ..  . Hix \ix−1 ( y) 6 kix−1 − kix , let the                define an [n, kix−1 − y, dix ,y ]q code, so we have dix−1 6 dix ,1 6 dix ,2 6 · · · 6 dix ,ki −kix = dix . From the properx−1 ties of generalized tensor product codes, it is easy to obtain the dimension K = ∑sj=1 (i j − i j−1 )ki j + (m − γ )k0 . From Theorem 4, the minimum distance satisfies { D > min (γ + 1)d0 , (γ + 1)di1 ,1 , . . . , (γ + 1)di1 ,ki . . . , . . . , (γ − is−1 + 1)dis−1 , 0 −ki1 −1 , (γ − i 1 + 1 ) di1 , (γ − is−1 + 1)dis ,1 , . . . , (γ − is−1 + 1)dis ,ki −kis −1 , dis s−1 { = min (γ + 1)d0 , (γ − i1 + 1)di1 , } . . . , (γ − is−1 + 1)dis−1 , dis . } Remark 4. In some prior works, we find that generalized tensor product codes are called generalized error-location (GEL) codes [4], [16]. Recently, in [24], the similarity between GII codes and GEL codes was observed. However, the exact relation between them was not studied. In [24], the author also proposed a new generalized integrated interleaving scheme over binary BCH codes, called GII-BCH codes. These codes can also be seen as a special case of generalized tensor product codes. VI. C ONCLUSION In this work, we presented a general construction for MELRCs over small fields. This construction yields optimal MELRCs with respect to an upper bound on the minimum distance for a wide range of code parameters. Then, an erasure decoder was proposed and corresponding correctable erasure patterns were identified. ME-LRCs based on Reed-Solomon codes were shown to be optimal among all codes having the same erasure-correcting capability. Finally, generalized integrated interleaving codes were proved to be a subclass of generalized tensor product codes, thus giving the exact relation between these two codes. ACKNOWLEDGMENT This work was supported by NSF Grants CCF-1405119 and CCF-1619053, BSF Grant 2015816, and Western Digital Corporation. R EFERENCES [1] A. Barg, I. Tamo, and S. Vladut, “Locally recoverable codes on algebraic curves,” arXiv preprint arXiv:1603.08876, 2016. [2] M. Blaum, J. L. Hafner, and S. Hetzler, “Partial-MDS codes and their application to RAID type of architectures,” IEEE Trans. Inf. Theory, vol. 59, no. 7, pp. 4510–4519, 2013. [3] M. Blaum and S. R. Hetzler, “Integrated interleaved codes as locally recoverable codes: properties and performance,” International Journal of Information and Coding Theory, vol. 3, no. 4, pp. 324–344, 2016. [4] M. Bossert, H. Grießer, J. Maucher, and V. V. Zyablov, “Some results on generalized concatenation of block codes,” in Proc. International Symposium on Applied Algebra, Algebraic Algorithms, and Error-Correcting Codes. Springer, 1999, pp. 181–190. [5] V. Cadambe and A. Mazumdar, “An upper bound on the size of locally recoverable codes,” in Proc. IEEE NetCod, 2013, pp. 1–5. [6] G. Calis and O. O. Koyluoglu, “A general construction for PMDS codes,” IEEE Communications Letters, vol. 21, no. 3, pp. 452–455, 2017. [7] R. Gabrys, E. Yaakobi, M. Blaum, and P. H. Siegel, “Constructions of partial MDS codes over small fields,” in Proc. IEEE ISIT, 2017. [8] G. A. Gibson, Redundant Disk Arrays: Reliable, Parallel Secondary Storage. MIT Press, 1992. [9] P. Gopalan, C. Huang, H. Simitci, and S. Yekhanin, “On the locality of codeword symbols,” IEEE Trans. Inf. Theory, vol. 58, no. 11, pp. 6925–6934, 2012. [10] S. Goparaju and R. Calderbank, “Binary cyclic codes that are locally repairable,” in Proc. IEEE ISIT, June 2014, pp. 676–680. [11] M. Hassner, K. Abdel-Ghaffar, A. Patel, R. Koetter, and B. Trager, “Integrated interleaving-a novel ECC architecture,” IEEE Trans. Magn., vol. 37, no. 2, pp. 773–775, 2001. [12] P. Huang, E. Yaakobi, H. Uchikawa, and P. H. Siegel, “Binary linear locally repairable codes,” IEEE Trans. Inf. Theory, vol. 62, no. 11, pp. 6268–6283, Nov 2016. [13] ——, “Cyclic linear binary locally repairable codes,” in Proc. IEEE ITW, 2015, pp. 1–5. [14] ——, “Linear locally repairable codes with availability,” in Proc. IEEE ISIT, 2015, pp. 1871–1875. [15] H. Imai and H. Fujiya, “Generalized tensor product codes,” IEEE Trans. Inf. Theory, vol. 27, no. 2, pp. 181–187, Mar 1981. [16] J. Maucher, V. V. Zyablov, and M. Bossert, “On the equivalence of generalized concatenated codes and generalized error location codes,” IEEE Trans. Inf. Theory, vol. 46, no. 2, pp. 642–649, 2000. [17] F. Oggier and A. Datta, “Self-repairing homomorphic codes for distributed storage systems,” in Proc. IEEE INFOCOM, 2011, pp. 1215– 1223. [18] D. S. Papailiopoulos and A. G. Dimakis, “Locally repairable codes,” in Proc. IEEE ISIT, 2012, pp. 2771–2775. [19] N. Prakash, G. Kamath, V. Lalitha, and P. Kumar, “Optimal linear codes with a local-error-correction property,” in Proc. IEEE ISIT, July 2012, pp. 2776–2780. [20] R. Roth, Introduction to Coding Theory. Cambridge University Press, 2006. [21] I. Tamo and A. Barg, “A family of optimal locally recoverable codes,” IEEE Trans. Inf. Theory, vol. 60, no. 8, pp. 4661–4676, Aug 2014. [22] X. Tang and R. Koetter, “A novel method for combining algebraic decoding and iterative processing,” in Proc. IEEE ISIT, 2006, pp. 474– 478. [23] J. Wolf, “On codes derivable from the tensor product of check matrices,” IEEE Trans. Inf. Theory, vol. 11, no. 2, pp. 281–284, Apr 1965. [24] Y. Wu, “Generalized integrated interleaved codes,” IEEE Trans. Inf. Theory, vol. 63, no. 2, pp. 1102–1119, Feb 2017. [25] K. Yang and P. V. Kumar, “On the true minimum distance of Hermitian codes,” in Coding theory and algebraic geometry. Springer, 1992, pp. 99–107. A PPENDIX A P ROOF OF L EMMA 2 Proof: For the case of x = 0, it is trivial. For 1 6 x 6 ⌈ kk∗ ⌉ − 1, x ∈ Z+ , let I represent the set of the coordinates of the first x rows in the array. Thus, |I| = xn0 . First, consider the code CI = {cI : c ∈ C} whose dimension is denoted by kI , which satisfies kI 6 xk∗ . Then, we consider the code CI0 = {c[ρn0 ]\I : cI = 0 and c ∈ C}. Since the code C is linear, the size of the code CI0 is qk−kI and it is a linear code as well. Moreover, the minimum distance dˆ of the code CI0 is at least d, i.e., dˆ > d. Thus, we get an upper bound on the minimum distance d, Proposition 12. If xH T = 0 and s1 = s2 = · · · = s j = 0, then wq ( xi ) > d′j for xi ̸= 0, i = 1, 2, . . . , ℓ. (q) d 6 dˆ 6dopt [ρn0 − |I|, k − kI ] (q) 6dopt [ρn0 − xn0 , k − xk∗ ]. k, Similarly, we also get an upper bound on the dimension (q) (q) k − kI 6 kopt [ρn0 − |I|, dˆ] 6 kopt [ρn0 − xn0 , d]. Therefore, we conclude that (q) (q) k 6 kopt [ρn0 − xn0 , d] + kI 6 kopt [ρn0 − xn0 , d] + xk∗ . Proof: The condition s1 = s2 = · · · = s j = 0 means that xi B Tj = 0 for i = 1, 2, . . . , ℓ; that is, xi is a codeword in the code defined by the parity-check matrix B j , whose minimum distance is d′j . Therefore, we have wq ( xi ) > d′j for xi ̸= 0, i = 1, 2, . . . , ℓ. Now, if s1 ̸= 0, then wq ( x) > wqv1 (s1 ) > δ1 > dm , which contradicts the assumption. Thus, we have s1 = 0. ( a) A PPENDIX B P ROOF OF L EMMA 3 Proof: We can construct a (ρ, n0 , k; > d0 , > d)q MELRC in two steps, and use the GV bound [20] twice. First, there exists a [ρ(n0 − r0 ), k, > d]q array code G1 of size ρ × (n0 − r0 ) where r0 is an integer 0 6 r0 < n0 , if it satisfies ) d−2 ( ρ(n0 − r0 ) − 1 (16) (q − 1)i < qρ(n0 −r0 )−k . ∑ i i =0 Second, there exists a length-n0 code G2 with minimum distance at least d0 , if its redundancy r0 satisfies ( d0 −2 ( n − 1 ) ) 0 (17) ( q − 1 )i . r0 > logq ∑ i i =0 Now, we encode each row of the code G1 using the code G2 by adding r0 more redundancy symbols. The resulting code is a (ρ, n0 , k; > d0 , > d)q ME-LRC. Let ( d −2 ) r0 = ⌈logq ∑i=0 0 (n0 i−1)(q − 1)i ⌉, and substitute it into (16), producing (3). A PPENDIX C P ROOF OF T HEOREM 4 µ Proof: A codeword x in CGTP is an n′ ℓ-dimensional vector over Fq , denoted by x = ( x1 , x2 , . . . , xℓ ), where xi in x is an n′ -dimensional vector, for i = 1, 2, . . . , ℓ. j We prove it by contradiction and induction. Assume that there exists a codeword x such that xH T = 0, wq ( x) < dm , and x ̸= 0. We first state a proposition which will be used in the following proof. ′T Let si = xi H j , for i = 1, 2, . . . , ℓ and j = 1, 2, . . . , µ . j Thus, si is a v j -dimensional vector over Fq , and is considj j j ered as an element in Fqv j . Let s j = (s1 , s2 , . . . , sℓ ), an j ℓ-dimensional vector over Fqv j , whose components are si , i = 1, 2, . . . , ℓ. To prove Theorem 4, we need to show that if xH T = 0 and wq ( x) < dm = min{δ1 , δ2 d′1 , δ3 d′2 , . . . , δµ dµ′ −1 , dµ′ }, then x must be the all-zero vector 0. Then, consider the second level. If s2 ̸= 0, then wq ( x) > wqv2 (s2 )d′1 > δ2 d′1 > dm , where step (a) is from Proposition 12. This contradicts the assumption, so we have s2 = 0. By induction, we must have s1 = s2 = · · · = sµ −1 = 0. For the last level, i.e., the µ th level, if sµ ̸= 0, then wq ( x) > wqvµ (sµ )dµ′ −1 > δµ dµ′ −1 > dm , which contradicts our assumption. Now, if s1 = s2 = · · · = sµ = 0, then wq ( x) > dµ′ > dm , which also contradicts our assumption. Thus, our assumption is violated. A PPENDIX D P ROOF OF T HEOREM 6 Proof: The proof follows from the decoding procedure of decoder D A . The ME-LRC C A has d0 = d′1 and d = dµ′ . For a received word y = ( y1 , y2 , . . . , yℓ ), each vector yi , 1 6 i 6 ℓ, corresponds to a row in the array. For the first level, since δ1 = ∞, the correct syndrome vector (s11 , . . . , s1ℓ ) is the all-zero vector, i.e., (s11 , . . . , s1ℓ ) = 0. Thus, the rows with number of erasures less than d′1 are corrected. For the second level, the remaining uncorrected row ĉ j , j ∈ F , has at least d′1 erasures. The total number of such uncorrected rows with indices in F is less than δ2 , because we require N1 6 δ2 − 1 in the condition. Thus, the correct syndrome vector (s21 , . . . , s2ℓ ) can be obtained. As a result, the rows with number of erasures less than d′2 are corrected. Similarly, by induction, if the decoder runs until the µ th level, the remaining uncorrected row ĉ j , j ∈ F , has at least dµ′ −1 erasures. The total number of such uncorrected rows with indices in F is less than δµ , because we require Nµ −1 6 δµ − 1 in the condition. Therefore, all the correct syndrome vectors (si1 , . . . , siℓ ), i = 1, 2, . . . , µ , are obtained. On the other hand, the remaining uncorrected row ĉ j , j ∈ F , has at most dµ′ − 1 erasures, since we also require Nµ 6 0 in the condition. Thus, all these uncorrected rows can be corrected in this step with all these correct syndromes. A PPENDIX E P ROOF OF C OROLLARY 7 Proof: The ME-LRC C A has d0 = d′1 and d = dµ′ . We only need to show that the received word y with any dµ′ − 1 erasures satisfies the condition in Theorem 6. We prove it by contradiction. If the condition is not satisfied, there is at least an integer i, 1 6 i 6 µ , such that Ni > δi+1 . Therefore, we have we ( y) > di′ δi+1 > dµ′ , where the last inequality is from the requirement of Construction A. Thus, we get a contradiction to the assumption that the received word y has dµ′ − 1 erasures.
7cs.IT
arXiv:1709.03908v1 [math.CO] 12 Sep 2017 A NEW FAMILY OF MRD CODES IN Fq2n×2n WITH RIGHT AND MIDDLE NUCLEI Fqn ROCCO TROMBETTI 1 AND YUE ZHOU 2,† Abstract. In this paper, we present a new family of maximum rank distance (MRD for short) codes in F2n×2n of minimum distance 2 ≤ d ≤ 2n. In q particular, when d = 2n, we can show that the corresponding semifield is exactly a Hughes-Kleinfeld semifield. The middle and right nuclei of these MRD codes are both equal to Fqn . We also prove that the MRD codes of minimum distance 2 < d < 2n in this family are inequivalent to all known ones. The equivalence between any two members of this new family is also determined. 1. Introduction Let K denote a field. The set of all m × n matrices over K form a K-vector space, which we denote by Km×n . For A, B ∈ Km×n , we define d(A, B) = rank(A − B), which is often called the rank metric or the rank distance on the Km×n . A subset C ⊆ Km×n with respect to the rank metric is called a rank-metric code or a rank-distance code. If C contains at least two elements, the minimum distance of C is given by d(C) = min {d(A, B)}. A,B∈C,A6=B When C is a K-linear subspace of Km×n , we say that C is a K-linear code and its dimension dimK (C) is defined to be the dimension of C as a subspace over K. with d(C) = d, it Let Fq denote the finite field of q elements. For any C ⊆ Fm×n q is well-known that #C ≤ q max{m,n}(min{m,n}−d+1) , which is the Singleton bound for the rank metric; see [14]. When the equality holds, we call C a maximum rank-distance (MRD for short) code. More properties of MRD codes can be found in [14], [21], [23], [41] and [44]. Rank-metric codes, in particular, MRD codes have been studied since 1970s and have seen much interest in recent years due to a wide range of applications including storage systems [45], cryptosystems [22], spacetime codes [40] and random linear network coding [32]. 1 Dipartimento di Mathematica e Applicazioni “R. Caccioppoli”, Università degli Studi di Napoli “Federico II”, I-80126 Napoli, Italy 2 College of Science, National University of Defense Technology, 410073 Changsha, China † Corresponding author E-mail addresses: [email protected], [email protected]. Date: September 13, 2017. 1 2 R. TROMBETTI AND Y. ZHOU In finite geometry, there are several interesting structures including quasifields, semifields, and splitting dimensional dual hyperovals can be equivalently described as special types of rank-metric codes; see [16], [17], [29], [48] and the references therein. In particular, a finite quasifield corresponds to an MRD code in Fqn×n of minimum distance n and a finite semifield corresponds to such an MRD code that is a subgroup of Fqn×n (see [13] for the precise relationship). Many essentially different families of finite quasifields and semifields are known [34], which yield many inequivalent MRD codes in Kn×n of minimum distance n. In contrast, it appears to be much more difficult to obtain inequivalent MRD codes in Fqn×n of minimum distance strictly less than n. For the relationship between MRD codes and other geometric objects such as linear sets and Segre varieties, we refer to [37]. Besides quasifields, there are only a few known constructions of MRD codes in Fqn×n . The first construction of MRD codes was given by Delsarte [14]. This construction was later rediscovered by Gabidulin [21] and generalized by Kshevetskiy and Gabidulin [33]. Today this family is usually called the generalized Gabidulin codes, sometimes it is also simply called the Gabidulin codes (see Section 2, for a precise definition). It is easy to show that a Gabidulin code is always Fqn -linear. Recently, another Fq -linear family was found by Sheekey [47] and we often call them (generalized) twisted Gabidulin codes. This family has been further generalized into additive MRD codes [43]. Given any 2 ≤ d ≤ n, these constructions can provide us MRD codes of minimum distance d. For MRD codes in Fqn×n of minimum distance d = n−1, there are a few more constructions. First, there is a nonlinear family constructed by Cossidente, Marino and Pavese [6] and later generalized by Durante and Siciliano [20]. Besides this family, there are other constructions associated with maximum scattered linear sets over PG(1, q 6 ) and PG(1, q 8 ) presented recently in [8] and [10]. For more results concerning maximum scattered linear sets and associated MRD codes, see [3], [7], [11] and [12]. with m < n, there are many different approach to For MRD codes in Fm×n q construct them. A canonical way to get them is puncturing (or projecting) MRD ′ codes in Fqm ×n with n ≥ m′ > m. In [25], a new criterion for the punctured Gabidulin codes is presented, and for small m and n, several constructions of MRD codes which are not equivalent to them are obtained. In [42], it is presented a generic construction of MRD codes by using algebraic geometry approaches, under the condition that n is large enough compared with d and m. In [9], an approach to derive MRD codes in Fm×n from linear sets is investigated. In [19], a nonlinear q construction is presented. Recently, Schmidt and the second author [46] showed that even in Gabidulin codes there are a huge subset of inequivalent MRD codes. In this paper, we present a new family of MRD codes in Fq2n×2n which exist for all minimum distance 2 ≤ d ≤ 2n. In particular, when d = 2n, we can show that the corresponding semifield is exactly the Hughes-Kleinfeld semifield [27] found in 1960. Through the investigation of their middle and right nuclei, we can prove that the MRD codes in this new family are inequivalent to all known constructions. The rest of this paper is organized as follows. In Section 2, we introduce semifields, describe rank-metric codes in Fqn×n via linearized polynomials and introduce the equivalence between rank-metric codes as well as their dual codes and adjoint codes. In Section 3, we present our new family of MRD codes and determine their middle and right nuclei. Based on these results, we show that they are inequivalent A NEW FAMILY OF MRD CODES 3 to all the known MRD codes except for one special case which is later excluded in Section 4. Another result in Section 4 is the complete answer to the equivalence problem between different members of this new family. 2. Preliminaries Roughly speaking, a semifield S is an algebraic structure satisfying all the axioms of a skewfield except (possibly) the associativity of its multiplication. A finite field is a trivial example of a semifield. Furthermore, if S does not necessarily have a multiplicative identity, then it is called a presemifield. For a presemifield S, (S, +) is necessarily abelian [31]. The first family of non-trivial semifields was constructed by Dickson [18] more than a century ago. In [31], Knuth showed that the additive group of a finite semifield S is an elementary abelian group, and the additive order of the nonzero elements in S is called the characteristic of S. Hence, any finite semifield can be represented by (Fq , +, ∗), where q is a power of a prime p. Here (Fq , +) is the additive P i j group of the finite field Fq and x ∗ y can be written as x ∗ y = i,j aij xp y p , which forms a map from Fq × Fq to Fq . We refer to [34] for a recent and comprehensive survey on finite semifields. Geometrically speaking, there is a well-known correspondence, via coordinatisation, between (pre)semifields and projective planes of Lenz-Barlotti type V.1, see [15, 28]. The most important equivalence relation defined on (pre)semifields is the isotopism. Given two (pre)semifields S1 = (Fnp , +, ∗) and S2 = (Fnp , +, ⋆). If there exist three bijective linear mappings L, M, N : Fnp → Fnp such that M (x) ⋆ N (y) = L(x ∗ y) for any x, y ∈ Fnp , then S1 and S2 are called isotopic, and the triple (M, N, L) is called an isotopism between S1 and S2 . In [1], Albert showed that two (pre)semifields coordinatize isomorphic planes if and only if they are isotopic. Every presemifield can be normalized into a semifield under an appropriate isotopism; see [4] and [34]. Given a semifield S with multiplication ∗, we define its left, middle and right nucleus by Nl (S) := {a ∈ S : a ∗ (x ∗ y) = (a ∗ x) ∗ y for all x, y ∈ S}, Nm (S) := {a ∈ S : x ∗ (a ∗ y) = (x ∗ a) ∗ y for all x, y ∈ S}, Nr (S) := {a ∈ S : x ∗ (y ∗ a) = (x ∗ y) ∗ a for all x, y ∈ S}. It is not difficult to prove that the semifield S can be viewed as a left vector space over its left nucleus. In particular, when S is finite, we can further show Nl (S) is actually a finite field Fq . Let us assume that S is of size q n . For every b ∈ S, the map x 7→ x ∗ b defines an n × n matrix Mb over its left nucleus Nl (S). Furthermore, all such matrices together form a rank metric code {Mb : b ∈ S} which is actually an MRD code, because the difference between any two distinct members Mb and Md in it equals Mb − Md = Mb−d which is always nonsingular. This MRD code is usually called the semifield spread set associated with S; see [34]. Next, let us turn to rank-metric codes. As we are working with rank-metric codes in Fqn×n rather than Fm×n with m < n in this paper, it is more convenient to q describe such a rank-metric code using the language of q-polynomials or linearized 4 R. TROMBETTI AND Y. ZHOU polynomials over Fqn which are the polynomials in the set o nX i L(n,q) [X] = ci X q : ci ∈ F q n . n In fact, there is a bijection between Fqn×n and L(n,q) [X]/(X q −X); for more results about linearized polynomials, we refer to [35]. As we mentioned in the introduction part, the most well-known family of MRD codes is called (generalized) Gabidulin codes. They can be described by the following subset of linearized polynomials. s Gk,s = {a0 x + a1 xq + . . . ak−1 xq s(k−1) : a0 , a1 , . . . , ak−1 ∈ Fqn }, where s is relatively prime with n. It is obvious that there are q kn polynomials in Gk,s and each polynomial in it has at most q k−1 roots which means its minimum distance d = n − k + 1. Hence its size meets the Singleton bound. For x ∈ Fqm , let Nqm /q (x) denote the norm from Fqm to Fq , i.e. Nqm /q (x) = 1+q+···qm−1 x . The following result follows from [24, Theorem 10]. Lemma 2.1. Let s and m be two relatively prime positive integers. Suppose that s sk f = f0 X + f1 X q + · · · + fk X q ∈ Fqm [X] is a linearized polynomial with fk 6= 0. If f has q k roots, then Nqsm /qs (f0 ) = (−1)km Nqsm /qs (fk ). In [47], Sheekey applied Lemma 2.1 and found a new family of MRD codes s Hk,s (η, h) = {a0 x + a1 xq + · · · + ak−1 xq s(k−1) h + ηaq0 xq sk : a0 , a1 , . . . , ak−1 ∈ Fqm }, km where η satisfies Nqsm /qs (η) 6= (−1) , i.e. Nqm /q (η) 6= (−1)km . Such an MRD code is usually called a (generalized) twisted Gabidulin code. It is clear that if we allow η equal 0, Gk,s can be viewed as a subfamily of the twisted Gabidulin codes. h Replacing the field automorphism a0 7→ aq0 in the coefficient of the last term of the elements in Hk,s (η, h) by an automorphism in Aut(Fqn ) \ Aut(Fqn /Fq ), Otal and Özbudak [43] generalized this family into an additive one. There are several slightly different definitions of equivalence of rank-metric codes. In this paper, we use the following notion of equivalence. Definition 2.2. Two rank-metric codes C1 and C2 in Km×n are equivalent if there exist A ∈ GLm (K), B ∈ GLn (K), C ∈ Km×n and ρ ∈ Aut(K) such that (1) C2 = {AM ρ B + C : M ∈ C1 }. For m = n, if C2 is equivalent to C1 or C1T := {M T : M ∈ C1 } where ( . )T means transposition, then we say C1 and C2 are isometrically equivalent. An equivalence map from a rank-metric code C to itself is also called an automorphism of C. When C1 and C2 are both additive and equivalent, it is not difficult to show that we can choose C = 0 in (1). In particular, when C1 and C2 are semifield spread sets, they are equivalent if and only if the associated semifields are isotopic [34, Theorem 7]. Back to the descriptions in linearized polynomials, given two rank-metric codes C1 and C2 which consist of linearized polynomials, they are equivalent if there exist ϕ1 , ϕ2 ∈ L(n,q) [X] permuting Fqn , ψ ∈ L(n,q) [X] and ρ ∈ Aut(Fq ) such that ϕ1 ◦ f ρ ◦ ϕ2 + ψ ∈ C2 for all f ∈ C1 , P P i i where ◦ stands for the composition of maps and f ρ = aρi X q for f = ai X q . A NEW FAMILY OF MRD CODES 5 In general, it is a difficult job to tell whether two given rank-metric codes are equivalent or not. There are several invariants which may help us distinguish them. Given a K-linear rank-metric code C ⊆ Km×n , its middle nucleus is defined as Nm (C) = {M ∈ Km×n : M C ∈ C for all C ∈ C}, and its right nucleus is defined as Nm (C) = {M ∈ Km×n : CM ∈ C for all C ∈ C}. These two concepts were introduced in [39] and they can be viewed as a natural generalization of the middle and right nucleus of semifields. In [36], they are called the left idealizer and the right idealizer of C, respectively. In general, we can also define the left nucleus of C. However, for MRD codes over K containing singular matrices, it is always K which means it is not a useful invariant; see [39]. For a rank-metric code C given by a set of linearized polynomials, its middle nucleus and right nucleus can also be written as sets of linearized polynomials. Precisely the middle nucleus of C is Nm (C) = {ϕ ∈ L(n,q) : f ◦ ϕ ∈ C for all f ∈ C}. It is defined by f ◦ ϕ rather than ϕ ◦ f because we always consider a row vector u multiplying a matrix C which is a member of a rank-metric code. This means that M ∈ Nm (C) only if uM C = uC ′ for some C ′ ∈ C. Similarly, the right nucleus of C is Nr (C) = {ϕ ∈ L(n,q) : ϕ ◦ f ∈ C for all f ∈ C}. They played an important role in [46] proving a lower bound of numbers of inequivalent Gabidulin codes in Fm×n . The middle and right nuclei of generalized twisted q Gabidulin codes together with a complete answer to the equivalence between members in this family can be found in [38]. by We define a symmetric bilinear form on the set Fm×n q hM, N i := Tr(M N T ), where N T is the transpose of N . The Delsarte dual code of an Fq -linear code C is C ⊥ := {M ∈ Fm×n : hM, N i = 0 for all N ∈ C}. q One important result proved by Delsarte [14] is that the Delsarte dual code of a linear MRD code is still MRD. As we are considering MRD codes using linearized polynomials, the Delsarte dual can also be interpreted in the following way [47]. We define the bilinear form b on q-polynomials by ! n−1 X ai b i , b (f, g) = Trqn /q i=0 Pn−1 qi where f (x) = i=0 ai x and g(x) = C ⊥ of a set of q-polynomials C is Pn−1 i=0 i bi xq ∈ Fqn [x]. The Delsarte dual code C ⊥ = {f : b(f, g) = 0 for all g ∈ C}. It is well-known and also not difficult to show directly that two linear rank-metric codes are equivalent if and only if their duals are equivalent. Let C be an MRD codes in Km×n . It is obvious that {M T : M ∈ C} is also an MRD codes, because the ranks of M T and M are the same. When K = Fq 6 R. TROMBETTI AND Y. ZHOU and m = n, we can also interpret the transposes of matrices into an operation on q-polynomials. Pn−1 i The adjoint of a q-polynomial f = i=0 ai xq is given by fˆ := n−1 X aiq n−i xq n−i . i=0 If C is a rank-metric codes consisting of q-polynomials, then the adjoint code of C is Cb := {fˆ : f ∈ C}. In fact, the adjoint of f is equivalent to the transpose of the matrix derived from f . This result can be found in [30]. Regarding the adjoint and Delsarte dual operation, we have ⊥ \ b =N Nm (C) r (C) = Nr (C ), (2) and \ b Nm (C ⊥ ) = N m (C) = Nr (C), (3) which are proved in [39, Proposition 4.2]. 3. A class of MRD codes In the rest of this paper, we write N (x) instead of Nq2n /q (x) for short. By applying Lemma 2.1, we can get another family of MRD codes. Theorem 3.1. Let s and n be two integers satisfying gcd(s, 2n) = 1. For γ ∈ Fq2n satisfying that N (γ) is a non-square in Fq , we define ) ( k−1 X qis qks ci X + γbX : c1 , · · · , ck−1 ∈ Fq2n , a, b ∈ Fqn . (4) Dk,s (γ) := aX + i=1 Then Dk,s (λ) is an MRD code. Proof. It is clear that #Dk,s (λ) = q 2nk . We need to show that for each polynomial f ∈ Dk,s (λ), it has at most q k−1 roots which means its minimum distance is d = 2n − k + 1. Hence Dk,s (λ) is an MRD code. Pk−1 is k By way of contradiction, let us assume that f = aX + i=1 ci X q + γbX q has q k roots which implies that a and b are both nonzero. By Lemma 2.1, Nq2sn /qs (a) = (−1)2nk Nq2sn /qs (γb). Hence Nq2sn /qs (a/b) = Nq2sn /qs (γ) = N (γ). As a, b ∈ Fqn ,  a   a 2(1+qs +···+q(n−1)s )  a 2 Nq2sn /qs = Nqn /q = b b b which is a square in Fq . However, this contradicts the assumption on N (γ).  Let us first look at the Delsarte dual code of Dk,s (γ). It is straightforward to compute that ) ( 2n−1 X ⊥ qks qis Dk,s (γ) = −γbX + aX + ci X : a, b ∈ Fqn , ci ∈ Fq2n . i=k+1 q2n−ks By applying (·) module X q2n − X on it, we get D2n−k,s (−γ q 2n−ks ). Proposition 3.2. The Delsarte dual code of Dk,s (γ) is equivalent to D2n−k,s (−γ q 2n−ks It can also be readily verified the following result of the adjoint code of Dk,s (γ). ). A NEW FAMILY OF MRD CODES 7 Proposition 3.3. The adjoint code of Dk,s (γ) is equivalent to Dk,s (1/γ). s By Theorem 3.1, it is clear that aX + γbX q defines a semifield multiplication. As γ ∈ / Fqn (otherwise N (γ) must be a square in Fq ), for every x ∈ Fq2n , we can write it as x = c + dγ for some c, d ∈ Fqn . s s Assume that γ q +1 = u + vγ for certain u, v ∈ Fqn . By expanding ax + γbxq , we have s s s s ax + γbxq = (ac + bdq u) + (ad + bcq + bdq v)γ. We view them as vectors in F2qn and define a semifield multiplication (5) s s s (c, d) ∗ (a, b) = (ac + bdq u, ad + bcq + bdq v), for a, b, c, d ∈ Fqn . By comparing with [28, Theorem 9.7], we see that (5) is exactly the multiplication of a Hughes-Kleinfeld semifield [27], which is also the multiplication of a Knuth semifield of type II [31]. By [28, Lemma 9.8], Fqn is the right and middle nucleus of H. In [26], a necessary and sufficient condition for x + yγ ∈ Nl (H) are derived. Proposition 3.4. Let ∗ be the multiplication defined by (5) and H denote the associated semifield (F2qn , +, ∗). (a) Nr (H) = Fqn . (b) Nm (H) = Fqn . (c) For x, y ∈ Fqn , x + yγ ∈ Nl (H) if and only if ( 2s 2s s s xq + y q v q = x + y q v, s s 2s s 2s 2s s yu + xq v + y q v 2 = y q uq + xq v + y q v q +1 . Next, let us investigate the middle and right nucleus of the MRD codes Dk,s (γ) defined in Theorem 3.1. They are very important invariants with respect to the equivalence of rank-metric codes. We will use them later to show that Dk,s (γ) contains MRD codes which are not equivalent to any known ones. Theorem 3.5. Let k be an integer satisfying 1 ≤ k ≤ 2n − 1. Then the right nucleus of Dk,s (γ) is (6) Nr (Dk,s (γ)) = {aX : a ∈ Fqn }, and its middle nucleus is (7) Nm (Dk,s (γ)) = {aX : a ∈ Fqn }. Proof. When k = 1, Dk,s (γ) is isotopic to a Hughes-Kleinfeld semifield H, and the result can be then derived from Proposition 3.4. By duality, we get the result for k = 2n − 1. P2n−1 is In the rest part, we assume that 2 ≤ k ≤ 2n− 2. Assume that ϕ = i=0 di X q s is an element in Nr (Dk,s (γ)). As ϕ(c1 X q ) ∈ Dk,s (γ), we see that di = 0 for k < i < 2n − 1. In fact, only d2n−1 , d0 and d1 can be nonzero. When k = 2, this is obvious. When k > 2, this statement can be directly verified by checking js ϕ(cj X q ) ∈ Dk,s (γ) for j = 2, · · · , k − 1. ks Next we consider ϕ(aX + γbX q ) which should also be in Dk,s (γ). As ϕ = (2n−1)s s ks d2n−1 X q + d0 X + d1 X q , in the expansion of ϕ(aX + γbX q ) the coefficient (2n−1)s (2n−1)s (k+1)s s of X q is d2n−1 aq and the coefficient of X q is d1 (γb)q . Since a and 8 R. TROMBETTI AND Y. ZHOU b can take any value in Fqn , by checking the elements in Dk,s (γ) we see that d2n−1 and d1 must be 0. Thus ϕ = d0 X. From ks ϕ(aX + γbX q ) = d0 aX + γd0 bX q ks ∈ Dk,s (γ) we derive d0 ∈ Fqn . d By D k,s (γ) = Dk,s (1/γ) in Proposition 3.3 and (2), we get    ⊥ !  1 1 = Nr D2n−k,s − q2n−ks Nm (Dk,s (γ)) = Nr Dk,s γ γ which equals {aX : a ∈ Fqn }.  Corollary 3.6. In Fq2n×2n , the MRD code Dk,s (γ) is not equivalent to any generalized Gabidulin code. When k 6= n or h 6= n, Dk,s (γ) is also not equivalent to any generalized twisted Gabidulin code Hk,t (η, h) with η 6= 0. Proof. When k = 1, a (generalized) Gabidulin code {aX : a ∈ Fq2n } is derived from the multiplication of the finite field Fq2n , which is never isotopic to a HughesKleinfeld semifield. Hence the corresponding MRD codes are not equivalent. Moreover, H1,s (η, h) associates a generalized twisted field, which is a presemifield. If s 6≡ h (mod 2n), it is isotopic to a semifield whose middle nucleus is of size q gcd(2n,s−h) and right nucleus is of size q gcd(2n,h) ; otherwise s ≡ h (mod 2n) and this presemifield is isotopic to the finite field Fq2n ; see [2] and [5]. Thus H1,s (η, h) cannot be equivalent to D1,s (γ). When k = 2n−1, we can simply consider the equivalences between their Delsarte dual codes which have been already determined above. In the rest, we only have to investigate the equivalence problem for 1 < k < 2n − 1. According to [39, Corollary 5.9 (a)], the middle (or right) nucleus of a generalized Gabidulin codes over Fq2n is always Fq2n . Hence it is different from the middle nucleus of Dk,s (γ) by Theorem 3.5, which means that they are not equivalent. According to [39, Corollary 5.9 (b)], Nm (Hk,t (η, h)) in Fq2n×2n is of size q gcd(2n,sk−h) and Nr (Hk,t (η, h)) is of size q gcd(2n,h) . Hence, Dk,s (γ) and Hk,t (η, h) are equivalent only if gcd(2n, h) = h and gcd(2n, sk − h) = n which means h = n = k.  For k = 2, as the middle nucleus of D2,s (γ) is not of size q 2n , it is also not equivalent to those MRD codes associated with maximum scattered linear sets constructed in [8] and [10]. 4. Equivalence In Section 3, we have shown that most members of the MRD codes Dk,s (γ) are new with respect to the equivalence of rank-metric codes. In the last part of this section, we will completely solve the last open case whether Dn,s (γ) is equivalent to Hn,t (θ, n) or not. First we investigate the equivalence between different members of this family. If we want to further determine the isometric equivalence between Dk,s (γ) and Dk,t (θ), the answer follows directly from our result about the equivalence map from Dk,s (γ) to Dk,t (θ) or its adjoint Dk,t (1/θ). By using our knowledge of the middle and right nucleus of Dk,s (γ), we can prove the following results. A NEW FAMILY OF MRD CODES 9 Lemma 4.1. Let n, s, t ∈ Z+ satisfying gcd(2n, s) = gcd(2n, t) = 1. Let γ and θ be in Fq2n satisfying that N (γ) and N (θ) are both non-square in Fq . Let (ϕ1 , ϕ2 , ρ) be an equivalence map between Dk,s (γ) and Dk,t (θ) for 1 < k < 2n − 1. If k 6= n or n ≥ 3, then ϕ1 and ϕ2 are both monomials. 2n−ks Proof. By Proposition 3.2, D2n−k,s (−γ q ) is equivalent to the Delsarte dual code of Dk,s (γ). As two MRD codes are equivalent if and only if there Delsarte duals are equivalent, we only have to prove the statement for k ≤ n. According to the definition of equivalence, ϕ1 ◦ f ρ ◦ ϕ2 ∈ Dk,t (θ) for every f ∈ Dk,s (γ). As f ρ ∈ Dk,s (γ ρ ), ϕ1 must be in the normalizer of Nr (Dk,s (γ ρ )) in GL(2n, q). By Theorem 3.5, the right nucleus Nr (Dk,s (γ ρ )) = {aX : a ∈ Fqn }. It follows that l l+n ϕ1 = cX q + dX q for a certain l ∈ {0, · · · , 2n − 1} and c, d ∈ Fq2n . This result is well-known and P qi can be verified directly as follows. Assume that ϕ1 = 2n−1 i=0 ai X . Then for each b ∈ Fqn , there always exists some b′ ∈ Fqn such that ϕ1 ◦ bX = 2n−1 X i i ai bq X q = b′ X ◦ ϕ1 = 2n−1 X i b ′ ai X q . i=0 i=0 qi ′ This implies that if ai 6= 0 then b = b , which means that there are at most two coefficients al and al+n are nonzero for a certain l. By the same argument, we can also show that j ϕ2 = gX q + hX q j+n for some j ∈ {0, · · · , 2n − 1} and g, h ∈ Fq2n . is Now let us look at the image of ci X q under the equivalence map (ϕ1 , ϕ2 , id). By calculation, is ϕ1 ◦ ci X q ◦ ϕ2 l j =ccqi (gX q + hX q j+n )q is+l + dciq l+n j (gX q + hX q j+n )q is+l+n (8) =(cg q is+l l cqi + dhq is+l+n cqi l+n is )X q j+is+l + (chq is+l l cqi + dg q is+l+n cqi l+n )X q j+is+l+n . j+is+l When k < n, as ϕ1 ◦ ci X q ◦ ϕ2 ∈ Dk,t (θ), one of the coefficients of X q and qj+is+l+n qj+is+l+tn X =X must be 0 for all ci ∈ Fq2n . Together with the condition that ϕ1 and ϕ2 are permutation polynomials, we show that c = h = 0 or d = g = 0 or c = g = 0 or d = h = 0, which means ϕ1 and ϕ2 are both monomials. is j+is+l When k = n, as ϕ1 ◦ ci X q ◦ ϕ2 ∈ Dk,t (θ), the coefficients of X q and qj+is+l+n qj+is+l+tn X = X are both nonzero only if exact one of j + is + l and j + is + l + tn equals 0. If n ≥ 3, i can be taken for at least two different values from {1, 2, · · · , k − 1}. Hence we can choose the value of i such that j + is + l 6= 0, n. As in the case k < n, we see that ϕ1 and ϕ2 must be both monomials.  The following lemma can be proved by using exactly the same argument and we omit its proof. Lemma 4.2. Let n, s, t ∈ Z+ satisfying gcd(2n, s) = gcd(2n, t) = 1. Let γ and θ be in F∗q2n satisfying that N (γ) is a non-square in Fq and N (η) 6= 1. Let (ϕ1 , ϕ2 , ρ) 10 R. TROMBETTI AND Y. ZHOU be an equivalence map between Dn,s (γ) and Hn,t (η, n). If n ≥ 3, then ϕ1 and ϕ2 are both monomials. Now we can determine the equivalence between Dk,s (γ) and Dk,t (θ). Theorem 4.3. Let n, s, t ∈ Z+ satisfying gcd(2n, s) = gcd(2n, t) = 1. Let γ and η be in Fq2n satisfying that N (γ) and N (θ) are both non-square in Fq . Let k be an integer satisfying 1 < k < 2n − 1. When k 6= n or n ≥ 3, the MRD code Dk,s (γ) is equivalent to Dk,t (θ) if and only if one of the following collections of conditions are satisfied. ks (a) s ≡ t (mod 2n), there exist σ ∈ Aut(Fq2n ) and h ∈ Fq2n such that γ σ hq −1 = θ. (b) s ≡ −t (mod 2n), there exist σ ∈ Aut(Fq2n ) and h ∈ Fq2n such that ks γ σ hq −1 = 1/θ. Proof. As in the proof of Lemma 4.1, we only have to handle the cases k ≤ n. Assume that (ϕ1 , ϕ2 , ρ) is an equivalence map between Dk,s (γ) and Dk,t (θ). l When k 6= n or n ≥ 3, by Lemma 4.1, we can assume that ϕ1 = dX q and j ϕ2 = gX q for some d, g ∈ F∗qn and l, j ∈ {0, · · · , 2n − 1}. For arbitrary ci ∈ Fq2n with i ∈ {1, · · · , k − 1}, l is ϕ1 ◦ ci X q ◦ ϕ2 = dcqi g q is+l Xq is+l+j . It follows that is + l + j ∈ {t, 2t, · · · , (k − 1)t}. As gcd(2n, s) = gcd(2n, t) = 1, we can assume that s = rt which means {irt + l + j : i = 1, 2, · · · , k − 1} = {t, 2t, · · · , (k − 1)t}. As k ≤ n, it is straightforward to see that either r = 1 and l + j ≡ 0 (mod 2n), or r = −1 and l + j ≡ kt (mod 2n). When r = 1, i.e. s ≡ t (mod 2n) and j ≡ −l (mod 2n), for a, b ∈ Fqn , applying ks (ϕ1 , ϕ2 , ρ) onto aX + γbX q , we obtain   l l l l ks+l ks ks ◦ ϕ2 = daρq g q X + dγ ρq bρq g q Xq , ϕ1 ◦ aρ X + γ ρ bρ X q l l ks+l which belongs to Dk,t (θ) if and only if dg q ∈ Fqn and dγ ρq g q ∈ θFqn . Let σ l l ρq q denote the automorphism of Fqn defined by x 7→ x . Let h = g . Then we see ks that there must be a solution of h such that γ σ hq −1 = θ. When r = −1, i.e. s ≡ −t (mod 2n) and j ≡ kt − l (mod 2n), for a, b ∈ Fqn , we ks apply (ϕ1 , ϕ2 , ρ) onto aX + γbX q and get   l l kt l l ks+l ks ◦ ϕ2 = daρq g q X q + dγ ρq bρq g q X, ϕ1 ◦ aρ X + γ ρ bρ X q l l ks+l ∈ Fqn . Let σ which belongs to Dk,t (θ) if and only if dg q ∈ θFqn and dγ ρq g q l l ρq q denote the automorphism of Fqn defined by x 7→ x . Let h = g . Then we see ks that there must be a solution of h such that γ σ hq −1 = 1/θ. Therefore we have proved the necessary condition in the statement for k 6= n or n ≥ 3. For sufficiency, it is routine to do a verification.  There are 3 cases which are not covered by Theorem 4.3: k = 1, k = 2n − 1 and k = n = 2. For k = 1, D1,s (γ) defines a Hughes-Kleinfeld semifield whose autotopism group has been completely determined in [26]. It appears that using the A NEW FAMILY OF MRD CODES 11 same approach, the equivalence between D1,s (γ) and D1,t (θ) can also be determined. Hence, in the rest of this section, we will skip the case k = 1. Moreover, for k = 2n − 1, the MRD code is the Delsarte dual code of a Hughes-Kleinfeld semifield by Proposition 3.2. We will also skip this case, because the equivalence problem for this case can be completely converted into the equivalence problem for HughesKleinfeld semifields. Next we investigate the last case in which k = n = 2. As gcd(2n, s) = gcd(2n, t) = 1 and n = 2, t ≡ ±s (mod 2n). In fact, t and s can only be 1 or −1 modulo 2n. Theorem 4.4. Let s, t ∈ Z+ satisfying gcd(4, s) = gcd(4, t) = 1. Let γ and θ be in Fq4 satisfying that Nq4 /q (γ) and Nq4 /q (θ) are both non-square in Fq . The MRD code D2,s (γ) is equivalent to D2,t (θ) if and only if one of the following collections of conditions are satisfied. 2s (a) s ≡ t (mod 4), there exists σ ∈ Aut(Fq4 ) and h ∈ Fq4 such that γ σ hq −1 = θ. (b) s ≡ t (mod 4), there exist c, d, g, h ∈ Fq4 , ρ ∈ Aut(Fq ) and l ∈ {0, 1, 2, 3} such that  s+l 2 s+l  cg q − dq hq = 0,   s+l 2 2 s+l  chq θq − dq g q θ = 0, l l+2  cg q + dhq = 0,   2s+l l l l+2  chq γ ρq + dg q γ ρq = 0. 2s (c) s ≡ −t (mod 4), there exists σ ∈ Aut(Fq4 ) and h ∈ Fq4 such that γ σ hq −1 = 1/θ. (d) s ≡ −t (mod 4), there exist c, d, g, h ∈ Fq4 , ρ ∈ Aut(Fq ) and l ∈ {0, 1, 2, 3} such that  s+l 2 s+l  cg q − dq hq = 0,   s+l 2 2 s+l  chq θq − dq g q θ = 0, l l+2  chq + dg q = 0,   2s+l l l l+2  q cg γ ρq + dhq γ ρq = 0. Proof. In this proof, we will still write n instead of 2 in some equations even though we have assumed that n = 2. s If cX q ∈ D2,s (γ) is always mapped to another monomial for all c ∈ Fq2n , then the same calculation in Theorem 4.3 shows the necessary and sufficient conditions (a) and (c). s In the rest of the proof, we always assume that cX q ∈ D2,s (γ) is mapped to a binomial for some c. Taking i = 1 in (8), we see that j + s + l can be taken for exact two possible value: j + s + l ≡ 0 (mod 2n) or j + s + l ≡ n (mod 2n). Let us consider the case s ≡ t (mod 2n). First we assume that j + s + l ≡ 0 s (mod 2n). From ϕ1 ◦ c1 X q ◦ ϕ2 ∈ D2,t (θ) and (8), we derive that the coefficient j+s+l+n j+s+l belongs to θFqn , which of X q belongs to Fqn and the coefficient of X q means that (9) cg q s+l l cq1 + dhq s+l+n l+n c1q n = cq g q s+l+n l+n cq1 n + dq hq s+l l cq1 and (10) (chq s+l l cq1 + dg q s+l+n l+n cq1 n n )θq = (cq hq s+l+n l+n c1q n + dq g q s+l l cq1 )θ 12 R. TROMBETTI AND Y. ZHOU hold for every c1 ∈ Fq2n . If we view (9) as a polynomial of c1 , by comparing the l+n l coefficients of cq1 (or those of c1q ) in it, we obtain cg q (11) s+l n = dq hq s+l . Similarly, from (10) we derive chq (12) s+l n n θq = dq g q s+l θ. Furthermore, from ϕ1 ◦ aX ◦ ϕ2 ∈ D2,t (θ) with a ∈ Fqn we can derive more j+l 2n−s conditions. As the coefficient of X q = Xq in it must be zero, by plugging a = ci and i = 0 into (8), we get l cg q + dhq (13) l+n = 0. Analogously, by checking the coefficient of X q 2s ρ γ bX q ◦ ϕ2 ∈ D2,t (θ) with b ∈ Fqn , we obtain (14) chq 2s+l l l γ ρq + dg q γ ρq j+2s+l+n l+n = Xq j+l = Xq 3s in ϕ1 ◦ = 0. For j + s + l ≡ n (mod 2n), the proof is similar. By checking the coefficients of n j+s+l+n s X = X q and X q = X in ϕ1 ◦ c1 X q ◦ ϕ2 ∈ D2,t (θ), we obtain qj+s+l chq (15) (16) cg s+l n = dq g q qs+l qn θ l =d h By checking the coefficient of X q b ∈ Fqn , we get (18) j+l+n chq + dg q cg q j+2s+l 2s+l l , qn qs+l Furthermore, as the coefficient of X q be 0 for every a ∈ Fqn , (17) s+l = Xq l+n = Xq l θ. 3s in ϕ1 ◦ aX ◦ ϕ2 ∈ D2,t (θ) must = 0. 3s in ϕ1 ◦ γ ρ bX q γ ρq + dhq γ ρq l+n 2s ◦ ϕ2 ∈ D2,t (θ) with = 0. Hence, (15), (17), (16) and (18) can be simply obtained by switching of g and h in (11), (13), (12) and (14), respectively. We finish the proof of the necessity part of (b). After a careful check of the previous calculations, we can see that if c, d, g, h, ρ and l satisfy (11), (12), (13) and (14) simultaneously, then the map (ϕ1 , ϕ2 , ρ) is indeed an equivalence map between D2,s (γ) and D2,t (θ). Therefore the condition (b) is also sufficient. For the case (d) in which s ≡ −t (mod 4), the proof is the same. For j +s+l ≡ 0 (mod 2n), we can also get the same equations (11) and (12). However, now g and h are switched in (13) and (14). We omit the details of these calculations.  Remark 1. It is possible that the conditions (b) and (d) hold. For instance, let q = 3, s = 1 and γ = θ = ω which is a root of X 4 + 2X 3 + 2 ∈ Fq [X]. Then l = 0, c = 1, d = ω 36 , g = ω 2 , h = ω 54 and ρ = id offer us an equivalence map from D2,1 (γ) to itself. Remark 2. By Theorem 4.3 (a) and Theorem 4.4 (a) (b), the automorphism group of an MRD codes Dk,s (γ) can also be determined. A NEW FAMILY OF MRD CODES 13 Recall that in Corollary 3.6, the equivalence between Dn,s (γ) and Hn,t (η, n) is the unique open case. Finally we will solve this problem by using the same approach appeared in the proofs of Theorems 4.3 and 4.4. Theorem 4.5. Let n, s, t ∈ Z+ satisfying gcd(2n, s) = gcd(2n, t) = 1. Let γ and η be in F∗q2n satisfying that N (γ) is a non-square in Fq and N (η) 6= 1. Then Dk,s (γ) and Hk,t (η, h) are not equivalent for all k and h. Proof. By Corollary 3.6, we only have to face the case k = n = h. Assume that (ϕ1 , ϕ2 , ρ) defines an equivalence map from Dn,s (γ) to Hn,t (η, n). As we are going to show that such a map never exist, without loss of generality, we assume that ρ = id; otherwise we consider the equivalence map from Dn,s (γ ρ ) to Hn,t (η, n). We separate our proof into two parts depending on the value of n. (a) When n ≥ 3, the proof is quite similar to that for Theorem 4.3. By Lemma l j 4.2, we can assume that ϕ1 = dX q and ϕ2 = gX q for some d, g ∈ F∗qn and l, j ∈ {0, · · · , 2n − 1}. For arbitrary ci ∈ Fq2n with i ∈ {1, · · · , n − 1}, l is ϕ1 ◦ ci X q ◦ ϕ2 = dcqi g q is+l Xq is+l+j , which should belong to Hn,t (η, n). It follows that is + l + j ∈ {t, 2t, · · · , (n − 1)t}. As gcd(2n, s) = gcd(2n, t) = 1, we can assume that s = rt whence {irt + l + j : i = 1, 2, · · · , n − 1} = {t, 2t, · · · , (n − 1)t}. It is straightforward to see that r = 1 and l + j ≡ 0 (mod 2n), or r = −1 and l + j ≡ nt (mod 2n). No matter r = 1 or −1, applying (ϕ1 , ϕ2 , id) to aX, we see that one of the kt coefficients of X and X q is zero and the other one is a function of a. This contradicts the assumption ϕ1 ◦ aX ◦ ϕ2 ∈ Hn,t (η, n) for every a ∈ Fqn . (b) When n = 2, it is clear that s and t are congruent to ±1 modulo 2n. In fact, it is sufficient to consider the case t = s, because Hn,−s (η, n) is equivalent to 2s Hn,s (1/η q , n). As the middle and right nuclei of Dn,s (γ) to Hn,t (η, n) are Fqn , we can assume l l+n j j+n that ϕ1 = cX q + dX q and ϕ2 = gX q + hX q . Our first goal is to show that ϕ1 and ϕ2 must be monomials. Assume, by way of contradiction, that c, d, g, h are all nonzero. Plugging i = 1 and ci = w into (8), we get s ϕ1 ◦ wX q ◦ ϕ2 (19) =(cg q s+l l wq + dhq s+l+n wq l+n )X q j+s+l + (chq s+l l wq + dg q s+l+n wq l+n )X q j+s+l+n , which should belong to Hn,s (η, n) for all w ∈ Fq2n . As in the proof of Theorem 4.4, we see that j + s + l can only take two possible value: j + s + l ≡ 0 (mod 2n) or j + s + l ≡ n (mod 2n). s If j + s + l ≡ 0 (mod 2n), from ϕ1 ◦ wX q ◦ ϕ2 ∈ Hn,t (η, n), we derive η(cg q s+l l wq + dhq s+l+n wq l+n n )q = chq s+l l wq + dg q s+l+n wq l+n 14 R. TROMBETTI AND Y. ZHOU for every w ∈ Fq2n , which means ( n 3s+l 3s+l ηcq g q = dg q , n s+l s+l ηdq hq = chq . As we have assumed that g and h are both nonzero, the two above equations n n n implies that ηcq = d and ηdq = c. Hence η q +1 = 1, which implies that N (η) = 2 3 2 2 ηη q η q η q = (ηη q )(ηη q )q = 1 contradicting the assumption that N (η) 6= 1. For j + s + l ≡ n (mod 2n), the proof is analogous and we omit it. l j Therefore we have proved that ϕ1 = dX q and ϕ2 = gX q for some d, g ∈ F∗q2n and l, j ∈ {0, · · · , 3}. As the case n ≥ 3 proved in part (a), it is routine to expand ϕ1 ◦ aX ◦ ϕ2 and to check that it cannot belong to Hn,t (η, n). Hence there is no equivalence map from Dn,s (γ) to Hn,t (η, n).  acknowledgment This work is supported by the Research Project of MIUR (Italian Office for University and Research) “Strutture geometriche, Combinatoria e loro Applicazioni” 2012. Yue Zhou is supported by the National Natural Science Foundation of China (No. 11401579, 11531002). References [1] A. A. Albert. Finite division algebras and finite planes. In Proc. Sympos. Appl. Math., Vol. 10, pages 53–70. American Mathematical Society, Providence, R.I., 1960. [2] A. A. Albert. Generalized twisted fields. Pacific Journal of Mathematics, 11:1–8, 1961. [3] D. Bartoli and Y. Zhou. Exceptional scattered polynomials. arXiv:1708.00349 [math], 2017. [4] J. Bierbrauer. Projective polynomials, a projection construction and a family of semifields. Designs, Codes and Cryptography, 79(1):183–200, Apr 2016. [5] M. Biliotti, V. Jha, and N. L. Johnson. The collineation groups of generalized twisted field planes. Geometriae Dedicata, 76:97–126, 1999. [6] A. Cossidente, G. Marino, and F. Pavese. Non-linear maximum rank distance codes. Designs, Codes and Cryptography, 79(3):597–609, June 2016. [7] B. Csajbók, G. Marino, and O. Polverino. Classes and equivalence of linear sets in PG(1, q n ). arXiv:1607.06962 [math], July 2016. [8] B. Csajbók, G. Marino, O. Polverino, and C. Zanella. A new family of MRD-codes. arXiv:1707.08487 [math], 2017. [9] B. Csajbók, G. Marino, O. Polverino, and F. Zullo. Maximum scattered linear sets and MRD-codes. arXiv:1701.06831 [math], Jan. 2017. [10] B. Csajbók, G. Marino, and F. Zullo. New maximum scattered linear sets of the projective line. arXiv:1709.00926 [math], 2017. [11] B. Csajbók and C. Zanella. On the equivalence of linear sets. Designs, Codes and Cryptography, 81(2):269–281, 2016. [12] B. Csajbók and C. Zanella. Maximum scattered Fq -linear sets of PG(1, q 4 ). arXiv:1705.00731 [math], May 2017. [13] J. de la Cruz, M. Kiermaier, A. Wassermann, and W. Willems. Algebraic structures of MRD codes. Advances in Mathematics of Communications, 10(3):499–510, 2016. [14] P. Delsarte. Bilinear forms over a finite field, with applications to coding theory. Journal of Combinatorial Theory, Series A, 25(3):226–241, Nov. 1978. [15] P. Dembowski. Finite Geometries. Springer, 1997. [16] U. Dempwolff and Y. Edel. Dimensional dual hyperovals and APN functions with translation groups. Journal of Algebraic Combinatorics, 39(2):457–496, June 2014. [17] U. Dempwolff and W. M. Kantor. Orthogonal dual hyperovals, symplectic spreads, and orthogonal spreads. Journal of Algebraic Combinatorics, 41(1):83–108, May 2015. [18] L. E. Dickson. On commutative linear algebras in which division is always uniquely possible. Transactions of the American Mathematical Society, 7(4):514–522, 1906. A NEW FAMILY OF MRD CODES 15 [19] G. Donati and N. Durante. A generalization of the normal rational curve in PG(d, q n ) and its associated non-linear MRD codes. Designs, Codes and Cryptography, Jul 2017. [20] N. Durante and A. Siciliano. Non-linear maximum rank distance codes in the cyclic model for the field reduction of finite geometries. The Electronic Journal of Combinatorics, 24:P2.33, 2017. [21] E. Gabidulin. Theory of codes with maximum rank distance. Problems of information transmission, 21:3–16, 1985. [22] E. M. Gabidulin. Public-key cryptosystems based on linear codes over large alphabets: efficiency and weakness. In Codes and Cyphers, pages 17–31. Formara Limited, 1995. [23] M. Gadouleau and Z. Yan. Properties of codes with the rank metric. In IEEE Global Telecommunications Conference 2006, pages 1–5, 2006. [24] R. Gow and R. Quinlan. Galois theory and linear algebra. Linear Algebra and its Applications, 430(7):1778–1789, Apr. 2009. [25] A.-L. Horlemann-Trautmann and K. Marshall. New criteria for MRD and Gabidulin codes and some rank metric code constructions. arXiv:1507.08641 [cs, math], July 2015. to appear in Advances in Mathematics of Communications. [26] D. R. Hughes. Collineation groups of non-Desarguesian planes II. some seminuclear division algebras. American Journal of Mathematics, 82(1):113–119, 1960. [27] D. R. Hughes and E. Kleinfield. Seminuclear extensions of Galois fields. American Journal of Mathematics, 82(3):pp. 389–392, 1960. [28] D. R. Hughes and F. C. Piper. Projective planes. Springer-Verlag, New York, 1973. Graduate Texts in Mathematics, Vol. 6. [29] N. L. Johnson, V. Jha, and M. Biliotti. Handbook of finite translation planes, volume 289 of Pure and Applied Mathematics (Boca Raton). Chapman & Hall/CRC, Boca Raton, FL, 2007. [30] W. M. Kantor. Commutative semifields and symplectic spreads. Journal of Algebra, 270(1):96–114, 2003. [31] D. E. Knuth. Finite semifields and projective planes. Journal of Algebra, 2:182–217, 1965. [32] R. Koetter and F. Kschischang. Coding for errors and erasure in random network coding. IEEE Transactions on Information Theory, 54(8):3579–3591, Aug. 2008. [33] A. Kshevetskiy and E. Gabidulin. The new construction of rank codes. In International Symposium on Information Theory, 2005. ISIT 2005. Proceedings, pages 2105–2108, Sept. 2005. [34] M. Lavrauw and O. Polverino. Finite semifields. In L. Storme and J. De Beule, editors, Current research topics in Galois Geometry, chapter 6, pages 131–160. NOVA Academic Publishers, 2011. [35] R. Lidl and H. Niederreiter. Finite fields, volume 20 of Encyclopedia of Mathematics and its Applications. Cambridge University Press, Cambridge, second edition, 1997. [36] D. Liebhold and G. Nebe. Automorphism groups of Gabidulin-like codes. Archiv der Mathematik, 107(4):355–366, Oct. 2016. [37] G. Lunardon. MRD-codes and linear sets. Journal of Combinatorial Theory, Series A, 149:1– 20, July 2017. [38] G. Lunardon, R. Trombetti, and Y. Zhou. Generalized twisted Gabidulin codes. arXiv:1507.07855 [cs, math], July 2015. [39] G. Lunardon, R. Trombetti, and Y. Zhou. On kernels and nuclei of rank metric codes. Journal of Algebraic Combinatorics, 46(2):313–340, Sep 2017. [40] P. Lusina, E. Gabidulin, and M. Bossert. Maximum rank distance codes as space-time codes. IEEE Transactions on Information Theory, 49(10):2757–2760, Oct 2003. [41] K. Morrison. Equivalence for rank-metric and matrix codes and automorphism groups of gabidulin codes. IEEE Transactions on Information Theory, 60(11):7035–7046, 2014. [42] A. Neri, A.-L. Horlemann-Trautmann, T. Randrianarisoa, and J. Rosenthal. On the genericity of maximum rank distance and gabidulin codes. Designs, Codes and Cryptography, pages 1– 23, 2017. Online First. [43] K. Otal and F. Özbudak. Additive rank metric codes. IEEE Transactions on Information Theory, 63(1):164–168, Jan 2017. [44] A. Ravagnani. Rank-metric codes and their duality theory. Designs, Codes and Cryptography, pages 1–20, Apr. 2015. 16 R. TROMBETTI AND Y. ZHOU [45] R. M. Roth. Maximum-rank array codes and their application to crisscross error correction. IEEE Transactions on Information Theory, 37(2):328–336, Mar 1991. [46] K.-U. Schmidt and Y. Zhou. On the number of inequivalent MRD codes. arXiv:1702.04582 [math], Feb. 2017. [47] J. Sheekey. A new family of linear maximum rank distance codes. Advances in Mathematics of Communications, 10(3):475–488, 2016. [48] H. Taniguchi and S. Yoshiara. A unified description of four simply connected dimensional dual hyperovals. European Journal of Combinatorics, 36:143–150, 2014.
7cs.IT
K-THEORY AND ACTIONS ON EUCLIDEAN RETRACTS arXiv:1801.00020v1 [math.KT] 29 Dec 2017 BARTELS, A. Abstract. This note surveys axiomatic results for the Farrell-Jones Conjecture in terms of actions on Euclidean retracts and applications of these to GLn (Z), relative hyperbolic groups and mapping class groups. Introduction Motivated by surgery theory Hsiang [43] made a number of influential conjectures about the K-theory of integral group rings Z[G] for torsion free groups G. These conjectures often have direct implications for the classification theory of manifolds of dimension ≥ 5. A good example is the following. An h-cobordism is a compact manifold W that has two boundary components M0 and M1 such that both inclusions Mi → W are homotopy equivalences. The Whitehead group Wh(G) is the quotient of K1 (Z[G]) by the subgroup generated by the canonical units ±g, g ∈ G. Associated to an h-cobordism is an invariant, the Whitehead torsion, in Wh(G), where G is the fundamental group of W . A consequence of the s-cobordism theorem is that for dim W ≥ 6, an h-cobordism W is trivial (i.e., isomorphic to a product M0 ×[0, 1]) iff its Whitehead torsion vanishes. Hsiang conjectured that for G torsion free Wh(G) = 0, and thus that in many cases h-cobordisms are products. The Borel conjecture asserts that closed aspherical manifolds are topologically rigid, i.e., that any homotopy equivalence to another closed manifold is homotopic to a homeomorphism. The last step in proofs of instances of this conjecture via surgery theory uses a vanishing result for Wh(G) to conclude that an h-cobordism is a product and that therefore the two boundary components are homeomorphic. Farrell-Jones [28] pioneered a method of using the geodesic flow on non-positively curved manifolds to study these conjectures. This created a beautiful connection between K-theory and dynamics that led Farrell-Jones [30], among many other results, to a proof of the Borel Conjecture for closed Riemannian manifolds of nonpositive curvature of dimension ≥ 5. Moreover, Farrell-Jones [29] formulated (and proved many instances of) a conjecture about the structure of the algebraic Ktheory (and L-theory) of group rings, even in the presence of torsion in the group. Roughly, the Farrell-Jones Conjecture states that the main building blocks for the K-theory of Z[G] is the K-theory of Z[V ] where V varies of the family of virtually cyclic subgroups of G. It implies a number of other conjectures, among them Hsiang’s conjectures, the Borel Conjecture in dimension ≥ 5, the Novikov Conjecture on the homotopy invariant of higher signatures, Kaplansky’s conjecture about idempotents in group rings, see [54] for a summary of these and other applications. My goal in this note is twofold. The first goal is to explain a condition formulated in terms of existence of certain actions of G on Euclidean retracts that implies the Farrell-Jones Conjecture for G. This condition was developed in joint work with Lück and Reich [8, 12] where the connection between K-theory and dynamics has Date: December 2017. 1991 Mathematics Subject Classification. 18F25, 19G24, 20F67, 20F65. Key words and phrases. Farrell-Jones Conjecture, K- and L-theory of group rings. 1 2 BARTELS, A. been extended beyond the context of Riemannian manifolds to prove the FarrellJones Conjecture for hyperbolic and CAT(0)-groups. The second goal is to outline how this condition has been used in joint work with Lück, Reich and Rüping and with Bestvina to prove the Farrell-Jones Conjecture for GLn (Z) and mapping class groups. A common difficulty for both families of groups is that their natural proper actions (on the associated symmetric space, respectively on Teichmüller space) is not cocompact. In both cases the solution depends on a good understanding of the action away from cocompact subsets and an induction on a complexity of the groups. As a preparation for mapping class groups we also discuss relatively hyperbolic groups. The Farrell-Jones Conjecture has a prominent relative, the Baum-Connes Conjecture for topological K-theory of group C ∗ -algebras [15, 16]. The two conjectures are formally very similar, but methods of proofs are different. In particular, the conditions discussed in Section 2 are not known to imply the Baum-Conjecture. The classes of groups for which the two conjectures are known differ. For example, by work of Kammeyer-Lück-Rüping [44] all lattice in Lie groups satisfy the Farrell-Jones Conjecture; despite Lafforgues [51] positive results for many property T groups, the Baum-Connes Conjecture is still a challenge for SL3 (Z). Wegner [77] proved the Farrell-Jones Conjecture for all solvable groups, but the case of amenable (or just elementary amenable) groups is open; in contrast Higson-Kasparov [41] proved the Baum-Connes Conjecture for all a-T-menable groups, a class of groups that contains all amenable groups. On the other hand, hyperbolic groups satisfy both conjectures. See Mineyev-Yu [62] and Lafforgue [52] for the Baum-Connes Conjecture and, as mentioned above, [8, 12] for the Farrell-Jones Conjecture. For a more comprehensive summary of the current status of the Farrell-Jones Conjecture the reader is directed to [55, 70]. Acknowledgement. It is a pleasure to thank my teachers, coauthors, and students for the many things they taught me. The work described here has been supported by the SFB 878 in Münster. 1. The Formulation of the Farrell-Jones Conjecture Classifying spaces for families. A family F of subgroups of a group G is a non-empty collection of subgroups that is closed under conjugation and subgroups. Examples are the family Fin of finite subgroups and the family VCyc of virtually cyclic subgroups (i.e., of subgroups containing a cyclic subgroup as a subgroup of finite index). For any family of subgroups of G there exists a G-CW -complex EF G with the following property: if E is any other G-CW -complex such that all isotropy groups of E belong to F , then there is a up to G-homotopy unique Gmap E → EF G. This space is not unique, but it is unique up to G-homotopy equivalence. Informally one may think about EF G as a space that encodes the group G relative to all subgroups from F . Often there are interesting geometric models for this space, in particular for F = Fin. More information about this space can be found ` for example in [53]. An easy way to construct EF G is as the infinite join ∗∞ ( i=0 F ∈F G/F ). If F is closed under supergroups of finite index ′ ′ (i.e., if F ∈ F is a subgroup ` of finite index in F , then also F ∈ F ), then the full simplicial complex on F ∈F G/F is also a model for EF G; we will denote this model later by ∆F (G). The formulation of the conjecture. The original formulation of the FarrellJones Conjecture [29] used homology with coefficients in stratified and twisted Ωspectra. Here we use the equivalent [38] formulation developed by Davis-Lück [23]. K-THEORY AND ACTIONS ON EUCLIDEAN RETRACTS 3 Given a ring R, Davis-Lück construct a homology theory X 7→ H∗G (X; KR ) for G-spaces with the property that H∗G (G/H; KR ) ∼ = K∗ (R[H]). Let F be a family of subgroups of the group G. Consider the projection map EF G → G/G to the one-point G-space G/G. It induces the F -assembly map G G ∼ αG F : H∗ (EF ; KR ) → H∗ (G/G; KR ) = K∗ (R[G]). Conjecture 1.1 (Farrell-Jones Conjecture). For any group G and any ring R the assembly map αG VCyc is an isomorphism. This version of the conjecture has been stated in [7]. The original formulation of Farrell and Jones [29] considered only the integral group ring Z[G]. Moreover, Farrell and Jones wrote that they regard this and related conjectures only as estimates which best fit the know data at this time. However, the conjecture is still open and does still fit with all known data today. Transitivity principle. Informally one can view the statement that the assembly map αG F is an isomorphism for a group G and a ring R as the statement that K∗ (R[G]) can be assembled from K∗ (R[F ]) for all F ∈ F (and group homology). If V is a family of subgroups of G that contains all subgroups from F , then one can apply this slogan in two steps, for G relative to V and for each V ∈ V relative to the F ∈ F with F ⊆ V . The implementation of this is the following transitivity principle. Theorem 1.2 ([29, 56]). For V ∈ V set FV := {F | F ∈ F , F ⊆ V }. Assume that for all V ∈ V the assembly map αVFV is an isomorphism. Then αG F is an isomorphism iff αG is an isomorphism. V Twisted coefficients. Often it is beneficial to study more flexible generalizations of Conjecture 1.1. Such a generalization is the Fibred Isomorphism Conjecture of Farrell-Jones [29]. An alternative is the Farrell-Jones Conjecture with coefficients in additive categories [14], here one allows additive categories with an action of a group G instead of just a ring as coefficients. This version of the conjecture applies in particular to twisted group rings. These generalizations of the Conjecture have better inheritance properties. Two of these inheritance property are stability under directed colimits of groups, and stability under taking subgroups. For a summary of the inheritance properties see [70, Thm. 27(2)]. Often proofs of cases of the Farrell-Jones Conjecture use these inheritance properties in inductions or to reduce to special cases. We will mean by the statement that G satisfies the Farrell-Jones Conjecture relative to F , that the assembly map αG F is bijective for all additive categories A with G-action. However, this as a technical point that can be safely ignored for the purpose of this note. Other theories. The Farrell-Jones Conjecture for K-theory discussed so far has an analog in L-theory, as it appears in surgery theory. For some of the applications mentioned before this is crucial. For example, the Borel Conjecture for a closed aspherical manifold M of dimensions ≥ 5 holds if the fundamental group of M satisfies both the K-and L-theoretic Farrell-Jones Conjecture. However, proofs of the Farrell-Jones Conjecture in K- and L-theory are by now very parallel. Recently, the techniques for the Farrell-Jones Conjecture in K- and L-theory have been extended to also cover Waldhausen’s A-theory [25, 47, 74]. In particular, the conditions we will discuss in Section 2 are now known to imply the Farrell-Jones Conjecture in all three theories. 4 BARTELS, A. 2. Actions on compact spaces Amenable actions and exact groups. Definition 2.1 (Almost invariant maps). Let X, E be G-spaces where E is equipped with a G-invariant metric d. We will say that a sequence of maps fn : X → E is almost G-equivariant if for any g ∈ G sup d(fn (gx), gfn (x)) → 0 as n → ∞. x∈X For a discrete group G we equip the space Prob(G) of probability measures on G with the metric it inherits as subspace of l1 (G). This metric generates the topology of point-wise convergence on Prob(G). We recall the following definition. Definition 2.2. An action of a group G on a compact space X is said to be amenable if there exists a sequence of almost equivariant maps X → Prob(G). A group is amenable iff its action on the one point space is amenable. Groups that admit an amenable action on a compact Hausdorff space are said to be exact or boundary amenable. The class of exact groups contains all amenable groups, hyperbolic groups [1], and all linear groups [35]. Other prominent groups that are known to be exact are mapping class groups [39, 48] and the group of outer automorphisms of free groups [18]. The Baum-Connes assembly map is split injective for all exact groups [40, 79]. This implies the Novikov conjecture for exact groups. This is an analytic result for the Novikov conjecture, in the sense that it has no known proof that avoids the Baum-Connes Conjecture. There is no corresponding injectivity result for assembly maps in algebraic K-theory. For a survey about amenable actions and exact groups see [66]. Finite asymptotic dimension. Results for assembly maps in algebraic K-theory and L-theory often depend on a finite dimensional setting; the space of probability measures has to be replaced with a finite dimensional space. We write ∆(G) for the full simplicial complex with vertex set G and ∆(N ) (G) for its N -skeleton. The space ∆(G) can be viewed as the space of probability measures on G with finite support. We equip ∆(G) with the l1 -metric; this is the metric it inherits from Prob(G). Definition 2.3 (N -amenable action). We will say that an action of a group G on a compact space X is N -amenable if there exists a sequence of almost equivariant maps X → ∆(N ) (G). The natural action of a countable group G on its Stone-Čech compactification βG is N -amenable iff the asymptotic dimension of G is at most N [37, Thm. 6.5]. This condition (for any N ) also implies exactness and therefore the Novikov conjecture [42]. For groups G of finite asymptotic dimension for which in addition the classifying space BG can be realized as a finite CW -complex, there is an alternative argument for the Novikov Conjecture [78] that has been translated to integral injectivity results for assembly maps in algebraic K-theory and L-theory [3, 22]. These injectivity results have seen far reaching generalizations to groups of finite decomposition complexity [36, 45, 69]. N -F -amenable actions. Constructions of transfer maps in algebraic K-theory and L-theory often depend on actions on spaces that are much nicer than βG. A good class of spaces to use for the Farrell-Jones Conjecture are Euclidean retracts, i.e., compact spaces that can be embedded as a retract in some Rn . Brouwer’s fixed point theorem implies that for an action of a group on an Euclidean retract any cyclic subgroup will have a fixed point. It is not difficult to check that this obstructs the existence of almost equivariant maps to ∆(N ) (assuming G contains K-THEORY AND ACTIONS ON EUCLIDEAN RETRACTS 5 an element of infinite order). Let F be a family of subgroups of G that is closed ` under taking supergroups of finite index. Let S := F ∈F G/F be the set of all left cosets to members of F . Let ∆F (G) be the full simplicial complex on S and (N ) ∆F (G) be its N -skeleton. We equip ∆F (G) with the l1 -metric. Definition 2.4 (N -F -amenable action). We will say that an action of G on a compact space X is N -F -amenable if there exists a sequence of almost equivariant (N ) maps X → ∆F (G). If an action is N -F -amenable for some N ∈ N, then we say that it is finitely F -amenable. Remark 2.5. Let X be a G-CW -complex with isotropy groups in F and of dimension (N ) ≤ N . As ∆F (G) is a model for EF G we obtain a cellular G-map f : X → ∆F (G); this map is also continuous for the l1 -metric. In particular, the constant sequence fn ≡ f is almost equivariant. Therefore one can view N -F -amenability for G-spaces as a relaxation of the property of being a G-CW -complex with isotropy in F and of dimension ≤ N . This relaxation is necessary to obtain compact examples and reasonably small F : If a G-CW -complex is compact, then it has only finitely many cells. In particular, for each cell the isotropy group has finite index in G, so F would have to contain subgroups of finite index in G. Theorem 2.6 ([8, 12]). Suppose that G admits a finitely F -amenable action on a Euclidean retract. Then G satisfies the Farrell-Jones Conjecture relative to F . Remark 2.7. The proof of Theorem 2.6 depends on methods from controlled topology/algebra that have a long history. An introduction to controlled algebra is given in [67]; an introduction to the proof of Theorem 2.6 can be found in [4]. Here we only sketch a very special case, where these methods are not needed. Assume that the Euclidean retract is a G-CW -complex X. As pointed out in Remark 2.5 this forces F to contain subgroups of finite index in G. As X is contractible, the cellular chain complex of X provides a finite resolution L C∗ over Z[G] of the trivial G-module Z. Note that in each degree Ck = Z[G/Fi ] is a finite sum of permutation modules with Fi ∈ F and Fi of finite index in G. For a finitely generated projective R[G]-module P we obtain a finite resolution C∗ ⊗Z P of P . Each module in the resolution is a finite sum of modules of the form Z[G/F ]⊗Z P with F ∈ F and of finite index in G. Here Z[G/F ]⊗Z P is equipped with the diagonal G-action and can be identified with the R[G]-module obtained by first restricting P to and then inducing back up from R[F ] to  an R[F ]-module  R[G]. In particular, Z[G/F ]⊗Z P ∈ K0 (R[G]) is in the image of the assembly P map relative to the family F . It follows that [P ] = k (−1)k [Ck ⊗Z P ] is also in the image. Therefore the assembly map H0 (EF G; KR ) → K0 (R[G]) is surjective. (This argument did not use that F is closed under supergroups of finite index.) Example 2.8. Let G be a hyperbolic group. Its Rips complex can be compactified to a Euclidean retract [19]. The natural action of G on this compactification is finitely VCyc-amenable [11]. To obtain further examples of finitely F -amenable actions on Euclidean retracts, it is helpful to replace VCyc with a larger family of subgroups F . Groups that act acylindrically hyperbolic on a tree admit finitely F -amenable actions on Euclidean retracts where F is the family of subgroups that is generated by the virtually cyclic subgroups and the isotropy groups for the original action on the tree [50]. Relative hyperbolic groups and mapping class groups are discussed in Section 4. Remark 2.9. A natural question is which groups admit finitely VCyc-amenable actions on Euclidean retracts. A necessary condition for an action to be finitely 6 BARTELS, A. VCyc-amenable is that all isotropy groups of the action are virtually cyclic. Therefore, a related question is which groups admit actions on Euclidean retracts such that all isotropy groups are virtually cyclic. The only groups admitting such actions that I am aware of are hyperbolic groups. In fact, I do not even know whether or not the group Z2 admits an action on a Euclidean retract (or on a disk) such that all isotropy groups are virtually cyclic. There are actions of Z2 on disks without a global fixed point. This is a consequence of Oliver’s analysis of actions of finite groups on disks [65]. On the other hand, there are finitely generated groups for which all actions on Euclidean retracts have a global fixed point [2]. Homotopy actions. There is a generalization of Theorem 2.6 using homotopy actions. In order to be applicable to higher K-theory these actions need to be homotopy coherent. The passage from strict actions to homotopy actions is already visible in the work of Farrell-Jones where it corresponds to the passage from the asymptotic transfer used for negatively curved manifolds [28] to the focal transfer used for non-positively curved manifolds [30]. Definition 2.10 ([75, 76]). A homotopy coherent action of a group G on a space X is a continuous map ∞ a ((G×[0, 1])j ×G×X) → X Γ: j=0 such that  Γ(gk , . . . , gj , Γ(gj−1 , . . . , x))      Γ(g  k , . . . , gj gj−1 , . . . , x)   Γ(g , . . . , t , g , x) k 2 1 Γ(gk , tk , . . . , t1 , g0 , x) =  Γ(g , . . . , t k j+1 tj , . . . , g0 , x)      Γ(gk−1 , . . . , t1 , g0 , x)    x tj = 0 tj = 1 g0 = e, 0 < k gj = e, 1 ≤ j < k gk = e, 0 < k g0 = e, k = 0 Here Γ(g, −) : X → X should be thought of the action of g on X, the map Γ(g, −, h, −) : [0, 1]×X → X is a homotopy from Γ(g, −) ◦ Γ(h, −) to Γ(gh, −) and the remaining data in Γ encodes higher coherences. In order to obtain sequences of almost equivariant maps for homotopy actions it is useful to also allow the homotopy action to vary. Definition 2.11 (N -F -amenability for homotopy coherent actions). A sequence of homotopy coherent actions (Γn , Xn ) of a group G is said to be N -F -amenable if (N ) there exists a sequence of continuous maps fn : Xn → ∆F (G) such that for all k and all gk , . . . , g0 ∈ G sup d(fn (Γ(gk , tk , . . . , t1 , g0 , x), gk · · · g0 fn (x)) → 0 as n → ∞. x∈X,tk ,...,t1 ∈[0,1] Theorem 2.12 ([8, 76]). Suppose that G admits a sequence of homotopy coherent actions on Euclidean retracts of uniformly bounded dimension that is finitely F amenable. Then G satisfies the Farrell-Jones Conjecture relative to F . Remark 2.13. Groups satisfying the assumptions of Theorem 2.12 are said to be homotopy transfer reducible in [25]. The original formulations of Theorems 2.6 and 2.12 were not in terms of almost equivariant maps, but in terms of certain open covers of G×X. We recall here the formulation used for actions. (The formulation for homotopy actions is more cumbersome.) A subset U of a G-space is said to be an F -subset if there is F ∈ F such that gU = U for all g ∈ F and U ∩ gU = ∅ for all g ∈ G \ F . K-THEORY AND ACTIONS ON EUCLIDEAN RETRACTS 7 A collection U of subsets is said to be G-invariant if gU ∈ U for all g ∈ G, U ∈ U. If no point is contained in more than N + 1 members of U, then U is said to be of order (or dimension) ≤ N . A G-invariant cover by open F -subsets is said to be an F -cover. For a compact G-space X we equip now G×X with the diagonal G-action. For S ⊆ G finite an F -cover of G×X is said to be S-wide (in the G-direction) if ∀ (g, x) ∈ G×X ∃ U ∈ U such that gS×{x} ⊆ U. Then the action of G on X is N -F -amenable iff for any S ⊂ G finite there exists an S-wide F -cover U for G×X of dimension at most N [37, Prop. 4.5]. A translation from covers to maps is also used in [8, 12, 76]. From the point of view of covers (and because of the connection to the asymptotic dimension) it is natural to think of the N in N -F -amenable as a kind of dimension for the action of G on X, see [37, 72]. A further difference between the formulations used above and in the references given is that the conditions on the topology of X are formulated differently, but certainly Euclidean retracts satisfy the condition from [8]. Example 2.14. Theorem 2.12 applies to CAT(0)-groups where F = VCyc is the family of virtually cyclic subgroups [10, 76]. An application of Theorem 2.12 to GLn (Z) will be discussed in Section 4. Remark 2.15 (The Farrell-Hsiang method). There are interesting groups for which one can deduce the Farrell-Jones Conjecture using Theorems 2.6 (or 2.12) and inheritance properties. However, it is not clear that these methods can account for all that is currently known. A third method, going back to work of FarrellHsiang [27], combines induction results for finite groups [24, 73] with controlled topology/algebra. An axiomatization of this method is given in [9]. In important part of the proof of the Farrell-Jones Conjecture for solvable groups [77] is a combination of this method with Theorem 2.12. Remark 2.16 (Trace methods). The K-theory Novikov Conjecture concerns injectivity of assembly maps in algebraic K-theory, i.e., lower bounds for the algebraic K-theory of group rings. For the integral group ring of groups, that are only required to satisfy a mild homological finiteness assumption, trace methods have been used by Bökstedt-Hsiang-Madsen [20] and Lück-Reich-Rognes-Varisco [57] to obtain rational injectivity results. The latter result in particular yields interesting lower bounds for Whitehead groups. For the group ring over the ring of Schatten class operators Yu [80] proved rational injectivity of the Farrell-Jones assembly map for all groups. This is the only result I am aware of for the Farrell-Jones Conjecture that applies to all groups! 3. Flow spaces The construction of almost equivariant maps often uses the dynamic of a flow associated to the situation. Definition 3.1. A flow space for a group G is a metric space FS equipped with a flow Φ and an isometric G-action where the flow and the G-action commute. For α > 0, δ > 0, c, c′ ∈ FS we write dfol (c, c′ ) < (α, δ) to mean that there is t ∈ [−α, α] such that d(Φt (c), c′ ) < δ. Example 3.2. Let G be the fundamental group of a Riemannian manifold M . Then the sphere bundle S M̃ equipped with the geodesic flow is a flow space for the fundamental group of M . For manifolds of negative or non-positive curvature this 8 BARTELS, A. flow space is at the heart of the connection between K-theory and dynamics used to great effect by Farrell-Jones. This example has generalizations to hyperbolic groups and CAT(0)-groups. For hyperbolic groups Mineyev’s symmetric join is a flow space [61]. Alternatively, it is possible to use a coarse flow space for hyperbolic groups, see Remark 3.7 below. For groups acting on a CAT(0)-space a flow space has been constructed in [10]. It consists of all parametrized geodesics in the CAT(0)-space (technically all generalized geodesics) and the flow acts by shifting the parametrization. Almost equivariant maps often arise as compositions ϕ ψ (N ) X− → FS − → ∆F (G), where the first map is almost equivariant in an (α, δ)-sense, and the second map is G-equivariant and contracts (α, δ)-distances to ε-distances. The following Lemma summarizes this strategy. Lemma 3.3. Let X be a G-space, where G is a countable group. Let N ∈ N. Assume that there exists a flow space FS satisfying the following two conditions. (A) For any finite subset S of G there is α > 0 such that for any δ > 0 there is a continuous map ϕ : X → FS such that for x ∈ X, g ∈ S we have dfol (ϕ(gx), gϕ(x)) < (α, δ). (B) For any α > 0, ε > 0 there are δ > 0 and a continuous G-map ψ : FS → (N ) ∆F (G) such that dfol (c, c′ ) < (α, δ) =⇒ d(ψ(c), ψ(c′ )) < ε holds for all c, c′ ∈ FS . Then the action of G on X is N -F -amenable. (N ) Proof. Let S ⊂ G be finite and ε > 0. We need to construct a map f : X → ∆F (G) for which d(f (gx), gf (x)) < ε for all x ∈ X, g ∈ S. Let α be as in (A) with respect (N ) to S. Choose now δ > 0 and a G-map ψ : FS → ∆F (G) as in (B). Next choose ϕ : X → FS as in (A) with respect to this δ > 0. Then f := ψ ◦ ϕ has the required property.  Remark 3.4 (On the constructions of ϕ). Maps ϕ : X → FS as in condition (A) in Lemma 3.3 can in negatively or non-positively curved situations often be constructed using dynamic properties of the flow. We briefly illustrate this in a case already considered by Farrell and Jones. Let G be the fundamental group of a closed Riemannian manifold of strict negative sectional curvature M . Let M̃ be its universal cover and S∞ the sphere at infinity for M̃ . The action of G on M̃ extends to S∞ . For each x ∈ M̃ there is a canonical identification between the unit tangent vectors at x and S∞ : every unit tangent vector v at x determines a geodesic ray c starting in x, the corresponding point ξ ∈ S∞ is c(∞). One say that v points to ξ. The geodesic flow Φt on S M̃ has the following property. Suppose that v and v ′ are unit tangent vectors at x and x′ pointing to the same point in S∞ . Then dfol (Φt (v), Φt (v ′ )) < (α, δt ) where α depends only on d(x, x0 ) and δt → 0 uniformly in v, v ′ (still depending on d(x, x0 )). This statement uses strict negative curvature. (For closed manifolds of non-positive sectional curvature the vector v ′ has to be chosen more carefully depending on v and t; this necessitates the use of the focal transfer [30] respectively the use of homotopy coherent actions.) This contracting property of the geodesic flow can be translated into the construction of maps as in (A). Fix a point x0 ∈ M̃ . Define ϕ0 : S∞ → S M̃ by sending K-THEORY AND ACTIONS ON EUCLIDEAN RETRACTS 9 ξ to the unit tangent vector at x0 pointing to ξ. For t ≥ 0 define ϕt (ξ) := Φt (ϕ0 (ξ)). Using the contracting property of the geodesic flow is not difficult to check that for any g ∈ G there is α > 0 (roughly α = d(gx0 , x0 )) such that for any δ > 0 there is t0 satisfying ∀t ≥ t0 , ∀ξ ∈ S∞ dfol (ϕt (gξ), gϕt (ξ)) < (α, δ). Remark 3.5. Of course the space S∞ used in Remark 3.4 is not contractible and therefore not a Euclidean retract. But the compactification M̃ ∪ S∞ of M̃ is a disk, in particular a Euclidean retract. As M̃ has the homotopy type of a free G-CW -complex, there is even a G-equivariant map M̃ → ∆(d) (G) where d is the dimension of M . In particular, the action of G on M̃ is d-amenable. It is not difficult to combine the two statements to deduce that the action of G on M̃ ∪ S∞ is finitely F -amenable. This is best done via the translation to open covers of G×(M̃ ∪ S∞ ) discussed in Remark 2.13, see for example [72]. An important point in the formulation of condition (B) is the presence of δ > 0 uniform over FS . If the action of G on FS is cocompact, then a version of the Lebesgue Lemma guarantees the existence of some uniform δ > 0, i.e., it suffices (N ) to construct ψ : FS → ∆F (G) such that d(ψ(c), ψ(Φt (c)) < ε for all t ∈ [−α, α], c ∈ FS . Remark 3.6 (Long thin covers of FS ). Maps ϕ as in condition (B) of Lemma 3.3 are best constructed as maps associated to long thin covers of the flow space. These long thin covers are an alternative to the long thin cell structures employed by Farrell-Jones [28]. An open cover U of FS is said to be an α-long cover for FS if for each c ∈ FS there is U ∈ U such that Φ[−α,α] (c) ⊆ U. It is said to be α-long and δ-thick if for each c ∈ FS there is U ∈ U containing (N ) the δ-neighborhood of Φ[−α,α] (c). The construction of maps FS → ∆F (G) as in condition (B) of Lemma 3.3 amounts to finding for given α an F -cover U of FS of dimension at most N that is α-long and δ-thick for some δ > 0 (depending on α). For cocompact flow spaces such covers can be constructed in relatively great generality [11, 46]. Cocompactness is used to guarantee δ-thickness. For not cocompact flow spaces on can still find α-long covers, but without a uniform thickness, they do not provide the maps needed in (B). Remark 3.7 (Coarse flow space). We outline the construction of the coarse flow space from [5] for a hyperbolic group G. Let Γ be a Cayley graph for G. The vertex set of Γ is G. Adding the Gromov boundary to G we obtain the compact space G = G ∪ ∂G. Assume that Γ is δ-hyperbolic. The coarse flow space CF consists of all triples (ξ− , v, ξ+ ) with ξ± ∈ G and v ∈ G such that there is some geodesic from ξ− to ξ+ in Γ that passes v within distance ≤ δ. Informally, v coarsely belongs to a geodesic from ξ− to ξ+ . The coarse flow space is the disjoint union of its coarse flow lines CF ξ− ,ξ+ := {ξ− }×Γ×{ξ+ } ∩ CF . The coarse flow lines are are quasi-isometric to R (with uniform constants depending on δ). There are versions of the long thin covers from Remark 3.6 for CF . For α > 0 these are VCyc-covers U of bounded dimension that are α-long in the direction of the coarse flow lines: for (ξ− , v, ξ+ ) ∈ CF there is U ∈ U such that {ξ− }×Bα (v)×{ξ+ }∩ CF ξ− ,ξ+ ⊆ U . There is also a coarse version of the map ϕt from Remark 3.4. To define it, fix a base point v0 ∈ G. For t ∈ N, ϕt sends ξ ∈ ∂G to (v0 , v, ξ) where d(v0 , v) = t and v belongs to a geodesic from v0 to ξ. It is convenient to extend ϕt to a map G×∂G → CF , with ϕt (g, ξ) := (gv0 , v, ξ) where now d(v0 , v) = t and v belongs to 10 BARTELS, A. a geodesic from gv0 to ξ. Of course v is only coarsely well defined. Nevertheless, ϕt can be used to pull long thin covers for CF back to G×∂G. For S ⊆ G finite there are then t > 0 and α > 0 such that this yields S-wide covers for G×∂G. The proof of this last statement uses a compactness argument and it is important at this point that Γ is locally finite and that G acts cocompactly on Γ. 4. Covers at infinity The Farrell-Jones Conjecture for GLn (Z). The group GLn (Z) is not a CAT(0)group, but it has a proper isometric action on a CAT(0)-space, the symmetric space X := GLn (R)/O(n). Fix a base point x0 ∈ X. For R ≥ 0 let BR be the closed ball of radius R around x0 . This ball is a retract of X (via the radial projection along geodesics to x0 ) and inherits a homotopy coherent action ΓR from the action of GLn (Z) on X. Let F be the family of subgroups generated by the virtually cyclic and the proper parabolic subgroups of GLn (Z). The key step in the proof of the Farrell-Jones Conjecture for GLn (Z) in [13] is, in the language of Section 2, the following. Theorem 4.1. The sequence of homotopy coherent actions (BR , ΓR ) is finitely F -amenable. In particular GLn (Z) satisfies the Farrell-Jones Conjecture relative to F by Theorem 2.12. Using the transitivity principle 1.2 the Farrell-Jones Conjecture for GLn (Z) can then be proven by induction on n. The induction step uses inheritance properties of the Conjecture and that virtually poly-cyclic groups satisfy the Conjecture. The verification of Theorem 4.1 follows the general strategy of Lemma 3.3 (in a variant for homotopy coherent actions). The additional difficulty in verifying assumption (B) is that, as the action of GLn (Z) on the symmetric space is not cocompact, the action on the flow space is not cocompact either. The general results reviewed in Section 3 can still be used to construct for any α > 0 an αlong cover U for the flow space. However it is not clear that the resulting cover is δ-thick, for a δ > 0 uniformly over FS . The remedy for this short-coming is a second collection of open subsets of FS . Its construction starts with an F -cover for X at ∞, meaning here, away from cocompact subsets. Points in the symmetric space can be viewed as inner products on Rn and moving towards ∞ corresponds to degeneration of inner products along direct summands W ⊂ Zn ⊂ Rn . This in turn can be used to define horoballs in the symmetric space, one for each W , forming the desired cover [33]. For each W the corresponding horoball is invariant for the parabolic subgroup {g ∈ GLn (Z) | gW = W }, more precisely, the horoballs are F -subsets, but not VCyc-subsets. The precise properties of the cover at ∞ are as follows. Lemma 4.2 ([13, 33]). For any α > 0 there exists a collection U∞ of open F subsets of X of order ≤ n that is of Lebesgue number ≥ α at ∞, i.e., there is K ⊂ X compact such that for any x ∈ X \ GLn (Z) · K there is U ∈ U∞ containing the α-ball Bα (x) in X around x. This cover can be pulled back to the flow space where it provides a cover at ∞ for the flow space that is both (roughly) α-long and α-thick at ∞. Then one is left with a cocompact subset of the flow space where the cover U constructed first is α-long and δ-thick. This argument for GLn (Z) has been generalized to GLn (F (t)) for finite fields F , and GLn (Z[S −1 ]), for S a finite set of primes [71] using suitable generalizations of the above covers at ∞. In this case the parabolic subgroups are slightly bigger, K-THEORY AND ACTIONS ON EUCLIDEAN RETRACTS 11 in particular the induction step (on n) here uses that the Farrell-Jones Conjecture holds for all solvable groups. Using inheritance properties and building on these results the Farrell-Jones Conjecture has been verified for all subgroups of GLn (Q) [71] and all lattices in virtually connected Lie groups [44]. Relatively hyperbolic groups. We use Bowditch’s characterization of relatively hyperbolic groups [21]. A graph is fine if there are only finitely many embedded loops of a given length containing a given edge. Let P be a collection of subgroups of the countable group G. Then G is hyperbolic relative to P if G admits a cocompact action on a fine hyperbolic graph Γ such that all edge stabilizers are finite and all vertex stabilizers belong to P. The subgroups from P are said to be peripheral or parabolic. The requirement that Γ is fine encodes Farb’s Bounded Coset Penetration property [26]. Bowditch assigned a compact boundary ∆ to G as follows. As a set ∆ is the union of the Gromov boundary ∂Γ with the set of all vertices of infinite valency in Γ. The topology is the observer topology; a sequence xn converges in this topology to x if given any finite set S of vertices (not including x), for almost all n there is a geodesic from xn to x that misses S. (For general hyperbolic graphs this topology is not Hausdorff, but for fine hyperbolic graphs it is.) The main result from [5] is that if G is hyperbolic relative to P, then G satisfies the Farrell-Jones Conjecture relative to the family of subgroups F generated by VCyc and P (P needs to be closed under index two supergroups here for this to include the L-theoretic version of the Farrell-Jones Conjecture). This result is obtained as an application of Theorem 2.6. The key step is the following. Theorem 4.3 ([5]). The action of G on ∆ is finitely F -amenable. This is a direct consequence of Propositions 4.4 and 4.5 below, using the characterization of N -F -amenability from Remark 2.13 by the existence of S-wide covers of G×∆. To outline the construction of these covers and to prepare for the mapping class group we introduce some notation. Pick a G-invariant proper metric on the set E of edges of Γ; this is possible as G and E are countable and the action of G on E has finite stabilizers. For each vertex v of Γ with infinite valency let Ev be the set of edges incident to v. Write dv for the restriction of the metric to Ev . For ξ ∈ ∆, ξ 6= v we define its projection πv (ξ) to Ev as the set of all edges of Γ that are appear as initial edges of geodesics from v to ξ. This is a finite subset of Ev (this depends again of fineness of Γ). Fix a vertex v0 of finite valence as a base point. For g ∈ G, ξ ∈ ∆ define their projection distance at v by dπv (g, ξ) := dv (πv (gv0 ), πv (ξ)). For ξ = v, set dπv (g, v) := ∞. (For relative hyperbolic groups a related quantity is often called an angle; the terminology here is chosen to align better with the case of the mapping class group.) If we vary g (in a finite set) and ξ (in an open neighborhood) then for fixed v the projection distance dπv (g, ξ) varies by a bounded amount. Useful is the following attraction property for projection distances: there is Θ0 such that if dπv (g, ξ) ≥ Θ0 , then any geodesic from gv0 to ξ in Γ passes through v. Conversely, if some geodesic from gv0 to ξ misses v, then dπv (g, ξ) < Θ′0 for some uniform Θ′0 . Projection distances are used to control the failure of Γ to be locally finite. In particular, provided all projection distances are bounded by a constant Θ, a variation of the argument for hyperbolic groups (using a coarse flow space), can be adapted to provide S-long covers for the Θ-small part of G×∆. The following is a precise statement. 12 BARTELS, A. Proposition 4.4. There is N (depending only on G and ∆) such that for any Θ > 0 and any S ⊆ G finite there exists a collection U of open VCyc-subsets of G×∆ that is S-wide on the Θ-small part, i.e., if (g, ξ) ∈ G×∆ satisfies dπv (g, ξ) ≤ Θ for all vertices v, then there is U ∈ U with gS×{ξ} ⊆ U . To deal with large projection distances an explicit construction can be used (similar to the case of GLn (Z)). For (g, ξ) ∈ G×∆ let VΘ (g, ξ) := {v | dπv (g, ξ) ≥ Θ}. As a consequence of the attraction property, for sufficiently large Θ, the set VΘ (g, ξ) consists of vertices that belong to any geodesic from gv0 to ξ. In particular, it can be linearly ordered by distance from gv0 . For a fixed vertex v and Θ > 0 define W (v, Θ) ⊂ G×∆ as the (interior of the) set of all pairs (g, ξ) for which v is minimal in VΘ (g, ξ), i.e., v is the vertex closest to gv0 for which dπv (g, ξ) ≥ Θ. Then W(Θ) := {W (v, Θ) | v ∈ V } is a collection of pairwise disjoint open P-subsets of G×∆. Proposition 4.5. Let S ⊂ G be finite. Then there are θ′′ ≫ θ′ ≫ θ ≫ 0 such that W(θ) ∪ W(θ′ ) is a G-invariant collection of open P-subsets of order ≤ 1 that is S-long on the θ′′ -large part of G×∆: if dπv (g, ξ) ≥ θ′′ for some vertex v, then there is W ∈ W(θ) ∪ W(θ′ ) such that gS×{ξ} ⊂ W . A difficulty in working with the W(v, Θ) is that it is for fixed Θ not possible to control exactly how VΘ (g, ξ) varies with g and ξ. In particular whether or not a vertex v is minimal in VΘ (g, ξ) can change under small variation in g or ξ. A consequence of the attraction property that is useful for the proof of Proposition 4.5 is the following: suppose there are vertices v0 and v1 with dπvi (g, ξ) > Θ′ ≫ Θ, then the segment between v0 and v1 in the linear order of VΘ (g, ξ) is unchanged under suitable variations of (g, ξ) depending on Θ′ . Remark 4.6. A motivating example of relatively hyperbolic groups are fundamental groups G of complete Riemannian manifolds M of pinched negative sectional curvature and finite volume. These are hyperbolic relative to their virtually finitely generated nilpotent subgroups [21, 26]. In this case we can work with the sphere at ∞ of the universal cover M̃ of M . The splitting of G×S∞ into a Θ-small part and a Θ-large part can be thought of as follows. Fix a base point x0 ∈ M̃ . Instead of a number Θ we choose a cocompact subset G · K of M̃ . The small part of G×∆ consists then of all pairs (g, ξ) for which the geodesic ray from gx0 to ξ is contained in X; the large part is the complement. Under this translation the cover from Proposition 4.5 can again be thought of as a cover at ∞ for M̃ . Moreover, the vertices of infinite valency in Γ correspond to horoballs in M̃ , and projection distances to time geodesic rays spend in horoballs. Note that the action of G on the graph Γ in the definition of relative hyperbolicity we used is cocompact, but Γ is not a proper metric space. Conversely, in the above example the action of G on M̃ is no longer cocompact, but now M̃ is a proper metric space. A similar trade off (cocompact action on non proper space versus non-cocompact action on proper space) is possible for all relatively hyperbolic groups [34], assuming the parabolic subgroups are finitely generated. The mapping class group. Let Σ be a closed orientable surface of genus g with a finite set P of p marked points. We will assume 6g + 2p − 6 > 0. The mapping class group Mod(Σ) of Σ is the group of components of the group of orientation preserving homeomorphisms of Σ that leave P invariant. Teichmüller space T is the space of marked complete hyperbolic structures of finite area on Σ \ P . The mapping class group acts on Teichmüller space by changing the marking. Thurston [32] defined K-THEORY AND ACTIONS ON EUCLIDEAN RETRACTS 13 an equivariant compactification of Teichmüller space T . As a space T is a closed disk, in particular it is an Euclidean retract. The boundary of the compactification PMF := T \ T is the space of projective measured foliations on Σ. The key step in the proof of the Farrell-Jones Conjecture for Mod(Σ) is the following. Theorem 4.7 ([6]). Let F be the family of subgroups of Mod(Σ) that virtually fix a point in PMF . The action of Mod(Σ) on PMF is finitely F -amenable. From this it follows quickly that the action on T is finitely F -amenable as well, and applying Theorem 2.6 we obtain the Farrell-Jones Conjecture for Mod(Σ) relative to F . Up to passing to finite index subgroups, the groups in F are central extensions of products of mapping class groups of smaller complexity. Using the transitivity principle and inheritance properties one then obtains the Farrell-Jones Conjecture for Mod(Σ) by induction on the complexity of Σ. The only additional input in this case is that the Farrell-Jones Conjecture holds for finitely generated free abelian groups. The proof of Theorem 4.7 uses the characterization of N -F -amenability from Remark 2.13 and provides suitable covers for Mod(Σ)×PMF . Similar to the relative hyperbolic case the construction of these covers is done by splitting Mod(Σ)×PMF into two parts. Here it is natural to refer to these parts as the thick part and the thin part. (The thick part corresponds to the Θ-small part in the relative hyperbolic case.) Teichmüller space has a natural filtration by cocompact subsets. For ε > 0 the ε-thick part T≥ε ⊆ T consist of all marked hyperbolic structures such that all closed geodesics have length ≥ ε. The action of Mod(Σ) on T≥ε is cocompact [64]. Fix a base point x0 ∈ T . Given a pair (g, ξ) ∈ Mod(Σ)×PMF there is a unique Teichmüller ray cg, that starts at g(x0 ) and is “pointing towards ξ” (technically, the vertical foliation of the quadratic differential is ξ). The ε-thick part of Mod(Σ)×PMF is defined as the set of all pairs (g, ξ) for which the Teichmüller ray cg,ξ stays in T≥ε . An important tool in covering both the thick and the thin part is the complex of curves C(Σ). A celebrated result of Mazur-Minsky is that C(Σ) is hyperbolic [59]. Klarreich [49] studied a coarse projection map π : T → C(Σ) and identified the Gromov boundary ∂C(Σ) of the curve complex. In particular, the projection map has an extension π : PMF → C(Σ) ∪ ∂C(Σ). (On the preimage of the Gromov boundary this extension is a continuous map; on the complement it is still only a coarse map.) Teichmüller space is not hyperbolic, but its thick part T≥ε has a number of hyperbolic properties: The Masur criterion [58] implies that for (g, ξ) in the thick part, cg,ξ (t) → ξ as t → ∞. Moreover, the restriction of Klarreichs projection map π : PMF → C(Σ) ∪ ∂C(Σ) to the space of all such ξ is injective. A result of Minsky [63] is that geodesics c that stay in T≥ε are contracting. This is a property they share with geodesics in hyperbolic spaces: the nearest point projection T → c maps balls disjoint from c to uniformly bounded subsets. Teichmüller geodesics in the thick part T≥ε project to quasi-geodesics in the curve complex with constants depending only on ε. All these properties eventually allow for the construction of suitable covers of any thick part using a coarse flow space and methods from the hyperbolic case. A precise statement is the following. Proposition 4.8. There is d such that for any ε > 0 and any S ⊂ Mod(Σ) finite there exists a Mod(Σ)-invariant collection U of F -subsets of Mod(Σ)×PMF of order ≤ d such for any (g, ξ) for which cg,ξ stays in T≥ε there is U ∈ U with gS×{ξ} ⊆ U . 14 BARTELS, A. The action of the mapping class group on the curve complex C(Σ) does not exhibit the mapping class group as a relative hyperbolic group in the sense discussed before; the 1-skeleton of C(Σ) is not fine. Nevertheless, there is an important replacement for the projections to links used in the relatively hyperbolic case, the subsurface projections of Masur-Minsky [60]. In this case the projections are not to links in the curve complex, but to curve complexes C(Y ) of subsurfaces Y of Σ. (On the other hand, often links in the curve complex are exactly curve complexes of subsurfaces.) The theory is however much more sophisticated than in the relatively hyperbolic case. Projections are not always defined; sometimes the projection is to points in the boundary of C(Y ) and the projection distance is ∞. BestvinaBromberg-Fujiwara [17] used subsurface projections to prove that the mapping class group has finite asymptotic dimension. In their work the subsurfaces of Σ are organized in a finite number N of families Y such that two subsurfaces in the same family will always intersect in an interesting way. This has the effect that the projections for subsurfaces in the same family interact in a controlled way with each other. Each family Y of subsurfaces is organized in [17] in an associated simplicial complex, called the projection complex. The vertices of the projection complex are the subsurfaces from Y. A perturbation of the projection distances can thought of as being measured along geodesics in the projection complex and now behaves very similar as in the relative hyperbolic case, in particular the attraction property is satisfied in each projection complex. This allows the application of a variant of the construction from Proposition 4.5 for each projection complex that eventually yield the following. Proposition 4.9. Let Y be any of the finitely many families of subsurfaces. For any S ⊆ G finite there exists Θ > 0 and a Mod(Σ)-invariant collection U of F subsets of Mod(Σ)×PMF of order ≤ 1 such for any (g, ξ) for which there is Y ∈ Y with dπY (g, ξ) ≥ Θ there is U ∈ U with gS×{ξ} ⊆ U . The final piece, needed to combine Propositions 4.8 and 4.9 to a proof of Theorem 4.7, is a consequence of Rafi’s analysis of short curves in Σ along Teichmüller rays [68]: for any ε > 0 there is Θ such that if some curve of Σ is ε-short on cg,ξ (i.e., if cg,ξ is not contained in T≥ε ) then there is a subsurface Y such that dπY (g, ξ) ≥ Θ. Remark 4.10. Farrell-Jones [31] proved topological rigidity results for fundamental groups of non-positively curved manifolds that are in addition A-regular. The latter condition bounds the curvature tensor and its covariant derivatives over the manifold. All torsion free discrete subgroups of GLn (R) are fundamental groups of such manifolds. Similar to the examples discussed in this section a key difficulty in [31] is that the action of the fundamental group G of the manifold on the universal cover is not cocompact. The general strategy employed by Farrell-Jones seems however different, in particular, it does not involve an induction over some kind of complexity of G. The only groups that are considered in an intermediate step are polycyclic groups, and the argument directly reduces from G to these and then uses computations of K-and L-theory for polycyclic groups. This raises the following question: Can the family of subgroups in Theorem 4.1 be replaced with the family of virtually polycyclic subgroups? Recall that the cover constructed for the flow space at ∞ is both α-long and α-thick, while only δ-thickness is needed. So it is plausible that there exist thinner covers at ∞ that work for the family of virtually polycyclic subgroups. For the mapping class group the family from Theorem 4.7 can not be chosen to be significantly smaller; all isotropy groups for the action have to appear in K-THEORY AND ACTIONS ON EUCLIDEAN RETRACTS 15 the family. But one might ask, whether there exist N -F -amenable actions (or N F -amenable sequences of homotopy coherent actions) of mapping class groups on Euclidean retracts, where F is smaller than the family used in Theorem 4.7. References [1] S. Adams. Boundary amenability for word hyperbolic groups and an application to smooth dynamics of simple groups. Topology, 33(4):765–783, 1994. [2] G. Arzhantseva, M. R. Bridson, T. Januszkiewicz, I. J. Leary, A. Minasyan, and J. Świtkowski. Infinite groups with fixed point properties. Geom. Topol., 13(3):1229–1263, 2009. [3] A. Bartels. Squeezing and higher algebraic K-theory. K-Theory, 28(1):19–37, 2003. [4] A. Bartels. On proofs of the Farrell-Jones conjecture. In Topology and geometric group theory, volume 184 of Springer Proc. Math. Stat., pages 1–31. Springer, [Cham], 2016. [5] A. Bartels. Coarse flow spaces for relatively hyperbolic groups. Compos. Math., 153(4):745– 779, 2017. [6] A. Bartels and M. Bestvina. The Farrell-Jones Conjecture for Mapping Class Groups. Preprint, arXiv:1606.02844 [math.GT], 2016. [7] A. Bartels, T. Farrell, L. Jones, and H. Reich. On the isomorphism conjecture in algebraic K-theory. Topology, 43(1):157–213, 2004. [8] A. Bartels and W. Lück. The Borel conjecture for hyperbolic and CAT(0)-groups. Ann. of Math. (2), 175:631–689, 2012. [9] A. Bartels and W. Lück. The Farrell-Hsiang method revisited. Math. Ann., 354(1):209–226, 2012. [10] A. Bartels and W. Lück. Geodesic flow for CAT (0)-groups. Geom. Topol., 16(3):1345–1391, 2012. [11] A. Bartels, W. Lück, and H. Reich. Equivariant covers for hyperbolic groups. Geom. Topol., 12(3):1799–1882, 2008. [12] A. Bartels, W. Lück, and H. Reich. The K-theoretic Farrell-Jones conjecture for hyperbolic groups. Invent. Math., 172(1):29–70, 2008. [13] A. Bartels, W. Lück, H. Reich, and H. Rüping. K- and L-theory of group rings over GLn (Z). Publ. Math. Inst. Hautes Études Sci., 119:97–125, 2014. [14] A. Bartels and H. Reich. Coefficients for the Farrell-Jones conjecture. Adv. Math., 209(1):337– 362, 2007. [15] P. Baum and A. Connes. Geometric K-theory for Lie groups and foliations. Enseign. Math. (2), 46(1-2):3–42, 2000. [16] P. Baum, A. Connes, and N. Higson. Classifying space for proper actions and K-theory of group C ∗ -algebras. In C ∗ -algebras: 1943–1993 (San Antonio, TX, 1993), volume 167 of Contemp. Math., pages 240–291. Amer. Math. Soc., Providence, RI, 1994. [17] M. Bestvina, K. Bromberg, and K. Fujiwara. Constructing group actions on quasi-trees and applications to mapping class groups. Publ. Math. Inst. Hautes Études Sci., 122:1–64, 2015. [18] M. Bestvina, V. Guirardel, and C. Horbez. Boundary amenability of Out(fn ). Preprint, arXiv:1705.07017 [math.GR], 2017. [19] M. Bestvina and G. Mess. The boundary of negatively curved groups. J. Amer. Math. Soc., 4(3):469–481, 1991. [20] M. Bökstedt, W. C. Hsiang, and I. Madsen. The cyclotomic trace and algebraic K-theory of spaces. Invent. Math., 111(3):465–539, 1993. [21] B. H. Bowditch. Relatively hyperbolic groups. Internat. J. Algebra Comput., 22(3):1250016, 66, 2012. [22] G. Carlsson and B. Goldfarb. The integral K-theoretic Novikov conjecture for groups with finite asymptotic dimension. Invent. Math., 157(2):405–418, 2004. [23] J. F. Davis and W. Lück. Spaces over a category and assembly maps in isomorphism conjectures in K- and L-theory. K-Theory, 15(3):201–252, 1998. [24] A. W. M. Dress. Induction and structure theorems for orthogonal representations of finite groups. Ann. of Math. (2), 102(2):291–325, 1975. [25] N.-E. Enkelmann, W. Lück, M. Pieper, M. Ullmann, and C. Winges. On the Farrell-Jones Conjecture for waldhausen’s a-Theory. Preprint, arXiv:1607.06395v2 [math.GT], 2016. [26] B. Farb. Relatively hyperbolic groups. Geom. Funct. Anal., 8(5):810–840, 1998. [27] F. T. Farrell and W. C. Hsiang. The topological-Euclidean space form problem. Invent. Math., 45(2):181–192, 1978. [28] F. T. Farrell and L. E. Jones. K-theory and dynamics. I. Ann. of Math. (2), 124(3):531–569, 1986. 16 BARTELS, A. [29] F. T. Farrell and L. E. Jones. Isomorphism conjectures in algebraic K-theory. J. Amer. Math. Soc., 6(2):249–297, 1993. [30] F. T. Farrell and L. E. Jones. Topological rigidity for compact non-positively curved manifolds. In Differential geometry: Riemannian geometry (Los Angeles, CA, 1990), volume 54 of Proc. Sympos. Pure Math., pages 229–274. Amer. Math. Soc., Providence, RI, 1993. [31] F. T. Farrell and L. E. Jones. Rigidity for aspherical manifolds with π1 ⊂ GLm (R). Asian J. Math., 2(2):215–262, 1998. [32] A. Fathi, F. Laudenbach, and V. Poénaru. Thurston’s work on surfaces, volume 48 of Mathematical Notes. Princeton University Press, Princeton, NJ, 2012. Translated from the 1979 French original by Djun M. Kim and Dan Margalit. [33] D. R. Grayson. Reduction theory using semistability. Comment. Math. Helv., 59(4):600–634, 1984. [34] D. Groves and J. F. Manning. Dehn filling in relatively hyperbolic groups. Israel J. Math., 168:317–429, 2008. [35] E. Guentner, N. Higson, and S. Weinberger. The Novikov conjecture for linear groups. Publ. Math. Inst. Hautes Études Sci., (101):243–268, 2005. [36] E. Guentner, R. Tessera, and G. Yu. A notion of geometric complexity and its application to topological rigidity. Invent. Math., 189(2):315–357, 2012. [37] E. Guentner, R. Willett, and G. Yu. Dynamic asymptotic dimension: relation to dynamics, topology, coarse geometry, and C ∗ -algebras. Math. Ann., 367(1-2):785–829, 2017. [38] I. Hambleton and E. K. Pedersen. Identifying assembly maps in K- and L-theory. Math. Ann., 328(1-2):27–57, 2004. [39] U. Hamenstädt. Geometry of the mapping class groups. I. Boundary amenability. Invent. Math., 175(3):545–609, 2009. [40] N. Higson. Bivariant K-theory and the Novikov conjecture. Geom. Funct. Anal., 10(3):563– 581, 2000. [41] N. Higson and G. Kasparov. E-theory and KK-theory for groups which act properly and isometrically on Hilbert space. Invent. Math., 144(1):23–74, 2001. [42] N. Higson and J. Roe. Amenable group actions and the Novikov conjecture. J. Reine Angew. Math., 519:143–153, 2000. [43] W. C. Hsiang. Geometric applications of algebraic K-theory. In Proceedings of the International Congress of Mathematicians, Vol. 1, 2 (Warsaw, 1983), pages 99–118. PWN, Warsaw, 1984. [44] H. Kammeyer, W. Lück, and H. Rüping. The Farrell-Jones conjecture for arbitrary lattices in virtually connected Lie groups. Geom. Topol., 20(3):1275–1287, 2016. [45] D. Kasprowski. On the K-theory of groups with finite decomposition complexity. Proc. Lond. Math. Soc. (3), 110(3):565–592, 2015. [46] D. Kasprowski and H. Rüping. Long and thin covers for cocompact flow spaces. Preprint, arXiv:1502.05001, 2015. [47] D. Kasprowski, M. Ullmann, C. Wegner, and C. Winges. The A-theoretic Farrell-Jones Conjecture for virtually solvable groups. Preprint, arXiv:1609.07468 [math.KT], 2016. [48] Y. Kida. The mapping class group from the viewpoint of measure equivalence theory. Mem. Amer. Math. Soc., 196(916):viii+190, 2008. [49] E. Klarreich. The boundary at infinity of the complex of curves and the relative Teichmüller space. preprint, 1999. [50] S. Knopf. Acylindrical Actions on Trees and the Farrell-Jones Conjecture. Preprint, arXiv:1704.05671, 2017. [51] V. Lafforgue. K-théorie bivariante pour les algèbres de Banach et conjecture de Baum-Connes. Invent. Math., 149(1):1–95, 2002. [52] V. Lafforgue. La conjecture de Baum-Connes à coefficients pour les groupes hyperboliques. J. Noncommut. Geom., 6(1):1–197, 2012. [53] W. Lück. Survey on classifying spaces for families of subgroups. In Infinite groups: geometric, combinatorial and dynamical aspects, volume 248 of Progr. Math., pages 269–322. Birkhäuser, Basel, 2005. [54] W. Lück. K- and L-theory of group rings. In Proceedings of the International Congress of Mathematicians. Volume II, pages 1071–1098. Hindustan Book Agency, New Delhi, 2010. [55] W. Lück. Slides for a talk in oberwolfach: The Farrell-Jones Conjecture. http://131.220.77.52/lueck/data/MFO-seminar IV 1710.pdf, 2017. [56] W. Lück and H. Reich. The Baum-Connes and the Farrell-Jones conjectures in K- and Ltheory. In Handbook of K-theory. Vol. 1, 2, pages 703–842. Springer, Berlin, 2005. [57] W. Lück, H. Reich, J. Rognes, and M. Varisco. Algebraic K-theory of group rings and the cyclotomic trace map. Adv. Math., 304:930–1020, 2017. K-THEORY AND ACTIONS ON EUCLIDEAN RETRACTS 17 [58] H. Masur. Hausdorff dimension of the set of nonergodic foliations of a quadratic differential. Duke Math. J., 66(3):387–442, 1992. [59] H. A. Masur and Y. N. Minsky. Geometry of the complex of curves. I. Hyperbolicity. Invent. Math., 138(1):103–149, 1999. [60] H. A. Masur and Y. N. Minsky. Geometry of the complex of curves. II. Hierarchical structure. Geom. Funct. Anal., 10(4):902–974, 2000. [61] I. Mineyev. Flows and joins of metric spaces. Geom. Topol., 9:403–482, 2005. [62] I. Mineyev and G. Yu. The Baum-Connes conjecture for hyperbolic groups. Invent. Math., 149(1):97–122, 2002. [63] Y. N. Minsky. Quasi-projections in Teichmüller space. J. Reine Angew. Math., 473:121–136, 1996. [64] D. Mumford. A remark on Mahler’s compactness theorem. Proc. Amer. Math. Soc., 28:289– 294, 1971. [65] R. Oliver. Fixed-point sets of group actions on finite acyclic complexes. Comment. Math. Helv., 50:155–177, 1975. [66] N. Ozawa. Amenable actions and applications. In International Congress of Mathematicians. Vol. II, pages 1563–1580. Eur. Math. Soc., Zürich, 2006. [67] E. K. Pedersen. Controlled algebraic K-theory, a survey. In Geometry and topology: Aarhus (1998), volume 258 of Contemp. Math., pages 351–368. Amer. Math. Soc., Providence, RI, 2000. [68] K. Rafi. A characterization of short curves of a Teichmüller geodesic. Geom. Topol., 9:179– 202, 2005. [69] D. A. Ramras, R. Tessera, and G. Yu. Finite decomposition complexity and the integral Novikov conjecture for higher algebraic K-theory. J. Reine Angew. Math., 694:129–178, 2014. [70] H. Reich and M. Varisco. Algebraic K-theory, assembly maps, controlled algebra, and trace methods. Preprint, arXiv:1702.02218, 2017. [71] H. Rüping. The Farrell-Jones conjecture for S-arithmetic groups. J. Topol., 9(1):51–90, 2016. [72] D. Sawicki. On equivariant asymptotic dimension. Groups Geom. Dyn., 11(3):977–1002, 2017. [73] R. G. Swan. Induced representations and projective modules. Ann. of Math. (2), 71:552–578, 1960. [74] M. Ullmann and C. Winges. On the Farrell-Jones Conjecture for algebraic K-theory of spaces: the Farrell-Hsiang method. Preprint, arXiv:1509.07363 [math.KT], 2015. [75] R. M. Vogt. Homotopy limits and colimits. Math. Z., 134:11–52, 1973. [76] C. Wegner. The K-theoretic Farrell-Jones conjecture for CAT(0)-groups. Proc. Amer. Math. Soc., 140(3):779–793, 2012. [77] C. Wegner. The Farrell-Jones conjecture for virtually solvable groups. J. Topol., 8(4):975– 1016, 2015. [78] G. Yu. The Novikov conjecture for groups with finite asymptotic dimension. Ann. of Math. (2), 147(2):325–355, 1998. [79] G. Yu. The coarse Baum-Connes conjecture for spaces which admit a uniform embedding into Hilbert space. Invent. Math., 139(1):201–240, 2000. [80] G. Yu. The Novikov conjecture for algebraic K-theory of the group algebra over the ring of Schatten class operators. Adv. Math., 307:727–753, 2017. WWU Münster, Mathematisches Institut, Einsteinstr. 62, D-48149 Münster, Germany E-mail address: [email protected] URL: http://www.math.uni-muenster.de/u/bartelsa
4math.GR
J. SALLY’S QUESTION AND A CONJECTURE OF Y. SHIMODA arXiv:1209.0672v1 [math.AC] 4 Sep 2012 SHIRO GOTO, LIAM O’CARROLL AND FRANCESC PLANAS-VILANOVA Abstract. In 2007, Y. Shimoda, in connection with a long-standing question of J. Sally, asked whether a Noetherian local ring, such that all its prime ideals different from the maximal ideal are complete intersections, has Krull dimension at most two. In this paper, having reduced the conjecture to the case of dimension three, if the ring is regular and local of dimension three, we explicitly describe a family of prime ideals of height two minimally generated by three elements. Weakening the hypothesis of regularity, we find that, to achieve the same end, we need to add extra hypotheses, such as completeness, infiniteness of the residue field and the multiplicity of the ring being at most three. In the second part of the paper we turn our attention to the category of standard graded algebras. A geometrical approach via a double use of a Bertini Theorem, together with a result of A. Simis, B. Ulrich and W.V. Vasconcelos, allows us to obtain a definitive answer in this setting. Finally, by adapting work of M. Miller on prime Bourbaki ideals in local rings, we detail some more technical results concerning the existence in standard graded algebras of homogeneous prime ideals with an “excessive”number of generators. 1. Introduction It is by now a classic result that in a Noetherian local ring, the existence of a uniform bound on the minimal number of generators of all its ideals is equivalent to its Krull dimension being at most 1. In 1978, J. Sally (see [Sal, p. 52]) extended this result in the following way. Let (R, m, k) be a Noetherian local ring. Then there exists an integer N ≥ 1 such that the minimal number of generators µ(I) of an ideal I is bounded above by N, for any ideal I of R such that m is not an associated prime of I, if and only if dim(R), the Krull dimension of R, is at most 2. In particular, if dim(R) ≤ 2, then there exists a bound on the minimal number of generators of all its prime ideals. She then remarked that it is an open question whether the converse is true (cf. [op cit., p. 53]). In other words, if (R, m, k) is a Noetherian local ring such that there exists an integer N ≥ 1 such that µ(p) ≤ N, for all prime ideals p of R, is then dim(R) ≤ 2? This question has remained open and not much progress has been made since that time. In 2007, Y. Shimoda ([Shi]) asked whether a Noetherian local ring, such that all its prime ideals different from the maximal ideal are complete intersections, has Krull dimension at most 2. Observe that, if all the primes other than the maximal ideal are complete intersections, then in particular the cardinalities of their sets of minimal generators are bounded above by the Krull dimension of the ring. Date: September 5, 2012. 1 Though the question of Y. Shimoda seems easier than that of J. Sally because its hypothesis is at first sight much stronger, it has proved to be difficult to answer. For the sake of simplicity we will call a Noetherian local ring (R, m, k) a Shimoda ring if every prime ideal in the punctured spectrum is of the principal class, i.e., the minimal number of generators µ(p) of every prime ideal p of R, p 6= m, is equal to height(p). We first observe that one can reduce the conjecture to the consideration of a UFD local domain R of Krull dimension at most 3. A positive answer to Shimoda’s conjecture then amounts to showing that either dim(R) = 2, or else dim(R) = 3 and then exhibiting a prime ideal of height 2 minimally generated by 3 (or more) elements. This question is attractive because of the combination of its simplicity and its seeming difficulty. Indeed, we are not able, up to now, to produce in full generality a non-complete intersection prime ideal of height 2 in a UFD local domain of Krull dimension 3. However, in Section 2, we are able to give some partial positive answers. Firstly, if the ring R is regular and local of Krull dimension 3, then we explicitly describe a family of prime ideals of height 2 minimally generated by 3 elements. These ideals are determinantal ideals of 2 × 3-matrices and, in the geometric case, they are precisely the defining ideals of irreducible affine space monomial curves. Next, when trying to weaken the hypothesis of regularity, we find that we need to add extra hypotheses, such as completeness, infiniteness of the residue field and the multiplicity of the ring being at most 3. In this case, if dim(R) = 3, we exhibit an ideal of height 2 minimally generated by 3 elements which has a minimal prime over it which is not Gorenstein. This will enable us to conclude that a Shimoda ring, with these extra hypotheses, has Krull dimension at most 2. In the second part of the paper we turn our attention to the category of standard graded algebras (with a graded definition of the notion of a Shimoda ring). In Section 3 we find that a geometrical approach via a double use of a Bertini Theorem, together with a result of A. Simis, B. Ulrich and W.V. Vasconcelos (see [SUV]), allows us to obtain a definitive result. In the final section, Section 4, we sketch an adaptation of M. Miller’s arguments in [Mil] to the case of standard graded rings of interest, that allows us, under somewhat more technical hypotheses, to produce homogeneous prime ideals requiring an arbitrarily large number of generators, and also allows us to present a mild generalization of the core result of Section 3. 2. The Shimoda Conjecture for rings of small multiplicity Let (R, m, k) be a Noetherian local ring, of Krull dimension d ≥ 1, and k = A/m, the residue class field of R. For the remainder of this section we fix the following notation: the ring R will be called a Shimoda ring if every prime ideal in the punctured spectrum is of the principal class, i.e., the minimal number of generators µ(p) of every prime ideal p of R, p 6= m, is equal to height(p). 2 Remark 2.1. Let (R, m, k) be a Noetherian local ring of Krull dimension d ≥ 1. If d = 1, R is Shimoda if and only if R is a domain. In particular, a Shimoda ring need not be Gorenstein since R = k[[t3 , t4 , t5 ]] is a one-dimensional local domain that is not Gorenstein ([Eis, Ex. 21.11] or [Mat2, Ex. 18.8, p. 152]). Moreover, the completion of a Shimoda ring is not necessarily a Shimoda ring, because A = (k[x, y]/(y 2 − x2 − x3 ))(x,y) is a one-dimensional Noetherian local domain whose completion is not a domain ([Eis, p. 185 and ff.]). Remark 2.2. Let (R, m, k) be a Noetherian local ring of Krull dimension d ≥ 2. If R is Shimoda, then R is a UFD, and the converse holds if d = 2. Suppose now that d ≥ 2 and R is Shimoda. Then R is also Cohen-Macaulay. Indeed, take p a prime ideal with height(p) = d − 1. Since µ(p) = height(p), then p is generated by a regular sequence x1 , . . . , xd−1 , say (see [Dav, Remark, p. 203]). Set y ∈ m, y 6∈ p. Since p is prime, x1 , . . . , xd−1 , y is a regular sequence of length d. The purpose of this section is to prove the following result, where e(R) stands for the multiplicity of R with respect to m. Theorem 2.3. Let (R, m, k) be a Shimoda ring of Krull dimension d ≥ 2. Suppose that, in addition, either (a) R is regular, or (b) R is complete, R ⊃ k, k is infinite and e(R) ≤ 3. Then d = 2. Note in passing that when we weaken the hypothesis from “regular” (e(R) = 1) to “e(R) ≤ 3”, we need to add the extra hypotheses that R is complete and contains its infinite residue field. Note that in either case, R is a local Cohen-Macaulay domain and we use properties of such domains without further mention. We first show how to reduce the dimension of R. Remark 2.4. Let (R, m, k) be a Shimoda ring of Krull dimension d ≥ 2. Let a ∈ m \ m2 . Then R/aR is a Shimoda ring of Krull dimension d − 1. Moreover, (a) If R is regular, then R/aR is regular. (b) If R is complete, then R/aR is complete. If k is infinite, then a can be chosen to be a superficial element and e(R/aR) = e(R). In particular, in Theorem 2.3, one can suppose that 2 ≤ d ≤ 3. Proof of the Remark 2.4. Take a ∈ m \ m2 . Given p, a prime ideal, p 6= m, such that a ∈ p, then a ∈ p \ mp. Therefore µ(p/aR) = µ(p) − 1 = height(p) − 1 = height(p/aR) and R/aR is Shimoda of Krull dimension d − 1. Clearly, if R is regular, R/aR is regular, and if R is complete, R/aR is complete. Moreover, if k is infinite, there exists a superficial element a ∈ m \ m2 and e(R/aR) = e(R) (see, e.g., [SH, Propositions 8.5.7 and 11.1.9]). Finally, suppose Theorem 2.3 true for 2 ≤ d ≤ 3 and suppose there exists a Shimoda ring R of Krull dimension d ≥ 4. By successively factoring out appropriate elements, one would get a 3 Shimoda ring of Krull dimension d = 3. But, by Theorem 2.3, in the case 2 ≤ d ≤ 3, one would deduce d = 2, a contradiction. Therefore, there can not exist Shimoda rings of Krull dimension d ≥ 4 and one has only to prove Theorem 2.3 for the case 2 ≤ d ≤ 3.  Let us fix now some notations that will hold for the rest of the section (see [OP]). Remark 2.5. Let (R, m, k) be a Cohen-Macaulay local ring and let x1 , x2 , x3 be a regular sequence in R. Take a = (a1 , a2 , a3 ) ∈ N3 and b = (b1 , b2 , b3 ) ∈ N3 . Let c = a + b, c = (c1 , c2 , c3 ). Let M be the matrix ! a1 a2 a3 x1 x2 x3 , M= xb22 xb33 xb11 and v1 = xc11 − xb22 xa33 , v2 = xc22 − xa11 xb33 and D = xc33 − xb11 xa22 , the 2 × 2 minors of M up to a change of sign. Consider I = I2 (M) = (v1 , v2 , D), the determinantal ideal generated by the 2 × 2 minors of M. Then I is a non-Gorenstein height-unmixed ideal of height two, minimally generated by three elements. For simplicity, I will be called the HN (from Herzog Northcott) ideal associated to x1 , x2 , x3 and a, b ∈ N3 . We will also set m1 = c2 c3 − a2 b3 , m2 = c1 c3 − a3 b1 , m3 = c1 c2 − a1 b2 and m(a, b) = (m1 , m2 , m3 ) ∈ N3 . Note that mi ≥ 3. Proof of Remark 2.5. That I is height-unmixed of height 2 follows from [OP, 4.2 and 2.2]. That I is minimally generated by 3 elements and is non-Gorenstein follows from [OP, 6.1] and its proof, where a non-symmetric minimal resolution of R/I is shown.  The proof of Theorem 2.3 is divided into two parts. We first state the regular case. Proposition 2.6. Let (R, m, k) be a regular local ring of Krull dimension 3. Let x1 , x2 , x3 be a regular system of parameters in R and a, b ∈ N3 . Let I be the HN ideal associated to x1 , x2 , x3 and a, b. If gcd(m(a, b)) = 1, then I is prime. In particular, R is not Shimoda. Taking a = (1, 1, 1) and b = (2, 1, 1) in Proposition 2.6, one obtains the following result. Corollary 2.7. Let (R, m, k) be a regular local ring of Krull dimension 3. Let x1 , x2 , x3 be a regular system of parameters in R. Then I = (x31 − x2 x3 , x22 − x1 x3 , x23 − x21 x2 ) is a height two prime ideal minimally generated by three elements. For the case of small multiplicity, we have the following. As above, we consider the HN ideal associated to a = (1, 1, 1) and b = (2, 1, 1). Proposition 2.8. Let (R, m, k) be a complete Gorenstein local domain of Krull dimension 3. Suppose that, in addition, R ⊃ k, k is infinite and e(R) ≤ 3. Let (x1 , x2 , x3 )R be a minimal reduction of m. Then there is a minimal prime over I = (x31 − x2 x3 , x22 − x1 x3 , x23 − x21 x2 ) which is not Gorenstein. In particular, R is not Shimoda. Let us give now a proof of Theorem 2.3 using Propositions 2.6 and 2.8. 4 Proof of Theorem 2.3 using Propositions 2.6, 2.8. Let R be a Shimoda ring of Krull dimension d ≥ 2 as in Theorem 2.3. By Remark 2.4, one can suppose that 2 ≤ d ≤ 3. If R is regular, by Proposition 2.6, d 6= 3. Suppose now that R is complete, R ⊃ k, k infinite and e(R) ≤ 3. By Remark 2.2, R is Cohen-Macaulay, and since R is complete, it admits a canonical module (see, e.g., [BH, 3.3.8]). By Remark 2.2 again, R is a UFD. Hence R is Gorenstein (see, e.g., [BH, 3.3.19]). Therefore, by Proposition 2.8, d 6= 3.  Before proving Proposition 2.6, we note the following reasonably elementary fact. Lemma 2.9. Let (S, n) be a Cohen-Macaulay local ring of Krull dimension 2. Let x, y be a system of parameters of S. Let p, q, r, s ∈ N, with 1 ≤ r < p and 1 ≤ s < q. Then lengthS (S/(xp , y q , xr y s )) = [pq − (p − r)(q − s)] · lengthS (S/(x, y)). Proof. Let I = (x, y). Since I is generated by a regular sequence, there is a natural graded isomorphism (S/I)[X, Y ] ∼ = G(I), sending X to x + I 2 and Y to y + I 2 , between the polynomial ring in two indeterminates X, Y over S/I and the associated graded ring G = G(I) = ⊕n≥0 I n /I n+1 of I. If z ∈ S, let z ∗ denote its initial form in G, i.e., z ∗ = z + I m , where z ∈ I m−1 \ I m , and 0∗ = 0. In particular, (x∗ )n = (xn )∗ and (y ∗)m = (y m )∗ , for all n, m ≥ 1. If J is an ideal of S, let J ∗ denote the homogeneous ideal of G generated by all the initial forms of elements of J. One has Jn∗ = (J ∩ I n ) + I n+1 /I n+1 for n ≥ 0 (see, e.g., [VV]). Now take J = (xp , y q , xr y s ) ⊂ I. Observe that, for n ≥ p + q, I n ⊆ J and (J ∗ )n = Gn . For each n ≥ 0, consider the short exact sequences of S-modules: 0 → (J + I n+1 )/J → (J + I n )/J → (J + I n )/(J + I n+1 ) → 0. P lengthS ((J + I n )/(J + I n+1 )). But Hence lengthS (S/J) = p+q−1 n=0 J + In ∼ In Gn In ∼ ∼ = = ∗. = n+1 n n+1 n n+1 J +I I ∩ (J + I ) (J ∩ I ) + I Jn Thus p+q−1 lengthS (S/J) = X n=0 lengthS (Gn /Jn∗ ) = X lengthS (Gn /Jn∗ ) = lengthS (G/J ∗ ). n≥0 Let L ⊆ J ∗ be the ideal generated by the initial forms of xp , y q and xr y s in G. Since the regular sequence x, y is permutable, by [KS, Remark 4(4)], L = J ∗ . Hence lengthS (G/J ∗ ) = lengthS (G/L) < ∞. Through the isomorphism (S/I)[X, Y ] ∼ = G(I), one sees that G/L is i j isomorphic to the free S/I-module with basis X Y , where (i, j) ∈ {0, . . . , p−1}×{0, . . . , s− 1} or (i, j) ∈ {0, . . . , r − 1} × {s, . . . , q − 1}. Therefore lengthS (G/L) = [pq − (p − r)(q − s)] · lengthS (S/I).  Remark 2.10. The result of K. Kiyek and J. Stückrad (cf. [KS, Remark 4(4)]) that we have used in the proof of Lemma 2.9, which deals only with the case where the regular sequence at hand is permutable, says precisely that xp , y q , xr y s is an I-standard base of 5 J = (xp , y q , xr y s ) (see, e.g., [HIO, § 13]). In this respect, one can prove the following fact, which uses a generalization of a classic theorem of Rees (cf. [BH, Theorem 1.1.7]). Theorem 2.11. Let R be a commutative ring and let z = z1 , . . . , zn be an R-sequence. Set I = (z1 , . . . , zn ). Let Z = Z1 , . . . , Zn be indeterminates over R. (a) If F (Z) ∈ R[Z] is homogeneous of degree d and F (z) ∈ I d+i , then F (Z) ∈ I i [Z]. (b) If m1 , . . . , ms is a finite set of monomials in z, then m1 , . . . , ms is an I-standard base of J = (m1 , . . . , ms ). Proof. If i = 0, the result is trivial. Rees’ Theorem is the case i = 1. Suppose i > 1 and that we have established the result in the case where i is replaced by i − 1. Suppose then that F (Z) ∈ R[Z] is homogeneous of degree d and F (z) ∈ I d+i . Then F (z) ∈ I d+i−1 , so by the inductive hypothesis the coefficients of F lie in I i−1 . Thus, with λ = (λ1 , . . . , λn ), X aλ Z1λ1 . . . Znλn , F (Z) = λ1 +...+λn =d for aλ ∈ I i−1 . Hence, for each such λ, aλ = X bλ,µ z1µ1 . . . znµn , µ1 +...+µn =i−1 for elements bλ,µ in R, with µ = (µ1 , . . . , µn ) and µ1 + . . . + µn = i − 1. Let ξ = (ξ1 , . . . , ξn ) = λ + µ, so that ξ1 + . . . + ξn = d + i − 1, and let X X bλ,µ Z1ξ1 . . . Znξn ∈ R[Z]. H(Z) = (λ1 +...+λn =d) (µ1 +...+µn =i−1) Then H is homogeneous of degree d + i − 1 and H(z) = F (z) ∈ I d+i . By Rees’ Theorem, the coefficients bλ,µ of H lie in I, and it follows that each aλ lies in I i , which proves (a). α α Let m1 , . . . , ms be monomials in z, where mi = z1 1,i . . . zn n,i , for i = 1, . . . , s. Let αi = α1,i + . . . + αn,i be the “degree” of mi , which is well-defined due to the standard isomorphism (R/IR)[Z1 , . . . , Zn ] → G(I) = ⊕d≥0 I d /d+1 , sending Zi to zi + I 2 . Let δ1 < . . . < δt be the degrees occurring in {α1 , . . . , αs }, t ≥ 1. Let Ji denote the ideal of R generated by those monomials among m1 , . . . , ms that have degree δi . We must show that, for d ≥ 0, J ∩ I d = m1 I d−α1 + . . . + ms I d−αs , or equivalently, (1) J ∩ I d = J1 I d−δ1 + . . . + Jt I d−δt , understanding that I r = R whenever r ≤ 0 and noting that I and J are not comaximal (see [VV, p. 94, lines 23-24]). We prove (1) by induction on t ≥ 1. Item (a) easily yields the case where t = 1, i.e., where all the monomials have the same degree. So we suppose that t ≥ 2 and that (1) holds when t is replaced by t − 1. Suppose first of all that d ≤ δt . Then, by induction, J ∩I d = (J1 +. . .+Jt )∩I d = ((J1 +. . .+Jt−1 )∩I d )+Jt = J1 I d−δ1 +. . .+Jt−1 I d−δt−1 +Jt I d−δt , 6 since I d−δt = R in this case. Finally, suppose on the other hand that d > δt . Then J ∩ I d = J ∩ I δt ∩ I d = (J1 + . . . + Jt ) ∩ I δt ∩ I d = (J1 I δt −δ1 + . . . + Jt ) ∩ I d , by the case just considered. But the ideal J1 I δt −δ1 + . . . + Jt is generated by monomials in z of degree δt , so by the case t = 1, (J1 I δt −δ1 + . . . + Jt ) ∩ I d = (J1 I δt −δ1 + . . . + Jt )I d−δt = J1 I d−δ1 + . . . + Jt I d−δt , as required.  To prove Proposition 2.6, we are going to use formulae from the theory of multiplicities (see, e.g., [BH, Chapter 4, Section 6] and [SH, Chapter 11, Sections 1 and 2]). b be the completion of R. Since R is regular and local with Proof of Proposition 2.6. Let R b is a maximal ideal m generated by the regular system of parameters x1 , x2 , x3 , then R b = (x1 , x2 , x3 )R, b generated by the regular system regular local ring with maximal ideal mR b Let I = (v1 , v2 , D) be the HN ideal associated to of parameters x1 , x2 , x3 (considered in R). b = (v1 , v2 , D)R b is the HN ideal associated to x1 , x2 , x3 x1 , x2 , x3 and a, b ∈ N3 in R. Then I R b Since I = I R b ∩ R, if we prove that I R b is prime, then I will be and a, b ∈ N3 , regarded in R. prime too. Therefore we can suppose that R is complete. Until further notice we assume only that R is a complete Cohen-Macaulay local ring of dimension 3 and that x1 , x2 , x3 generates a reduction of m. We do this so that we can avoid repeating core aspects of the argument when we come to prove Proposition 2.8. Take p any associated prime of I, hence of height 2. Let D = R/p. Then D is a onedimensional Noetherian domain. Let V be its integral closure in its quotient field K. Since R is a complete local ring, R is a Nagata ring (see, e.g., [Mat1, p. 234]) and hence V is a finite D-module and, in particular, V is Noetherian as a ring. Since D is a Noetherian complete local domain, its integral closure V is a local ring (see, e.g., [Die, Corollary 6.1, p. 116]). Therefore, V is a one-dimensional Noetherian local integrally closed ring, hence a DVR. Let ν denote the corresponding valuation on K. By abuse of notation, let xi denote the image in V of each xi and set νi = ν(xi ). In V , xc11 = xb22 xa33 , xc22 = xa11 xb33 and xc33 = xb11 xa22 . Applying ν to these equalities, one gets the following system of equations:  c1 ν1 = b2 ν2 + a3 ν3   , c2 ν2 = a1 ν1 + b3 ν3   c3 ν3 = b1 ν1 + a2 ν2 the third equation (say) being the sum of the first two. So we reduce to a system of two linearly independent equations, considered over Q, whose solution is (ν1 , ν2 , ν3 ) = q · m(a, b), for some q ∈ Q, q = u/v, u, v ∈ N (see [OP, 4.4]). Set l = gcd(ν1 , ν2 , ν3 ) ∈ N. Since v(ν1 , ν2 , ν3 ) = u(m1 , m2 , m3 ) and gcd(m1 , m2 , m3 ) = 1, then vl = u and l = q ∈ N. Hence νi = lmi . 7 Clearly x1 R + I = (x1 , xc22 , xc33 , xb22 xa33 ) is an m-primary ideal of R. Set S = R/x1 R, and (by abuse of notation) consider x2 , x3 as a regular sequence in S and a system of parameters of S. Since R/(x1 R + I) ∼ = S/(xc22 , xc33 , xb22 xa33 )S (and x1 R ⊂ AnnR (S)), then lengthR (R/(x1 R + I)) = lengthS (S/(xc22 , xc33 , xb22 xa33 )S), which by Lemma 2.9, is equal to m1 · lengthS (S/(x2 , x3 )S). But S/(x2 , x3 )S ∼ = R/(x1 , x2 , x3 ), so lengthS (S/(x2 , x3 )S) = lengthR (R/(x1 , x2 , x3 )) = eR (x1 , x2 , x3 ), because R is Cohen-Macaulay of Krull dimension d = 3 and (x1 , x2 , x3 ) is an m-primary ideal ([SH, 11.1.10]). Moreover, as (x1 , x2 , x3 ) is a reduction of m, then eR (x1 , x2 , x3 ) = e(R) ([SH, 1.2.5 and 11.2.1]). Since x1 R + I ⊆ x1 R + p, then lengthR (R/(x1 R + I)) ≥ lengthR (R/(x1 R + p)). But lengthR (R/(x1 R + p)) = lengthR/p ((R/p)/(x1 · R/p)) = lengthD (D/x1 D). Observe that xc22 , xc33 ∈ x1 D and so x1 D is an m/p-primary ideal of the one-dimensional Cohen-Macaulay local domain (D, m/p, k). By [SH, 11.1.10], lengthD (D/x1 D) = eD (x1 ; D). On the other hand, V is a finite Cohen-Macaulay D-module of rankD (V ) = 1, and so eD (x1 ; D) = eD (x1 ; D) · rankD (V ) = lengthD (V /x1 V ) ([BH, 4.6.11]). Set r = [kV : kD ], the degree of the extension of the residue fields of V and of D. Then lengthD (V /x1 V ) = r · lengthV (V /x1 V ). Finally, since V is a DVR, lengthV (V /x1 V ) = ν(x1 ) = ν1 = lm1 . Therefore, putting together all the (in)equalities, we have: (2) m1 · e(R) = lengthR (R/(x1 R + I)) ≥ lengthR (R/(x1 R + p)) = rlm1 . Observe that in all this reasoning, we have only used that R is a complete Cohen-Macaulay local ring of dimension 3 and that x1 , x2 , x3 generates a reduction of m. Now, using that R is regular, we have e(R) = 1 and we deduce from (2) that lengthR (R/(x1 R + I)) = lengthR (R/(x1 R + p)). On tensoring the exact sequence 0 → p/I → R/I → R/p → 0 by R/x1 R, one obtains the exact sequence 0 → L/x1 L → R/(x1 R + I) → R/(x1 R + p) → 0, where L = p/I, because x1 is not in p so x1 R ∩ p = x1 p. Then, from the equality lengthR (R/(x1 R + I)) = lengthR (R/(x1 R + p)), one deduces that lengthR (L/x1 L) = 0 and hence L = x1 L, which by the Lemma of Nakayama implies that L = 0 and I = p.  Now we turn to the proof of Proposition 2.8. Proof of Proposition 2.8. First observe that since k is infinite, there exist x1 , x2 , x3 in m such that (x1 , x2 , x3 ) is a minimal reduction of m. In particular, x1 , x2 , x3 is a system of parameters in R. Since (R, m, k) is a complete Noetherian local domain and R ⊃ k, there exists a k-algebra homomorphism ϕ : k[[X1 , X2 , X3 ]] → R, from the power series 8 ring in the three indeterminates X1 , X2 , X3 over k to R, with ϕ(Xi ) = xi , and such that if S = im(ϕ) = k[[x1 , x2 , x3 ]], then S ∼ = k[[X1 , X2 , X3 ]] via ϕ, S is a complete regular local ring, and R is a finite extension of S (see, e.g., [Mat2, 29.4] and its proof). Let ψ : k[[X1 , X2 , X3 ]] → k[[t]] be defined as ψ(X1 ) = t3 , ψ(X2 ) = t4 and ψ(X3 ) = t5 , where t is an indeterminate over k. Then ker(ψ) = (X13 − X2 X3 , X22 − X1 X3 , X32 − X12 X2 ) (see e.g., [OP, 7.8], where a proof for the polynomial case is given). Set J = ϕ(ker(ψ)), which is a prime ideal of S and set I = JR = (x31 − x2 x3 , x22 − x1 x3 , x23 − x21 x2 ), which is the HN ideal of R associated to x1 , x2 , x3 and a = (1, 1, 1) and b = (2, 1, 1). Take now any minimal prime p over I and suppose that p is Gorenstein. We will reach a contradiction. Since JR = I ⊆ p and S ⊂ R is an integral extension, then J ⊆ I ∩ S ⊆ p ∩ S and S/(p ∩ S) ⊂ R/p is an integral extension. Hence p ∩ S is a prime ideal of S of height 2 and J = p ∩ S. Set A = S/J ∼ = k[[t3 , t4 , t5 ]] ⊂ k[[t]] and D = R/p. By the Auslander-Buchsbaum formula, R is a free S-module of rankS (R) = e(S) · rankS (R) = e(R) =: e (see, e.g., [BH, 4.6.9]). By base change, it follows that R/I is a free A-module of rank e. Hence D is a torsion-free A-module of rank e′ , where 1 ≤ e′ ≤ e. We claim that e′ > 1. Indeed, suppose that e′ = 1. Then the quotient field of D can be identified with the quotient field of A. By [Eis, Ex. 21.11], the quotient field of A is k((t)) and the integral closure V of A is k[[t]]. Since A and D have the same quotient field with A ⊂ D a finite and hence integral extension, V is also the integral closure of D. By [Eis, op. cit.], the conductor A :A V of V into A equals (t3 , t4 , t5 )A = mA , the maximal ideal of A. Since A :A V is also an ideal in V and t3 ∈ A :A V , we therefore have t3 V ⊆ A :A V = mA = (t3 , t4 , t5 )A ⊆ (t3 , t4 , t5 )V = t3 V. Therefore t3 V = A :A V = mA . Now x1 D is a reduction of the maximal ideal mD of D, since (x1 , x2 , x3 )R is a reduction of m. Hence the integral closure x1 D of the ideal x1 D equals mD , so x1 V = (x1 D)V = mD V (see, e.g., [SH, 6.8.1]). Since x1 V = t3 V and A ⊂ D is integral, it follows that mD ⊆ mD V = t3 V = mA ⊆ mD . Hence mD = mD V = t3 V . In particular, D 6= V and mD is also an ideal of V . Hence mD ⊆ D :D V ( D so mD = D :D V = t3 V . Using the fact that D is Gorenstein and [SH, 12.2.2], one has 2 = 2 · lengthD (D/mD ) = 2 · lengthD (D/(D :D V )) = lengthD (V /(D :D V )) = lengthD (V /t3 V ) = r · lengthV (V /t3 V ) = 3r, where r = [kV : k], which is a contradiction. Thus e′ ≥ 2. Now, R/I and A are local rings with the same residue field. Hence, since x1 A is an mA primary ideal of the Noetherian local domain A and rankA (D) = e′ , by [SH, 11.2.6], and 9 regarding D as an R/I-module, we have the following: eR/I (x1 · R/I; D) = eA (x1 A; D) = eA (x1 A; A) · rankA (D) = 3e′ ≥ 6. Analogously, since rankA (R/I) = e ≤ 3 by hypothesis, then eR/I (x1 · R/I; R/I) = eA (x1 A; R/I) = eA (x1 A; A) · rankA (R/I) = 3e ≤ 9. But, by the associativity formula (see, e.g., [SH, 11.2.4]), letting p vary through the minimal primes in R over I, we have X 9 ≥ eR/I (x1 ; R/I) = eR/I (x1 ; R/p) · lengthRp (Rp /IRp ) ≥ 6 · (the number of such p). p Hence the number of such p equals 1 and for this unique p, lengthRp (Rp /IRp ) = 1. Therefore I = p, which is a contradiction, since I is not Gorenstein and p is Gorenstein.  3. The Shimoda Conjecture in the setting of a standard graded algebra In connection with the Shimoda property for affine rings, we note the following result. Proposition 3.1. Let S be an affine domain of Krull dimension d at least 3. Then S contains a prime ideal P that requires more than height(P ) generators. Proof. Since S is an excellent domain, the regular locus Reg(S) is a non-empty Zariski-open subset of Spec(S). Hence there exists a non-zero element s ∈ S such that Ss is a regular ring. Since S is a Hilbert ring with (0) a prime ideal, there exists a maximal ideal M ∈ S such that s∈ / M, so SM is a regular local ring. Now height(M) = d ≥ 3 and so, by Theorem 2.3(a), and its proof, there exists p ∈ Spec(SM ) having dimension 1 such that µ(p) = d. Set P = p ∩ S. Since p = PM it follows that height(P ) = d − 1, yet P requires at least d generators.  Remark 3.2. Note however that if S is also a standard graded algebra, the preceding result does not provide any information as to whether the resulting prime ideal P is or is not homogeneous. We now consider a different but analogous set-up to the local case considered in Section 2, influenced by the well-known similarities between the theories of local and standard graded rings. So let A = k[x1 , . . . , xn ] be a standard graded algebra over the field k, i.e., A is graded by the non-negative integers N0 with k sitting in degree 0 and each of the xi having degree 1. For simplicity we suppose that k is infinite (in fact, eventually we will suppose that k has characteristic 0 or is algebraically closed). Let M := (x1 , ..., xn )A denote the irrelevant ideal of A, and we take n to be the minimal number of homogeneous generators µ(M) of M (see [BH, 1.5.15(a)], noting that µ(M) = µ(MM )). We fix this notation. We also suppose that A is a Shimoda ring in the graded sense, or a gr-Shimoda ring for short, meaning that each relevant homogeneous prime ideal p of A is generated by height(p) elements, i.e., is of the principal class: note that these generators can be chosen to be homogeneous (once again, see [BH, 1.5.15(a)]). 10 Finally, we suppose throughout that A has Krull dimension at least 2. Proposition 3.3. Suppose that the gr-Shimoda ring A has Krull dimension d (with d ≥ 2). Then A is a Gorenstein domain that satisfies the Serre condition (Rd−1 ). Proof. On taking p to be a minimal prime (necessarily homogeneous) in the above condition, we see that p = (0) so that A is a domain. Next let p be a homogeneous prime that is maximal with respect to the property of being relevant. By Davis’ result [Dav, Remark p. 203], p is generated by a regular sequence of length height(p). It follows that M contains a regular sequence of length height(M). Hence A is Cohen-Macaulay (see [BH, Ex. 2.1.27(c)]). Next, since A is a homomorphic image of a polynomial ring (with the standard grading) by a homogeneous morphism of degree 0, A has a graded canonical module C (see [BH, 3.6.10 and 3.6.12(b)]). By [Eis, Ex. 21.18(b)], we may take C to be a homogeneous ideal of A. If C = A, then A is Gorenstein ([BH, 3.6.11]). Suppose on the other hand that C is a proper ideal in A. By [BH, 3.6.9 and 3.3.18], C is an unmixed ideal of height 1 whose (necessarily homogeneous) associated primes are therefore principal, by the gr-Shimoda property. Let p be any one of these associated primes, with p = (p) say. Then each Ap is a DVR, so that each Q localitzation Cp is a principal ideal, Cp = ptp Ap , say. It is easily seen that C = ( p ptp )A, since the latter ideal has the same associated primes as C and agrees with C locally at each of these primes. Hence C is principal. That A is Gorenstein now follows from [BH, 3.6.11]. Finally, to show that A satisfies (Rd−1 ), we examine Ap for each prime ideal p of height h, with h ≤ d − 1. Clearly we need only consider the case where p is non-homogeneous, so that h > 0. By the first paragraph of the proof of [BH, 1.5.8], or by [Fle, (5.1) Lemma, (b)], we see that for such p, Ap is a regular local ring, and the result follows.  We next consider the effect of Noether normalization applied to A, noting that k is infinite. This result highlights another way that the present situation is to an extent analogous to that of Section 2. We distinguish between the Krull dimension of A, denoted k-dim(A), and the dimension of A as a projective scheme, denoted pr-dim(A). Thus k-dim(A) = pr-dim(A) + 1 (see [Eis, pp. 286-287]). Proposition 3.4. Let d = k-dim(A). Then there exists a regular sequence of homogeneous elements of degree 1 in A, which we relabel as x1 , . . . , xd , such that A = k[x1 , . . . , xn ] with A a free finitely generated graded module over the subring B := k[x1 , ..., xd ], B being isomorphic to the polynomial ring over k in d variables (with the standard grading) under the natural homogeneous mapping of degree 0. Proof. See [BH, 1.5.17(c)], together with the discussion in [Sta, § 3, particularly that on p. 63]. Note also that n remains invariant under this re-writing of A, as n = µ(M) = µ(MM ).  We now show how to reduce dimension for gr-Shimoda rings. 11 Proposition 3.5. Let a be a homogeneous element of the gr-Shimoda ring A, a lying in M \ M2 . Then A/aA is again a gr-Shimoda ring. Proof. First of all we note that such an element a exists. By the graded version of Nakayama’s Lemma, M = 6 M2 since A is not a field. Since M is a homogeneous ideal, it has a set of homogeneous generators not all of which can lie in M2 , by the previous observation. Pick a to be a suitable member of this set of generators. Clearly A/aA is a standard graded ring (with k-dim(A/aA) at least 1) having irrelevant ideal M/aA. Consider a relevant homogeneous prime ideal p of A/aA. Then p = P/aA for a unique relevant homogeneous prime ideal P of A that contains a. Since a ∈ M \ M2 , a fortiori a ∈ P \ MP. Hence a is a minimal homogeneous generator of P, so that µ(p) = µ(P) − 1 = height(P) − 1 = height(p), and the result follows.  We can now give our main result. Theorem 3.6. Suppose that in the gr-Shimoda ring A, the base field k has characteristic 0. Then pr-dim(A) ≤ 2. Proof. We suppose that pr-dim(A) ≥ 3 and deduce a contradiction by finding a homogeneous ideal of height pr-dim(A) − 1 that requires a generating set of cardinality pr-dim(A). By Proposition 3.5 we may suppose that pr-dim(A) = 3. We use without mention the fact that A is a standard graded Cohen-Macaulay affine domain, together with the well-known properties of such domains, such as their being catenary. Note that these are the only hypotheses (together with the fact that char(k) = 0) used in the remainder of the argument. That is, we prove the following. “A standard graded Cohen-Macaulay affine domain over a field of characteristic zero and of Krull dimension 4 has a relevant homogeneous prime ideal p of height 2 with µ(p) = 3.” Step 1. Constructing the prototype of the prime ideal we seek. Since k-dim(A) = 4, using Proposition 3.4 we can choose a regular sequence x, y, z of homogeneous elements of degree 1 in A. Set u = x2 − yz, v = y 2 − xz and w = z 2 − xy, and let I = (u, v, w)A; note that u, v, w are homogeneous of degree 2. It is easily seen that u, v, w are all non-zero elements: for example, note that rad((u, y)A) = rad((x, y)A); hence (u, y)A has height 2, so u 6= 0. By way of motivation, we note in passing that I is an almost complete intersection Northcott ideal ([Nor], [Vas]) and it is also a determinantal ideal. (The ideal analyzed in Section 2 also had these properties, and they will prove crucial in the analysis of the ideal K that we will focus on below.) We sketch the details as follows. Setting ! x z , Φ= −z −y 12 and letting (3) ⊤ denote matrix transpose, we have that  ⊤  ⊤ Φ · x −y = u v and that w = det(Φ). Note that u, v is a regular sequence, since u is a non-zero element in the domain A and v cannot lie in any associated prime of u (which is necessarily at height 1), because (u, v, z)A has height 3 as a consequence of the equality rad((u, v, z)A) = rad((x, y, z)A). It follows from [OP, §2], since (u, v)A has grade 2, that I is grade-unmixed of height 2. Note also that ! x y z I = I2 , y z x so, given the properties of I stated above, the graded version of the Hilbert-Burch Theorem provides us with the resolution ϕ2 ϕ1 0 → A2 (−2) → A3 (−1) → A → A/I → 0, where ϕ1 =  w u v  and ϕ⊤ 2 = x y z y z x ! , so this resolution is then minimal. Hence µ(I) = 3. We also see, using the AuslanderBuchsbaum formula, that IM and hence I is a Cohen-Macaulay ideal. (For further properties of such ideals, see [OP, §2 and 5].) We now use and develop [SUV, 4.6] along similar lines, in the present setting. Consider the polynomial extension D := A[X, Y ]; note that D is a standard graded CohenMacaulay domain, on giving each of X and Y the weighting 1. Set K = (u + xX, v + xY )D :D x. In (3), on premultiplying by adj(Φ), one obtains the relationship xw = −yu − zv. One easily checks that, as a result, w − yX − zY ∈ K. We wish to show that K is a prime ideal of D, that K = L := (u + xX, v + xY, w − yX − zY )D and that µ(K) = 3 (note that K is then a homogeneous ideal of D, being generated by quadrics). The ideal K is the prototype of the prime ideal in A that we seek so as to establish our contradiction. We first show that u + xX is a prime element in D. Note that, as remarked above, u is a regular element in the domain A, since u 6= 0. Next, x is regular modulo uA, otherwise (u, x)A would have height 1 which would contradict the fact that rad((u, x)A) (= rad((x, y)A) ∩ rad((x, z)A)) has height 2. That u + xX is a prime element in A[X] follows from [Kap, Ex. 3, p. 102]; hence u + xX is also a prime element in D. Now v + xY ∈ / (u + xX)D, since v ∈ / uA, and we deduce that u + xX, v + xY is a regular sequence in D. Note also that K is a proper ideal in D, since otherwise it would follow that x ∈ (u, v)A, which would contradict the fact that x is actually regular modulo the larger ideal I ≡ (u, v, w)A. Hence K 13 is unmixed of height 2, by basic properties of the colon operation vis-à-vis finite intersections and primary ideals. Since w − yX − zY ∈ K, it is easily seen that, up to radical, K + xD contains the elements x, y, z(z − Y ). Since each of the ideals (x, y, z)D and (x, y, z − Y )D has height 3, it follows that K + xD has height 3 and hence that x is regular modulo K, since K is unmixed of height 2. But it is clear that (D/K)x is isomorphic to Ax [ ux , xv ], i.e., to the domain Ax . Hence D/K is a domain, so K is indeed a prime ideal. Recall the ideal L := (u + xX, v + xY, w − yX − zY )D introduced above. Let ! x + X −z Ψ= . −z + Y y  ⊤  ⊤ Note that Ψ · x y = u + xX v + xY and that det(Ψ) = −(w − yX − zY ). We have seen above that grade((u + xX, v + xY )D) = 2 and that K (and so L) is a proper ideal. Hence by [Nor, Theorem 2], the projective dimension of D/L is 2 and L is grade and hence height unmixed, with all its associated primes having height 2. The argument above that established that K + xD had height 3 actually showed that L + xD had height 3. Hence x is regular modulo L. Now (u + xX, v + xY )D ⊆ L ⊆ K ≡ (u + xX, v + xY )D :D x, so that on localizing at each associated prime p of L we have (u + xX, v + xY )Dp ⊆ Lp ⊆ Kp ≡ (u + xX, v + xY )Dp . Hence Lp = Kp for all such p, so L = K. Thus K = (u + xX, v + xY, w − yX − zY )D. So x z −y y x + X −z + Y K = I2 ! , and again by the Hilbert-Burch Theorem, on setting ψ1 =  w − yX − zY v + xY −(u + xX)  , ψ2⊤ = x z −y y x + X −z + Y ! , we have the minimal free resolution (4) ψ2 ψ1 0 → D 2 (−2) → D 3 (−1) → D → D/K → 0 of D/K. In particular, µ(K) = 3, and as before, an application of the Auslander-Buchsbaum formula shows that K is a Cohen-Macaulay ideal. So K has the features that we are seeking - a homogeneous prime ideal of height 2, minimally generated by 3 elements - except that the projective variety V+ (K) (consisting set-theoretically of all relevant homogeneous prime ideals containing K) lies not in Proj(A) 14 but in Proj(A[X, Y ]). We now employ a double use of a Bertini Theorem to project back into Proj(A). Step 2. Retracting back into A via the double use of a Bertini Theorem. For a standard graded Cohen-Macaulay algebra S of positive Krull dimension, note that the scheme Proj(S) is integral if and only if S is an integral domain. We use this fact below without further comment. In our setting then, the scheme Proj(D/K) is integral. Note that pr-dim(D/K) = 3. Now apply [Fle, (5.5) Satz], with the role of the fi there being played by the images of x1 , ..., xn , X, Y in D/K, each of which has grade 1. We abuse notation by continuing to write x1 , ..., xn , X, Y for these respective images. Hence, in applying [Fle, (5.5) Satz], D+ (f0 , ..., fn ) there in fact equals all of Proj(D/K). (Recall that D+ (f0 , ..., fn ) consists set-theoretically of all relevant homogeneous prime ideals that do not contain {f0 , ..., fn }.) We now make the elementary observation that if R is a ring and r is an element of R, then, for an indeterminate Z over R, we have a natural induced isomorphism R[Z]/(Z − r) ≈ R arising from the natural retraction on R[Z] that maps Z to r (and is the identity map on R). We apply this observation to a generic hyperplane section H, say, of Proj(D/K) and then to a further generic hyperplane section H ′ , say, of Proj(D/K) ∩ H, resulting finally in an integral subscheme of Proj(A). Hence, by [Fle, (5.5) Satz], for all α := (α1 , ..., αn+2 ) in a non-empty Zariski-open set in n+2 k , and following the standard notation in [Fle, Section 5], V+ (fα ) is an integral subscheme of Proj(D/K), where fα := α1 x1 + ... + αn xn + αn+1 X + αn+2 Y . Without loss of generality, we may suppose that αn+2 6= 0, using the intersection of non-empty Zariski-open sets in k n+2 . Note that K is generated by quadrics, so that the variety V+ (K) is nondegenerate (i.e., is not contained in a hyperplane). Set Y ′ = −(α1 x1 + ... + αn xn + αn+1 X)/αn+2 , D ′ = A[X]. By our elementary observation therefore, K ′ := (u + xX, v + xY ′ , w − yX − zY ′ )D ′ is a homogeneous prime ideal which is generated by quadrics, and pr-dim(D ′ /K ′ ) = 2. Fix such an α in k n+2 . Proceeding as before then, there exists a non-empty Zariski-open set in k n+1 , depending on α, such that for all β := (β1 , ..., βn+1 ) in this set, we have βn+1 6= 0, and on setting X ′ = −(β1 x1 + ... + βn xn )/βn+1 , and Y ′′ = −(α1 x1 + ... + αn xn + αn+1 X ′ )/αn+2, K ′′ := (u + xX ′ , v + xY ′′ , w − yX ′ − zY ′′ )A is a prime ideal with pr-dim(A/K ′′ ) = 1, so grade(K ′′ ) = height(K ′′ ) = 2. Now the presentation ! x z −y K = I2 y x + X −z + Y 15 carries over to give the equality x z −y ′ y x + X −z + Y ′′ K ′′ = I2 ! , and hence the resolution (4) carries over to give the minimal presentation ψ′ ψ′ 0 → A2 (−2) →2 A3 (−1) →1 A → A/K ′′ → 0, where ψ1′ =  w − yX ′ − zY ′′ v + xY ′′ −(u + xX ′ )  , (ψ2′ )⊤ = x z −y ′ y x + X −z + Y ′′ ! . In particular, µ(K ′′ ) = 3, so K ′′ is the prime ideal we seek, and this yields the desired contradiction.  Remark 3.7. Note that this result is optimal. Consider the ring A where A is the polynomial ring k[X, Y, Z], with k algebraically closed, under the natural grading. We see from the projective Nullstellensatz and the fact that A is a UFD that A is a gr-Shimoda ring, and pr-dim(A) = 2. 4. More technical results in the standard graded case We can adapt arguments used in M. Miller’s paper [Mil] to prove the following results. We provide only some additional details, as necessary. Theorem 4.1. (cf. [Mil, Corollary 2.7 (and Theorem 2.1)]). Let S be a standard graded algebra over a field k of characteristic 0, such that S is a domain, k-dim(S) ≥ 4 and S satisfies the Serre conditions (R2 ) and (S3 ). Suppose further that homogeneous prime ideals of S of height 1 are principal. Then S possesses height 2 homogeneous prime ideals requiring an arbitrarily large number of generators. Sketch of Proof. Let M be the graded second syzygy in a minimal graded free resolution of a finitely generated graded S-module of finite projective dimension at least 3, so that M itself is not free. Note that by the Evans-Griffith Syzygy Theorem (cf. [BH, Corollary 9.5.6 and subsequent remark]) rank(M) ≥ 2. Let d = rank(M) − 1. α Suppose that we have the graded presentation S c → S b → M → 0 (ignoring twists). Set A = S[Yi,j | 1 ≤ i ≤ b, 1 ≤ j ≤ d], giving each indeterminate Yi,j over S weight 1. Define φ : Ad → Ab by the matrix (Yi,j ). Let J = coker(α ⊕ φ). Hence we have the presentation α⊕φ Ac ⊕ Ad → Ab → J → 0. Clearly rank(J) = 1. Note that N := M ⊗ A ≃ Ab /α(Ac ) as graded modules, so there is an exact sequence ψ 0 → Ad → N → J → 0, 16 with the graded homomorphism ψ induced by φ. A straightforward adaptation of Miller’s argument ([op. cit., p. 31]) then shows that the graded module J is a torsion-free A-module, and so, as is easily seen, a homogeneous ideal (up to graded isomorphism). Next we show that we may suppose that the homogeneous ideal J has height 2. Suppose that J has height 1. By an obvious ‘homogeneous’ analogue of the proof of [Kap, Theorem 5(b)], given our hypotheses, a non-zero homogeneous element of A that is not a unit is a unique product of homogeneous elements each of which generates a homogeneous prime ideal (we note that, since A is a domain, a product of non-zero elements of A is homogeneous if and only if each element is homogeneous). Hence, taking a finite set of homogeneous generators of J, we can find a smallest homogeneous principal ideal aA, with a a homogeneous element, that contains J. Clearly a−1 J is a homogeneous ideal. We claim that a−1 J is of height at least 2. For otherwise, a−1 J is contained in a homogeneous prime P of height 1, and by hypothesis, P = bA for some homogeneous element b ∈ A; it would then follow that J ⊆ abA ⊆ aA, so abA = aA, i.e., bA = A, a contradiction. Hence we may suppose that J has height at least 2. Another straightforward adaptation of Miller’s argument ([op. cit., pp. 31-32]) shows that J has precisely height 2 and that J is in fact prime. Finally, on replacing A by its localization at its irrelevant maximal ideal M, we can directly apply the argument in [Mil] to prove his Corollary 2.7 with reference to the ideal J := JM + (Z1 , . . . , Zbd )AM considered in SN , where N denotes the irrelevant maximal ideal in S (cf. [op. cit., p. 33 and top of p. 34]): note that if x1 , . . . , xn are the degree 1 generators of N , then the Zl are general k-linear combinations of x1 , . . . , xn and the Yi,j , and so are homogeneous forms of degree 1. We next observe that a standard graded algebra D with irrelevant maximal ideal P is a domain if DP is a domain. It follows that the homogeneous ideal Je := J + (Z1 , . . . , Zbd )A, considered in S, is a prime ideal. Since Je requires at least as many generators as J, the result follows.  The arguments in the proof of Step 1 of the proof of Theorem 3.6 can easily be adapted for use in the following more general situation. We adopt the notation of Section 3. Theorem 4.2. Let A be a standard graded algebra over a field of characteristic 0, with A a domain. Suppose that k-dim(A) ≥ 4 and that A satisfies the Serre condition (S3 ). Then A possesses a homogeneous prime ideal P of height 2 with µ(P ) = 3. Sketch of Proof. In the light of [BH, Propositions 1.5.15(e), 1.5.11 and 1.5.12], a straightforward adaptation of the original argument means that we need only comment on how to adapt the details of Step 2, viz. the application of the double use of a Bertini Theorem, in order to finish the proof. Replace D by its localization at its irrelevant ideal, M′ say, and A by its localization at its irrelevant ideal M. 17 As before, D/K has projective dimension 2. Since A satisfies the condition (S3 ) and X, Y are indeterminates over A, depth(D) ≥ 5. By the Auslander-Buchsbaum Formula therefore, depth(D/K) ≥ 3. We can now mimic the argument in Part 2 of the proof of [Mil, Theorem 2.1, p. 33ff.], here cutting with 2 generic hyperplanes Z1 and Z2 , in the notation on p. 33 of [Mil]. By H. Flenner’s Bertini Theorem (cf. [Mil, p. 32]), D/(K +Z1 D) is analytically irreducible. Note that depth(D/(K + Z1 D)) ≥ 2. Hence another application of Bertini’s Theorem is allowed and we have that D/(K + (Z1 , Z2 )D) is again analytically irreducible. Z From the short exact sequence 0 → D →1 D → D/Z1 D → 0, we quickly establish that TorD i (D/K, D/Z1D) = 0 for i ≥ 1. Hence the minimal free resolution of D/K over D afforded by the Hilbert-Burch Theorem descends via tensoring with D/Z1 D to give the minimal free resolution over D/Z1 D of D/(K + Z1 D), this being in Hilbert-Burch form. Repeating the argument on further factoring out Z2 · D/Z1D shows that µD/(Z1 ,Z2 )D ((K + (Z1 , Z2 )D)/(Z1, Z2 )D) = 3, and the full result now follows.  Remark 4.3. Suppose we have a standard graded domain S over a field k such that every homogeneous prime ideal of height 1, respectively 2, is generated by 1, respectively 2, elements. It easily follows from the first part of the proof of [BH, Theorem 1.5.9] that S satisfies the condition (R2 ) . Hence if we further suppose that pr-dim(S) ≥ 3 and that k has characteristic 0, then S satisfies the hypotheses of the Theorem 4.1. It follows that Theorem 3.6 is a consequence of Theorem 4.1. On the other hand, it is clear that Theorem 3.6 is also a consequence of Theorem 4.2. Acknowledgment We would like to thank J.-M. Giral for drawing the paper [Mil] to our attention and to thank Abdó Roig for useful discussions. References [BH] W. Bruns, J. Herzog, Cohen-Macaulay rings, Cambridge studies in advanced mathematics 39, Cambridge University Press 1993. [Dav] E.D. Davis, Ideals of the principal class, R-sequences and a certain monoidal transformation. Pacific J. Math. 20 (1967), 197-205. [Die] J. Dieudonné, Topics in local algebra. Notre Dame Mathematical Lectures, No. 10 University of Notre Dame Press, Notre Dame, Ind. 1967. [Eis] D. Eisenbud, Commutative algebra with a view toward algebraic geometry. Graduate Texts in Mathematics, 150. Springer-Verlag, New York, 1995. [Fle] H. Flenner, Die Sätze von Bertini für lokale Ringe. Math. Ann. 229 (1977), no. 2, 97-111. [HIO] M. Herrmann, S. Ikeda, U. Orbanz, Equimultiplicity and Blowing up. An Algebraic Study. With an Appendix by B. Moonen. Springer-Verlag, Berlin, 1988. 18 [Kap] I. Kaplansky, Commutative rings. Revised edition. The University of Chicago Press, Chicago, Ill.London, 1974. [KS] K. Kiyek, J. Stückrad, Integral closure of monomial ideals on regular sequences, Rev. Mat. Iberoamericana 19 (2003), 483-508. [Mat1] H. Matsumura, Commutative algebra. Second edition. Mathematics Lecture Note Series, 56. Benjamin/Cummings Publishing Co., Inc., Reading, Mass., 1980. [Mat2] H. Matsumura, Commutative ring theory. Translated from the Japanese by M. Reid. Cambridge Studies in Advanced Mathematics, 8. Cambridge University Press, Cambridge, 1986. [Mil] M. Miller, Bourbaki’s theorem and prime ideals. J. Algebra 64 (1980), 29-36. [Nor] D.G. Northcott, A homological investigation of a certain residual ideal, Math. Ann. 150 (1963), 99-110. [OP] L. O’Carroll, F. Planas-Vilanova, Ideals of Herzog-Northcott type. Proc. Edinb. Math. Soc. (2) 54 (2011), no. 1, 161-186. [Sal] J.D. Sally, Numbers of generators of ideals in local rings, Lecture Notes in Pure and Applied Mathematics 35, Marcel Dekker, Inc., New York and Basel, 1978. [Shi] Y. Shimoda, On power stable ideals and symbolic power ideals, The Proceedings of the 3rd Japan Vietnam Joint Seminar on Commutative Algebra, 90-95, Hanoi (Vietnam), December 2-7, 2007. [Sta] R. P. Stanley, Hilbert functions of graded algebras. Advances in Math. 28 (1978), no. 1, 57-83. [SUV] A. Simis, B. Ulrich, W.V. Vasconcelos, Jacobian dual fibrations. Amer. J. Math. 115 (1993), no. 1, 47-75. [SH] I. Swanson, C. Huneke, Integral closure of ideals, rings, and modules. London Mathematical Society Lecture Note Series, 336. Cambridge University Press, Cambridge, 2006. [VV] P. Valabrega, G. Valla, Form rings and regular sequences. Nagoya Math. J. 72 (1978), 93-101. [Vas] W.V. Vasconcelos, Computational methods in commutative algebra and algebraic geometry. Algorithms and Computation in Mathematics, 2. Springer-Verlag, Berlin, 1998. Shiro Goto Department of Mathematics School of Science and Technology Meiji University Tama, Kawasaki, KANAG 214, JAPAN [email protected] Liam O’Carroll Maxwell Institute for Mathematical Sciences School of Mathematics University of Edinburgh EH9 3JZ, Edinburgh, Scotland L.O’[email protected] Francesc Planas-Vilanova Departament de Matemàtica Aplicada 1 Universitat Politècnica de Catalunya Diagonal 647, ETSEIB 08028 Barcelona, Catalunya [email protected] 19
0math.AC
Statistical and computational phase transitions in spiked tensor estimation arXiv:1701.08010v2 [math.ST] 16 Dec 2017 Thibault Lesieur† , Léo Miolane♦ , Marc Lelarge♦ , Florent Krzakala? & Lenka ZdeborovᆠAbstract We consider tensor factorization using a generative model and a Bayesian approach. We compute rigorously the mutual information, the Minimal Mean Squared Error (MMSE), and unveil informationtheoretic phase transitions. In addition, we study the performance of Approximate Message Passing (AMP) and show that it achieves the MMSE for a large set of parameters, and that factorization is algorithmically “easy” in a much wider region than previously believed. It exists, however, a “hard” region where AMP fails to reach the MMSE and we conjecture that no polynomial algorithm will improve on AMP. This study inscribes into the line of research on low-rank tensor decomposition, a problem with many applications ranging from signal processing to machine learning [2, 6, 25]. We consider the model of [24] where the tensor is a noisy version of a r-dimensional randomly generated spike and analyze the Bayesoptimal inference of the spike, compute the associated mutual information and the minimum mean-squared error (MMSE). We also investigate whether the MMSE is achievable with some known efficient algorithms, and most particularly by approximate message passing (AMP). 1 The spiked tensor model One observes an order-p tensor Y ∈ Np RN created as Y = p r (p − 1)! X N p−1 2 (Xk0 )⊗p + V , (1) k=1 Np N R is a symmetric where X10 , . . . , Xr0 ∈ RN are r unknown vectors to be inferred from Y , and V ∈ tensor accounting for noise. We denote by X the N ×r matrix that collects the r vectors Xk . The observed tensor Y can thus be seen as a rank r perturbation of a random symmetric tensor V . Consider now the setting where the X 0 is generated at random from a known prior distribution. The core question considered in this paper is: What is the best possible reconstruction of X 0 one can hope for? In fact, we can look at even more general noise than just additive one as in (1). Denote for i = 1, . . . , N , xi = (xi,1 , . . . , xi,r ) ∈ Rr the r-dimensional vector created by aggregating the ith coordinates of the r vectors This paper was presented at the IEEE International Symposium on Information Theory (ISIT) 2017 in Aachen, Germany. † Institut de Physique Théorique, CNRS & CEA & Université Paris-Saclay, Saclay, France. ? Laboratoire de Physique Statistique, CNRS & Université Pierre et Marie Curie & École Normale Supérieure & PSL Université, Paris, France. ♦ Département d’Informatique de l’ENS, École Normale Supérieure & CNRS & PSL Research University & Inria, Paris, France. 1 Xk . Assume that for 1 ≤ i ≤ N the x0i are generated independently from a probability distribution PX over Rr . We denote, for (i1 , i2 , . . . , ip ) ∈ {1, . . . , N }p Wi01 ,i2 ,...,ip = p r (p − 1)! X N p−1 2 x0i1 ,k x0i2 ,k · · · x0ip ,k . (2) k=1 For simplicity, we will assume to only observe the extra-diagonal elements of Y , i.e. the coefficients Yi1 ,i2 ,...,ip for 1 ≤ i1 < · · · < ip ≤ N . The case where all coefficients are observed can be directly deduced from this case. The observed tensor Y is generated from W 0 using a noisy component-wise output channel Pout so that   Y (3) P (Y |X 0 ) = Pout Yi1 ,i2 ,...,ip Wi01 ,i2 ,...,ip . i1 <i2 <···<ip The simplest situation corresponds to eq. (1) with additive white Gaussian noise (AWGN), i.e. Pout ( · | w) = N (w, ∆). Given the above generative model and assuming that both the prior distribution PX and the output channel Pout are known we can write the Bayes-optimal estimator of X 0 as marginalization of the following posterior distribution P (X|Y ) = N Y  1 Y PX (xi ) Pout Yi1 ,i2 ,...,ip Wi1 ,i2 ,...,ip , ZN i=1 (4) i1 <i2 <···<ip where ZN is a normalization constant depending of the observed tensor Y , Wi1 ,i2 ,...ip is defined analogously to (2) (with X instead of X 0 ). We will study this tensor estimation problem in the limit where the dimension N → ∞ while the rank r p−1 remains constant. The factor N 2 is here to ensure that information-theoretically the inference problem is neither trivially hard nor trivially easy when one deals with signals such that kxi k and the noise magnitude p are of order 1. The factor (p − 1)! is used for convenient rescaling of the signal-to-noise ratio. 2 Related work and summary of results Recently there have been numerous works on the matrix (p = 2) version of the above setting. In particular an explicit single-letter characterization of the mutual information and of the optimal Bayesian reconstruction error have been rigorously established [4, 9, 12–14]. A large part of these results rely on the approximate message passing (AMP) algorithm. For the rank-one matrix estimation problems AMP has been introduced by [23], who also derived the state evolution (SE) formula to analyze its performance, generalizing techniques developed by [5]. In [15–17] the generalization to larger rank, and general output channel, was considered. Following the theorem proven in [4, 9, 14], we know that indeed AMP is Bayes-optimal and achieves the minimum mean-squared error (MMSE) for a large set of parameters of the problem. There, however, might exist a region denoted as hard, where this is not the case, and polynomial algorithms improving on AMP are not known. In comparison, there has been much less work on Bayesian low-rank tensor estimation. In statistical physics, the measure (4) was considered for Y with random i.i.d. components. For a Gaussian PX , it is called the spherical p-spin glass [8], while for Rademacher PX it is the Ising p-spin glass [19]. AMP for tensor estimation is actually equivalent to the so-called Thouless-Anderson-Palmer equations in spin glass theory [7, 12, 27]. In the context of tensor PCA these equations have been studied by Richard and Montanari [24] for the maximum likelihood estimation. Interestingly, they showed that the hard phase was particularly large in the tensor estimation case and that, with side information (such that for each component xi,k we have its direct noisy 2 observation), the estimation problem becomes easier. However, such a kind of component-wise side information is very strong and rarely available in applications. The tight statistical limits for the present tensor model were also studied in [12] for the special case of the Rademacher (Ising) prior. For more generic priors only upper and lower bounds are known rigorously [22]. Summary of results: In this contribution, we aim to bridge the gap between what is known for the general r, Px , Pout Bayesian estimation for low-rank matrices and what is known for low-rank tensors. We present the following contributions: (A) The AMP algorithm and its state evolution analysis for the Bayes-optimal tensor estimation, see sections 3 and 4. (B) The so-called channel universality result that allows us to map any generic channel Pout on a model with additive Gaussian noise, see section 3. (C) Rigorous formula for the asymptotic mutual information and the MMSE, thus generalizing the matrix results of [4, 14], see section 5. (D) The identification of statistical and computational phase transitions. In fact, we show that as soon as the effect of a non-zero-mean prior is taken properly into account, the hard region shrinks considerably, making the tensor decomposition problem much easier than hitherto believed, at least for algorithms that do take the prior information into account. Having a reliable prior information on the distribution of xi (not on each of the components as in [24]) is rather realistic in applications, for instance when constraints of negativity or membership to clusters are imposed. This is presented in sections 4 and 6. 3 AMP algorithm & channel universality We discuss in this section the Approximate Message Passing (AMP) algorithm for the Bayesian version of the problem. This is a relatively straightforward generalization of what has been done for the low-rank matrix estimation in e.g. [17, 18, 23], i.e. p = 2 case of the present setting. In general, AMP is derived from belief propagation by taking into account that every variable in the corresponding graphical model has a large number of neighbors. Since the incoming messages are considered independent one can use the central limit theorem and represent each message as a Gaussian with a given mean x̂i ∈ Rr and covariance σi ∈ Rr×r . A crucial property, called channel universality, that the tensor-AMP shares with the low-rank matrix estimation, allows to drastically simplify the problem of tensor estimation with generic output channel Pout . The justification of this property follows closely the low-rank matrix estimation case, and we refer the reader to [13, 15, 17]. First, we define the Fisher score tensor S associated to the output channel Pout and its Fisher information ∆ as ∂ log Pout (Y, w) , ∂w w=0 "  # ∂ log Pout (Y, w) 2 ≡ EY ∼Pout (· | 0) . ∂w w=0 S ≡ (5) 1 ∆ (6) out (y,w) where it is understood in (5) that the function y 7→ ∂ log P∂w acts component-wise on Y . Informally w=0 speaking, the channel universality property states that the mutual information of the problem defined by the output channel Pout is the same as the one of a AWGN (1) with variance ∆, and that the AMP algorithm written for the Bayes-optimal inference of low-rank tensors then depends on the data tensor Y and the output channel Pout only trough the tensor S and the effective noise ∆. 3 AMP involves an auxiliary function that depends explicitly on the prior as follows. Define the probability distribution x> Ax 1 > PX (x)eB x− 2 , M(x) = (7) ZX (A, B) where ZX (A, B) is a normalization factor. Then AMP uses the function fin (A, B) ∈ Rr , A ∈ Rr×r , B ∈ Rr defined by the expectation fin (A, B) = EM(x) [x] as well as the covariance matrix ∂B fin (A, B). We shall denote the overlap of u = (u1 , . . . uN ), v = (v1 , . . . vN ) ∈ (Rr )N by u·v = n 1 X uj vj> ∈ Rr×r . N j=1 AMP is then written as an iterative update procedure on the estimates of the posterior means and co-variances x̂i and σi that uses auxiliary variables Bi ∈ Rr and A ∈ Rr×r : Bit = p (p − 1)! X N Si,i2 ,i3 ···ip x̂ti2 i2 <i3 <···<ip p−1 2 ◦ x̂ti3 ◦ ··· ◦ x̂tip   N (p − 1) 1 X t t t−1 ◦(p−2) − x̂t−1 σj ◦ (x̂ · x̂ ) (8) i ∆ N j=1 1 t t ◦(p−1) (x̂ · x̂ ) ∆ = fin (At , Bit ) At = x̂t+1 i σit+1 t = ∂B fin (A (9) (10) , Bit ) , (11) where ◦ denotes a component-wise (Hadamard) product of matrices, and x◦p the corresponding componentwise power. 4 4.1 Theoretical analysis State evolution of AMP The evolution of the AMP algorithm in the limit of large systems N → ∞ can be tracked via a low-dimensional set of equations called the state evolution (SE). For maximum-likelihood estimation the state evolution have been used in [24]. Its heuristic derivation for the present case of general rank r, prior PX , and output Pout follows line by line the matrix estimation case detailed in [17]. For the Bayes-optimal inference, SE is written in terms of an order parameter M t ∈ Rr×r describing the overlap between x̂t (the AMP estimator at iteration t) and the ground truth x0 defined as M t = x̂t · x0 , and reads h   i  ct , M ct x0 + M ct 1/2 Z x> , M t+1 = EZ,x0 fin M (12) 0 ct = (M t )◦(p−1) /∆ , M (13) where Z ∼ N (0, Ir ) and x0 ∼ PX are independent. M ◦(n) is again the n-th Hadamard power of a matrix M. We shall not present a rigorous proof of the SE for tensor estimation and rely instead on standard arguments from statistical physics. The performance of the AMP algorithm can be understood by initializing the SE at M t=0 = 0. Or when M = 0 is a fixed point of SE we initialize as M t=0 = , an infinitesimally small number (accounting for the fact that a random initialization of AMP will —due to finite size fluctuations— be infinitesimally correlated with the ground truth). We denote MAMP the fixed point of the state evolution resulting from iterations of (12-13) from this initialization. The mean-squared error achieved by tensor-AMP 4 is then MSEAMP = Tr [ΣX − MAMP ] . (14) where ΣX = Ex [xx> ]. When PX has zero mean, this is the covariance matrix of PX . 4.2 Information-theoretically optimal inference Our next goal is to analyze the performance of (possibly intractable) Bayes-optimal inference that evaluates the marginals of the posterior probability distribution (4). The error achieved by this procedure will be denoted the minimum mean-squared error (MMSE) and is formally defined as    i 2 1 1 h 2 0 MMSEN = inf E X − θ̂(Y ) = E X 0 − E[X 0 |Y ] , N N θ̂ where the infimum is taken over all measurable functions θ̂ of Y . In order to compute the MMSE it is instrumental to compute the mutual information I(X 0 ; Y ). This quantity is related to the free energy from statistical physics (see section 5 and [13]). To compute the limit of such quantities, one traditionally applies the replica method stemming from statistical physics [19]. We take advantage of the fact that for the Bayesoptimal inference the so-called replica symmetric version of this method yields the correct free energy [28]. The replica method yields r 1 X 1 I(X 0 ; Y ) −−−−→ (ΣX )pk,k0 − sup φRS (M ) , N →∞ 2p∆ N M ∈Sr+ k,k0 =1   r  1/2  p − 1 X p c c c φRS (M ) = E log ZX M , M x0 + M Mkk Z − 0 2p∆ Z,x0 0 (15) (16) k,k =1 c = M ◦(p−1) /∆, ZX (A, B) is defined in eq. (7), x0 ∼ PX and Z ∼ N (0, Ir ) are independent random where M variables. Sr+ denotes the set of r × r symmetric positive semi-definite matrices. In section 5 we prove this result for the rank-one case (r = 1). The replica free energy (16) not only provides the limit of the mutual information I(X 0 ; Y ), but thanks to an “I-MMSE Theorem” (similar to [11]) it yields the value of the MMSE for tensor estimation, see sec. 5. Denoting M ∗ = argmaxM φRS (M ) we get MMSE = lim MMSEN = Tr [ΣX − M ∗ ] . N →∞ (17) We proved (17) rigorously, but only in the rank-one case and for odd values of p, see again sec. 5. Notice that when r ≥ 2 the estimation problem is symmetric under permutations of the r columns of X 0 : (17) is not expected to be true without further assumptions. 4.3 Statistical and computational trade-off By evaluation of the derivative of (16) with respect to M one can check that critical points of (16) are fixed points of the state evolution equations (12-13) allowing all the results to be read of the curve φRS (M ): The global maximum of (16) gives the MMSE while the (possibly local) maximum reached by iteration of (12-13) from the uninformative initialization yields the MSEAMP . We now discuss the interplay between the MMSE and MSEAMP . The working hypothesis in this paper is that AMP yields lowest MSE among known polynomial algorithms. Depending on the parameters of model (4), i.e. the order of the tensor p, rank r, prior distribution PX , and output channel Pout that appears in the SE 5 only via its Fisher information ∆, we can distinguish between two cases: the easy phase where asymptotically AMP is Bayes optimal so that MMSE = MSEAMP , and the hard phase where MMSE < MSEAMP . Given both the MMSE and MSEAMP are non-decreasing in ∆ we denote the borders of the hard phase (when it exists) as follows: Information theoretic threshold ∆IT as the (limsup of the) highest ∆ for which MMSE < MSEAMP . Algorithmic threshold ∆Alg as the (liminf of the) lowest ∆ for which MMSE < MSEAMP . Another threshold used in this paper is that of a critical value ∆c defined as smallest ∆ such that for ∆ > ∆c one has MAMP = M ∗ (∆ = +∞) (the estimate one can do when the noise is infinite), and for ∆ < ∆c one has MAMP > M ∗ (∆ = +∞). Note that from the definition we must have ∆c ≥ ∆Alg . In cases where the hard phase does not exist, but ∆c < ∞ we will consider that ∆c = ∆IT = ∆Alg . Existing results on maximum likelihood estimation [24] suggest that for tensor decomposition p ≥ 3 we have ∆Alg = ∆c = 0 in the limit N → 0 considered in this paper. This means that the spiked model of low-rank tensor decomposition is algorithmically very hard, compared to matrix p = 2 case. The authors of [24] give a good account on how ∆ needs to scale with N for known polynomial algorithms to work. For the Bayes-optimal estimation the situation seems at first sight similar. Indeed, whenever the prior PX has a zero mean, for p ≥ 3 we get ∆Alg = ∆c = 0 and the hard phase is consequently huge. This can be seen as follows. Indeed if the mean of the prior PX is zero then the state evolution equations (12-13) have a fixed point M = 0. Expanding the state evolution around this fixed point we find M t+1 = h i 1 ◦(p−1) ΣX (M t ) ΣX . ∆ (18) Whenever p ≥ 3 the fixed point M = 0 is stable for all ∆ > 0. Hence ∆Alg = ∆c = 0 for priors of zero mean. A closer look, however, shows that the situation is not so pessimistic. Indeed, as soon as the mean of the prior PX is non-zero, M = 0 is no longer a fixed point of the state evolution and once we solve the state evolution equations we observe either ∆Alg > 0 (with AMP performing optimally for ∆ < ∆Alg ) or the hard phase is completely absent and AMP has information-theoretically optimal performance for all ∆. We give examples of such priors in section 6. 5 Rigorous results We present in this section rigorous results for the rank-one case (r = 1). As mentioned above, the universality property [13, 15] reduces the computation of the mutual information to the case of additive white Gaussian noise. Consider a probability distribution PX over R that admits a finite second moment ΣX . The observation model (1) reduces in the rank-one case to p (p − 1)! 0 Yi1 ,...,ip = x . . . x0ip + Vi1 ,...,ip for 1 ≤ i1 < · · · < ip ≤ N, N (p−1)/2 i1 i.i.d. where X 0 = (x01 , . . . , x0N ) i.i.d. ∼ PX and (Vi1 ,...,ip )i1 <···<ip ∼ N (0, ∆) are independent. We define the Hamiltonian p X 2 (p − 1)! (p − 1)! −1 HN (X) = ∆ Yi1 ,...,ip xi1 . . . xip − xi1 . . . xip , (19) p−1 (p−1)/2 2N N i <···<i 1 p for X = (x1 , . . . , xN ) ∈ RN . We also write dPX (X) = 6 QN i=1 dPX (xi ). The posterior distribution (4) of X 0 given Y reads then: dP (X 0 = X|Y ) = 1 dPX (X)eHN (X) , ZN (20) where ZN is the appropriate normalizing factor. Then the free energy is defined as (minus) the logarithm of ZN of the Boltzmann probability divided by N and averaged over Y . This is of particular interest since it is related to the mutual information (see [13]): I(X 0 ; Y ) = N p Σ − E[log ZN ] + O(1) . 2p∆ X In the rank-one case, the expression (16) of φRS simplifies, so that we will use in this section " Z !# r mp−1 mp−1 0 mp−1 2 p−1 p φRS : m ≥ 0 7→ E log dPX (x) exp Zx + xx − x − m , ∆ ∆ 2∆ 2p∆ (21) where E is the expectation with respect to the independent random variables x0 ∼ PX and Z ∼ N (0, 1). The proof of (15) reduces then to the following Theorem. Theorem 1 (Replica-symmetric formula for the free energy). Let PX be a probability distribution over R, with finite second moment. Then, for all ∆ > 0 FN ≡ 1 E [log ZN ] −−−−→ sup φRS (m) ≡ FRS (∆) . N →∞ m≥0 N We now define the tensor-MMSE, T-MMSEN by   p! X T-MMSEN (∆) = inf p θ̂  N  x0i1 . . . x0ip − θ̂(Y )i1 ...ip i1 <···<ip (22)  2  ,  where the infimum is taken over all measurable functions θ̂ of the observations Y . 1 . Using an “I-MMSE Theorem” (see [11]) and the fact that the tensor MMSE is achieved Let us write λ = ∆ by the posterior mean of (X 0 )⊗p given Y , it is not difficult to verify that  ∂FN N (N − 1) . . . (N − p + 1) p = Σ − T-MMSE (∆) . N X ∂λ 2pN p The arguments are the same than in the matrix (p = 2) case, see [14] Corollary 17. T-MMSE(∆) increases ∂ with the noise level ∆, so that ∂λ FN is a non-decreasing function of λ. FN is thus a convex function of λ, and ∂ ∂ so is FRS its pointwise limit. Consequently, ∂λ FN → ∂λ FRS at all values of λ at which FRS is differentiable, that is for almost every ∆ > 0. For these values of ∆, one can also verify that the maximizer m∗ of φRS is unique: we refer to [14] for a detailed proof in the matrix case p = 2. We thus obtain the following theorem: Theorem 2 (Tensor-MMSE). For almost every ∆ > 0, φRS admits a unique maximizer m∗ (∆) over R+ and T-MMSEN −−−−→ ΣpX − m∗ (∆)p . N →∞ The information-theoretic threshold ∆IT is the maximal value of ∆ such that lim T-MMSEN < ΣpX − EPX [x]2p (which is the asymptotic performance achieved by random guess). We obtain thus the precise location of the information-theoretic threshold:  ∆IT = sup ∆ > 0 m∗ (∆) > EPX [x]2 . 7 Let X = (x1 , . . . , xN ) be a sample from the posterior (4), independently of everything else. An extension of Theorem 2 of [14] (that was derived for priors PX with bounded support) to the tensor case, gives that for almost every ∆ > 0, !p N 1 X 0 E (23) xi xi − m∗ (∆)p −−−−→ 0 , N →∞ N i=1 i.e. the pth -power of the overlap X · X0 concentrates around m∗ . This leads to Theorem 3 (Vector-MMSE for odd p). Suppose that PX has a bounded support. If p is odd, then for almost every ∆ > 0 MMSEN −−−−→ ΣX − m∗ (∆). N →∞ Before showing how (23) implies Theorem 3 we need to introduce a fundamental property of Bayesian inference: the Nishimori identity. Proposition 1 (Nishimori identity). Let (X, Y ) be a couple of random variables on a polish space. Let k ≥ 1 and let X (1) , . . . , X (k) be k i.i.d. samples (given Y ) from the distribution P (X = · | Y ), independently of every other random variables. Let us denote h·i the expectation with respect to P (X = · | Y ) and E the expectation with respect to (X, Y ). Then, for all continuous bounded function f Ehf (Y, X (1) , . . . , X (k) )i = Ehf (Y, X (1) , . . . , X (k−1) , X)i . Proof. It is equivalent to sample the couple (X, Y ) according to its joint distribution or to sample first Y according to its marginal distribution and then to sample X conditionally to Y from its conditional distribution P (X = · | Y ). Thus the (k + 1)-tuple (Y, X (1) , . . . , X (k) ) is equal in law to (Y, X (1) , . . . , X (k−1) , X). We will now use Proposition 1 to prove Theorem 3. Proof of Theorem 3. Let h·i denote the expectation with respect to the posterior distribution P (X 0 = · | Y ), and let X be a sample from this distribution, independently of everything else. The best estimator of X 0 in term of mean-squared error is the posterior mean hXi = (hx1 i, . . . , hxN i). Therefore MMSEN "N "N # # X X 1 1 0 2 0 2 2 0 = E (xi − hxi i) = E (xi ) + hxi i − 2hxi xi i N N i=1 i=1 0 0 = ΣX + EhX · X i − 2EhX · Xi , where X 0 is another sample from h·i, independently of everything else. We apply now the Nishimori identity (Proposition 1) to obtain EhX · X 0 i = EhX 0 · Xi. This gives MMSEN = ΣX − EhX · X 0 i . We then deduce from (23) that EhX · X 0 i −−−−→ m∗ , because p is here supposed to be odd. This concludes N →∞ the proof. We will now prove Theorem 1. For the matrix case (p = 2), this has been proved in [4, 13, 14] and we explain here how this can be adapted to the case p ≥ 2. To prove the limit (22), one shows successively an upper bound on lim sup FN and the matching lower bound on lim inf FN . As shown in [14] (Section 6.2.2) one only need to prove Theorem 1 for input distributions PX with finite support S. We now assume to be in this situation. 8 5.1 Adding a small perturbation One of the key ingredient of the proof is the introduction of a small perturbation of our model, that takes the form of a small amount of side information. This kind of techniques are frequently used for the study of spin glasses, where these small perturbations forces the Gibbs measure to verify some crucial identities, see [21]. In our context of Bayesian inference, we will see that small quantities of side information “breaks” the correlations of the signal variables under the posterior distribution. Let us fix  ∈ [0, 1], and suppose we have access to the additional information, for 1 ≤ i ≤ N Yi0 = ( x0i if Li = 1, ∗ if Li = 0, (24) where Li i.i.d. ∼ Ber() and ∗ is a symbol that does not belong to R. The posterior distribution of X is now    Y Y 1  P (X|Y, Y 0 ) = 1(xi = Yi0 )  PX (xi ) eHN (X) , ZN, 0 0 i|Yi 6=∗ i|Yi =∗ where ZN, is the appropriate normalization constant. For X = (x1 , . . . , xN ) ∈ RN we will use the notation X̄ = (L1 x01 + (1 − L1 )x1 , . . . , LN x0N + (1 − LN )xN ) . (25) X̄ is thus obtained by replacing the coordinates of X that are revealed by Y 0 by their revealed values. The notation X̄ will allow us to obtain a convenient expression for the free energy of the perturbed model FN, = i X 1 h 1 E log ZN, = E log PX (X)eHN (X̄) . N N N X∈S The next lemma shows that the perturbation does not change the free energy up to the order of . Recall that we supposed the support S of PX to be finite, so we can find a constant K such that S ⊂ [−K, K]. Lemma 1. For all n ≥ 1 and all , 0 ∈ [0, 1], we have |FN, − FN,0 | ≤ K 2p | − 0 |. ∆ Lemma 1 follows from a direct adaptation of Proposition 23 from [14] to the tensor case. Consequently, if we suppose  ∼ U([0, 1]) and define N = N −1/2  and Li i.i.d. ∼ Ber(N ), independently of everything else, we have |FN − E [FN,N ]| −−−−→ 0 , (26) N →∞ where E denotes the expectation with respect to  only. It remains therefore to compute the limit of the free energy under a small perturbation. As shown in [20], the perturbation (24) forces the correlations to vanish asymptotically. Lemma 2 (Lemma 3.1 from [20]).  E   1 N2 X I(x0i ; x0j | Y, Y 0 ) ≤ 1≤i,j≤N 2H(PX ) √ . N Let us write h·i the expectation with respect to P (X = · | Y, Y 0 ), and let X (1) , X (2) be two independents samples from P (X = · | Y, Y 0 ), independently of everything else. We define Q = hX (1) · X (2) i. Notice 9 that Q is a non-negative random variable. As a consequence of Lemma 2, the overlaps under the posterior distribution concentrates around Q: Lemma 3.  E X (1) ·X (2) 2  −−−−→ 0 and −Q N →∞  E X (1) 0 ·X −Q 2  −−−−→ 0 , N →∞ (27) where E denotes the expectation with respect all random variables. Lemma 3 follows from the arguments of section 4.4 from [14]. The arguments presented in this section are robust and apply to a large class of Hamiltonians. In particular, we will be able to apply in the sequel Lemmas 1 and 3 to other Hamiltonians and posterior distributions (and corresponding free energies). 5.2 Guerra’s interpolation scheme The lower bound is obtained by extending the bound derived for p = 2 in [13], using a Guerra-type interpolation [10] as was already done for tensors by Korada and Macris in [12] (who consider tensors in the special case of Rademacher PX ). Lemma 4. lim inf FN ≥ sup φRS (m) . N →∞ m≥0 Proof. We use a Guerra-type interpolation [10]: Let 0 ≤ t ≤ 1 and m ∈ R+ . We suppose to observe Y and Ỹ given by  p  t(p − 1)! 0   Yi1 ,...,ip = x . . . x0ip + Vi1 ,...,ip for 1 ≤ i1 < · · · < ip ≤ N N (p−1)/2 i1 p    Ỹj = (1 − t)mp−1 x0 + Ṽj for 1 ≤ j ≤ N j where the variables Vi1 ,...,ip and Ṽj are i.i.d. N (0, ∆) random variables. We define the interpolating Hamiltonian p X t(p − 1)! t(p − 1)! −1 HN,t (X) = ∆ Yi1 ,...,ip xi1 . . . xip − (xi1 . . . xip )2 (p−1)/2 2N p−1 N i1 <···<ip + ∆−1 N p X 1 (1 − t)mp−1 Ỹj xj − (1 − t)mp−1 x2j . 2 j=1 Then, the posterior distribution of X 0 given Y and Ỹ reads P (X 0 = X|Y, Ỹ ) = 1 PX (X) exp(HN,t (X)) , ZN,t (28) where ZN,t is the appropriate normalization. Let ψN (t) = N1 E[log ZN,t ] be the corresponding free energy. Notice that ( ψN (1) = FN , p ψN (0) = φRS (m) − (1−p)m . 2∆p Let h·it denote the expectation with respect to the posterior (28) and let X be a sampled from (28), independently of everything else. 10 Using Gaussian integration by parts and the Nishimori identity of Proposition 1 one can show (see [13,14]) that for all 0 ≤ t ≤ 1 0 ψN (t) = 1 E 2∆p X · X0 p − pmp−1 X · X 0  + oN (1) . t By convexity of the function a 7→ ap on R+ we have, for all a, b ≥ 0: ap − pbp−1 a ≥ (1 − p)bp . We would p 0 (t) ≥ (1−p)m . This would conclude like to use this inequality with a = X · X 0 and b = m to obtain that ψN 2∆p the proof of the lower bound because  Z lim inf FN = lim inf ψN (1) = lim inf ψN (0) + N →∞ N →∞ N →∞ 1  0 ψN (t)dt 0 ≥ φRS (m) . However, we do not know that X · X 0 ≥ 0 almost surely. To bypass this issue we can add, as in sec. 5.1, a small perturbation (24) that forces X · X 0 concentrates around a non-negative value (Lemma 3), without affecting the “interpolating free energy” ψN (t) in the N → ∞ limit, see (26). The arguments are the same than in sec. 5.1, so we omit the details and the rewriting of the previous calculations with the perturbation term. This concludes the proof. 5.3 Proving the upper-bound: Aizenman-Sims-Starr scheme We are now going to show how the arguments of [14] for the upper bound —using cavity computations with an Aizenman-Sims-Starr approach [1]— can be extended to the tensor case. Lemma 5. lim sup FN ≤ sup φRS (m) . N →∞ m≥0 Proof. We are going to compare the system with N variables to the system with N + 1 variables. Define P −1 AN = E[log ZN +1 ] − E[log ZN ]. Consequently, FN = N1 N k=0 Ak and lim sup FN ≤ lim sup AN . N We are thus going to upper-bound AN . Let X ∈ S be the N -first variables and σ ∈ S the (N + 1)th 0 (X) + σz(X) + σ 2 s(X) where variable. We decompose HN +1 (X, σ) = HN p ∆−1 (p − 1)! ∆−1 (p − 1)! = Y x . . . x − (xi . . . xip )2 , ip (p−1)/2 i1 ...ip i1 2(N + 1)p−1 1 (N + 1) i1 <···<ip p X (p − 1)! −1 z(X) = ∆ Y x . . . xip−1 , (p−1)/2 i1 ...ip−1 ,n+1 i1 (N + 1) i <···<i ≤n 0 HN (X) X 1 p−1 s(X) = −∆−1 X i1 <···<ip−1 ≤n (p − 1)! (xi . . . xip−1 )2 . 2(N + 1)p−1 1 0 (X) + y(X) in law, where One can also decompose HN (X) = HN −1 y(X) = ∆  1/2 X p p−1 (p − 1)! + rn Vi01 ...ip xi1 . . . xip Np i1 <···<ip    p−1 1 0 0 2 + (p − 1)! + rn xi1 . . . xip xi1 . . . xip − (xi1 . . . xip ) . Np 2 In the above definition, the V 0 are i.i.d. N (0, ∆) random variables, independent of everything else, and rn = 0 we o(N −p ). If we denote by h·i0 the Gibbs measure on S N corresponding to the Hamiltonian log PX + HN 11 can rewrite AN = E log DX PX (σ)eσz(X)+σ 2 s(X) E0 0 − E log ey(X) , (29) σ∈S where X is a sample from h·i0 , independently of everything else. AN is thus a difference of two terms that will correspond exactly to the terms of (21). As in sec. 5.1, we can show that under a small perturbation of the system, the overlap X · X 0 with the planted configuration concentrates around a non-negative value Q0 . This leads to simplifications in (29): lim sup AN ≤ lim sup E[φRS (Q0 )] ≤ FRS . N →∞ (30) N →∞ For a precise derivation of (30), the reader is invited to report to the matrix case (see [14], sec. 4.6), since there is no major difference with the tensor case on this point. The arguments presented there are commonly used in the study of spin glasses and are the analog of cavity computations in the SK model developed in [26], sec. 1.5. This concludes the proof. 6 Examples of phase transitions 0.5 1.0 0.4 SE stable SE unstable Uninformative Informative 0.4 ∆Dyn ∆Alg ∆IT 0.2 0.0 0.0 0.1 0.2 0.3 ∆ 0.4 0.5 0.3 6 ∆/ρ4 0.6 0.2 0.1 0.6 0.0 0.0 ∆Alg ∆IT 8 ∆ MSE 0.8 10 ∆Alg ∆IT 4 2 0.1 0.2 0.3 µ 0.4 0.5 0 0.00 0.05 0.10 0.15 0.20 ρ Figure 1: Left panel: Comparison between the AMP fixed point reached from uninformative (marked with crosses) or informative (i.e. strongly correlated with the ground truth, marked with pluses) initialization and the fixed point of the SE equations (stable fixed point in blue, unstable in red). The data are for the Gaussian prior with mean µ = 0.2, unit variance, p = 3, r = 1. The AMP runs are done on a system of size N = 1000. Central panel: Phase diagram for the order p = 3 tensor factorization, rank r = 1, Gaussian prior of mean µ (x-axes) and unit variance. In the green-shaded zone AMP matches the information-theoretically optimal performance, MMSE = MSE < MSEAMP . The tri-critical point is √ AMP . In the orange-shaded zone MMSE p−1 located at µTri = (p − 2)/(2 p − 1) and ∆Tri = xp−2 where xTri = (p − 2)(3p − 4)/p2 . Right Tri /(1 + xTri ) panel: Phase diagram for the order p = 3 tensor factorization, rank r = 1, the Bernoulli prior as a function of ρ and ∆/ρ4 . The tri-critical point is located at ρTri = 0.178 and ∆Tri /ρ4 = 2.60. As ρ → 0 we observed ∆Alg /ρ4 → 2e. Compare to Fig. 5 in [17] where the same phase diagram is presented for the matrix factorization p = 2 case. We used the state evolution eqs. (12-13), and the free energy (16), to compute the values of the thresholds ∆c , ∆IT and ∆Alg for several examples of the prior distributions: Gaussian (spherical spins), PX (x) = N (µ, 1); Rademacher (Ising spins), PX (x) = 21 [δ(x − 1) + δ(x + 1)]; Bernoulli (sub-tensor localization), r P PX (x) = ρδ(x − 1) + (1 − ρ)δ(x); and clustering (tensor stochastic block model), PX (x) = 1r δ(x − ~ek ), k=1 where ~ek ∈ Rr is a vector with a 1 at coordinate k and 0 elsewhere. Examples of values of the thresholds for the above priors are given in Table 1. For the zero mean Gaussian and the Rademacher prior our results for ∆IT indeed agree with those presented in [12, 22]. Central and right part of Fig. 1 present the thresholds for 12 0.25 the Gaussian and Bernoulli prior as a function of the mean µ and density ρ, respectively. Left part of Fig. 1 illustrates that indeed the fixed points of the state evolution agree with the fixed points of the AMP algorithm. HH Prior H HH p H Gaussian N (0, 1) Rademacher Bernoulli ρ = 0.1 3 clusters ∆IT p log(p) ∆Alg ∆IT ∆Alg ∆IT ρ−p ∆Alg ρ−2p+2 ∆IT ∆Alg ∆Alg r2p−2 p−1 2 2 log 2 1 1 1 − − 1 1 3 0.754 0 0.2828 0 0.577 3.738 1 1 4 0.701 0 0.1902 0 0.398 6.017 1.18 1 5 0.685 0 0.1473 0 0.311 8.251 1.62 1 10 0.677 0 0.07216 0 0.154 19.30 6.59 1 Table 1: Examples of the information-theoretic ∆IT and algorithmic ∆Alg thresholds for order-p tensor decomposition for different priors on the factors. For the Gaussian case ∆IT p log(p) converges to 1 at large p. For the Bernoulli case the rescaling in power of ρ is for convenience to present quantities of order one, we did not check if it describes the large p limit. 6.1 Results for Gaussian prior In this section we detail the analysis of the state evolution for rank r = 1 Gaussian prior of mean µ and variance 1. PXGauss = N (µ, 1) . (31) Using (12) one gets for the SE equation M t+1 = ∆µ2 + (M t )p−1 (1 + µ2 ) , ∆ + (M t )p−1 (32) where M is a scalar, and ∆ is the inverse Fisher information of the output channel. It turns out that as soon as p ≥ 3 the SE equation exhibits multiple stable fixed points. For the zero mean µ = 0 case one gets p−1 M t+1 = (M t ) . ∆ + (M t )p−1 (33) Here the fixed point M = 0 is stable whatever the noise ∆ > 0 and therefore AMP will not achieve performance better than random guessing for any ∆ > 0. Ref. [24] studies the scaling of ∆ with N for which AMP and other algorithms succeed. For positive mean µ > 0, however, the AMP algorithm is able to recover the signal for values of ∆ < ∆Alg with ∆Alg (µ) = xp−2 Alg (1 + xAlg )p−1 , ∆Dyn (µ) = p − 2 + 2µ2 − xp−2 Dyn (1 + xDyn )p−1 , p (p − 2)2 − 4µ2 (p − 1) , 2(1 + µ2 ) p p − 2 + 2µ2 + (p − 2)2 − 4µ2 (p − 1) xDyn (µ) = , 2(1 + µ2 ) xAlg (µ) = (34) (35) (36) where we defined a new threshold ∆Dyn as the smallest such that for ∆ > ∆Dyn the state evolution has a 13 unique fixed point. We know of no analytical formula for ∆IT and for Figure 1 we computed it numerically. The tri-critical point where all these curve meet is located at p−2 µTri = √ . 2 p−1 (37) Using the above expressions we derive that ∆Dyn (µ = 0) ∆Alg (µ) = ∼ µ→0   1 p − 2 p−1 1 ∼ , p→∞ ep p−2 p−1  2 p−2 µ . p−2 (38) (39) We can also compute the limit of the ∆IT (µ = 0, p) as p → ∞ and get ∆IT (µ = 0, p) ∼ p→∞ 1 . p log(p) (40) This scaling agrees with the large p behavior derived in [24] and [22]. 6.2 Results for clustering prior An interesting example of the prior for rank r tensor estimation is PXClusters (x) = 1 X δ(x − ~ek ) . r (41) 1≤k≤r This describes a model of r non-overlapping clusters. Due to the channel universality, this prior also describes the stochastic block model on dense hyper-graphs as considered for sparse hyper-graph in e.g. [3]. This model was considered in detail for p = 2 in [17]. The above clustering prior has non-zero mean, and it also exhibits the transition ∆c from a phase where recovery of clusters better than chance is not possible, to a phase where it is. To analyze the SE equations we first notice that the stable fixed point will be of the form M= bIr (1 − b)Jr + ∈ Rr×r , b ∈ [0; 1] , r r2 (42) where Ir is the identity matrix and Jr is a matrix filled with ones. b = 0 means that the estimate of the marginals does not carry any information. b = 1 means perfect reconstruction. The state evolution now becomes   p−1  t p−1  bt 1−bt + − 1−b r2 r2  r  t+1 b = Mr r (43) , ∆ where Mr is a function that was defined and studied in [15]. Its Taylor expansion is Mr (x) =  x 2r − 4 3 + x + O x . r2 2r4 (44) We further notice that b = 0 is always a fixed point of (43). By expanding (43) to first order one gets bt+1 = bt   p−1 2 + O bt . 2p−2 ∆r 14 (45) This fixed point therefore becomes unstable when ∆ < ∆c ≡ p−1 . r2p−2 (46) By analyzing eq. (43) further we can prove that ∀x ∈ R+ m(x) = Mr (x)  ∆(x) = r Mr (x) r + 1−Mr (x) r2 p−1 − (47)  1−Mr (x) r2 p−1 . (48) x m is a fixed point of (43) when m = m(x) and ∆ = ∆(x). Rather than finding the fixed point iteratively, the above equations allow us to draw all the fixed point of (43), be it stable or unstable We have that m(x) is a stable fixed point of (43) if and only if ∂∆(x) < 0. ∂x (49) The next question is whether there is a first or second order phase transition at ∆c . To answer this, one needs to analyze whether the fixed point close to b = 0 is stable or unstable. For this we to compute ∂∆(x) ∂x at x = 0 to get using (44) that ∂∆(x) p−1 = (−2p − r + pr) . ∂x 2r2p (50) Therefore if −2p − r + pr > 0 there will be no stable fixed point close to b = 0 and the system must have a first order phase transition (discontinuity in the MSEAMP ) at ∆c = ∆Alg . For two clusters r = 2, there is a second order phase transition at ∆c for all p ≥ 2. However, analyzing the state evolution numerically we observed that for p ≥ 5 there is a discontinuity later at some ∆Alg < ∆c . For three and more clusters r ≥ 3 we always have ∆Alg = ∆c , and for −2p−r +pr ≤ 0 we have not detected any other discontinuities. Values of ∆IT for three clusters and some values of p are given in Table 1. Acknowledgment This work has been supported by the ERC under the European Union’s FP7 Grant Agreement 307087-SPARCS. References [1] Michael Aizenman, Robert Sims, and Shannon L Starr. Extended variational principle for the Sherrington-Kirkpatrick spin-glass model. Physical Review B, 68(21):214403, 2003. [2] Animashree Anandkumar, Rong Ge, Daniel Hsu, Sham M Kakade, and Matus Telgarsky. Tensor decompositions for learning latent variable models. Journal of Machine Learning Research, 15(1):2773–2832, 2014. [3] Maria Chiara Angelini, Francesco Caltagirone, Florent Krzakala, and Lenka Zdeborová. Spectral detection on sparse hypergraphs. In 2015 53rd Annual Allerton Conference on Communication, Control, and Computing, page 66, 2015. [4] Jean Barbier, Mohamad Dia, Nicolas Macris, Florent Krzakala, Thibault Lesieur, and Lenka Zdeborová. Mutual information for symmetric rank-one matrix estimation: A proof of the replica formula. In Advances in Neural Inf. Proc. Systems, page 424, 2016. 15 [5] Mohsen Bayati and Andrea Montanari. The dynamics of message passing on dense graphs, with applications to compressed sensing. Information Theory, IEEE Transactions on, 57(2):764–785, 2011. [6] Andrzej Cichocki, Danilo Mandic, Lieven De Lathauwer, Guoxu Zhou, Qibin Zhao, Cesar Caiafa, and Huy Anh Phan. Tensor decompositions for signal processing applications: From two-way to multiway component analysis. IEEE Signal Processing Magazine, 32(2):145–163, 2015. [7] A Crisanti and H-J Sommers. Thouless-anderson-palmer approach to the spherical p-spin spin glass model. J. de Phys. I, 5:805, 1995. [8] Andrea Crisanti and H-J Sommers. The spherical p-spin interaction spin glass model: The statics. Zeitschrift für Physik B Condensed Matter, 87(3):341–354, 1992. [9] Y. Deshpande and A. Montanari. Information-theoretically optimal sparse PCA. In 2014 IEEE International Symposium on Information Theory, pages 2197–2201, 2014. [10] Francesco Guerra. Broken replica symmetry bounds in the mean field spin glass model. Commun. Math. Phys., 233:1, 2003. [11] Dongning Guo, Shlomo Shamai, and Sergio Verdú. Mutual information and minimum mean-square error in gaussian channels. IEEE Transactions on Information Theory, 51(4):1261–1282, 2005. [12] Satish Babu Korada and Nicolas Macris. Exact solution of the gauge symmetric p-spin glass model on a complete graph. Journal of Statistical Physics, 136(2):205–230, 2009. [13] F. Krzakala, J. Xu, and L. Zdeborová. Mutual information in rank-one matrix estimation. In IEEE Inf. Theory Workshop, pages 71–75, 2016. [14] Marc Lelarge and Léo Miolane. arXiv:1611.03888, 2016. Fundamental limits of symmetric low-rank matrix estimation. [15] T. Lesieur, F. Krzakala, and L. Zdeborová. MMSE of probabilistic low-rank matrix estimation: Universality with respect to the output channel. In 53rd Annual Allerton Conf. on Comm., Control, and Comp., 2015. [16] T. Lesieur, F. Krzakala, and L. Zdeborová. Phase transitions in sparse PCA. In IEEE Int. Symp. on Inf. Theory, page 1635, 2015. [17] Thibault Lesieur, Florent Krzakala, and Lenka Zdeborová. Constrained low-rank matrix estimation: phase transitions, approximate message passing and applications. Journal of Statistical Mechanics: Theory and Experiment, 2017(7):073403, 2017. [18] Ryosuke Matsushita and Toshiyuki Tanaka. Low-rank matrix reconstruction and clustering via approximate message passing. In Advances in Neural Information Processing Systems, page 917. 2013. [19] M. Mézard, G. Parisi, and M. A. Virasoro. Spin-Glass Theory and Beyond, volume 9. World Scientific, 1987. [20] Andrea Montanari. Estimating random variables from random sparse observations. Trans. Emerging Tel. Tech., 19(4):385, 2008. [21] Dmitry Panchenko. The Sherrington-Kirkpatrick model. Springer Science & Business Media, 2013. [22] Amelia Perry, Alexander S Wein, and Afonso S Bandeira. Statistical limits of spiked tensor models. arXiv preprint arXiv:1612.07728, 2016. [23] Sundeep Rangan and Alyson K Fletcher. Iterative estimation of constrained rank-one matrices in noise. In IEEE Int. Symp. on Inf. Theory, pages 1246–1250, 2012. 16 [24] Emile Richard and Andrea Montanari. A statistical model for tensor PCA. In Advances in Neural Information Processing Systems, page 2897, 2014. [25] Nicholas D Sidiropoulos, Lieven De Lathauwer, Xiao Fu, Kejun Huang, Evangelos E Papalexakis, and Christos Faloutsos. Tensor decomposition for signal processing and machine learning. IEEE Transactions on Signal Processing, 65(13):3551–3582, 2017. [26] Michel Talagrand. Mean field models for spin glasses: Volume I: Basic examples, volume 54. Springer Science & Business Media, 2010. [27] D. J. Thouless, P. W. Anderson, and R. G. Palmer. Solution of ’solvable model of a spin glass’. Phil. Magazine, 35(3):593, 1977. [28] Lenka Zdeborová and Florent Krzakala. Statistical physics of inference: Thresholds and algorithms. Advances in Physics, 65:453–552, 2016. 17
10math.ST
1 Rate region boundary of the SISO Z-interference channel with improper signaling arXiv:1605.04201v2 [cs.IT] 23 Jun 2017 Christian Lameiro∗ , Member, IEEE, Ignacio Santamarı́a†, Senior Member, IEEE, and Peter J. Schreier∗, Senior Member, IEEE Abstract—This paper provides a complete characterization of the boundary of an achievable rate region, called the Pareto boundary, of the single-antenna Z interference channel (Z-IC), when interference is treated as noise and users transmit complex Gaussian signals that are allowed to be improper. By considering the augmented complex formulation, we derive a necessary and sufficient condition for improper signaling to be optimal. This condition is stated as a threshold on the interference channel coefficient, which is a function of the interfered user rate and which allows insightful interpretations into the behavior of the achievable rates in terms of the circularity coefficient (i.e., degree of impropriety). Furthermore, the optimal circularity coefficient is provided in closed form. The simplicity of the obtained characterization permits interesting insights into when and how improper signaling outperforms proper signaling in the singleantenna Z-IC. We also provide an in-depth discussion on the optimal strategies and the properties of the Pareto boundary. Keywords—Improper signaling, Z interference channel, Pareto boundary. I. I NTRODUCTION It is widely known that proper Gaussian signals are capacityachieving in different wireless communication networks, such as the point-to-point, broadcast and multiple-access channels. Because of that, the use of such a signaling scheme is generally assumed in the study of multiuser wireless networks. The capacity-achieving property of proper signaling stems from the maximum entropy theorem, which states that the entropy of a random variable under a power constraint is maximized for a proper Gaussian distribution [1]. However, in networks where interference presents the major limiting factor, proper Gaussian signaling has recently been shown to be suboptimal, and improper Gaussian signaling, also known as asymmetric complex signaling, has been proved to outperform proper signaling in different interference networks [2]–[19]. An improper complex random variable differs from its proper counterpart in that its real and imaginary parts are correlated or have unequal variance, or, in other words, the random variable is correlated with its complex conjugate [20]. Such signals arise naturally in communications, e.g., due to gain imbalance between the in-phase and in-quadrature branches, or due to the use of specific digital modulations, ∗ C. Lameiro and P. J. Schreier are with the Signal & System Theory Group, Universität Paderborn, Germany (email: {christian.lameiro, peter.schreier}@sst.upb.de). † I. Santamarı́a is with the Department of Communications Engineering, University of Cantabria, Spain (e-mail: [email protected]). such as binary phase shift keying (BPSK) or Gaussian minimum shift keying (GMSK). Whenever the received signal is improper, linear operations must be replaced by widely linear operations, which are linear in both the random variable and its complex conjugate, in order to fully exploit the correlation between the signal and its complex conjugate [20]–[22]. The design of such widely linear receivers has been extensively studied in the literature (see, e.g., [23]–[26] and references therein). However, the transmission of improper signals to handle interference more effectively is a rather new line of research. We would like to add that digital modulation schemes yield cyclostationary signals, i.e., the mean, autocovariance and complementary autocovariance funtions are periodic. It has been shown that exploiting this property along with the impropriety leads to an improved performance [27]–[31]. For example, [27] shows that the optimal transmit signal must also be cyclostationary if the receiver is corrupted by a cyclostationary Gaussian noise. The first study on the benefits of improper signaling for interference management was carried out in the 3-user interference channel (IC) [2]. That work showed an improvement in terms of degrees-of-freedom (DoF), which represent the maximum number of interference-free streams and characterize the asymptotic sum-capacity. Similar DoF results were derived for the 4-user IC [3], the 3-user multiple-input multiple-output (MIMO) IC [4], the interfering broadcast channel [5], and the MIMO X-channel [6]. However, improper signaling not only increases the achievable DoF, but also the achievable rates in interference-limited networks. The optimal rate region boundary for maximally improper transmissions (i.e., perfect correlation between real and imaginary parts or either zero real or imagniary part) was derived for the 2-user IC in [7], showing substantial improvements over proper signaling. Additionally, [8] proposed a suboptimal design of the improper transmit parameters, which outperforms the proper and the maximally improper scheme. A similar suboptimal design was also proposed in [9] for the K-user multiple-input singleoutput (MISO) IC. Improper signaling in the IC has also been applied to reduce the symbol error rate [10] and as a mixed improper/proper approach in the MIMO-IC [11]. In addition to the IC, the use of improper Gaussian signaling has also been shown beneficial for other multiuser scenarios, such as the broadcast channel with linear precoding [12], relay-assisted communications [13], or underlay cognitive radio networks [14]–[16]. A particular case of the 2-user IC is the Z-IC, also known 2 as one-sided IC [32]. The difference with respect to the 2-user IC is the fact that only one of the receivers is affected by interference.1 The capacity region of the Z-IC is only known in the strong and very strong interference regimes [32], [34], and it is achievable by a non-linear operation at the receiver. Less complex non-linear techniques have also been studied (see, e.g., the characterization of the rate region boundary of the 2-user IC with successive interference cancellation at the receivers [35]). Even for such techniques, it is not known whether proper signaling is optimal. Nevertheless, it is more convenient to perform linear (or widely-linear) operations while treating the interference as noise, so that the complexity can be reduced. Restricted to linear operations, improper signaling presents a useful tool to improve the performance over proper signaling. Improper signaling for the Z-IC has recently been considered in [17], where the sum-rate maximizing scheme was derived in closed form. To that end, [17] considered the so-called real-composite model, where complex signals are regarded as real signals of double dimension. However, despite some remarkable efforts [36], the real-composite model is not as insightful as the augmented complex model, which works with the signal and its complex conjugate. For example, the circularity coefficient, which measures the degree of impropriety, is a quantity easily derived in the augmented complex formulation, but is much more difficult to express through its real-composite counterpart. The multi-antenna ZIC with improper signaling has been addressed in [18] and its journal version [19]. The inclusion of the spatial dimension makes a complete analytical assessment intractable, which is why the authors proposed an heuristic scheme to optimize the widely linear operation at the transmitter, which permits a trade-off between the rates of both users. This way, [19] obtained an achievable rate region that is larger than that obtained by proper signaling. Although the authors of [18], [19] mainly focused on the real-composite representation, they also considered the augmented complex formulation. Thus, the optimal transmission scheme for the interfered user is obtained using the former, whereas that of the interfering user through the latter. In our work, we adopt the augmented complex model to provide a complete and insightful characterization of the optimal rate region boundary, called the Pareto boundary, of the single-antenna Z-IC, when users may transmit improper Gaussian signals, assuming that interference is treated as noise. Our main contributions are summarized next. • We extend the results of [17], where only one point of the rate region boundary is derived, to provide a complete characterization of the Pareto optimal boundary in closed form. We show that the rate region boundary can be described by a threshold on the interference channel coefficient, which determines when improper signaling is optimal. • By adopting the augmented complex formulation, we provide, for each point of the boundary, closed-form 1 Notice that the Z-IC is different from the Z-channel, where the cross-link also conveys a desired message (see, e.g., [33] and references therein) expressions for the transmit powers and circularity coefficients, which are a direct measure of the degree of impropriety of the transmit signals. This permits insightful conclusions and a full assessment of the improvements of improper signaling over proper signaling in the single-antenna Z-IC. Thus, we analyze how the degree of impropriety affects the rate in the different boundary points, and we investigate the conditions that must be fulfilled for improper signaling to outperform proper signaling. The connection between the optimal circularity coefficients of both users and a further indepth discussion of our characterization is also provided. The rest of the paper is organized as follows. Section II provides some preliminaries of improper random variables and describes the system model. The characterization of the rate region boundary is derived in Section III, and a discussion on the results is presented in Section IV along with several numerical examples illustrating our findings. Finally, Section V concludes the paper. II. S YSTEM MODEL A. Preliminaries of improper complex random variables We first provide some definitions and results for improper random variables that will be used throughout the paper. For a comprehensive treatment of the subject, we refer the reader to [20]. The variance of a zero-mean complex random variable x is defined as σ 2 = E[|x|2 ], where | · | is the absolute value and E[·] is the expectation operator. The complementary variance of a zero-mean complex random variable x is defined as σ̃ 2 = E[x2 ]. If σ̃ 2 = 0, then x is called proper, otherwise improper. Furthermore, σ 2 and σ̃ 2 are a valid pair of variance and complementary variance if and only if σ 2 ≥ 0 and |σ̃ 2 | ≤ σ 2 . The circularity coefficient of a complex random variable x is defined as the absolute value of the quotient of its complementary variance and its variance, i.e., σ̃ 2 . (1) σ2 The circularity coefficient satisfies 0 ≤ κ ≤ 1 and thus measures the degree of impropriety of x. If κ = 0, then x is proper, otherwise improper. If κ = 1 we call x maximally improper. κ= B. System description We consider the single-input single-output (SISO) Z-IC with no symbols extensions. Without loss of generality and for the sake of exposition, we adhere√to the standard form, as depicted in Fig. 1. Denoting by a12 the real channel coefficient between transmitter 2 and receiver 1, the signal at both receivers can be modeled by √ y1 = s1 + a12 s2 + n1 , (2) y 2 = s2 + n 2 , (3) 3 1 s1 √ a 12 + n1 ∼ CN (0, 1) 1 s2 y1 + y2 n2 ∼ CN (0, 1) Fig. 1. SISO Z-IC in standard form. This model is described by three parameters: the power budgets P1 and P2 , and the interference channel coefficient a12 . where si and ni are the transmitted signal and noise of the ith user, respectively. The additive white Gaussian noise (AWGN) has variance 1 and is assumed to be proper, whereas the transmitted signals are complex Gaussian random variables with variance E[|si |2 ] = pi and complementary variance E[s2i ] = p̃i . Thus, the rate achieved by each user, as a function of the design parameters pi and p̃i , i = 1, 2, is given by [8]   p1 R1 (p1 , p̃1 , p2 , p̃2 ) = log2 1 + 1 + p2 a12 ! 2 1 − c−2 1 y1 |c̃y1 | , (4) + log2 2 2 1 − c−2 z1 |c̃z1 | ! 2 1 − c−2 1 y2 |c̃y2 | R2 (p2 , p̃2 ) = log2 (1 + p2 ) + log2 , (5) 2 2 1 − c−2 z2 |c̃z2 | where cy1 = p1 + p2 a12 + 1 , cy 2 = p 2 + 1 , (6) (7) are the variances of the received signals, c̃y1 = p̃1 + p̃2 a12 , c̃y2 = p̃2 , (8) (9) the complementary variances of the received signals, cz1 = p2 a12 + 1 , cz 2 = 1 , (10) (11) are the variances of the interference-plus-noise signals, z1 = √ a12 s2 + n1 , z2 = n2 , and c̃z1 = p̃2 a12 , c̃z2 = 0 , (12) (13) the complementary variances of z1 and z2 . Assuming that the power budget of the ith user is Pi , the achievable rate region with improper Gaussian signaling is then the union of all achievable rate tuples, i.e., [ (R1 (p1 , p̃1 , p2 , p̃2 ) , R2 (p2 , p̃2 )) . (14) R= 0≤pi ≤Pi , ∀i |p̃i |≤pi Notice that we have included the constraint |p̃i | ≤ pi in (14). As stated in Section II-A, this condition must be fulfilled for pi and p̃i to be a valid pair of variance and complementary variance. III. PARETO BOUNDARY OF THE RATE REGION The Pareto boundary of the rate region described by (14) comprises all Pareto optimal points, which are defined as follows [37]. Definition 1. We call the rate pair (R1 , R2 ) Pareto-optimal if (R1′ , R2 ) and (R1 , R2′ ), with R1′ > R1 and R2′ > R2 , are not achievable.2 In this section we characterize this boundary by deriving the optimal transmission parameters, pi and p̃i , i = 1, 2, that achieve each point of the boundary. First, we notice that, since user 1 does not interfere with user 2, its optimal transmit strategy maximizes its own achievable rate. Consequently, its transmit power must be maximized, which implies p1 = P1 . Second, pi and p̃i are a valid pair of variance and complementary variance if and only if pi ≥ 0 and |p̃i | ≤ pi . Consequently, the complementary variance can be expressed as p̃i = pi κi eφi , where κi is the circularity coefficient, which measures the degree of impropriety. Hence, |p̃i | ≤ pi is equivalent to 0 ≤ κi ≤ 1. With these considerations, R1 can then be expressed as " 2 1 (p2 a12 + P1 + 1) R1 = log2 2 2 2 (p2 a12 + 1) − |p2 eφ2 κ2 a12 | # 2 p2 eφ2 κ2 a12 + P1 eφ1 κ1 − . (15) 2 2 (p2 a12 + 1) − |p2 eφ2 κ2 a12 | Through (15) it is clear that R1 is maximized when 2 p2 eφ2 κ2 a12 + P1 eφ1 κ1 is minimized, which yields   p2 κ2 a12 κ1 = min ,1 , (16) P1 φ1 = φ2 + π . (17) From (16) we observe that, if κ2 = 0, i.e., user 2 transmits a proper signal, then user 1 must also transmit a proper signal by setting κ1 = 0. Similarly, if user 2 transmits an improper signal (κ2 > 0), then the signal transmitted by user 1 must also be improper. According to (17), the difference between the phases of the complementary variances of the desired and interference signals at receiver 1 is π. Such a phase difference can be interpreted by looking at the joint distribution of the real and imaginary parts of the desired signal and interference at receiver 1. The level contours of their distributions are ellipses whose major axes are rotated by π/2 with respect to each other [20], so that the signal and interference power are concentrated along orthogonal dimensions. Now we observe the following. With the optimal choice of φ1 , given by (17), the effect of φ2 is compensated at receiver 2 For the sake of brevity, we omit the dependence of R and R on the 1 2 design parameters when it is self-evident or not relevant. 4 1. Thus the achievable rate of user 1 is independent of the specific value of φ2 . Furthermore, since user 2 is not affected by interference, φ2 also has no impact on its achievable rate. Hence, without loss of generality, we can take φ2 = 0. With these considerations, the design parameters are reduced to the transmit power p2 and circularity coefficient κ2 of user 2. After some manipulations of (5) and (15), the achievable rates of user 1 and user 2, as a function of the design parameters, are given by     (p a +P +1)2  if κ < 1  12 log2 1+p a 2 p12 a 1(1−κ2 )+2 )  1 2 12 ( 2 12 2  R1 (p2 , κ2 )= ,  2P (p a (1+κ2 )+1)  if κ = 1  21 log2 1 + 1+p a1 2 p12a (1−κ 1 2 )+2 ) 2 12 ( 2 12 2 (18)   1 2 (19) R2 (p2 , κ2 ) = log2 1 + p2 p2 (1 − κ2 ) + 2 , 2 and the achievable rate region defined in (14) can then be expressed as [ R= (R1 (p2 , κ2 ) , R2 (p2 , κ2 )) . (20) 0≤p2 ≤P2 0≤κ2 ≤1 In order to characterize the boundary of the region defined in (20), we notice that the achievable rate of user 1 is bounded as 0 ≤ R1 (p2 , κ2 ) ≤ log2 (1 + P1 ) . (21) For each achievable rate of user 1, the corresponding Pareto optimal point is given by the one maximizing the rate of user 2, R2 (p2 , κ2 ), which can be cast as the following optimization problem P: maximize R2 (p2 , κ2 ) , subject to 0 ≤ p2 ≤ P2 , 0 ≤ κ2 ≤ 1 , R1 (p2 , κ2 ) ≥ α log2 (1 + P1 ) , p2 ,κ2 16 14 12 P2 10 8 q(κ2 ) 6 4 2 κmax 0 0 0.2 0.4 0.6 0.8 1 κ2 Fig. 2. Example of the transmit power and power constraints of user 2, for P1 = 20, P2 = 10, a12 = 0.5 and α = 0.7. The shaded area is the set of possible transmit powers, and κmax is the value of κ2 such that q(κmax ) = P2 . constraint p2 ≤ q (κ2 ), where q(κ2 ) is given by  q 2 2 2   (P1 −γ2R̄ )+ (γ2R̄ +1)[P1 (1−κ2 )+(γ2R̄ −2P1 )κ2 ] if κ < 1 1 2 −1 a (γ +1) 1−κ [ ( ) ] 12 2R̄ 2 . q(κ2 )=    2P1 1  − 1 if κ = 1 1 a12 (1−κ2 ) γ2R̄ (23) Proof: Please refer to Appendix A. As a result of Lemma 1, we can equivalently state problem P as (22) for a given α ∈ [0, 1]. Thus, we can compute every point of the rate region boundary by varying α between 0 and 1 and solving problem P. The set of constraints of problem P, which defines the feasibility set of our design parameters, consists of two constraints affecting the design parameters independently, namely, the power budget constraint and the bounds on the circularity coefficient, and an additional one that jointly constrains p2 and κ2 . The latter expresses a rate constraint on user 1, so that a specific point of the region boundary, determined by α, is computed. For a given κ2 , this constraint essentially limits the transmit power of user 2, p2 . Consequently, we can rewrite it in a more convenient form, as we express in the following lemma. Lemma 1. Let γx = 2x − 1 and R̄ = α log2 (1 + P1 ). The rate constraint R1 (p2 , κ2 ) ≥ R̄ is then equivalent to the power P: maximize R2 (p2 , κ2 ) , subject to 0 ≤ p2 ≤ min [q(κ2 ), P2 ] , 0 ≤ κ2 ≤ 1 . p2 ,κ2 For the sake of illustration, we plot in Fig. 2 an example of the two constraints affecting the transmit power of user 2, namely, q(κ2 ) and P2 . Obviously, q(κ2 ) is increasing in κ2 , since an interference with a higher degree of impropriety is less harmful, which is why user 1 tolerates a higher amount of interference power without reducing its achievable rate. To achieve the global maximum of problem P, it is clear that we must set p2 (κ2 ) = min[q(κ2 ), P2 ], where we have explicitly expressed its dependence on κ2 . Consequently, the number of design parameters is reduced to one, κ2 , and P is further simplified to P: maximize R2 (κ2 ) , subject to 0 ≤ κ2 ≤ 1 , κ2 5 where R2 (κ2 ) can now be expressed as     1 R2 (κ2 ) = log2 1 + p2 (κ2 ) p2 (κ2 ) 1 − κ22 + 2 2 Lemma 2. Let p2 (κ2 ) = q(κ2 ) and assume that there exists 2 (κ2 ) 2 (κ2 ) κ̂2 such that ∂R∂κ ≥ 0. Then ∂R∂κ > 0 for all 2 2 2 2 κ2 =κ̂2 κ2 > κ̂2 . Proof: Please refer to Appendix B. Lemma 2 leads to the following key result. Lemma 3. Let p2 (κ2 ) = q(κ2 ) and P1 , µ(α) = 1 − γ2R̄ − γR̄  (P1 −γR̄ )2  √ 2 ι(α) = (γ2R̄ −γR̄ ) γ2R̄ −2P1 −γR̄  0 (25) if 2P1 < γ2R̄ . (26) otherwise Then, the dependency of R2 (κ2 ) on κ2 can be described as follows. • If a12 ≥ µ(α), then R2 (κ2 ) increases monotonically in κ2 . • If µ(α) > a12 > ι(α), then there exists κ̃2 > 0 such that R2 (κ2 ) ≤ R2 (0) for κ2 ≤ κ̃2 and R2 (κ2 ) > R2 (0) otherwise. Furthermore, R2 (κ2 ) increases monotonically for κ2 ≥ κ̃2 . • If ι(α) ≥ a12 , then R2 (κ2 ) ≤ R2 (0) for all values of κ2 . Proof: Please refer to Appendix C. To illustrate the characterization provided in Lemma 3, we plot in Fig. 3 the rate of user 2 normalized by the proper signaling rate with examples for each of the three cases described in Lemma 3. We use α = 0.7 and P1 = 10, which gives µ(0.7) = 0.571 and ι(0.7) = 0.545. Notice that the three curves vary only slightly with κ2 because the three chosen values of a12 (0.5, 0.55 and 0.6) are all close to the thresholds. Lemma 3 describes how the interference constraint q(κ2 ) shapes the dependency of the rate of user 2 on the degree of impropriety of its transmitted signal. Thus, if we 1.02 R2 (κ2 )/R2 (0) . (24) Notice that, by expressing p2 as a function of κ2 , R2 (κ2 ) is now also a function of κ2 only. That is, the key task now is to determine the optimal circularity coefficient of the second user, or, in other words, the degree of impropriety of its transmit signal such that its achievable rate, given by (24), is maximized. In the forthcoming lines we will analyze when R2 (κ2 ) is maximized by an improper signal, i.e., κ2 > 0, and then we will derive the optimal value of κ2 in those cases. That is, we want to determine the conditions that must be fulfilled for improper signaling to outperform conventional proper signaling. Since there are two different power constraints affecting p2 , namely, the power budget of user 2 and the interference power created at user 1, we will start by dropping the power budget constraint to analyze how the interference constraint, q(κ2 ), affects the rate of user 2 as a function of κ2 . We first present the following lemma. 1.01 a12 = 0.6 1 a12 = 0.55 0.99 a12 = 0.5 κ̃2 0.98 0 0.2 0.4 0.6 0.8 1 κ2 Fig. 3. Illustration of Lemma 3. We set α = 0.7 and P1 = 10, which yields µ(0.7) = 0.571 and ι(0.7) = 0.545. drop the power budget constraint or, alternatively, if the power budget is sufficiently high, improper signaling outperforms proper signaling for all values of κ2 if a12 ≥ µ(α), but only for some values of κ2 if µ(α) > a12 > ι(α), and it is always suboptimal if a12 ≤ ι(α). It can be clearly seen now what the impact of the power budget constraint is. For the cases a12 ≥ µ(α) and µ(α) > a12 > ι(α), κ2 = 1 maximizes R2 (κ2 ). However, increasing κ2 is only meaningful if it also permits increasing the transmit power. When the power budget constraint is considered, there may exist κmax such that q(κmax ) = P2 , hence increasing κ2 beyond that point does not permit a further increase in the transmit power (see Fig. 2). Therefore, when µ(α) > a12 > ι(α) improper signaling will be optimal if κmax > κ̃2 , where κ̃2 > 0 is such that R2 (κ̃2 ) = R2 (0) (see Fig. 3). On the other hand, improper signaling will be optimal for a12 ≥ µ(α) only if q(0) < P2 . This condition means that proper signaling does not permit transmitting at maximum power, thus there is still power left over, which can be exploited by improper signaling to improve the achievable rate. With all these ingredients, we can derive a complete characterization of the optimality of improper signaling for this scenario and, consequently, of the Pareto-optimal region. Our main result is presented in the following theorem. Theorem 1. Let us define ι(α) as in (26) and     1 P1 − 1 , ν(α) , ρ(α) = max P2 γR̄ (27) where ν(α) = ( η(α) γR̄ (2γ2R̄ +1)−P1 (2γR̄ +1) γR̄ [P2 +2(γ2R̄ +1)] 1 if η(α) ≥ P P2 + aI otherwise , (28) 6 and Then, improper signaling is required to maximize R2 (κ2 ) if and only if a12 > max [ι(α), ρ(α)] . (32) Furthermore, if this expression holds, the optimal circularity coefficient is  1 if q(1) ≤ P2 κ2 = (33) κmax otherwise , where κmax is the minimum value of κ2 such that P2 ≤ q(κmax ). Proof: Please refer to Appendix D. Before concluding this section, we introduce the following definition, which will be useful to describe the properties of the Pareto boundary. Definition 2. We call power-limited region all the points of the rate region boundary for which ι(α) = 0 or, alternatively, 2P1 ≥ γ2R̄ . We call the remaining points of the rate region the interference-limited region, i.e., those for which ι(α) > 0 or, alternatively, 2P1 < γ2R̄ . IV. D ISCUSSION AND NUMERICAL EXAMPLES This section provides a discussion on the derived characterization along with some numerical examples illustrating the most remarkable features of improper signaling in the Z-IC. Afterwards, the connection to related works in the literature is presented. A. Optimal strategies 1) Optimality of proper signaling: As pointed out at the beginning of Section III, if proper signaling is the optimal strategy for one of the users, then it is also optimal for the other one. This means that any point of the region boundary is achieved by both users employing the same signaling scheme, i.e., either proper or improper, but not a combination of both. 2) Maximally improper signaling for both users is optimal at most at one boundary point: It can be noticed that there is at most one boundary point where both users simultaneously transmit a maximally improper signal, i.e., κ1 = κ2 = 1 happens for no more than one Pareto-optimal point. This is due to the fact that, if the rate constraint (22) can be fulfilled for κ1 = 1 (which corresponds to the case 2P1 ≥ γ2R̄ ), then user 1 tolerates an infinite amount of maximally improper interference along the orthogonal direction (see second equation in (23)). However, user 2 may only increase its rate by increasing κ2 if it is operating below its power budget, since maximum sum-rate 1 8 0.8 6 p2 1 P2 1 aI = P2 aP = 10 circularity coefficients κ1 , κ2 η(α) = p (aP − P2 aI )2 + 2P2 (a2I + a2P ) , 2 (29)   P1 −1 , (30) γ  R̄  2P1 −1 . (31) γ2R̄ aP − P2 aI + p2 0.6 4 2 0.4 0.2 0 0 κ1 κ2 0 0.5 1 1.5 2 R1 [b/s/Hz] 2.5 3 3.5 Fig. 4. Dependency of the optimal circularity coefficients and the transmit power p2 on R1 for a12 = 2. the only purpose of increasing κ2 is to increase p2 as well. Because of this and because q(κ2 ) is a continuous function for 2P1 > γ2R̄ , setting κ1 = κ2 = 1 is always suboptimal when 2P1 > γ2R̄ . However, this reasoning is not applicable when 2P1 = γ2R̄ , since q(κ2 ) turns into a discontinuous function for κ1 = 1. This is because, in that case, q(κ2 < 1) = 0 and q(κ2 = 1) = ∞. The intuition behind this behavior is that, when 2P1 = γ2R̄ , the first user achieves its corresponding rate with a maximally improper signal, i.e., κ1 = 1, only if the interference is orthogonal to the signal subspace, i.e., only when the second user also transmits a maximally improper signal. Consequently, κ2 = 1 must hold if 2P1 = γ2R̄ and κ1 = 1. Furthermore, since p2 = P2 when κ2 = 1, the condition κ1 = 1 is equivalent, by (16), to P2 a12 ≥ P1 . That is, both users transmit maximally improper signals at the boundary point for which R1 = 12 log2 (1 + 2P1 ) only if P2 a12 ≥ P1 . To illustrate this property, we provide two simulation examples. As a first example we consider the channel coefficient a12 = 2 and the power budgets P1 = P2 = 10. Figure 4 shows the optimal circularity coefficients and the transmit power of user 2. In this example P2 a12 ≥ P1 holds and because of that both users transmit maximally improper signals for R1 = 2.2 b/s/Hz. We can observe a discontinuity in the maximum transmit power of user 2, which will be explained later. Notice that, as R1 approaches its maximum value of 3.46 b/s/Hz, κ1 goes towards 0 while κ2 remains static at 1. Although this might seem to violate our statement in Section IV-A1, κ1 is only equal to zero when p2 is, which only happens at R1 = 3.46 b/s/Hz. In this case, the circularity coefficient is no longer meaningful since the transmit power is zero. Therefore, all the boundary points satisfy our claim, which is also clear according to the optimal value of κ1 in (16). 7 10 1 8 0.8 6 p2 circularity coefficients κ1 , κ2 maximum sum-rate p2 0.6 4 2 0.4 0.2 0 0 κ1 κ2 0 0.5 1 1.5 2 R1 [b/s/Hz] 2.5 3 3.5 Fig. 5. Dependency of the optimal circularity coefficients and the transmit power p2 on R1 for a12 = 0.8. As a second example we consider a12 = 0.8. In this case P2 a12 < P1 , hence it is expected that there are no boundary points that are achieved by both users transmitting maximally improper signals. This can be observed in Fig. 5, which depicts the dependency of the circularity coefficients and transmit power on the rate, R1 . At the boundary point for R1 = 2.2 b/s/Hz, user 2 chooses the transmit signal as maximally improper, but the optimal circularity coefficient of user 1 equals 0.8. In Fig. 5 we can also observe a discontinuity in the maximum transmit power and in the circularity coefficients at approximately R1 = 2.9 b/s/Hz. This discontinuity is different from the one observed in Fig. 4 and can be explained as follows. At the point of the discontinuity we have a12 = max[ι(α), ρ(α)]. Furthermore, at this point q(1) < P2 . This means that maximally improper signaling achieves the same rate as proper signaling. Proper signaling starts outperforming improper signaling once R1 increases beyond that point, in which case the optimal circularity coefficient jumps from κ2 = 1 to κ2 = 0, and the maximum transmit power jumps then from q(1) to q(0). 3) Relationship between κ1 and κ2 : Expression (16) also permits drawing insightful conclusions about the relationship between the circularity coefficients of both users. According to Theorem 1, 0 < κ2 < 1 implies p2 = P2 . Hence, κ1 < κ2 holds if κ2 > 0 and P2 a12 < P1 , i.e., when the signal-tointerference ratio (SIR) is greater than one. In such a case, it can be noticed that the signal transmitted by the first user is never chosen as maximally improper at any point of the Pareto boundary. This behavior can be clearly observed in Fig. 5. On the other hand, when P2 a12 ≥ P1 , or, alternatively, when the SIR is equal to or lower than one, κ1 ≥ κ2 holds whenever 0 < κ2 < 1, as can be seen in Fig. 4. If the signal transmitted by user 1 is chosen as maximally improper for some points of the boundary, it will remain so as κ2 → 1, or, in other words, as γ2R̄ → 2P1 . This is because κ2 < 1 implies p2 = P2 , so that, by (16), the first user will not decrease its circularity coefficient. However, once κ2 equals 1, which corresponds to the point 2P1 = γ2R̄ , the degree of impropriety of the first user will then start decreasing with R1 , since these rates are not achievable for κ1 = 1. B. Properties of the Pareto boundary 1) Behavior in the power-limited and interference-limited regions: The power-limited region, which refers to the power limitation of the second user, comprises all boundary points satisfying 2P1 ≥ γ2R̄ , or, alternatively, R1 ≤ 12 log2 (1 + 2P1). In this region improper signaling is always optimal as long as the power budget P2 is sufficiently high. In other words, the optimality of improper signaling is only determined by ρ(α), which depends on P2 and goes towards 0 as P2 increases. Furthermore, user 1 can achieve its required rate with a maximally improper signal, i.e., with κ1 = 1. In this case, the interference along the unused dimension does not have any impact on its achievable rate, therefore R2 → ∞ as P2 → ∞. In the interference-limited region, user 1 must choose a circularity coefficient smaller than one to achieve the desired rate, and hence the tolerated interference is finite for all values of κ2 . As a result, the transmit power of the second user is eventually limited by q(κ2 ) as P2 grows, which bounds the achievable rate. In other words, the optimality of improper signaling is eventually determined by ι(α), which is independent of the power budget P2 . 2) Transition between power-limited and interferencelimited regions: An interesting feature is that there may be abrupt changes in the achievable rate of user 2 when we shift from one region to the other, which are due to a jump in the tolerable interference power (as observed in Fig. 4). This can be explained as follows. In the power-limited region the transmit power always equals the power budget when improper signaling is optimal, i.e., p2 (κ2 ) = P2 if κ2 > 0. However, in the interference-limited region the transmit power is dominated by the function q(κ2 ) when the power budget exceeds the value of that function, i.e., when P2 > q(1). Hence, by using (16), it can be easily seen that limγ2R̄ →2P + q2 (1) = aP121 , so that there 1 may be a jump in the maximum transmit power, i.e., a discontinuity in the maximum transmit power of user 2 as a function of R1 , from p2 = P2 to p2 = aP121 whenever P2 > aP121 . That is, the lower the SIR, the more prominent the power jump is, whereas no jump will be observed when the SIR is equal to or greater than 1. This discontinuity in the maximum transmit power implies a similar jump in the maximum achievable rate of user 2, making limγ2R̄ →2P + R2 (κ2 ) 6= limγ2R̄ →2P − R2 (κ2 ). This 1 1 is illustrated in Figs. 6 and 7, where the Pareto boundary of the rate region is depicted for the previously considered examples, i.e., a12 = 2 and a12 = 0.8, respectively. We also depict in the figures the rate region boundary for proper signaling, and the enlargement of the rate region due to time sharing. Figure 6 corresponds to a scenario where the SIR is below one, therefore we can observe that the Pareto boundary is discontinuous. As explained earlier, this is the result of a jump 8 3.5 3.5 power-limited interf.-limited 3 3 2.5 maximum sum-rate R2 [b/s/Hz] R2 [b/s/Hz] 2.5 2 1.5 Proper signaling Proper signaling (time sharing) Improper signaling Improper signaling (time sharing) 1 0.5 0 0 0.5 1 1.5 2 R1 [b/s/Hz] 2 power-limited 1.5 Proper signaling Proper signaling (time sharing) Improper signaling Improper signaling (time sharing) 1 0.5 2.5 3 interf.-limited maximum sum-rate 3.5 0 0 0.5 1 1.5 2 R1 [b/s/Hz] 2.5 3 3.5 Fig. 6. Pareto boundary for proper and improper transmissions, with P1 = P2 = 10 and a12 = 2. Fig. 7. Pareto boundary for proper and improper transmissions with P1 = P2 = 10 and a12 = 0.8. in the maximum transmit power of user 2, p2 , which can also be observed in Fig. 4. Specifically, the maximum transmit power in the interference limited region equals p2 = aP121 = 5, so there is a power jump from p2 = P2 = 10 to p2 = 5. Notice that at the discontinuity, which approximately corresponds to R1 = 2.2 b/s/Hz, all the rates in the half-open interval R2 ∈ [1.7, 2.3) b/s/Hz are achievable but do not belong to the Pareto boundary as defined at the beginning of Section III. Similarly, all the rate pairs R2 = 3.4 b/s/Hz and R1 ∈ (0, 0.6) b/s/Hz are achievable but do not belong to the Pareto boundary. Therefore, these two line segments are plotted in dashed lines in Fig. 6. The scenario corresponding to Fig. 7 presents an SIR greater than one. Because of that, the transition from the power-limited to the interference-limited regions presents no discontinuity as there is no jump in the transmit power at 2P1 = γ2R̄ (see Fig. 5). Notice that the discontinuity in the maximum transmit power at R1 = 2.9 b/s/Hz, which is due to a12 falling below the threshold, does not cause a discontinuity in the Pareto boundary. As previously explained, this is because at that point a12 = max[ι(α), ρ(α)] holds, so proper and improper signaling achieve the same rate. Therefore, R2 changes smoothly even though the transmit power and circularity coefficients jump. Figures 6 and 7 also illustrate that the dependency of R2 on R1 in the power-limited region is different from the interference-limited region. Specifically, R2 decreases more slowly with R1 in the former. This is due to the fact that in these examples p2 = P2 holds in the entire power-limited region, and therefore the decrease in R2 is due to an increase in κ2 . In the interference-limited region, however, p2 decreases towards 0 as R1 increases, which has a stronger effect on the achievable rate. The transition between these two different behaviors leads to a sharp bend in the Pareto boundary, as observed in Figs. 6 and 7. It corresponds to the transition between the power-limited and interference-limited regions only if P2 > q(1) holds in the entire interference-limited region, and improper signaling is optimal in the power-limited region. This is because these conditions imply that κ2 = 1 is optimal at the transition point 2P1 = γ2R̄ . Therefore, to satisfy the rate of the first user p2 must decrease once we move into the interference-limited region, which changes the behavior of R2 . If P2 < q(1) for some points in the interference-limited region, then κ2 < 1 is optimal at 2P1 = γ2R̄ . Increasing κ2 while keeping p2 constant permits achieving the required R1 when we move into the interference-limited region until κ2 = 1 is reached. Hence the dependency of R2 on R1 will change at the value of R1 such that P2 = q(1), which is slightly shifted to the right. Figures 6 and 7 also show the enlargement of the rate region due to improper signaling. In these examples, the interference level is significant, so the achievable rate region by improper signaling is substantially larger than that of proper signaling, especially in the scenario shown in Fig. 6. It is also worth highlighting that time sharing provides a substantial enlargement of the rate region for both proper and improper signaling, but there is still a significant gap between the performance of both strategies. In Fig. 6, the rate region achieved by improper signaling is convexified by time sharing between the extreme points of the Pareto boundary, which correspond to proper signaling transmissions, and the point at the transition between the power-limited and the interferencelimited regions, where both users transmit maximally improper signals. By contrast, all the Pareto boundary points in the power-limited region belong to the convex hull in Fig. 7, and time sharing improves the performance only in the 9 interference-limited region. 3) Operation regimes: Considering the possibly suboptimal approach of always treating interference as noise, we may distinguish the following operation regimes that account for the optimality of improper signaling. 1 • Strictly improper regime: If a12 ≥ PP , we say the 1 +1 Z-IC is in the strictly improper regime. In this regime and treating interference as noise, all the Pareto-optimal points satisfying R1 > log2 (1 + 1+PP21a12 ) are achieved by improper signaling, i.e., κ2 > 0. 1 > a12 > • Selective improper/proper regime: If PP 1 +1 minα max[ι(α), ρ(α)], we say the Z-IC is in the selective improper/proper regime. In this regime and treating interference as noise, only a subset of the Pareto-optimal points satisfying R1 > log2 (1 + 1+PP21a12 ) is achieved by improper signaling, i.e., κ2 > 0. • Strictly proper regime: If a12 ≤ minα max[ι(α), ρ(α)], we say the Z-IC is in the strictly proper regime. In this regime and treating interference as noise, all the Pareto-optimal points are achieved by proper signaling, i.e., κ2 = 0. Note that (R1 = log2 (1 + 1+PP21a12 ), R2 = log2 (1 + P2 )) always belongs to the Pareto boundary and can only be achieved by proper signaling. Because of that, the operation regimes described above explain the optimal strategies for the boundary points satisfying R1 > log2 (1 + 1+PP21a12 ). The strictly improper regime is obtained making use of Lemma 3, 1 which establishes that a12 ≥ µ(α) = 1− γ P−γ is a sufficient R̄ 2R̄ condition for improper signaling to be optimal if proper signaling does not allow maximum power transmission, i.e., if R1 > log2 (1 + 1+PP21a12 ). Therefore, µ(α) ≥ max[ι(α), ρ(α)] holds for all boundary points satisfying R1 > log2 (1 + 1+PP21a12 ). Furthermore, it can be easily checked that µ(1) = ι(1) = P1 P1 ⇒ max [ι(1), ρ(1)] = . (34) P1 + 1 P1 + 1 Since µ(α) is an increasing function, the condition for the optimality of improper signaling is fulfilled in the strictly improper regime for all Pareto-optimal points except for R1 = log2 (1 + 1+PP21a12 ). An example of the operation in the strictly improper regime is given in Fig. 6 (and the corresponding Fig. 4). Figures 5 and 7 correspond to the selective improper/proper regime, and we can observe that improper signaling is optimal only in the interval R1 ∈ (1.1, 2.9) b/s/Hz. When we move from the power-limited region to the interference-limited region we have   1 γR̄ 1 1 lim+ ι(α) = 1− √ = 4 γR̄ + 1 4 1 + 2P1 α→α0   1 1 1− √ , (35) ⇒ lim+ max [ι(α), ρ(α)] ≥ 4 1 + 2P1 α→α0 where α0 is such that 2P1 = γ2R̄ . Therefore, proper signaling is the optimal strategy in the whole interference-limited region 1 if a12 ≤ 14 (1 − √1+2P ). In such a case, benefits of improper 1 signaling, if any, are limited to the power-limited region. Notice that, as P1 increases, this condition converges to a12 ≤ 41 . C. Relationship to previous work Finally, we would like to connect our results to related works in the literature. In our previous work [14], we studied a similar scenario in the context of underlay cognitive radio. We considered the Z-IC with the restriction that the first user transmits only proper signals, i.e., κ1 = 0. A characterization of the maximum achievable rate of user 2 was derived in terms of a threshold in a12 . Specifically, improper signaling was shown to be optimal if a12 > 1 − γP1 . This threshold is 2R̄ strictly higher than the one obtained for the general Z-IC (it can easily be seen that 1 − γP1 ≥ µ(α) ≥ ι(α)). This is in 2R̄ agreement with the fact that, if we let the first user optimize its circularity coefficient, the rate achieved by the second user can only increase. The Z-IC was also considered in [17], and the transmit strategy that maximizes the sum-rate was derived in closed form based on the real-composite model. Although such a model is usually more convenient from an optimization pointof-view, it is not as insightful as the augmented complex model since some of the features of the improper signal are not easily captured. This is the case for the degree of impropriety, which is elegantly given by the circularity coefficient. We would like to stress that in [17] only one point of the rate region was characterized, whereas in this work we completely characterized the boundary of the rate region. Furthermore, since we consider the augmented complex model, we provided closedform formulas for the circularity coefficients, thus providing a more insightful description of how improper signaling behaves in this scenario. Nevertheless, some of the conclusions drawn in [17] fall within our characterization of the rate region boundary. By looking at the structure of the sum-rate maximizing transmit strategies in [17, Eq. (31)], we observe that improper signaling is chosen when a12 > 1. This condition belongs to the strictly improper regime, and the solution presented in [17, Eq. (31)] can be seen as a special case of (33). Furthermore, according to [17], when improper signaling is preferred for the sum-rate maximization, the signal transmitted by one of the users is chosen as maximally improper, and the selection of the user depends on whether or not the condition P2 a12 ≤ P1 holds. This is in agreement with our previous discussion, where we pointed out that the circularity coefficient of the first user is always equal to or smaller than that of the second user whenever the aforementioned condition is fulfilled. In such a case, the Pareto optimal point that corresponds to the sumrate maximization is then the point that satisfies q(1) = P2 . Otherwise, if P2 a12 > P1 , the sum-rate is maximized when the first user transmits a maximally improper signal, and in this case the circularity coefficient of the second user is smaller than 1 except for R1 = 12 log2 (1 + 2P1 ). The MIMO Z-IC with improper signaling has been considered in [18], [19]. Due to the inclusion of the spatial dimension, an analytical characterization of improper signaling similar to that of Theorem 1 becomes intractable. Notice that in the SISO Z-IC, the impropriety is completely described by the circularity coefficient. However, in the MIMO case, the description of the impropriety is by means of the complementary covariance matrix, which is characterized by a set of circularity coeffi- 10 cients and a unitary matrix [20]. Because of that, the authors proposed a heuristic scheme to design the complementary covariance matrix of the transmitted signal of user 2 such that it permits easy control of the degree of impropriety. An interesting aspect of [18], [19] is the fact that it does not completely stick to a particular model for the representation of improper signals. Thus, they use the real-composite model to optimize the transmission scheme of user 1, while they design the scheme of the second user by means of the augmentedcomplex formulation. Interestingly, the achievable rate regions obtained in [18], [19] have a shape similar to those in Figs. 6 and 7. This suggests that some of our conclusions for the single-antenna Z-IC might be extended to the more general MIMO Z-IC. V. C ONCLUSION We have analyzed the benefits of improper signaling in the single-antenna Z-IC. Under the assumption that interference is treated as Gaussian noise, we have derived a complete and insightful characterization of the Pareto boundary of the rate region, and the corresponding transmit powers and circularity coefficients in closed-form. This characterization has been derived by analyzing how the circularity coefficients affect the performance at the different boundary points. Specifically, we have shown that improper signaling is optimal when the interference coefficient exceeds a given threshold that depends on the rate achieved by the interfered user. We have shown that the rate region can be substantially enlarged by using improper signaling, especially when the relative level of interference is high. ACKNOWLEDGEMENTS The work of C. Lameiro and P. J. Schreier was supported by the German Research Foundation (DFG) under grant SCHR 1384/6-1. The work of I. Santamaria was supported by the Ministerio de Economia y Competitividad (MINECO), Spain, under projects RACHEL (TEC2013-47141-C4-3-R) and CARMEN (TEC2016-75067-C4-4-R). The authors would like to thank the anonymous reviewers for their valuable comments, which have helped improve the quality of the paper. A PPENDIX A PROOF OF L EMMA 1 Let us first consider κ1 = 1. In this case R1 (p2 , κ2 ) ≥ R̄ is equivalent, because of (18), to the quadratic inequality  p22 a212 γ2R̄ 1 − κ22 + p2 2a12 [γ2R̄ − P1 (1 + κ2 )] + γ2R̄ − 2P1 ≤ 0 , (36) which is convex and only has one positive root given by P1 (1 + κ2 ) − γ2R̄ + |P1 (1 + κ2 ) − γ2R̄ κ2 | p2 = . a12 γ2R̄ (1 − κ22 ) the real or imaginary part. This means that 2P1 ≥ γ2R̄ holds whenever κ1 = 1, which implies P1 (1 + κ2 ) − γ2R̄ κ2 ≥ 0. Hence, (37) can be simplified to p2 = q(κ2 ), where q(κ2 ) is given by the second branch of (23). If κ1 < 1, R1 (p2 , κ2 ) ≥ R̄ is equivalent to the quadratic inequality    p22 a212 (γ2R̄ + 1) 1 − κ22 − 1 − p2 2a12 (P1 − γ2R̄ ) 2 + γ2R̄ + 1 − (1 + P1 ) ≤ 0 . (38) The equivalent power constraint will then be given by one of the roots of this equation. In order to determine which one of the two roots must be considered we can use the fact that R1 (p2 , κ2 ) is decreasing in p2 . This means that at least one of the roots will violate the condition 0 ≤ κ1 < 1, with κ1 given by (16). The roots of (38) are " 1 p2 = (P1 − γ2R̄ ) ± a12 [(γ2R̄ + 1) (1 − κ22 ) − 1] r h # i 2 2 (P1 − γ2R̄ ) + [(γ2R̄ + 1) (1 − κ22 ) − 1] (1 + P1 ) − γ2R̄ − 1 . (39) Notice that (1 + P1 )2 − γ2R̄ − 1 = (1 + P1 )2 − (1 + P1 )2α ≥ 0. Therefore, if (38) is convex, i.e., if [(γ2R̄ + 1)(1 − κ22 ) − 1 ≥ 0, (38) has one negative and one positive root, and the equivalent power constraint will be given by the latter if it satisfies κ1 < 1. Otherwise, the rate expression for κ1 = 1 must be considered. If (38) is concave, both roots can be either positive, negative, or complex. For the last two cases, none of them satisfies 0 ≤ κ1 < 1, so the equivalent power constraint must be obtained through the expression for κ1 = 1. If the two roots are positive, the monotonicity of R1 (p2 , κ2 ) in p2 is not fulfilled for the largest root, hence the equivalent power constraint is determined by the smallest root. For all these cases, the root that must be considered is obtained by taking the positive square root in (39) which, after some manipulations, yields p2 = q(κ2 ), with q(κ2 ) being given by the first branch of (23). This concludes the proof. A PPENDIX B P ROOF OF L EMMA 2 Through (24), the derivative of R2 (κ2 ) with respect to κ22 is non-negative if  ∂R2 (κ2 ) ∂q(κ2 )  ≥0 ⇔ 2 q(κ2 )(1 − κ22 ) + 1 ≥ q(κ2 )2 . 2 2 ∂κ2 ∂κ2 (40) Let us first consider κ1 = 1. In this case we have q(κ2 ) ∂q(κ2 ) = . ∂κ22 2κ2 (1 − κ2 ) (41) Plugging (41) into (40) we obtain (37) Notice that the rate constraint of user 1 can only be fulfilled with a maximally improper signal if R̄ ≤ 21 log2 (1+2P1 ), that is, if user 1 can achieve its rate by transmitting all power over ∂R2 (κ2 ) ≥ 0 ⇔ q(κ2 )(1 − κ2 ) + 1 ≥ 0 , ∂κ22 (42) which holds for all values of κ2 . If κ1 < 1, (38) holds with equality for p2 = q(κ2 ). Therefore, evaluating (38) at p2 = 11 q(κ2 ) and taking the derivative with respect to κ22 yields Using this expression and (24), R2 (1) > R2 (0) holds if 2 q(κ2 ) a12 (γ2R̄ + 1) ∂q(κ2 ) . = ∂κ22 2q(κ2 )a12 [(γ2R̄ + 1) (1 − κ22 ) − 1] − 2 (P1 − γ2R̄ ) (43) Plugging (43) into (40) we obtain ∂R2 (κ2 ) ≥0 ⇔ ∂κ22   a12 (γ2R̄ + 1) q(κ2 )(1 − κ22 ) + 1 ≥ 1 . (44) q(κ2 )a12 [(γ2R̄ + 1) (1 − κ22 ) − 1] − (P1 − γ2R̄ ) The denominator in this expression can be shown to be positive as follows. Since we are assuming that κ1 < 1 holds, we have, 2 through (23), p that q(κ2 )a12 [(γ2R̄ + 1)(1 − κ2 ) − 1] − (P1 − γ2R̄ ) = (γ2R̄ + 1)[P12 (1 − κ22 ) + (γ2R̄ − 2P1 )κ22 ], hence it is positive. Consequently, we obtain ∂R2 (κ2 ) γ − P1 − q̄(κ2 ) ≥ 0 ⇔ a12 ≥ 2R̄ , 2 ∂κ2 γ2R̄ + 1 (45) where q̄(κ2 ) = q(κ2 )a12 , which does not depend on a12 as can be seen from (23). Let κ̂2 be such that the right-hand side 2) of (45) holds with equality. Since ∂q(κ > 0 for q(κ2 ) > 0, ∂κ22 q̄(κ2 ) increases with κ2 and, consequently, the right-hand side of (45) holds with strict inequality when κ2 > κ̂2 . As a result, the derivative of R2 (κ2 ) is positive whenever κ2 > κ̂2 , which concludes the proof. A PPENDIX C P ROOF OF L EMMA 3 By Lemma 2, R2 (κ2 ) increases monotonically in κ2 as long 2 (κ2 ) as ∂R∂κ ≥ 0 for κ2 = 0. Clearly, κ1 is also zero at this 2 2 point, thus we can apply (45), yielding   P1 γ − P − − 1 1 2 R̄ γR̄ γ − P1 − q̄(0) = a12 ≥ 2R̄ γ2R̄ + 1 γ2R̄ + 1 P1 =1− = µ1 (α) , (46) γ2R̄ − γR̄ which proves the first case. In the second case, improper signaling outperforms proper signaling only for κ2 > κ̃2 . By Lemma 2 we also know that, if improper signaling outperforms proper signaling for κ2 = κ′2 , then the rate improvement will be strictly positive for κ2 > κ′2 . Therefore, when R2 (1) > R2 (0) holds, there must be κ̃2 such that R2 (κ̃2 ) = R2 (0) and R2 (κ2 ) > R2 (0) for all κ2 > κ̃2 . To evaluate when R2 (1) > R2 (0) holds, we first consider the boundary points satisfying 2P1 ≥ γ2R̄ . Since at these points user 1 can achieve its required rate with a maximally improper signal, i.e., with κ1 = 1, we have that q(1) = ∞. Consequently, R2 (1) > R2 (0) is always fulfilled for a12 > 0, which yields the second branch of (26). For the boundary points satisfying 2P1 < γ2R̄ , κ1 < 1 holds. Therefore, using the corresponding branch in (23) for κ2 = 1 yields i p 1 h q(1) = γ2R̄ − P1 − (γ2R̄ + 1) (γ2R̄ − 2P1 ) . (47) a12 i p 2 h γ2R̄ − P1 − (γ2R̄ + 1) (γ2R̄ − 2P1 ) a12      P1 P1 1 1 −1 −1 +2 , > a12 γR̄ a12 γR̄ (48) which yields  P1 γR̄ −1 2 i. +1 (49) 2 Taking into account that γ2R̄ + 1 = (γR̄ + 1) , we have a12 > h p 2 γ2R̄ − P1 − (γ2R̄ + 1) (γ2R̄ − 2P1 ) − P1 γR̄ 2 a12 > (P1 − γR̄ ) p  γR̄ (γR̄ + 1) (γ2R̄ − 2P1 ) + γ2R̄ − 2γR̄ − 2γR̄ γ2R̄ − 2P1 =  2 (P1 − γR̄ ) p 2 = ι(α) , (γ2R̄ − γR̄ ) γ2R̄ − 2P1 − γR̄ (50) which yields the first branch of (26). Finally, if a12 ≤ ι(α) then we have R2 (1) ≤ R2 (0) and, by Lemma 2, R2 (κ2 ) ≤ R2 (0) for all κ2 . This yields the third case and concludes the proof. A PPENDIX D P ROOF OF T HEOREM 1 By Lemma 3, improper signaling can only be optimal if a12 > ι(α). However, this condition is not sufficient, since it does not take the power budget constraint into account. Firstly, since increasing κ2 can improve the rate only if this permits increasing the transmit power as well, we must have that q(0) < P2 . This is equivalent to     1 1 P1 P1 q(0) = − 1 < P2 ⇔ a12 > −1 . a12 γR̄ P2 γR̄ (51) If a12 is greater than this quantity and also than ι(α), we know by Lemma 3 that κ2 = 1 is optimal if q(1) ≤ P2 . Otherwise, it is clear that, if improper signaling is optimal, then the optimal circularity coefficient κ2 will satisfy p(κ2 ) = P2 . Let us denote this circularity coefficient as κ2 = κmax . Then, using (24) we have R2 (κmax ) ≥ R2 (0) ⇔ κ2max ≤ 1 −   2 1 1 P1 2P2 + 1 1 + . − 1 + P22 a12 γR̄ P22 (52) Therefore, we have to evaluate the condition for κmax , with p(κmax ) = P2 , fulfilling the above expression. Let us first consider the case κ1 = 1. From (23), we have that   1 2P1 q(κmax ) = P2 ⇔ κmax = 1 − − 1 . (53) a12 P2 γ2R̄ 12 Plugging this expression into (52) we have the condition   2 1 2P1 1− −1 ≤ a12 P2 γ2R̄   2 1 P1 2P2 + 1 1 , (54) −1 + 1− 2 1+ P2 a12 γR̄ P22 which, after some manipulations, yields the quadratic inequality  1 a212 + a12 (P2 aI − aP ) − P2 a2I + a2P ≥ 0 , 2 (55) where aP and aI are respectively given by (30) and (31). This quadratic expression is convex and has at most one positive root. Hence, the condition on a12 is given by the largest root, which is p aP − P2 aI + (aP − P2 aI )2 + 2P2 (a2I + a2P ) . a12 > 2 (56) This expression, however, is only applicable if κ1 = 1 holds for the above threshold. Using (16), κ1 = 1 holds if p aP − P2 aI + (aP − P2 aI )2 + 2P2 (a2I + a2P ) P1 + aI . ≥ 2 P2 (57) Otherwise we have κ1 < 1, thus we have to obtain κmax using the first branch in (23) (or, alternatively, taking equality in (38) with p2 = P2 ), which yields q(κmax ) = P2 ⇔ κ2max = P2 a12 [P2 a12 γ2R̄ − 2 (P1 − γ2R̄ )] + γ2R̄ + 1 − (1 + P1 )2 . P22 a212 (γ2R̄ + 1) (58) Plugging this expression into (52) we obtain 2 P2 a12 [P2 a12 γ2R̄ − 2 (P1 − γ2R̄ )] + γ2R̄ + 1 − (P1 + 1) ≤ P22 a212 (γ2R̄ + 1)   2 1 P1 2P2 + 1 1 , (59) −1 + 1− 2 1+ P2 a12 γR̄ P22 which, after some manipulations, yields   1 P1 − γ2R̄ 2 P2 + 2P2 +1 γ2R̄ + 1 a12 (γ2R̄ + 1)   2 2 1 P1 (1 + P1 ) − γ2R̄ − 1 + 1 − 1 + − 1 ≥0. + a212 (γ2R̄ + 1) a12 γR̄ (60) Since the above quadratic expression is convex, we have to consider the largest root. As expected, one of the roots is   P1 1 −1 , (61) P2 = a12 γR̄ which corresponds to the maximum power allowed by proper signaling. Since we have already considered this threshold in (51), we have to use the second root, which can be simplified to    1 P1 (2γR̄ + 1) 2 (γ2R̄ + 1) (1 − a12 ) − P2 = +1 . a12 γR̄ (62) If the above root is the largest one, the inequality (60) is satisfied when P2 exceeds this value, which yields the condition a12 > γR̄ (2γ2R̄ + 1) − P1 (2γR̄ + 1) . γR̄ [P2 + 2 (γ2R̄ + 1)] (63) Combining (51)-(63) we obtain that a12 > ρ(α) must hold for improper signaling to be optimal, with ρ(α) given by (27). However, this condition is only valid if a12 > ι(α). Consequently, improper signaling will be optimal if and only if a12 is greater than the dominating threshold, i.e., if and only if a12 > max[ι(α), ρ(α)]. If this expression is satisfied, then by Lemma 3 κ2 must be increased until p(κ2 ) = P2 , thus the optimal circularity coefficient is given by (33). This concludes the proof. R EFERENCES [1] F. D. Neeser and J. L. Massey, “Proper complex random processes with applications to information theory,” IEEE Transactions on Information Theory, vol. 39, no. 4, pp. 1293–1302, Jul. 1993. [2] V. Cadambe, S. Jafar, and C. Wang, “Interference alignment with asymmetric complex signaling—settling the Høst-Madsen-Nosratinia conjecture,” IEEE Transactions on Information Theory, vol. 56, no. 9, pp. 4552–4565, Sep. 2010. [3] C. Lameiro and I. Santamarı́a, “Degrees-of-freedom for the 4-user SISO interference channel with improper signaling,” in Proc. IEEE Int. Conf. Comm., Budapest, Hungary, Jun. 2013. [4] C. Wang, T. Gou, and S. A. Jafar, “On optimality of linear interference alignment for the three-user MIMO interference channel with constant channel coefficients,” [Online]. Available: http://escholarship.org/uc/item/6t14c361, 2011. [5] H.-Y. Shin, S.-H. Park, H. Park, and I. Lee, “A new approach of interference alignment through asymmetric complex signaling and multiuser diversity,” IEEE Transactions on Wireless Communications, vol. 11, no. 3, pp. 880–884, Mar. 2012. [6] L. Yang and W. Zhang, “Interference alignment with asymmetric complex signaling on MIMO X channels,” IEEE Transactions on Communications, vol. 62, no. 10, pp. 3560–3570, Oct. 2014. [7] Z. Ho and E. Jorswieck, “Improper Gaussian signaling on the two-user SISO interference channel,” IEEE Transactions on Wireless Communications, vol. 11, no. 9, pp. 3194–3203, Sep. 2012. [8] Y. Zeng, C. Yetis, E. Gunawan, Y. Guan, and R. Zhang, “Transmit optimization with improper Gaussian signaling for interference channels,” IEEE Transactions on Signal Processing, vol. 61, no. 11, pp. 2899– 2913, Jun. 2013. [9] Y. Zeng, R. Zhang, E. Gunawan, and Y. Guan, “Optimized transmission with improper Gaussian signaling in the K-user MISO interference channel,” IEEE Transactions on Wireless Communications, vol. 12, no. 12, pp. 6303–6313, Dec. 2013. [10] H. Nguyen, R. Zhang, and S. Sun, “Improper signaling for symbol error rate minimization in K-user interference channel,” IEEE Transactions on Communications, vol. 63, no. 3, pp. 857–869, Mar. 2015. [11] S. Lagen, A. Agustin, and J. Vidal, “Coexisting linear and widely linear transceivers in the MIMO interference channel,” IEEE Transactions on Signal Processing, vol. 64, no. 3, pp. 652–664, Feb. 2016. [12] C. Hellings, M. Joham, and W. Utschick, “QoS feasibility in MIMO broadcast channels with widely linear transceivers,” IEEE Signal Processing Letters, vol. 20, no. 11, pp. 1134–1137, Nov. 2013. 13 [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] C. Kim, E.-R. Jeong, Y. Sung, and Y. H. Lee, “Asymmetric complex signaling for full-duplex decode-and-forward relay channels,” in Proceedings of the International Conference on ICT Convergence (ICTC), Oct. 2012, pp. 28–29. C. Lameiro, I. Santamarı́a, and P. J. Schreier, “Benefits of improper signaling for underlay cognitive radio,” IEEE Wireless Communications Letters, vol. 4, pp. 22—25, Feb. 2015. C. Lameiro, I. Santamarı́a, and P. Schreier, “Analysis of maximally improper signalling schemes for underlay cognitive radio,” in Proceedings of the IEEE International Conference on Communications, London, UK, Jun. 2015. O. Amin, W. Abediseid, and M.-S. Alouini, “Outage performance of cognitive radio systems with improper Gaussian signaling,” in Proceedings of the IEEE International Symposium on Information Theory (ISIT), Hong Kong, China, Jun. 2015, pp. 1851–1855. E. Kurniawan and S. Sun, “Improper Gaussian signaling scheme for the Z-interference channel,” IEEE Transactions on Wireless Communications, vol. 14, no. 7, pp. 3912–3923, Jul. 2015. S. Lagen, A. Agustin, and J. Vidal, “Improper Gaussian signaling for the Z-interference channel,” in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Florence, Italy, May 2014, pp. 1145–1149. ——, “On the superiority of improper Gaussian signaling in wireless interference MIMO scenarios,” IEEE Transactions on Communications, vol. 64, no. 8, pp. 3350–3368, Aug. 2016. P. J. Schreier and L. L. Scharf, Statistical signal processing of complexvalued data: the theory of improper and noncircular signals. Cambridge, U.K.: Cambridge Univ. Press, 2010. T. Adali, P. J. Schreier, and L. L. Scharf, “Complex-valued signal processing: The proper way to deal with impropriety,” IEEE Transactions on Signal Processing, vol. 59, no. 11, pp. 5101–5125, Nov. 2011. G. Tauböck, “Complex-valued random vectors and channels: Entropy, divergence, and capacity,” IEEE Transactions on Information Theory, vol. 58, no. 5, pp. 2729–2744, May 2012. W. Gerstacker, R. Schober, and A. Lampe, “Receivers with widely linear processing for frequency-selective channels,” IEEE Transactions on Communications, vol. 51, no. 9, pp. 1512–1523, Sep. 2003. S. Buzzi, M. Lops, and S. Sardellitti, “Widely linear reception strategies for layered space-time wireless communications,” IEEE Transactions on Signal Processing, vol. 54, no. 6, pp. 2252–2262, Jun. 2006. P. Chevalier and F. Pipon, “New insights into optimal widely linear array receivers for the demodulation of BPSK, MSK, and GMSK signals corrupted by noncircular interferences—application to SAIC,” IEEE Transactions on Signal Processing, vol. 54, no. 3, pp. 870–883, Mar. 2006. R. Schober, W. Gerstacker, and L.-J. Lampe, “Data-aided and blind stochastic gradient algorithms for widely linear MMSE MAI suppression for DS-CDMA,” IEEE Transactions on Signal Processing, vol. 52, no. 3, pp. 746–756, Mar. 2004. B. W. Han and J. H. Cho, “Capacity of second-order cyclostationary complex Gaussian noise channels,” IEEE Transactions on Communications, vol. 60, no. 1, pp. 89–100, Jan. 2012. J. Yeo and J. H. Cho, “Asymptotic frequency-shift properizer for block processing of improper-complex second-order cyclostationary random processes,” IEEE Transactions on Information Theory, vol. 60, no. 7, pp. 4083–4100, Jul. 2014. J. Yeo, J. H. Cho, and J. S. Lehnert, “Joint transmitter and receiver optimization for improper-complex second-order stationary data sequence,” Journal of Communications and Networks, vol. 17, no. 1, pp. 1–11, Feb. 2015. J. Yeo, B. W. Han, J. H. Cho, and J. S. Lehnert, “Capacity of an orthogonal overlay channel,” IEEE Transactions on Wireless Communications, vol. 14, no. 11, pp. 6111–6124, Nov. 2015. M. Kim, J. H. Cho, and J. S. Lehnert, “Asymptotically optimal low-complexity SC-FDE with noise prediction in data-like improper- [32] [33] [34] [35] [36] [37] complex interference,” IEEE Transactions on Wireless Communications, vol. 15, no. 3, pp. 2090–2103, Mar. 2016. M. Costa, “On the Gaussian interference channel,” IEEE Transactions on Information Theory, vol. 31, no. 5, pp. 607–615, Sep. 1985. R. Prasad, S. Bhashyam, and A. Chockalingam, “On the sum-rate of the Gaussian MIMO Z channel and the Gaussian MIMO X channel,” IEEE Transactions on Communications, vol. 63, no. 2, pp. 487–497, Feb. 2015. H. Sato, “The capacity of the Gaussian interference channel under strong interference,” IEEE Transactions on Information Theory, vol. 27, no. 6, pp. 786–788, Nov. 1981. H. Song, J. Y. Ryu, and W. Choi, “Characterization of the Pareto boundary for the two-user symmetric Gaussian interference channel,” IEEE Transactions on Communications, vol. 62, no. 8, pp. 2812–2824, Aug. 2014. C. Hellings and W. Utschick, “Block-skew-circulant matrices in complex-valued signal processing,” IEEE Transactions on Signal Processing, vol. 63, no. 8, pp. 2093–2107, Apr. 2015. E. A. Jorswieck, E. G. Larsson, and D. Danev, “Complete characterization of the pareto boundary for the MISO interference channel,” IEEE Transactions on Signal Processing, vol. 56, no. 10, pp. 5292–5296, Oct. 2008.
7cs.IT
International Journal of Computer Trends and Technology (IJCTT) – volume 10 number 4 – Apr 2014 Application of Machine Learning Techniques in Aquaculture 1 Akhlaqur Rahman1 and Sumaira Tasnim2 Department of Electrical and Electronic Engineering Uttara University, Bangladesh 2 School of Engineering, Deakin University, Australia ABSTRACT: In this paper we present applications of different machine learning algorithms in aquaculture. Machine learning algorithms learn models from historical data. In aquaculture historical data are obtained from farm practices, yields, and environmental data sources. Associations between these different variables can be obtained by applying machine learning algorithms to historical data. In this paper we present applications of different machine learning algorithms in aquaculture applications. Keywords: aquaculture, machine learning, decision support system 1. INTRODUCTION Aquaculture refers to the farming of aquatic organisms such as fish and aquatic plants. It involves cultivating freshwater and saltwater populations under controlled conditions. Use of sensor technologies to monitor the environment where aquaculture operations take place is a recent trend. The sensors collect data about the aquaculture environment that are using by farm managers for decision making purposes. The literature states a number of activities related to decision support systems in Aquaculture farm operations. A number of decision support systems have been developed for. Some of them use machine learning and methods and other do not. For the sake of completeness we briefly discuss these other methods first and detail the machine learning based methods in the following section. Bourke et al. [1] developed a framework where real-time water quality indicators, as well as operational information were displayed and their impact on survival rate, biomass and production failure of aquaculture species were evaluated. Wang et al. [2] developed an early warning system for dangerous growing conditions. Padala and Zilber [3] used expert system generated rules to reduce stock loss and increase size and quality of yields. Ernst et al. [4] focused on managing hatchery production using rules and calculations of physical, chemical and biological processes. Silvert [5] developed a scientific model to evaluate the environmental impact. Halide et al. [6] used rules that are hand-crafted from domain experts. ISSN: 2231-2803 2. MACHINE LEARNING METHODS IN AQUACULTURE APPLICATIONS In this section we discuss the different applications where machine learning methods are applied in the aquaculture domain. 2.1 Shellfish Farm Closure Prediction Consumption of contaminated shellfish can cause severe illness and even death in humans. The authors in [7]–[9] developed a sensor network based approach to predict the contamination event using machine learning methods. In [7] the authors presented a machine learning methods to obtain a balance between farm closure and farm opening events. The authors in [8] presented a feature ranking algorithm to identify the most influential cause of closure. In [9] the authors adopted time series machine learning approaches like PCA (Principal Component Analysis) and ACF (Auto Correlation Function) to predict the closure event. 2.2 Algae Bloom Prediction Algae organisms grow widely throughout in the world. They provide food and shelter to other organisms. When the growth of algae is excessive, it can cause oxygen depletion in water and kill fish. The authors in [10]–[12] developed machine learning based methods to predict algae growth/bloom. The authors in [10] extracted a set of rules from data gathered by sensor networks to find associations between environmental variables and algae growth. The authors in [11] designed an ensemble method to find the relevant environmental variables responsible for algae growth and predict the growth. The authors in [12] developed machine learning methods to predict the propagation of algae patches along the waterway. 2.3 Missing Values Estimation Sensor networks deployed in the field to monitor aquaculture environments often suffer failure due to sensor bio-fouling, communication failure etc. This results in missing sensor readings that is required by the machine learning based decision making systems. This results in a requirement to deal with missing sensor values. The authors in [13] and [14] designed a multiple classifier based method to deal with missing sensor data. Instead of imputing missing data, prediction http://www.ijcttjournal.org Page214 International Journal of Computer Trends and Technology (IJCTT) – volume 10 number 4 – Apr 2014 of events is based on available sensor readings. Ensemble approach is shown to perform better than imputation methods. The authors assume equal weights for all features which may not be true in real world scenarios [15][16]. 2.4 Model Relocation In order to train machine learning methods it is required to provide historical labeled data. However when a farmer aims to setup a new aquaculture firm, historical farming data is not available for that location. The authors in [17] provided a guideline to utilize machine learning models trained for a different location to a new location based on similarity between the two locations. Results show that model relocation can significantly reduce the shortcoming generated from data unavailability for a particular location. 2.5 Benthic Habitat Mapping In order to get an understanding of the aquaculture habitats on seafloor the researchers sometimes send Autonomous Underwater Vehicles to collect seafloor images. The images are later visually analyzed by the researchers to produce the habitat map in a region. The authors in [18] developed an image processing and machine learning based method to automatically produce habitat maps from seafloor images. The results presented in the paper showed significant accuracy in obtaining correct habitat maps. 2.6 Sensor Data Quality Assessment Data produced by the sensors are sometimes faulty. The decisions based on faulty sensor reading will result in wrong conclusion. In [19][20] the authors have presented a novel ensemble classifier approach for assessing the quality of sensor data. The base classifiers are constructed by random under–sampling of the training data where the sampling process is guided by clustering. The inclusion of cluster based under–sampling and multi–classifier learning has shown to improve the accuracy of quality assessment. 3. CONCLUSIONS In this paper we have presented a review of machine learning methods in the aquaculture space. In future we aim to extend some of these methods n time series machine learning methods. REFERENCES [1] Bourke, G., Stagnitti, F., and Mitchell, B. 1993. A decision support system for aquaculture research and management. Aquacultural Engineering. 12, 2, 111–123. [2] Wang, R., Chen, D. Q., and Fu, Z. 2006. AWQEE-DSS: A Decision Support System for Aquaculture Water Quality Evaluation and Early-warning. International Conference on Computational Intelligence and Security. 2, 959–962. ISSN: 2231-2803 [3] Padala, A., Zilber, S. 1991. Expert Systems and Their Use in Aquaculture. Rotifer and Microalgae Culture Systems, Proceedings of a US-Asia Workshop, Honolulu, Hawaii ( 1991). [4] Ernst, D. H., Bolte, J. P., and Nath, S. S. 2000. AquaFarm: simulation and decision support for aquaculture facility design and management planning. Aquacultural Engineering. 23, 1–3 (September 2000), 121–179. [5] Silvert, W. 1994. Decision support systems for aquaculture licensing. Journal of Applied Ichthyology. 10, 307–311. [6] Halide, H. Stigebrandt, A., Rehbein, M., and McKinnon, A. D. 2009. Developing a decision support system for sustainable cage aquaculture. Environmental Modelling & Software. 24, 6 (June 2009), 694–702. [7] C. D’ Este, A. Rahman, and A. Turnbull, “Predicting Shellfish Farm Closures with Class Balancing Methods,” AAI 2012: Advances in Artificial Intelligence, Lecture Notes in Computer Science, pp. 39–48, 2012. [8] A. Rahman, C. D'Este, and J. McCulloch, “Ensemble Feature Ranking for Shellfish Farm Closure Cause Identification,” Proc. Workshop on Machine Learning for Sensory Data Analysis in conjunction with Australian AI conference, DOI: 10.1145/2542652.2542655 http://doi.acm.org/10.1145/2542652.2542655, 2013. [9] M. S. Shahriar, A. Rahman, and J. McCulloch, “Predicting Shellfish Farm Closures using Time Series Classification for Aquaculture Decision Support,” Elsevier Computer and Electronics in Agriculture. [10] M. S. Shahriar, C. D. Este and A. Rahman, “On Detecting and Predicting Harmful Algal Blooms in Coastal Information Systems,” Proc. IEEE Oceans, 2012. [11] A. Rahman and MS Shahriar, “Algae Bloom Prediction through Identification of Influential Environmental Variables: A Machine Learning Approach,” International Journal of Computational Intelligence and Applications, vol. 12, no. 2, 2013. [12] M. S. Shahriar and A. Rahman, “Spatio–temporal Prediction of Algal Bloom,” Proc. IEEE International Conference on Natural Computation (ICNC), pp. 968–972, China, 2013. [13] Q. Zhang, A. Rahman, and C. D'Este, “Impute vs. Ignore: Missing Values for Prediction,” Proc. IEEE International Joint Conference on Neural Networks (IJCNN), pp. 2193– 2200, Dallas, Texas, 2013. [14] A. Rahman, Claire D’ Este, and G. Timms, “Dealing with Missing Sensor Values in Predicting Shellfish Farm Closure,” Proceedings IEEE Intelligent Sensors, Sensor Networks and Information Processing (ISSNIP), pp. 351– 356, Melbourne, 2013. [15] A. Rahman and M. Murshed, “Feature weighting methods for abstract features applicable to motion based video indexing,” IEEE International Conference on Information Technology: Coding and Computing (ITCC), vol. 1, pp. 676–680, USA, 2004. [16] A. Rahman and M. Murshed, “Feature Weighting and Retrieval Methods for Dynamic Texture Motion Features,” International Journal of Computational Intelligence Systems, vol. 2, no. 1, pp. 27–38, March 2009. [17] C. D’ Este and A. Rahman, “Similarity Weighted Ensembles for Relocating Models of Rare Events,” Proc. International Workshop on Multiple Classifier Systems (MCS), Lecture Notes in Computer Science, pp. 25–36, Nanjing, China, May 15-17, 2013. [18] A. Rahman, “Benthic Habitat Mapping from Seabed Images using Ensemble of Color, Texture, and Edge Features,” International Journal of Computational Intelligence Systems, vol. 6, issue 6, pp. 1072–1081, DOI: 10.1080/18756891.2013.816055, 2013. [19] A. Rahman, D. Smith, and G. Timms, “A Novel Machine Learning Approach towards Quality Assessment of Sensor Data,” IEEE Sensors Journal, DOI: 10.1109/JSEN.2013.2291855. [20] A. Rahman, D. Smith, and G. Timms “Multiple Classifier System for Automated Quality Assessment of Marine Sensor Data,” Proceedings IEEE Intelligent Sensors, Sensor Networks and Information Processing (ISSNIP), pp. 362–367, Melbourne, 2013. http://www.ijcttjournal.org Page215
5cs.CE
ExTASY: Scalable and Flexible Coupling of MD Simulations and Advanced Sampling Techniques Vivekanandan Balasubramanian∗ , Iain Bethune† , Ardita Shkurti‡ , Elena Breitmoser† Eugen Hruska¶k , Cecilia Clementi§k , Charles Laughton‡ and Shantenu Jha∗ ∗ Department of Electrical and Computer Engineering, Rutgers University, Piscataway, NJ 08854, USA The University of Edinburgh, James Clerk Maxwell Building, Peter Guthrie Tait Road, Edinburgh, UK, EH9 3FD ‡ School of Pharmacy and Centre for Biomolecular Sciences, University of Nottingham, University Park, Nottingham NG7 2RD, UK § Department of Chemistry, Rice University, Houston, TX 77005, USA ¶ Department of Physics, Rice University, Houston, TX 77005, USA k Center for Theoretical Biological Physics, Rice University, Houston, TX 77005, USA arXiv:1606.00093v1 [cs.CE] 1 Jun 2016 † EPCC, of trajectories is required to build a statistically meaningful model of the dynamical process. The high dimensionality of these macromolecular systems and the complexity of the associated potential energy surfaces (creating multiple metastable regions connected by high free energy barriers) pose significant challenges to adequately sample the relevant regions of the configurational space. In other words, beside the “curse of dimensionality” associated with the large number of degrees of freedom, MD trajectories can easily get “trapped” in a low free energy state and fail to explore other biologically relevant states. The waiting time to escape from a local free energy minimum increases exponentially with the height of the free energy barrier that needs to be crossed to reach another state. Metastable states separated by free energy barriers of several tens of kB T (where kB is the Boltzmann constant and T is the physiological temperature) are not uncommon in biologically relevant systems, but can not at present be routinely sampled with standard MD simulations. In practice, better sampling of the relevant regions of a macromolecule configurational space can be achieved through methodologies able to bias the sampling towards scarcely visited regions, reducing the waiting time inside a metastable state by artificially flattening the energy barriers between states - e.g. Metadynamics [3] or Accelerated Dynamics [4]. Although the results can be usually reweighed to reproduce the correct Boltzmann statistics, kinetics properties are not easily recovered from biased simulations (unless used in combinations with unbiased simulations, see e.g. [5]). In addition, the design of an effective bias usually requires some a priori information on the system of interest, for instance on a suitable choice of collective variables to describe slow timescale processes. An alternative approach to tackle the sampling problem is the development of ensemble or swarm simulation strategies, where data from large numbers of simulations, which may be weakly coupled or not coupled at all, are integrated (e.g. Replica Exchange [6] and Markov State Models (MSMs) [7]). This last class of methods is of increasing interest for a variety of reasons. Firstly, the hardware roadmap is now Abstract—For many macromolecular systems the accurate sampling of the relevant regions on the potential energy surface cannot be obtained by a single, long Molecular Dynamics (MD) trajectory. New approaches are required to promote more efficient sampling. We present the design and implementation of the Extensible Toolkit for Advanced Sampling and analYsis (ExTASY) for building and executing advanced sampling workflows on HPC systems. ExTASY provides Python based “templated scripts” that interface to an interoperable and high-performance pilot-based run time system, which abstracts the complexity of managing multiple simulations. ExTASY supports the use of existing highly-optimised parallel MD code and their coupling to analysis tools based upon collective coordinates which do not require a priori knowledge of the system to bias. We describe two workflows which both couple large “ensembles” of relatively short MD simulations with analysis tools to automatically analyse the generated trajectories and identify molecular conformational structures that will be used on-the-fly as new starting points for further “simulation-analysis” iterations. One of the workflows leverages the Locally Scaled Diffusion Maps technique; the other makes use of Complementary Coordinates techniques to enhance sampling and generate start-points for the next generation of MD simulations. We show that the ExTASY tools have been deployed on a range of HPC systems including ARCHER (Cray CX30), Blue Waters (Cray XE6/XK7), and Stampede (Linux cluster), and that good strong scaling can be obtained up to 1000s of MD simulations, independent of the size of each simulation. We discuss how ExTASY can be easily extended or modified by endusers to build their own workflows, and ongoing work to improve the usability and robustness of ExTASY. I. I NTRODUCTION AND M OTIVATION Approximately 30-40% of compute cycles on US XSEDE [1], [2] is devoted to research on biomolecular systems using Molecular Dynamics (MD) simulations. Much of the computational cost comes from the need for an adequate sampling of the conformational space accessible to these complex and flexible systems in order to answer a particular research question. For example, to calculate free energies one needs an adequate sample from the Boltzmann weighted ensemble of states for the system in order to estimate the thermodynamic quantity of interest. Another example is the study of kinetic processes such as self-assembly or drug-target association, where the integration of data from large numbers 1 based almost entirely on increasing core counts, rather than clock speeds. On the face of it, these developments favour weak scaling problems (larger and larger molecular systems to be simulated) over strong scaling problems (getting more data faster on a system of fixed size). However, by running ensembles of simulations over these cores and integrating the data using, e.g. MSM approaches, timescales far in excess of those sampled by any individual simulation are effectively accessed. In the last few years several studies have been published [8]–[12] where, using MSM methods, processes such as protein folding or ligand binding have been completely and quantitatively characterized (thermodynamically and kinetically) from simulations orders of magnitude shorter than the process of interest. It is becoming increasingly clear that the application of ensemble simulation strategies on state-of-the-art computational facilities has an unparalleled potential to permit accurate simulations of the largest, most challenging, and generally most biologically relevant, biomolecular systems. The main challenge in the development of the ensemble approach for faster sampling of complex macromolecular systems is on the design of strategies to adaptively distribute the trajectories over the relevant regions of the systems configurational space, without using any a priori information on the system global properties. The definition of smart “adaptive sampling” approaches that can redirect computational resources towards unexplored yet relevant regions is currently of great interest. In light of the challenge posed by trends in computer architecture, the need to improve sampling, and the range of existing MD codes and analysis tools, we have designed and implemented the Extensible Toolkit for Advanced Sampling and analYsis (ExTASY). ExTASY provides three key features within a single framework to enable the development and applications requiring advanced sampling in a flexible and highly scalable environment. Firstly, as an extensible toolkit, ExTASY allows a wide range of existing software to be integrated, leveraging the significant community investment in highly optimised and well-tested software packages and enabling users to continue to work with tools that they are familiar with. Support for specific MD codes and analysis tools is provided in order to demonstrate how ExTASY may be used, but users can easily add other tools as needed. Secondly, ExTASY is flexible, providing a programmable interface to link individual software components together and construct sampling workflows. Workflows capture a sequence of execution of individual tools, and the data transfers and dependencies between them. First class support is provided for defining large ensembles of independent simulations. Thus complex calculations may be scripted and then executed without the need for user intervention. Thirdly, ExTASY workflows may be executed either locally, or on remote High Performance Computing systems. Complexities such as the batch queueing system and data transfer are abstracted, making it easy for users to make use of the most appropriate compute resources they have access to. In addition, this abstraction allows workflows to be executed without exposing each component to queue waiting time, and respecting the dependencies defined in the workflow, for many simulations to be scheduled and executed in parallel. The rest of the paper is organized as follows: In the next section we discuss the design and implementation of ExTASY. After a brief discussion in Section III of two distinct applications that have been used to design and validate ExTASY, Section IV provides a careful analysis of the performance and scalability of ExTASY. Given the complex interplay between functionality and performance when designing an extensible and production tool, we perform a wide range of experiments aimed to investigate strong and weak scaling properties, inter alia over a set of heterogeneous HPC platforms. We conclude with a discussion of the scientific impact as well as the lessons for sustainable software development. II. R ELATED W ORK The need for better sampling has driven developments in methodology (algorithm), hardware, and software for (bio)molecular simulation. One of the features of the popular Metadynamics method [3] or Accelerated Dynamics [4] is that a constant analysis of what has been sampled so far is used to bias future sampling into unexplored regions. A range of alternative approaches are now emerging that do likewise, but where alternating segments of data-gathering, and analysis to inform the direction of future sampling, are more coarsely grained. This iterative approach has the advantage over the metadynamics method that the identity of “interesting” directions for enhanced sampling does not need to be defined a priori, but can emerge and respond flexibly to the developing ensemble. Another advantage is that the MD-based sampling process and analysis method do not have to be implemented within the same executable, or in two tightly-coupled executables, permitting greater flexibility. Many such methods make use of collective variables (CVs) to define directions in which to promote sampling. A variety of novel, and established, algorithms for the unsupervised and adaptive construction of CVs exist. In addition to the work of Preto and Clementi [13], interleaving cycles of MD simulation with data analysis through Locally Scaled Diffusion Maps [14], related methods include the non-targeted PaCSMD method of Harada and Kitao [15], variants thereof [16], and the PCA-based method of Peng and Zhang [17]. Better sampling can also come from faster sampling, which has been enabled through hardware developments such as ANTON [18] and MD-GRAPE [19]. These special purpose computers enable much faster calculations of the different contributions to the forces along the trajectories, thus speeding up the clock time required to perform a time integration step in the MD simulation and allowing execution of significantly longer MD trajectories. The cost of, and access to such special purpose computers ensure that in spite of their potential, they will not be as accessible for the wider scientific community as general purpose approaches. Further ANTON requires a customized ecosystem, from bespoke MD engines and ANTON specific data analysis middleware (e.g., 2 HiMach). Thus ANTON style special-purpose approaches to bio-molecular simulation science cannot take advantage of the rich community driven advances and eco-system. Methods such as Replica Exchange and Metadynamics require a tight coupling between the simulation and analysis processes, and are thus typically implemented as additional facilities within the core MD code (e.g. replica exchange methods are implemented in AMBER [20], CHARMM [21], GROMACS [22], LAMMPS [23], and NAMD [24], for example), or are provided by a separate package that communicates in a fine-grained manner with the running MD executable, generally though specially-installed “hooks”; an example of this approach is the PLUMED package [25] which provides metadynamics capabilities (amongst others) to AMBER, GROMACS, LAMMPS, NAMD and also Quantum ESPRESSO [26]. In contrast, there is, to our knowledge, so far no established and generally available package to support the types of coarser-grained, adaptive workflows described above. of cores available or utilized. On the other hand, if access to a large number of cores is available, the framework should be able to use them effectively. In this regard, the strong and weak scalability of the framework is to be investigated. 3) Usability: Depending on the application, the user might need to change to using a larger/smaller set of input data, modify simulation parameters or replace any of the simulation or analysis tools. The framework should offer easy application setup, minimizing the user’s time and effort in the process. The user should only be concerned with decisions on “what” the workflow is and “where” it is being executed. The details as to “how” the deployment and execution occurs should be hidden by the underlying software. Thus the framework should use tools that abstract complexities of deployment and execution from the user. Workflow users and developers should be able to concentrate their efforts on the application logic and expect the underlying software to provide a level of transparent and automation of aspects such as data movement and job submission. III. E X TASY: R EQUIREMENTS , D ESIGN AND I MPLEMENTATION B. Design From these requirements, we identify the following as the primary design objectives of ExTASY. 1) Support range of HPC systems, abstract task execution, data movement from the user. 2) Flexible resource management and coupling capabilities between different stages as well as within a stage. 3) Provide the users with an easy method to specify or change workload parameters without delving into the application itself. These design objectives put together lead to the following simple software architecture: 1) Middleware for resource and execution management: The ExTASY framework is aimed to provide an easy method to run advanced sampling algorithms on HPC systems. In this process, the ExTASY framework should abstract users from the complexity of application composition, workload execution and resource management. There is need for a middleware that provides such resource and execution management that provides many of the required functionalities and performance. The details of how tasks are mapped or executed on to resources is abstracted from the ExTASY user. The user is only required to provide details of the workload and identify the resource(s) that are accessible. This design choice thus acknowledges the separation of concerns: workload description from its execution on HPC systems. 2) Configuration files: Composing the advanced sampling algorithms discussed in the previous sections using the components provided by a particular middleware stack, requires specific knowledge of the middleware itself. The ExTASY framework bypasses this complexity by adding one more level of abstraction. It provides ready-to-use scripts in accordance with advanced sampling methods discussed in this paper that are configurable via configuration files. In these configuration files, the user is exposed to only application-level, meaningful parameters that can be modified as necessary. In this section we first present the requirements that have been considered in the design and implementation of ExTASY, which we then go on to discuss. A. Requirements Consistent with the design of many new software systems and tools, we analyze the functionality, performance and usability requirements of ExTASY. 1) Functionality: Specific to sampling, there is a need to couple two very distinct computational stages; each stage can be short-lived when compared to the typical duration of monolithic simulation. Furthermore, the two stages differ in their resource requirements significantly: one stage is characterized by multiple compute intensive MD simulations, the other by a single analysis program that operates on data aggregated from multiple simulations. The “Ex” in ExTASY is a reference to the extensible nature of the framework, and thus any coupling must be between the abstraction of stages and not specific codes for fixed time duration. Scientists may have access to multiple systems, and wish to submit jobs on each, or may be forced to migrate from using one system to another due to CPU time allocation, or system end-of-life. It is imperative that any software system support interoperability, i.e. use of heterogeneous systems with minimal changes. MD simulations may be executed over multiple nodes, depending on the system size, thus any software system should also support the ability to execute tasks over multiple nodes. 2) Performance: In order to obtain good sampling, a large number of simulations must be executed. In many cases, the aggregate number of cores required by the set of simulations (“ensemble”) is much higher than the total number of cores that are available at a given instance or that can be allocated on a system. The framework must decouple the aggregate (or peak) resource requirement of the workload from the number 3 Fig. 2. The SAL pattern common to both sampling algorithms. The crux of the pattern is an iteration over 2 stages: simulation and analysis, where the number of simulation and analysis instances can be different. The pattern may also contain pre- and post- processing stages. Fig. 1. Design of the ExTASY framework using Ensemble Toolkit as middleware. The ExTASY framework provides ready-to-use scripts created using components provided by Ensemble Toolkit. Parameters related to the resource and workload are exposed via configuration files, which alone are the files that users interact with. Within Ensemble Toolkit, the workload is converted into executable units by the execution plugins and submitted to the resource using RADICAL-Pilot. IV. A PPLICATIONS We illustrate the capabilities of the ExTASY approach via two exemplar applications. The two different advanced sampling algorithms implemented with ExTASY are the Diffusion Map-directed-MD (DM-d-MD) and CoCo-MD techniques. Both these algorithms have a common execution pattern: an ensemble of simulation tasks followed by an analysis stage, performed for multiple iterations following the pattern shown in Figure 2. In the case of the DM-d-MD algorithm, the simulation stage consists of Gromacs and LSDMap in the analysis stage. Whereas, in the CoCo algorithm, the simulation stage consists Gromacs runs and trajectory conversions and analysis consists of CoCo. The individual simulation or analysis tools might differ depending on the algorithm chosen but the overall pattern is observed to be the same. C. Implementation 1) Ensemble Toolkit: As mentioned previously, ExTASY requires a middleware for resource and execution management. We chose to use Ensemble Toolkit [27] as the middleware component as it provides several relevant features such as the ability to support MPI tasks, dynamic resource management – one type of which is to be able to execute more tasks than the resources available, support for heterogeneous HPC systems and strong and weak scalability guarantees. Ensemble Toolkit has been tested upto O(1,000) tasks with short and long term plans to support O(10,000) and O(100,000) tasks [27]. Ensemble Toolkit is in turn based upon the pilot abstraction (and the RADICAL-Pilot [28] implementation of the pilot abstraction) to provide much of the flexible and scalable resource management capabilities. Ensemble Toolkit exposes three components to the user that can be used to express many applications: Kernel Plugins, Execution Patterns, and Resource Handle. Scripts that are part of ExTASY framework use these components to describe the application logic. 2) Configuration files: The application logic is expressed via components of Ensemble Toolkit. The resource and the workload specifications are exposed via configuration files. The ExTASY framework has two types of configuration files: (i) resource configuration, which consist of details of the resource where the application will to be executed such as the resource name, the runtime, and the username and account details used to access the resource, and (ii) kernel configuration, which defines the workload parameters such as the location of input files for the Molecular Dynamcis simulation and analysis tools, parameters for the tools, and workflow parameters such as the number of simulations. A. Diffusion Map-directed-MD The Diffusion Map-directed-MD (DM-d-MD) technique [13] improves the efficiency of computational resources by choosing which replicas of the protein are used to run MD. When replicas are too close to each other, the MD trajectories will be similar. The information gain from simulating MD with close replicas is small. Part of the replicas which are too close to each other are deleted. To hold the total number of replicas constant, replicas which are too far apart from each other are duplicated. In DM-d-MD, a non-linear dimensionality reduction technique, the locally scaled diffusion map (LSDMap) [14] is used to calculate the distance between different replicas. The deletion or duplication of replicas would destroy the correct sampling of the protein. By changing the weights of individual replicas in the reweighting step, the correct sampling of the protein is obtained. The DM-d-MD technique requires only the protein starting structure. No additional information about the protein is necessary. The user can fine tune the sampling mainly by varying the total number of replicas and the way how the local scale 4 in LSDMap is calculated. At the begin on the method, the replicas are generated from the protein starting structure. After the MD step, the LSDMap is calculated. LSDMap requires only the final structure for each replica from the MD step. Based on the LSDMap results new replicas for the next iteration of DM-d-MD are chosen from the current replicas. The reweighting ensures that the It was shown that DM-d-MD technique is, at least, one order of magnitude faster compared to plain MD [13]. This comparison was done for alanine dipeptide and a 12-aminoacid model system, Ala12. 2) HPC systems used: One of the requirements of ExTASY as that it should be interoperable, so we have used several different HPC systems for our experiments, and characterised the performance of ExTASY on each. Stampede is a Dell Linux cluster located at the Texas Advanced Computing Center, and is part of the Extreme Science and Engineering Discovery Environment (XSEDE). It consists of 6400 compute nodes, each with 2 Intel Xeon ‘Sandy Bridge’ processors, for a total of 16 CPU cores per node, as well as an Intel Xeon Phi co-processor (not used in our experiments). Stampede uses the SLURM batch scheduler for job submission. ARCHER is a Cray XC30 supercomputer hosted by EPCC, and operated on behalf of the Engineering and Physical Sciences Research Council (EPSRC) and the Natural Enviroment Research Council (NERC). It has 4920 compute nodes, each with 2 Intel Xeon ‘Ivy Bridge’ processors, giving 24 cores per node. ARCHER uses the Portable Batch System (PBS) for job submission. Blue Waters is a Cray XE6/XK7 operated by the National Centre for Supercomputing Applications on behalf of the National Science Foundation and the University of Illinois. The XE6 partition used in this work consists of 22640 compute nodes with 2 AMD ‘Interlagos’ processors, giving 32 cores per node. Blue Waters uses the TORQUE/Moab workload manager for job submission. B. The CoCo-MD workflow The CoCo (Complementary Coordinates) technique [29] was designed originally as a method to enhance the diversity of ensembles of molecular structures of the type produced by NMR structure determination. The method involves the use of PCA [30]–[32] in Cartesian space to map the distribution of the ensemble in a low (typically 2-4 dimensional) space, and then the identification of un-sampled regions. CoCo generates new conformations for the molecule that would correspond to these un-sampled regions. The number of new structures generated is under the users control the algorithm divides the space into bins at a chosen resolution, marks bins as sampled or not, first returns a structure corresponding to the centre of the un-sampled bin furthest from any sampled one, marks this bin as now sampled, and iterates as many times as desired. In the CoCo-MD workflow, an ensemble of structures from MD simulations are analysed using the CoCo method; new conformations become the start points for a new round of MD simulations. The latest MD data is added to the previous set, and CoCo repeated. The method is agglomerative - all MD data generated so far is used for each analysis; but also adaptive - a fresh PCA is performed each time. Applied to simulations of the alanine pentapeptide, the CoCo-MD workflow is able to reduce mean first passage times from the extended state to other local minimum states by factors of ten or greater compared to conventional simulations [33]. B. Evaluation of individual components Since the performance of the entire workflow depends on the performance of each of the component parts, we investigate the scaling of both the simulation code (Gromacs) and the analysis tools in isolation on each of the three target platforms, using the NTL9 system used for the full sampling workflows. 1) Simulation tools: The parallel efficiency of Gromacs with respect to a single core on each machine is shown in Figure 3. While efficincies of 69% (ARCHER, 24 cores), 78% (Stampede, 16 cores) and 46% (Blue Waters, 32 cores) suggest that while the scaling for such a relatively small simulation is not ideal, using a single node per simulation is a good use of the available hardware. Beyond a single node, the efficiency drops off so although multiple node simulation tasks are supported by Ensemble Toolkit they are not useful for this benchmark case. 2) Analysis tools: Due to the nature of the two workflows, there are many parallel simulation tasks, but only a single analysis task. Therefore, the analysis task may be configured to run on as many cores as are available to the simulations. Both CoCo and LSDMap are parallelised using MPI, and consist of parts which are independent, e.g., reading of trajectory files in CoCo, and involve communication e.g. diagonalisation of the covariance matrix in CoCo and the diffusion matrix in LSDMap, so the parallel scaling is expected to be sub-linear. The performance of CoCo is also strongly dependent on I/O since it reads the entire trajectory file rather than just the final configurations like LSDMap. V. P ERFORMANCE E VALUATION A. Experiment setup 1) Physical system: The 39-residue mixed α/β protein NTL9(1-39) (pdb code 2HBA, 14,100 atoms including water) is chosen as the physical system for our experiments. NTL9 has an experimentally measured folding time of around 1.5 ms [34], and its folding process has been extensively studied by experiment and all-atom MD simulations, both by means of the Folding@Home distributed computing platform coupled with MSM analysis [9], and by Anton supercomputer [35]. The relatively small size of NTL9, and the existence of previous MD simulation results over long timescales, make this protein an ideal candidate for testing and benchmarking our approach. Albeit small, NTL9 is much larger than a simple peptide, and exhibits a folding process with two competing routes [35], thus presenting a non-trivial test for adaptive sampling. 5 Fig. 6. Measurement of overhead from ExTASY on ARCHER, Blue Waters and Stampede. Using only the simulation stage of the SAL pattern, overhead from ExTASY is measured on the three machines at various core counts. Fig. 3. Gromacs parallel efficiency on ARCHER, Blue Waters and Stampede. A single 20ps gromacs simulation of the NTL9 system is performed using various core counts on the three machines and the execution time is measured. 1) Characterization of overheads: In addition to the necessity of characterizing performance overhead introduced by a new system approach and implementation, in order to instill confidence in potential users of ExTASY it is important to measure the overheads imposed by ExTASY. The objective is to discern the contributions from the different aspects of the ExTASY framework as opposed to MD and analysis components. The time taken to create, launch and terminate all of the simulations or the instantiate ExTASY framework itself on the resource are examples of the former overhead. All of the experiments use Ensemble Toolkit version 0.3.14. We ran a single iteration of the workflow with null workloads, i.e., where the task did no work (/bin/sleep 0), but otherwise was configured as a “normal” simulation task, launched using MPI and taking a whole node on each of the machines. The number of tasks ranged from 16 to 128, and they were all run concurrently. Figure 6 shows that the overheads on Stampede and Blue Waters are relatively small, growing from <5s to around 15s for 128 tasks. On ARCHER the overheads are much larger (70-350s), which after further investigation is due to the aprun job launcher, which performs poorly when multiple concurrent tasks are launched. 2) Strong scaling test: To test the strong scaling of the ExTASY workflows, we fix the number of executing instances (independent MD simulations) to 128 and vary the total number of CPU cores used to execute the workflow. The largest experiments use enough CPU cores that all of the MD simulations execute concurrently. For example, on ARCHER, 128 instances, each executing on a single node (24 cores) gives a maximum of 3072 cores. Since the MD simulations are independent and may be executed concurrently by ExTASY, we expect that as the number of cores is increased, the simulation time should decrease proportionally. The time spent in the analysis part is expected to be constant since the total amount of MD data is constant, and despite the parallelisation of the analysis tools as can be seen from Figure 5, the time to completion plateaus at fairly low core counts. Figure 7 (bottom) shows the results of this experiment for the DM-d-MD workflow executed on Blue Waters. The Fig. 4. Strong scaling of CoCo analysis tool on ARCHER, Blue Waters and Stampede. A total of 256 simulations are analyzed using various core counts on the three machines and execution time is measured. Fig. 5. Parallel efficiency of LSDMap on ARCHER, Blue Waters and Stampede. A total of 2449 structures are analyzed using various core counts on the three machines and execution time is measured. Figures 4 show the strong scaling of CoCo for a fixed input of 256 simulations. We see that CoCo is able to scale to at least 256 cores on ARCHER and Blue Waters, and to around 32 cores on Stampede, thus for our following experiments we configure the workflow to run CoCo with as many cores as there are input trajectories. LSDMap (Figure 5) however, does not scale efficiently much beyond a single node on each machine, even with over 2000 input structures. Nevertheless, we run LSDMap on as many cores as are available, even though it cannot use them fully. Due to the structure of the workflow, those cores would otherwise sit idle during the analysis phase. C. Evaluation of ExTASY 6 Fig. 7. Strong scaling of DM-d-MD workflow on ARCHER (top) and Blue Waters (bottom). The number of simulations is held constant at 128, number of cores per simulation at 24 on ARCHER and 32 on Bluewaters. The total number of cores used is varied with a constant workload, hence measuring strong scaling performance of the framework. Fig. 8. Strong scaling of CoCo-MD workflow on ARCHER (top) and Stampede (bottom). The number of simulations is held constant at 128, number of cores per simulation at 24 on ARCHER and 16 on Stampede. The total number of cores used is varied with a constant workload, hence measuring strong scale performance of the framework. simulation time decreases from 395.7s on 512 cores to 79.52s on 4096 cores, a speedup of 4.97x with 8x as many cores – yielding a scaling efficiency of 62%. The analysis time is essentially constant at around 100s, as expected. The loss of scaling efficiency for the simulation part comes from two sources. Firstly, there is the fixed overhead discussed in Section V-C1 associated with the execution of 128 concurrent tasks, which is approximately 15s. Secondly, the actual computation which occurs within each task take longer when more simulations are run concurrently, due to the fact that they all write to the same shared filesystem. For example, when 16 instances are run concurrently on 512 cores, the MD simulations take an average of 45.6s each. When all 128 instances are run concurrently, each takes 49.0s, or 3.4s slower. If these effects are removed, the effective scaling efficiency on 4096 cores rises to 77%. Similar results are obtained on ARCHER (Figure 7, top), although the scaling of the simulation part tails off on 3072 cores, and the LSDMap analysis takes somewhat longer, with higher variability than Blue Waters. Both of these are due to the fact that both the MD and analysis involve significant I/O, and it is known that opening many small files concurrently is slow as the metadata servers of the parallel Lustre filesystem become a bottleneck [36]. CoCo-MD on Stampede (Figure 8, bottom) has similar strong scaling for the simulation part as DM-d-MD. The simulation time decreases from 363s on 256 cores to 83.7s on 2048 cores – a speedup of 4.3x for a 8-fold increase in the number of cores (54% efficiency). However, the analysis time (CoCo) does not scale due to the fact that the parallelisation in CoCo is limited to the number of input trajectories, which is 128 in this case, even if more cores are available. The CoCo-MD workflow on ARCHER (Figure8, top) does not show as good scaling as DM-d-MD, or CoCo-MD on the other platforms. The reason for this lies in the fact that after the actual molecular dynamics calculation, a ‘trajectory conversion’ step is required to prepare the data for CoCo. This step only takes a fraction of a second to execute, but there is a very large overhead caused by aprun, which allocates resources to and launches each individual task. This does not occur on Blue Waters, which uses the ORTE [37] implementation of RADICAL-Pilot that is not yet default on ARCHER. 3) Weak scaling test: To investigate the weak scaling properties of ExTASY, we fix the ratio of number of instances to CPU cores, and vary the number of instances with the constraint that all simulations can execute concurrently. For example, on ARCHER 16 instances are executed on one node each (24 cores) giving a total of 384 cores, and the number of instances is increased to 128, i.e., the number of cores is 3072. Since all simulations run concurrently, and the length of each simulation does not change, we expect the simulation time to be a constant. However, the analysis part will increase since the performance of the analysis tools is a function both of the input data size (depending on the number of instances), as well as the number of cores available, and even though the number of cores is proportional to the data size, the amount of work grows faster than linearly with the data size. For the DM-d-MD workflow on Blue Waters (Figure 9, top) we observe a small increase of 21.8s in the simulation part as we scale from 512 to 4096 cores. Similar to the strong scaling results, this is combination of the overhead due to the increased number of tasks with a slowdown of the individual tasks themselves. The analysis is found to increase sub-linearly. As discussed in section V-B2 the LSDMap computation consists of parts which are both linear and quadratic in the size of the input data. Combined with an increasing number of cores 7 Fig. 10. Weak scaling of CoCo-MD workflow on ARCHER (top) and Stampede (bottom). The number of cores per simulation is held constant at 24 on ARCHER and 16 on Stampede. The total number of instances is varied from 16-128 and the cores used are increased proportionally. By keeping the ratio of the workload to the number of resources constant, we observe the weak scaling performance of the framework. Fig. 9. Weak scaling of DM-d-MD workflow on Blue Waters (top) and Stampede (bottom). The number of cores per simulation is held constant at 32 on Blue Waters and 16 on Stampede. The total number of simulations is varied from 16-128 and the cores used are increased proportionally. By keeping the ratio of the workload to the number of resources constant, we observe the weak scaling performance of the framework. available for the tool, sub-linear scaling is the result. Similar behaviour is observed on Stampede (Figure 9, bottom), although with a different weighting of the simulation and analysis part, reflecting the fact that the performance of each kernel depends on how well optimised the application binary is on the execution platform. The weak-scaling of CoCo-MD on ARCHER (Figure 10, top) shows very clearly the aprun bottleneck discussion in Section V-C2, and the effect increases as the number of concurrent tasks grows. However, the analysis part scales better than linearly, which is to be expected since CoCo consists of parts which weak scale ideally (independent operations per trajectory file) and parts such as the construction and diagonalisation of the covariance matrix which grow as the data size squared or more. On Stampede, the weak scaling of the simulation part of the CoCo-MD workflow (Figure 10, bottom) is much better than ARCHER. The simulation time grows only by around 50s compared to over 700s on ARCHER over the range of cores that we tested. CoCo scales almost identically to ARCHER. 4) Effect of larger ensembles: To distinguish the effects caused by strong scaling (increasing parallelism with a fixed amount of work) and weak scaling (increasing parallelism proportionally to the amount of work), we also measured the effect of increasing the amount of work with a fixed number of compute cores available. Figure 11 shows the results for the DM-d-MD workflow running on Blue Waters as we vary the number of MD instances from 128 to 1024, keeping the total number of cores available at 4096. Since each task runs on a single node (32 cores per instance), only 128 simulation tasks can run concurrently. Ideally, we would expect that the simulation time should increase linearly with the number of instances. In practice, we see that the time taken grows by Fig. 11. DM-d-MD workflow on Stampede. Workload is increased from 128 instances to 1024, keeping the number of cores constant at 4096. Within the overheads, the increase in execution time is in proportion with increase in the workload. only 7.4x as the number of instances increases from 128 and 1024 i.e., a factor of 8. This is due to the fact that some of the overheads related to managing the tasks that occur before or after execution in the 128 task case are one-time overheads, i.e., those overheads are hidden as they are done concurrently (in the RADICAL-Pilot Agent) with the execution of the remaining tasks when the number of instances is greater than 128. The scaling of the analysis part is consistent with that discussed in Section V-B2, that there is a close to linear scaling since the larger the ensemble size, the more parallelism is available in LSDMap 5) Dynamic simulations: An important characteristic of the LSDMap and CoCo based workflows is that the number of instances typically changes after each simulation-analysis iteration. Thanks to the pilot-abstraction, the ExTASY framework supports flexible mapping between the number of concurrent instances and the total number of cores, while being agnostic of the number of cores per instance. This functionality is used by the DM-d-MD workflow, where, depending on the progress 8 ExTASY is designed and implemented to provide a significant step in this direction. ExTASY allows to simulate many parallel MD trajectories (by means of standard MD engines), to extract long timescale information from the trajectories (by means of different dimensionality reduction methods), and to use the information extracted from the data for adaptively improving sampling. ExTASY has been utilized by different MD engines and analysis algorithms, with only pre-defined and localized changes. In Section III, we formally identified the functional, performance and usability requirements to support the coupling of MD simulations with advanced sampling algorithms. We then presented the design and implementation of ExTASY, an extensible, portable and scalable Python framework for building advanced sampling workflow applications to achieve these requirements. After establishing accurate estimates of the overhead of using ExTASY, Section V consisted of experiments designed to validate the design objectives; we performed experiments that characterized ExTASY along traditional scaling metrics, but also investigated ExTASY beyond single weak and strong scaling performance. With the exception of some machine specific reasons,ExTASY displayed linear scaling for both strong and weak scaling tests on various machines up to O(1000) simulation instances on up to O(1000) nodes for both DM-d-MD and CoCo-MD workflows. In order to keep the footprint of new software small, ExTASY builds upon well-defined and understood abstractions and their efficient and interoperable implementation (Ensemble Toolkit, RADICAL-Pilot). This provides double duty: The core functionality of ExTASY can be provided by simple higher level extensions of complex system software, while allowing it to build upon the performance and optimization of the underlying system software layers. This also allows ExTASY to employ good systems engineering practice: welldefined and good base performance, while being amenable to platform-specific optimizations (e.g. using ORTE on Blue Waters [37]). The design of ExTASY to reuse existing capabilities, for extensibility to different MD codes and sampling algorithms while providing well defined functionality and performance are essential features to ensure the sustainability of ExTASY. Compared to existing software tools and libraries for advanced sampling, ExTASY provides a much more flexible approach that is agnostic of individual tools and compute platforms, is architected to enable efficient and scalable performance and has a simple but general user interface. The ExTASY toolkit is freely available from http://www.extasy-project.org. The ExTASY toolkit has been used to deliver two handson computational science training exercises and tutorials to the bio-molecular simulations community with a focus on advanced sampling. Participants were given the opportunity to utilize HPC systems in real time for advanced sampling problems of their own. Details of both events can be found at http://extasy-project.org/events.html#epccmay2016. A link to the lessons and experience from the first workshop can be found at: https://goo.gl/nMSd27. Fig. 12. Support for dynamic workload in ExTASY: The DM-d-MD algorithm dictates the number of instances at every iteration. The number of instances in each iteration (for a total of 5 iterations) when starting with 32, 64, 128 instances is presented. through the conformation space of the system being explored, LSDMap may decide to spawn more (or less) trajectories for the next iteration of sampling. Figure 12 illustrates this capability. We ran the DM-d-MD workflow on Blue Waters for three configurations with 32, 64 and 128 initial instances. We can see that after an initial growth phase the number of instances seems to stablise for the remaining iterations, although the difference from the starting configuration and the number of iterations taken to stabilise is not algorithmically or systematically predictable. The flexible resource utilization capabilities that ExTASY is built upon prove critical. D. Summary of Experiments We have shown illustrative performance data for two different applications – CoCo-MD and DM-d-MD – based on different analysis methods, on three distinct HPC platforms - ARCHER, Blue Waters and Stampede. The overall scaling to O(1000) simulations is clearly demonstated, and we have analysed the scaling behaviour of the ExTASY framework itself (overheads) and the individual simulation and analysis programs which constitute the workflow. VI. D ISCUSSION AND C ONCLUSION State-of the-art computational biophysics approaches aim to balance three different requirements: force-fields accuracy, advanced sampling capabilities, and rigorous and fast data analysis. These points are strongly interconnected. In particular, it is becoming clear that advanced sampling and data analysis need to be tightly coupled to work efficiently and accurately, as the configurational space that has already been sampled needs to be analyzed on-the-fly to inform on how to proceed with further sampling. Furthermore, many advanced sampling algorithms for biomolecular simulations require flexible and scalable support for multiple simulations. As no final solution yet exists on the best strategy for adaptive sampling (and different physical systems may require a combination of strategies), there is a need to allow combination of different MD engines with different analysis tools, and can be easily extended or modified by end-users to build their own workflows, both for development of new strategies, and for applications to the study of complex biomolecular systems. 9 VII. ACKNOWLEDGMENTS This work was funded by the NSF SSI Awards (CHE-1265788 and CHE1265929) and EPSRC (EP/K039490/1). This work used the ARCHER UK National Supercomputing Service (http://www.archer.ac.uk). We acknowledge access to XSEDE computational facilities via TG-MCB090174 and Blue Waters via NSF-1516469. We gratefully acknowledge the input from various people who have helped the development of the ExTASY workflows: everyone else involved the ExTASY project, the attendees at ExTASY tutorials and beta testing sessions, and particularly David Dotson and Gareth Shannon who provided in-depth comments and suggestions. [16] [17] [18] R EFERENCES [1] “NSF XSEDE Annual Report (2012),” page 382, Figure 32 https://www.xsede.org/documents/10157/169907/2012+Q2+Annual+ Report.pdf. [2] Using XDMoD to facilitate XSEDE operations, planning and analysis XSEDE ’13 Proceedings of the Conference on Extreme Science and Engineering Discovery Environment: Gateway to Discovery, doi 10.1145/2484762.2484763. [3] A. Barducci, M. Bonomi, and M. Parrinello, “Metadynamics,” Wiley Interdisciplinary Reviews: Computational Molecular Science, vol. 1, no. 5, pp. 826–843, 2011. [Online]. Available: http: //dx.doi.org/10.1002/wcms.31 [4] L. C. Pierce, R. Salomon-Ferrer, C. A. F. de Oliveira, J. A. McCammon, and R. C. Walker, “Routine access to millisecond time scale events with accelerated molecular dynamics,” Journal of Chemical Theory and Computation, vol. 8, no. 9, pp. 2997–3002, 2012, pMID: 22984356. [Online]. Available: http://dx.doi.org/10.1021/ct300284c [5] H. Wu, F. Paul, C. Wehmeyer, and F. Noé, “Multi-ensemble Markov models of molecular thermodynamics and kinetics,” Proc. Natl. Acad. Sci. USA, vol. in press, 2016. [6] Y. Sugita and Y. Okamoto, “Replica-exchange molecular dynamics method for protein folding,” Chemical Physics Letters, vol. 314, no. 12, pp. 141 – 151, 1999. [Online]. Available: http://www.sciencedirect. com/science/article/pii/S0009261499011239 [7] J. D. Chodera and F. Noé, “Markov state models of biomolecular conformational dynamics,” Current Opinion in Structural Biology, vol. 25, pp. 135 – 144, 2014, theory and simulation / Macromolecular machines. [Online]. Available: http://www.sciencedirect.com/science/ article/pii/S0959440X14000426 [8] F. Noé, C. Schtte, E. Vanden-Eijnden, L. Reich, and T. R. Weikl, “Constructing the equilibrium ensemble of folding pathways from short off-equilibrium simulations,” Proceedings of the National Academy of Sciences, vol. 106, no. 45, pp. 19 011–19 016, 2009. [Online]. Available: http://www.pnas.org/content/106/45/19011.abstract [9] V. A. Voelz, G. R. Bowman, K. Beauchamp, and V. S. Pande, “Molecular simulation of ab initio protein folding for a millisecond folder NTL9(139),” J. Amer. Chem. Soc., vol. 132, no. 5, pp. 1526–1528, 2010. [10] I. Buch, T. Giorgino, and G. De Fabritiis, “Complete reconstruction of an enzyme-inhibitor binding process by molecular dynamics simulations,” Proc. Natl. Acad. Sci. USA, vol. 108, no. 25, pp. 10 184–10 189, 2011. [Online]. Available: http://www.pnas.org/content/108/25/10184.abstract [11] S. Gu, D.-A. Silva, L. Meng, A. Yue, and X. Huang, “Quantitatively characterizing the ligand binding mechanisms of choline binding protein using markov state model analysis,” PLoS Comput. Biol., vol. 10, no. 8, pp. 1–11, 08 2014. [Online]. Available: http: //dx.doi.org/10.1371%2Fjournal.pcbi.1003767 [12] N. Plattner and F. Noé, “Protein conformational plasticity and complex ligand-binding kinetics explored by atomistic simulations and markov models,” Nat. Commun., vol. 6, 2015. [13] J. Preto and C. Clementi, “Fast recovery of free energy landscapes via diffusion-map-directed molecular dynamics,” Phys. Chem. Chem. Phys., vol. 16, no. 36, p. 19181, 2014. [14] M. A. Rohrdanz, W. Zheng, M. Maggioni, and C. Clementi, “Determination of reaction coordinates via locally scaled diffusion map,” J. Chem. Phys., vol. 134, March 2011. [15] X. D. Guo et al., “Computational studies on self-assembled paclitaxel structures: Templates for hierarchical block copolymer assemblies and sustained drug release,” Biomaterials, vol. 30, no. 33, pp. 6556 – 6563, [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] 10 2009. [Online]. Available: http://www.sciencedirect.com/science/article/ pii/S0142961209008503 R. Harada, Y. Takano, T. Baba, and Y. Shigeta, “Simple, yet powerful methodologies for conformational sampling of proteins,” Phys. Chem. Chem. Phys., vol. 17, pp. 6155–6173, 2015. [Online]. Available: http://dx.doi.org/10.1039/C4CP05262E J. Peng and Z. Zhang, “Simulating large-scale conformational changes of proteins by accelerating collective motions obtained from principal component analysis,” Journal of Chemical Theory and Computation, vol. 10, no. 8, pp. 3449–3458, 2014, pMID: 26588312. [Online]. Available: http://dx.doi.org/10.1021/ct5000988 D. E. Shaw et al., “Anton, a special-purpose machine for molecular dynamics simulation,” Commun. ACM, vol. 51, no. 7, pp. 91–97, Jul. 2008. I. Ohmura et al., “MDGRAPE-4: a special-purpose computer system for molecular dynamics simulations,” Philosophical Transactions of the Royal Society of London A: Mathematical, Physical and Engineering Sciences, vol. 372, no. 2021, 2014. [Online]. Available: http://rsta.royalsocietypublishing.org/content/372/2021/20130387 R. Salomon-Ferrer et al., “An overview of the Amber biomolecular simulation package,” WIREs Comput. Mol. Sci., vol. 3, no. 2, pp. 198– 210, 2013. B. R. Brooks et al., “CHARMM: A program for macromolecular energy, minimization, and dynamics calculations,” J. Comput. Chem., vol. 4, no. 2, pp. 187–217, 1983. M. J. Abraham et al., “GROMACS: High performance molecular simulations through multi-level parallelism from laptops to supercomputers,” SoftwareX, vol. 12, pp. 19 – 25, 2015. S. Plimpton, “Fast parallel algorithms for short-range molecular dynamics,” J. Comp. Phys., vol. 117, no. 1, pp. 1–19, 1995. J. C. Phillips et al., “Scalable molecular dynamics with NAMD,” J. Comput. Chem., vol. 26, no. 16, pp. 1781–1802, 2005. G. A. Tribello, M. Bonomi, D. Branduardi, C. Camilloni, and G. Bussi, “PLUMED 2: New feathers for an old bird,” Comp. Phys. Comm., vol. 185, no. 2, pp. 604–613, 2014. P. Giannozzi et al., “Quantum ESPRESSO: a modular and opensource software project for quantum simulations of materials,” J. Phys. Condens. Matter, vol. 21, no. 39, p. 395502, 2009. V. Balasubramanian, A. Treikalis, O. Weidner, and S. Jha, “Ensemble Toolkit: Scalable and Flexible Execution of Ensembles of Tasks,” 2016, (accepted ICPP 2016) http://arxiv.org/abs/1602.00678. A. Merzky et al., “Executing Dynamic and Heterogeneous Workloads on Super Computers,” 2016, http://arxiv.org/abs/1512.08194. C. A. Laughton, M. Orozco, and W. Vranken, “COCO: A simple tool to enrich the representation of conformational variability in NMR structures,” Proteins, vol. 75, no. 1, pp. 206–216, April 2009. I. T. Jolliffe, Principal component analysis. New York: Springer Verlag, 2002. E. C. Sherer, S. A. Harris, R. Soliva, M. Orozco, and C. A. Laughton, “Molecular dynamics studies of DNA A-tract structure and flexibility.” J Am Chem Soc, vol. 121, pp. 5981–5991, 1999. S. T. Wlodek, T. W. Clark, L. R. Scott, and J. A. McCammon, “Molecular dynamics of acetylcholinesterase dimer complexed with tacrine.” J Am Chem Soc, vol. 119, pp. 9513–9522, 1997. A. Shkurti, E. Rosta, V. Balasubramanian, S. Jha, and C. Laughton, “CoCo-MD: A Simple and Effective Method for the Enhanced Sampling of Conformational Space,” 2016, unpublished, manuscript in preparation. J.-C. Horng, V. Moroz, and D. P. Raleigh, “Rapid cooperative two-state folding of a miniature αβ protein and design of a thermostable variant,” J. Mol. Biol., vol. 326, no. 4, pp. 1261–1270, 2003. K. Lindorff-Larsen, S. Piana, R. O. Dror, and D. E. Shaw, “How fastfolding proteins fold,” Science, vol. 334, no. 6055, pp. 517–520, 2011. D. Henty, A. Jackson, C. Moulinec, and V. Szeremi, “Performance of Parallel IO on ARCHER,” 2015. [Online]. Available: http://archer.ac.uk/ documentation/white-papers/parallelIO/ARCHER wp parallelIO.pdf M. Santcroos et al., “Executing dynamic heterogeneous workloads on Blue Waters with RADICAL-Pilot,” Proceedings of the Cray User Group (CUG) 2016, 2016. [Online]. Available: http://www2.epcc.ed.ac. uk/∼ibethune/files/RP-ORTE-CUG2016.pdf
5cs.CE
Published as a conference paper at ICLR 2017 M ETACONTROL O PTIMIZATION FOR arXiv:1705.02670v1 [cs.LG] 7 May 2017 Jessica B. Hamrick UC Berkeley & DeepMind [email protected] Oriol Vinyals DeepMind [email protected] A DAPTIVE I MAGINATION -BASED Andrew J. Ballard DeepMind [email protected] Nicolas Heess DeepMind [email protected] Razvan Pascanu DeepMind [email protected] Peter W. Battaglia DeepMind [email protected] A BSTRACT Many machine learning systems are built to solve the hardest examples of a particular task, which often makes them large and expensive to run—especially with respect to the easier examples, which might require much less computation. For an agent with a limited computational budget, this “one-size-fits-all” approach may result in the agent wasting valuable computation on easy examples, while not spending enough on hard examples. Rather than learning a single, fixed policy for solving all instances of a task, we introduce a metacontroller which learns to optimize a sequence of “imagined” internal simulations over predictive models of the world in order to construct a more informed, and more economical, solution. The metacontroller component is a model-free reinforcement learning agent, which decides both how many iterations of the optimization procedure to run, as well as which model to consult on each iteration. The models (which we call “experts”) can be state transition models, action-value functions, or any other mechanism that provides information useful for solving the task, and can be learned on-policy or off-policy in parallel with the metacontroller. When the metacontroller, controller, and experts were trained with “interaction networks” (Battaglia et al., 2016) as expert models, our approach was able to solve a challenging decision-making problem under complex non-linear dynamics. The metacontroller learned to adapt the amount of computation it performed to the difficulty of the task, and learned how to choose which experts to consult by factoring in both their reliability and individual computational resource costs. This allowed the metacontroller to achieve a lower overall cost (task loss plus computational cost) than more traditional fixed policy approaches. These results demonstrate that our approach is a powerful framework for using rich forward models for efficient model-based reinforcement learning. 1 I NTRODUCTION While there have been significant recent advances in deep reinforcement learning (Mnih et al., 2015; Silver et al., 2016) and control (Lillicrap et al., 2015; Levine et al., 2016), most efforts train a network that performs a fixed sequence of computations. Here we introduce an alternative in which an agent uses a metacontroller to choose which, and how many, computations to perform. It “imagines” the consequences of potential actions proposed by an actor module, and refines them internally, before executing them in the world. The metacontroller adaptively decides which expert models to use to evaluate candidate actions, and when it is time to stop imagining and act. The learned experts may be state transition models, action-value functions, or any other function that is relevant to the task, and can vary in their accuracy and computational costs. Our metacontroller’s learned policy can exploit the diversity of its pool of experts by trading off between their costs and reliability, allowing it to automatically identify which expert is most worthwhile. 1 Published as a conference paper at ICLR 2017 We draw inspiration from research in cognitive science and neuroscience which has studied how people use a meta-level of reasoning in order to control the use of their internal models and allocation of their computational resources. Evidence suggests that humans rely on rich generative models of the world for planning (Gläscher et al., 2010), control (Wolpert & Kawato, 1998), and reasoning (Hegarty, 2004; Johnson-Laird, 2010; Battaglia et al., 2013), that they adapt the amount of computation they perform with their model to the demands of the task (Hamrick et al., 2015), and that they trade off between multiple strategies of varying quality (Lee et al., 2014; Lieder et al., 2014; Lieder & Griffiths, in revision; Kool et al., in press). Our imagination-based optimization approach is related to classic artificial intelligence research on bounded-rational metareasoning (Horvitz, 1988; Russell & Wefald, 1991; Hay et al., 2012), which formulates a meta-level MDP for selecting computations to perform, where the computations have a known cost. We also build on classic work by Schmidhuber (1990a;b), which used an RL controller with a recurrent neural network (RNN) world model to evaluate and improve upon candidate controls online. Recently Andrychowicz et al. (2016) used a fully differentiable deep network to learn to perform gradient descent optimization, and Tamar et al. (2016) used a convolutional neural network for performing value iteration online in a deep learning setting. In other similar work, Fragkiadaki et al. (2015) made use of “visual imaginations” for action planning. Our work is also related to recent notions of “conditional computation” (Bengio, 2013; Bengio et al., 2015), which adaptively modifies network structure online, and “adaptive computation time” (Graves, 2016) which allows for variable numbers of internal “pondering” iterations to optimize computational cost. Our work’s key contribution is a framework for learning to optimize via a metacontroller which manages an adaptive, imagination-based optimization loop. This represents a hybrid RL system where a model-free metacontroller constructs its decisions using an actor policy to manage model-free and model-based experts. Our experimental results demonstrate that a metacontroller can flexibly allocate its computational resources on a case-by-case basis to achieve greater performance than more rigid fixed policy approaches, using more computation when it is required by a more difficult task. 2 M ODEL We consider a class of fully observed, one-shot decision-making tasks (i.e., continuous, contextual bandits). The performance objective is to find a control c ∈ C which, given an initial state x ∈ X , minimizes some loss function L between a known future goal state x∗ and the result of a forward process, f (x, c). The performance loss LP is the (negative) utility of executing the control in the world, and is related to the optimal solution c∗ ∈ C as follows: LP (x∗ , x, c) = L(x∗ , f (x, c)), c∗ = arg min LP (x∗ , x, c). (1) (2) c However, (2) defines only the optimal solution—not how to achieve it. 2.1 O PTIMIZING P ERFORMANCE We consider an iterative optimization procedure that takes x∗ and x as input and returns an approximation of c∗ in order to minimize (1). The optimization procedure consists of a controller, which iteratively proposes controls, and an expert, which evaluates how good those controls are. On the nth iteration, the controller π C : X × X × H → C takes as input, x∗ , x, and information about the history of previously proposed controls and evaluations hn−1 ∈ H, and returns a proposed control cn that aims to improve on previously proposed controls. An expert E : X × X × C → E takes the proposed control and provides some information en ∈ E about the quality of the control, which we call an opinion. This opinion is added to the history, which is passed back to the controller, and the loop continues for N steps, after which a final control cN is proposed. Standard optimization methods use principled heuristics for proposing controls. In gradient descent, for example, controls are proposed by adjusting cn in the direction of the gradient of the reward with respect to the control. In Bayesian optimization, controls are proposed based on selection criteria such as “probability of improvement”, or a meta-selection criterion for choosing among 2 Published as a conference paper at ICLR 2017 A B Scene ( ) History ( Manager ) Controller Action ( ) Control ( ) Switch World ( ) Expert Expert Opinion ( Outcome ( ) Performance loss ( ) Resource loss ( ) ... ) Memory History ( C Expert Scene ( ) History ( ) ) Metacontroller agent Figure 1: Metacontroller architecture and task. A: All components are part of the metacontroller agent (box) except the scene and the world, which are part of the agent’s environment. The manager takes the scene and history and determines which action to take (i.e., whether to execute or ponder, and with what expert to ponder with), denoted by the orange lines. The controller takes the scene and history and computes a control (e.g., the force to apply to a spaceship), denoted by the blue lines. The orange line ending with a circle at the switch reflects the fact that the manager’s action affects the behavior of the switch, which routes the controller’s control to either an expert (e.g., a simulation model of the spaceship’s trajectory, an action-value function, etc.) or the world. The outcome and reward from the expert, along with the history, action, and control, are fed into the memory, which produces the next history. The history is fed back to the controller on the next iteration in order to allow it to propose controls based on what it has already tried. B-C: Scenes consisted of a number of planets (depicted here by colored circles) of different masses as well as a spaceship (also with a variable mass). The task was to apply a force to the spaceship for one time step of simulation (depicted here as a solid red arrow) such that the resulting trajectory (dotted red arrow) would put the spaceship at a target (bullseye) after 11 steps of simulation. The white ring of the bullseye corresponds to a performance loss of 0.12-0.15, the black ring to a loss of 0.09-0.12, the blue ring to a loss of 0.06-0.09, the red ring to a loss of 0.03-0.06, and the yellow center to a loss of 0.03 or less. B depicts an easy, 1-planet scene, while C depicts a very difficult 5-planet scene. several basic selection criteria Hoffman et al. (2011); Shahriari et al. (2014). Rather than choosing one of several controllers, our work learns a single controller and instead focuses on selecting from multiple experts (see Sec. 2.2). In some cases f is known and inexpensive to compute, and thus the optimization procedure sets E ≡ f . However, in many real-world settings, f is expensive or non-stationary and so it can be advantageous to use an approximation of f (e.g., a state transition model), LP (e.g., an action-value function), or any other quantity that gives some information about f or LP . 2.2 O PTIMIZING C OMPUTATIONAL C OST Given a controller and one or more experts, there are two important decisions to be made. First, how many optimization iterations should be performed? The approximate solution usually improves 3 Published as a conference paper at ICLR 2017 with more iterations, but each iteration costs computational resources. However, most traditional optimizers either ignore the cost of computation or select the number of iterations using simple heuristics. Because they do not balance the cost of computation against the performance loss, the overall effectiveness of these approaches is subject to the skill and preferences of the practitioners who use them. Second, which expert should be used on each step of the optimization? Some experts may be accurate but expensive to compute in terms of time, energy and/or money, while others may be crude, yet cheap. Moreover, the reliability of the experts may not be known a priori, further limiting the effectiveness of the optimization procedure. Our use of a metacontroller address these issues by jointly optimizing over the choices of how many steps to take and which experts to use. We consider a family of optimizers which use the same controller, π C , but vary in their expert evaluators, {E1 , . . . , EK }. Assuming that the controller and experts are deterministic functions, the number of iterations N and the sequences of experts k = (k1 , . . . , kN −1 ) exactly determine the final control and performance loss LP . This means we have transformed the performance optimization over c into an optimization over N and k: (N, k)∗ = arg mink,n LP (x∗ , x, c(N, k, x, x∗ )), where the notation c(N, k, x, x∗ ) is used to emphasize that the control is a function N , k, x, and x∗ . If each optimizer has an associated computational cost τk , then N and k also exactly determine the PN −1 computational resource loss of the optimization run, LR (N, k) = n=1 τkn . The total loss is then the sum of LP and LR , each of which are functions of N and k, LT (x∗ , x, N, k) = LP (x∗ , x, c(N, k, x, x∗ )) + LR (N, k) (3) ∗ C ∗ = L(x , f (x, π (x , x, hN −1 ))) + N −1 X τkn , (4) n=1 and the optimal solution is defined as (N, k)∗ = arg minN,k LT (x∗ , x, N, k). Optimizing LT is difficult because of the recursive dependency on the history, hN −1 , and because the discrete choices of N and k mean LT is not differentiable. To optimize LT we recast it as an RL problem where the objective is to jointly optimize task performance and computational cost. As shown in Figure 1a, the metacontroller agent aM is comprised of a controller π C , a pool of experts {E1 , . . . , EK }, a manager π M , and a memory µ. The manager is a meta-level policy (Russell & Wefald, 1991; Hay et al., 2012) over actions indexed by k, which determine whether to terminate the optimization procedure (k = 0) or to perform another iteration of the optimization procedure with the k th expert. Specifically, on the nth iteration the controller produces a new control cn based on the history of controls, experts, and evaluations. The manager, also relying on this history, independently decides whether to end the optimization procedure (i.e., to execute the control in the world) or to perform another iteration and evaluate the proposed control with the knth expert (i.e., to ponder, after Graves (2016)). The memory then updates the history hn by concatenating k, cn , and en with the previous history hn−1 . Coming back to the notion of imagination-based optimization, we suggest that this iterative optimization process is analogous to imagining what will happen (using one or more approximate world models) before actually executing that action in the world. For further details, see Appendix A, and for an algorithmic illustration of the metacontroller agent, see Algorithm 1 in the appendix. We also define two special cases of the metacontroller for baseline comparisons. The iterative agent aI does not have a manager and uses only a single expert. Its number of iterations are pre-set to a single N . The reactive agent, a0 , is a special case of the iterative agent, where the number of iterations is fixed to N = 0. This implies that proposed controls are executed immediately in the world, and are not evaluated by an expert. For algorithmic illustrations of the iterative and reactive agents, see Algorithms 2 and 3 in the appendix. 2.3 N EURAL N ETWORK I MPLEMENTATION We use standard deep learning building blocks, e.g., multi-layer perceptrons (MLPs), RNNs, etc., to implement the controller, experts, manager, and memory, because they are effective at approximating complex functions via gradient-based and reinforcement learning, but other approaches could be used as well. In particular, we constructed our implementation to be able to make control decisions in complex dynamical systems, such as controlling the movement of a spaceship (Figure 1b-c), though we note that our approach is not limited to such physical reasoning tasks. Here we used mean-squared error (MSE) for our L and Adam (Kingma & Ba, 2014) as the training optimizer. 4 Published as a conference paper at ICLR 2017 Experts We implemented the experts as MLPs and “interaction networks” (INs) (Battaglia et al., 2016), which are well-suited to predicting complex dynamical systems like those in our experiments below. Each expert has parameters θEk , i.e. en = Ek (x∗ , x, cn ; θEk ), and may be trained either on-policy using the outputs of the controller (as is the case in this paper), or off-policy by any data that pairs states and controls with future states or reward outcomes. The objective LEk for each expert may be different depending on what the expert outputs. For example, the objective could be  the loss between the goal and future states, LEk = L f (x, c), Ek (x∗ , x, c; θEk ) , which is what we use in our experiments. Or, it could be the loss between LP and an action-value function that predicts LP directly, LEk = L LP (x∗ , x, c), Ek (x∗ , x, c; θEk ) . See Appendix B.1 for details. Controller and Memory We implemented the controller as an MLP with parameters θC , i.e. cn = π C (x∗ , x, hn−1 ; θC ), and we implemented the memory as a Long Short-Term Memory (LSTM) (Hochreiter & Schmidhuber, 1997) with parameters θµ . The memory embeds the history as a fixedlength vector, i.e. hn = µ(hn−1 , kn , cn , Ekn (x∗ , x, cn ); θµ ). The controller and memory were trained jointly to optimize (1). However, this objective includes f , which is often unknown or not differentiable. We overcame this by approximating LP with a differentiable critic analogous to those used in policy gradient methods (e.g. Silver et al., 2014; Lillicrap et al., 2015; Heess et al., 2015). See Appendices B.2 and B.3 for details. Manager We implemented the manager as a stochastic policy that samples from a categorical distribution whose weights are produced by an MLP with parameters θM , i.e. kn ∼ Categorical(k; π M (x∗ , x, hn−1 ; θM )). We trained the manager to minimize (3) using R EINFORCE (Williams, 1992), but other deep RL algorithms could be used instead. See Appendix B.4 for details. 3 E XPERIMENTS To evaluate our metacontroller agent, we measured its ability to learn to solve a class of physicsbased tasks that are surprisingly challenging. Each episode consisted of a scene which contained a spaceship and multiple planets (Figure 1b-c). The spaceship’s goal was to rendezvous with its mothership near the center of the system in exactly 11 time steps, but it only had enough fuel to fire its thrusters once. The planets were static but the gravitational force they exerted on the spacecraft induced complex non-linear dynamics on the motion over the 11 steps. The spacecraft’s action space was continuous, up to some maximum magnitude, and represented the instantaneous Cartesian velocity vector imparted by its thrusters. Further details are in Appendix C. We trained the reactive, iterative, and metacontroller agents on five versions of the spaceship task involving different numbers of planets.1 The iterative agent was trained to take anywhere from zero (i.e., the reactive agent) to ten ponder steps. The metacontroller was allowed to take a maximum of ten ponder steps. We considered three different experts which were all differentiable: an MLP expert which used an MLP to predict the final location of the spaceship, an IN expert which used an interaction network (Battaglia et al., 2016) to predict the full trajectory of the spaceship, and a true simulation expert which was the same as the world model. In some conditions the metacontroller could use exactly one expert and in others it was allowed to select between the MLP and IN experts. For experiments with the true simulation expert, we used it to backpropagate gradients to the controller and memory. For experiments with an MLP as the only expert, we used a learned IN as the critic. For experiments with an IN as one of its experts, the critic was an IN with shared parameters. We trained the metacontroller on a range of different ponder costs, τk , for the different experts. Further details of the training procedure are available in Appendix D. 3.1 R EACTIVE AND ITERATIVE AGENTS Figure 2 shows the performance on the test set of the reactive and iterative agents for different numbers of ponder steps. The reactive agent performed poorly on the task, especially when the task was more difficult. With the five planets dataset, it was only able to achieve a performance loss of 0.583 on average (see Figure 1 for a depiction of the magnitude of the loss). In contrast, the iterative agent with the true simulation expert performed much better, reaching ceiling performance on the 1 Available from: https://www.github.com/deepmind/spaceship dataset 5 Task loss Published as a conference paper at ICLR 2017 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 MLP expert 0 2 4 6 8 Number of ponder steps 10 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 Int. Net. expert 0 2 4 6 8 Number of ponder steps 10 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 True simulation expert one planet two planets three planets four planets five planets 0 2 4 6 8 Number of ponder steps 10 Figure 2: Test performance of the reactive and iterative agents. Each line corresponds to the performance of an iterative agent (either the true simulation expert, the MLP expert, or the interaction net expert) trained for a fixed number of ponder steps on one of the five datasets; the line color indicates which dataset the controller was trained on. In all cases, performance refers to the performance loss, LP . Left: the MLP expert struggles with the task due to its limited expressivity, but still benefits from pondering. Middle: the IN expert performs almost as well as the true simulation expert, even though it is not a perfect model. Right: The true simulation expert does quite well on the task, especially with multiple ponder steps. datasets with one and two planets, and achieving a performance loss of 0.0683 on the five planets dataset. The IN and MLP experts also improve over the reactive agent, with a minimum performance loss of 0.117 and 0.375 on the five planets dataset, respectively. Figure 2 also highlights how important the choice of expert is. When using the true simulation and IN experts, the iterative agent performs well. With the MLP expert, however, performance is substantially diminished. But despite the poor performance of the MLP expert, there is still some benefit of pondering with it. With even just a few steps, the MLP iterative agent outperforms its reactive counterpart. However comparing the reactive agent with the N = 1 iterative agent is somewhat unfair because the iterative agent has more parameters due to the expert and the memory. However, given that there tends to also be an increase in performance between one and two ponder steps (and beyond), it is clear that pondering—even with a highly inaccurate model—can still lead to better performance than a model-free reactive approach. 3.2 M ETACONTROLLER WITH O NE E XPERT Though the iterative agents achieve impressive results, they expend more computation than necessary. For example, in the one and two planet conditions, the performances of the IN and true simulation iterative agents received little performance benefit from pondering more than two or three steps, while for the four and five planet conditions they required at least five to eight steps before their performance converged. When computational resources have no cost, the number of steps are of no concern, but when they have some cost it is important to be economical. Because the metacontroller learns to choose its number of pondering steps, it can balance its performance loss against the cost of computation. Figure 3 (top row, middle and right subplots) shows that the IN and true simulation expert metacontroller take fewer ponder steps as τ increases, tracking closely the minimum of the iterative agent’s cost curve (i.e., the metacontroller points are always near the iterative agent curves’ minima). This adaptive behavior emerges automatically from the manager’s learned policy, and avoids the need to perform a hyperparameter search to find the best number of iterations for a given τ . The metacontroller does not simply choose an average number of ponder steps to take per episode: it actually tailors this choice to the difficulty of each episode. Figure 4 shows how the number of ponder steps the IN metacontroller chooses in each episode depends on that episode’s difficulty, as measured by the episode’s loss under the reactive agent. For more difficult episodes, the metacontroller tends to take more ponder steps, as indicated by the positive slopes of the best fit lines, and this proportionality persists across the different levels of τ in each subplot. 6 Published as a conference paper at ICLR 2017 Task Loss Total Cost 1.0 MLP expert 1.0 Int. Net. expert 1.0 0.8 0.8 0.8 0.6 0.6 0.6 0.4 0.4 0.4 0.2 0.2 0.2 0.0 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 0.0 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 0.0 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 0 1 2 3 4 5 6 7 8 9 10 Number of ponder steps 0 1 2 3 4 5 6 7 8 9 10 Number of ponder steps True simulation expert 0 1 2 3 4 5 6 7 8 9 10 Number of ponder steps Figure 3: Test performance of the metacontroller with a single expert on the five planets dataset. Each column corresponds to a different experts. The lines indicate the performance of the iterative agents for different numbers of ponder steps. The points indicate the performance of the metacontroller, with each point corresponding to a different value of τ . The x-coordinate of each point is an average across the number of ponder steps, and the y-coordinate is the average loss. Top row: Here we show total cost rather than just performance on the task (i.e., including computation cost). Different colors show the result for different τ , with the different lines showing the cost for the same iterative controller under different values of τ . The error bars (for the metacontroller) indicate 2.5% and 97.5% confidence intervals. When the point is below its corresponding curve, it means that the metacontroller was able to achieve a better speed-accuracy trade-off than that achievable by the iterative agent. Line colors of increasing brightness correspond to increasing τ , with τ values taken from [0, 0.0134, 0.0354, 0.0576, 0.0934, 0.152, 0.246]. Bottom row: Here we show just the performance loss (i.e., without computational cost). Each point corresponds to a different value of τ . The fact that the points are below the curve means the metacontroller agent learns to perform better than the iterative agent with the equivalent number of ponder steps. The ability to adapt its choice of number of ponder steps on a per-episode basis is very valuable because it allows the metacontroller to spend additional computation only on those episodes which require it. The total costs of the IN and true simulation metacontrollers’ are 11% and 15% lower (median) than the best achievable costs of their corresponding iterative agents, respectively, across the range of τ values we tested (see Figure 7 in the Appendix for details). There can even be a benefit to using a metacontroller when there are no computational resource costs. Consider the rightmost points in Figure 3 (bottom row, middle and right subplots), which show the performance loss for the IN and true simulation metacontrollers when τ is low. Remarkably, these points still outperform the best achievable iterative agents. This suggests that there can be an advantage to stopping pondering once a good solution is found, and more generally demonstrates that the metacontroller’s learning process can lead to strategies that are superior to those available to less flexible agents. The metacontroller with the MLP expert had very poor average performance and high variance on the five planet condition (Figure 3, top left subplot), which is why we restricted our focus in this section to how the metacontrollers with IN and true simulation experts behaved. The MLP’s poor performance is crucial, however, for the following section (3.3) which analyzes how a multipleexpert metacontroller manages experts which vary greater in their reliability. 7 Published as a conference paper at ICLR 2017 3.3 M ETACONTROLLER WITH T WO E XPERTS When we allow the manager to additionally choose between two experts, rather than only relying on a single expert, we find a similar pattern of results in terms of the number of ponder steps (Figure 5, left). Additionally, the metacontroller is successfully able to identify the more reliable IN network and consequently uses it a majority of the time, except in a few cases where the cost of the IN network is extremely high relative to the cost of the MLP network (Figure 5, right). This pattern of results makes sense given the good performance (described in the previous section) of the metacontroller with the IN expert compared to the poor performance of the metacontroller with the MLP expert. The manager should not generally rely on the MLP expert because it is simply not a reliable source of information. However, the metacontroller has more difficulty finding an optimal balance between the two experts on a step-by-step basis: the addition of a second expert did not yield much of an improvement over the single-expert metacontroller, with only 9% of the different versions (trained with different τ values for the two experts) achieving a lower loss than the best iterative controller. We believe the mixed performance of the metacontroller with multiple experts is partially due to an entropy term which we used to encourage the manager’s policy to be non-deterministic (see Appendix B.4). In particular, for high values of τ , the optimal thing to do is to always execute immediately without pondering. However, because of the entropy term, the manager is encourage to have a non-deterministic policy and therefore is likely to ponder more than it should—and to use experts that are more unreliable— even when this is suboptimal in terms of the total loss (3). Despite the fact that the metacontroller with multiple experts does not result in a substantial improvement over that which uses a single expert, we emphasize that the manager is able to identify and use the more reliable expert the majority of the time. And, it is still able to choose a variable number of steps according to how difficult the task is (Figure 5, left). This, in and of itself, is an improvement over more traditional optimization methods which would require that the expert is hand-picked ahead of time and that the number of steps are determined heuristically. Low ponder cost (τ = 0. 01) Medium ponder cost (τ = 0. 06) High ponder cost (τ = 0. 25) # Metacontroller ponder steps 10 8 6 4 2 slope = 1.214 [0.992, 1.450] correlation = 0.278 [0.227, 0.326] 0 0 1 2 Reactive controller loss slope = 1.533 [1.316, 1.767] correlation = 0.482 [0.428, 0.536] 3 0 1 2 Reactive controller loss slope = 0.456 [0.398, 0.515] correlation = 0.526 [0.471, 0.581] 3 0 1 2 Reactive controller loss 3 Figure 4: Relationship between the number of ponder steps and per-episode difficulty for the IN metacontroller. Each subplot’s x-axis represents the episode difficulty, as measured by the reactive controller’s loss. Each y-axis represents the number of ponder steps the metacontroller took. The points are individual episodes, and the line is the best fit regression line and 95% confidence intervals. The different subplots show different values of τ (labeled in the title). In each case, there is a clear positive relationship between the difficulty of the task and the number of ponder steps, suggesting that the metacontroller learns to spend more time on hard problems and less time on easier problems. At the bottom of each plot are the fitted slope and correlation coefficient values, along with their 95% confidence intervals in brackets. 8 Published as a conference paper at ICLR 2017 10.0 5.6 4.7 5.8 10.0 10.0 10.0 10.0 4.4 6.2 4.7 5.7 10.0 10.0 6.1 5.6 4.6 6.3 4.1 4.1 4.9 3.1 3.2 3.3 2.9 2.1 2.0 2.2 1.7 1.3 1.5 1.4 2.5 2.3 1.2 1.1 1.1 1.2 1.2 1.2 1.7 1.3 1.3 0.7 1.1 1.1 0.000 0.001 0.007 0.036 0.200 0.300 0.400 Cost of MLP expert Cost of Int. Net. expert 0.300 0.200 0.036 0.007 0.001 6.7 0.000 Fraction of samples using MLP expert 10.0 0.400 0.400 Cost of Int. Net. expert 0.300 0.200 0.036 0.007 0.001 0.000 Total number of samples 10.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.2 0.2 0.0 0.0 0.0 0.0 0.2 0.2 0.1 0.0 0.0 0.0 0.0 0.8 0.4 0.0 0.0 0.0 0.0 0.0 0.1 0.9 0.2 1.0 0.0 0.0 0.0 0.000 0.001 0.007 0.036 0.200 0.300 0.400 Cost of MLP expert Figure 5: Test performance of the metacontroller with multiple experts on the five planets dataset. Left: The average number of total ponder steps, for different values of τ . As with the single-expert metacontrollers, fewer ponder steps are taken when the cost is very high, and more are taken when the cost is low. Right: The fraction of ponder steps taken by the MLP expert relative to the IN expert. In the majority of cases, the metacontroller favors using the IN expert as it is much more reliable. The few exceptions (red squares) are cases when the cost of the IN expert is much higher relative to the cost of the MLP expert. 4 D ISCUSSION In this paper, we have presented an approach to adaptive, imagination-based optimization in neural networks. Our approach is able to flexibly choose which computations to perform as well as how many computations need to be performed, approximately solving a speed-accuracy trade-off that depends on the difficulty of the task. In this way, our approach learns to rely on whatever source of information is most useful and most efficient. Additionally, by consulting the experts on-the-fly, our approach allows agents to test out actions to ensure that their consequences are not disastrous before actually executing them. While the experiments in this paper involve a one-shot decision task, our approach lays a foundation that can be built upon to support more complex situations. For example, rather than applying a force only on the first time step, we could turn the problem into one of trajectory optimization for continuous control by asking the controller to produce a sequence of forces. In the case of planning, our approach could potentially be combined with methods like Monte Carlo Tree-Search (MCTS) (Coulom, 2006), where our experts would be akin to having several different rollout policies to choose from, and our controller would be akin to the tree policy. While most MCTS implementations will run rollouts until a fixed amount of time has passed, our approach would allow the manager to adaptively choose the number of rollouts to perform and which policies to perform the rollouts with. Our method could also be used to naturally augment existing model-free approaches such as DQN (Mnih et al., 2015) with online model-based optimization by using the model-free policy as a controller and adding additional experts in the form of state-transition models. An interesting extension would be to compare our metacontroller architecture with a naı̈ve model-based controller that performs gradient-based optimization to produce the final control. We expect our metacontroller architecture might require fewer model evaluations and to be more robust to model inaccuracies compared to the gradient-based method, because our method has access to the full history of proposed controls and evaluations whereas traditional gradient-based methods do not. Although we rely on differentiable experts in our metacontroller architecture, we do not utilize the gradient information from these experts. An interesting extension to our work would be to pass this gradient information through to the manager and controller (as in Andrychowicz et al. (2016)), which would likely improve performance further, especially in the more complex situations discussed here. Another possibility is to train some or all of the experts inline with the controller and metacontroller, rather than independently, which could allow their learned functionality to be more tightly integrated with the rest of the optimization loop, at the expense of their generality and ability to be repurposed for other uses. To conclude, we have demonstrated how neural network-based agents can use metareasoning to adaptively choose what to think about, how to think about it, and for how long to think for. Our 9 Published as a conference paper at ICLR 2017 method is directly inspired by human cognition and suggests a way to make agents much more flexible and adaptive than they currently are, both in decision making tasks such as the one described here, as well as in planning and control settings more broadly. ACKNOWLEDGMENTS We would like to thank Matt Hoffman, Andrea Tacchetti, Tom Erez, Nando de Freitas, Guillaume Desjardins, Joseph Modayil, Hubert Soyer, Alex Graves, David Reichert, Theo Weber, Jon Scholz, Will Dabney, and others on the DeepMind team for helpful discussions and feedback. R EFERENCES Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, and Nando de Freitas. Learning to learn by gradient descent by gradient descent. arXiv:1606.04474, 2016. Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, and Koray Kavukcuoglu. Interaction networks for learning about objects, relations and physics. Advances in Neural Information Processing Systems, 2016. Peter W. Battaglia, Jessica B. Hamrick, and Joshua B. Tenenbaum. Simulation as an engine of physical scene understanding. Proceedings of the National Academy of Sciences, 110(45):18327–18332, 2013. Emmanuel Bengio, Pierre-Luc Bacon, Joelle Pineau, and Doina Precup. Conditional computation in neural networks for faster models. arXiv:1511.06297, 2015. Yoshua Bengio. Deep learning of representations: Looking forward. arXiv:1305.0445, 2013. Rémi Coulom. Efficient selectivity and backup operators in monte-carlo tree search. In International Conference on Computers and Games, pp. 72–83. Springer, 2006. Katerina Fragkiadaki, Pulkit Agrawal, Sergey Levine, and Jitendra Malik. Learning Visual Predictive Models of Physics for Playing Billiards. Proceedings of the International Conference on Learning Representations (ICLR 2016), pp. 1–12, 2015. URL http://arxiv.org/abs/1511.07404. Jan Gläscher, Nathaniel Daw, Peter Dayan, and John P. O’Doherty. States versus rewards: Dissociable neural prediction error signals underlying model-based and model-free reinforcement learning. Neuron, 66(4):585 – 595, 2010. Alex Graves. Adaptive computation time for recurrent neural networks. arXiv:1603.08983, 2016. Jessica B. Hamrick, Kevin A. Smith, Thomas L. Griffiths, and Edward Vul. Think again? the amount of mental simulation tracks uncertainty in the outcome. In Proceedings of the 37th Annual Conference of the Cognitive Science Society, 2015. Nicholas Hay, Stuart J. Russell, David Tolpin, and Solomon Eyal Shimony. Selecting computations: Theory and applications. Proceedings of the 28th Conference on Uncertainty in Artificial Intelligence, 2012. Nicolas Heess, Gregory Wayne, David Silver, Tim Lillicrap, Tom Erez, and Yuval Tassa. Learning continuous control policies by stochastic value gradients. Advances in Neural Information Processing Systems, 2015. Mary Hegarty. Mechanical reasoning by mental simulation. Trends in Cognitive Sciences, 8(6):280 – 285, 2004. Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997. Matthew W Hoffman, Eric Brochu, and Nando de Freitas. Portfolio allocation for Bayesian optimization. In Proceedings of the 27th Conference on Uncertainty in Artificial Intelligence, pp. 327–336, 2011. Eric J. Horvitz. Reasoning about beliefs and actions under computational resource constraints. In Uncertainty in Artificial Intelligence, Vol. 3, 1988. Philip N Johnson-Laird. Mental models and human reasoning. Proceedings of the National Academy of Sciences, 107(43):18243–18250, 2010. Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv:1412.6980, 2014. Wouter Kool, Fiery A. Cushman, and Samuel J. Gershman. When does model-based control pay off? PLOS Computational Biology, in press. Sang Wan Lee, Shinsuke Shimojo, and John P. O’Doherty. Neural computations underlying arbitration between model-based and model-free learning. Neuron, 81:687–699, 2014. Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. Journal of Machine Learning Research, 17:1–40, 2016. Falk Lieder and Thomas L. Griffiths. Strategy selection as rational metareasoning. in revision. Falk Lieder, Dillon Plunkett, Jessica B. Hamrick, Stuart J. Russell, Nicholas J. Hay, and Thomas L. Griffiths. Algorithm selection by rational metareasoning as a model of human strategy selection. 27:2870–2878, 2014. 10 Published as a conference paper at ICLR 2017 Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv:1509.02971, 2015. Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015. Stuart Russell and Eric Wefald. Principles of metareasoning. Artificial Intelligence, 49(1):361 – 395, 1991. Jürgen Schmidhuber. An on-line algorithm for dynamic reinforcement learning and planning in reactive environments. In Proceedings of the International Joint Conference on Neural Networks (IJCNN), 1990a. Jürgen Schmidhuber. Reinforcement learning in Markovian and non-Markovian environments. Advances in Neural Information Processing Systems, 1990b. Bobak Shahriari, Ziyu Wang, Matthew W Hoffman, Alexandre Bouchard-Côté, and Nando de Freitas. An entropy search portfolio for Bayesian optimization. arXiv:1406.4625, 2014. David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. Deterministic policy gradient algorithms. Proceedings of the 31st International Conference on Machine Learning, 2014. David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016. Aviv Tamar, Sergey Levine, and Pieter Abbeel. Value Iteration Networks. Advances in Neural Information Processing Systems, 2016. URL http://arxiv.org/abs/1602.02867. Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8(3-4):229–256, 1992. D.M. Wolpert and M. Kawato. Multiple paired forward and inverse models for motor control. Neural Networks, 11(78):1317 – 1329, 1998. 11 Published as a conference paper at ICLR 2017 A M ETACONTROLLER D ETAILS Here, we give the precise definitions of the metacontroller agent. As described in the main text, the iterative and reactive agents are special cases of the metacontroller agent, and are therefore not discussed here. The metacontroller agent aM is comprised of the following components: • A history-sensitive controller, π C : X × X × H → C, which is a policy that maps goal and initial states, and a history, h ∈ H, to controls, whose aim is to minimize (1). • A pool of experts {E1 , . . . , EK }. Each expert E : X × X × C → E maps goal states, input states, and actions to opinions. Opinions can be either states-only (E = X ), states and rewards (E = X × R), or rewards-only (E = R). The expert corresponds to the evaluator for the optimization routine, i.e., an approximation of the forward process f . • A manager, π M : X × X × Hn → {0, . . . , K}, which is a policy which decides whether to send a proposed control to the world (k = 0) or to the k th expert for evaluation, in order to minimize (3). This formulation is based on that used by metareasoning systems (Russell & Wefald, 1991; Hay et al., 2012). Details on the corresponding MDP are given in Appendix A.1. • A memory, µ : Hn−1 × Z → Hn , which is a function that maps the prior history hn−1 ∈ Hn−1 , as well as the most recent manager choice, proposed control, and expert evaluation (k, c, e) ∈ {0, . . . , K} × C × E = Z, to an updated history hn ∈ Hn , which is then made available to the manager and controller on subsequent iterations. The history at step n is a recursively defined tuple which is the concatenation of the prior history with the most recently proposed control, expert evaluation, and expert identity: hn = hn−1 ∩ ((kn , cn , Ekn (x∗ , x, cn ))) = ((k1 , c1 , Ek1 (x∗ , x, c1 )), . . . , (kn , cn , Ekn (x∗ , x, cn ))) where h0 = () represents an empty initial history. Similarly, the finite set of histories up to step n is: Hn = Hn−1 × Z = Z n where H0 = {()}. The metacontroller produces: aM (x∗ , x) = π C (x∗ , x, hN −1 ) = cN (5) where N = n s.t. kn = 0. This function is summarized in Algorithm 1. The other agents (iterative and reactive), as mentioned in the main text, are simpler versions of the metacontroller agent and are summarized in Algorithms 2 and 3. A.1 M ETA -L EVEL MDP To implement the manager for the metacontroller agent, we draw inspiration from the metareasoning literature (Russell & Wefald, 1991; Hay et al., 2012) and formulate the problem as a finite-horizon Markov Decision Process (MDP) hS, A, P, Ri over the decision of whether to perform another iteration of the optimization procedure or to execute a control in the world. • The state space S consists of goal states, external states, and internal histories, S = X × X × H. • The action space A contains K + 1 discrete actions, {0, . . . , K}, which correspond to execute (k = 0) and ponder (k ∈ {1, . . . , K}), where ponder (after Graves (2016)) refers to performing an iteration of the optimization procedure with the k th expert. • The (deterministic) state transition model P : S × C × S → [0, 1] is,  P (x0 |x∗ , x, hn−1 , k) if k = 0 0 ∗ P (x , hn |x , x, hn−1 , k) = ∗ P (hn |x , x, hn−1 , k) otherwise where x0 = f (x, c) and c = π C (x∗ , x, hn−1 ) and,  1 if x0 = f (x, c) 0 ∗ P (x |x , x, hn−1 , k) = 0 otherwise  1 if hn = hn−1 ∪ {(k, c, Ek (x∗ , x, c))} P (hn |x∗ , x, hn−1 , k) = 0 otherwise 12 Published as a conference paper at ICLR 2017 Algorithm 1 Metacontroller agent. x is the scene and x∗ is the target. 1: function aM (x, x∗ ) 2: h0 ← () . Initial empty history 3: k0 ← π M (x, x∗ , h0 ) . Get an action from the manager 4: c0 ← π C (x, x∗ , h0 ) . Propose a control with the controller 5: n←0 6: while kn 6= 0 do . When k 6= 0, ponder with an expert 7: en ← Ekn (x, x∗ , cn ) . Get an expert’s opinion 8: hn+1 ← µ(hn , kn , cn , en ) . Update the history 9: n←n+1 10: kn ← π M (x, x∗ , hn ) . Choose the next action 11: cn ← π C (x, x∗ , hn ) . Propose the next control 12: end while 13: return cn 14: end function Algorithm 2 Iterative agent. x is the scene, x∗ is the target, and N is the number of ponder steps. 1: function aI (x, x∗ , N ) 2: h0 ← () . Initial empty history 3: c0 ← π C (x, x∗ , h0 ) . Propose a control with the controller 4: n←0 5: while n < N do . Ponder with an expert for N steps 6: en ← E(x, x∗ , cn ) . Get the expert’s opinion 7: hn+1 ← µ(hn , kn , cn , en ) . Update the history 8: n←n+1 9: cn ← π C (x, x∗ , hn ) . Propose the next control 10: end while 11: return cn 12: end function Algorithm 3 Reactive agent. x is the scene and x∗ is the target. 1: function a0 (x, x∗ ) 2: c0 ← π C (x, x∗ , ()) . Propose a control with the controller 3: return c0 4: end function • The (deterministic) reward function R : S × A × S → R maps the current state, current action, and next state to real-valued loss: ∗ 0 R(x , x, hn−1 , k, x ) =  L(x∗ , x0 ) τk if k = 0 (see Eq. 1) otherwise (see Eq. 3)  where x0 = f x, π C (x∗ , x, hn−1 ) . We approximate the solution to this MDP with a stochastic manager policy π M . The manager chooses actions proportional to the immediate reward for taking action k in state sn plus the expected sum of future rewards. This construction imposes a trade-off between accuracy and resources, incentivizing the agent to ponder longer and with more accurate (and potentially expensive) experts when the problem is harder. 13 Published as a conference paper at ICLR 2017 A B History (n-1) Scene Manager Controller REINFORCE Control Action History (n-1) Scene Manager Controller Control Action Predicted performance loss Switch Switch Critic World Expert 1 Expert 2 ... Expert K World Expert 1 Opinion Expert 2 ... Expert K Opinion Performance loss Resource loss Memory Memory Scene Scene History (n-1) History (n-1) History (n) History (n) C D History (n-1) Scene Manager Scene Manager Controller Control Action History (n-1) Predicted performance loss Controller Action Control Switch Switch Critic World Expert 1 Expert 2 ... Expert K World Expert 1 Expert 2 ... Expert K Opinion Outcome Performance loss Outcome Performance loss Opinions Memory Memory Scene Scene History (n-1) History (n-1) History (n) History (n) Figure 6: Training each part of the network. In each subplot, red arrows depict gradients. Dotted arrows indicate backward connections that are not part of the forward pass. Colored nodes indicate weights that are being updated. All backpropagation occurs at the very end of a full forward pass (i.e., after the control has been executed in the world). A: Training the controller and memory with backpropagation-through-time (BPTT), beginning with the critic, and flowing to the controller, through the memory, through the relevant expert, through the controller again, and so on. B: Training the manager using R EINFORCE (Williams, 1992). C: Training the experts (note that each expert may have a different loss with respect to the outcome from the world). D: Training the critic. B B.1 G RADIENTS E XPERTS Training the experts is a straightforward supervised learning problem (Figure 6c). The gradient is: ∂LEk ∂LEk ∂Ek = , E k ∂θ ∂Ek ∂θEk 14 (6) Published as a conference paper at ICLR 2017 where Ek is the k th expert and LEk is the loss function for the k th expert. For example, in the case of an action-value function expert, this loss function might be LEk (f, Ek ) = L(x∗ , f (x, c)) − Ek (x∗ , x, c; θEk ) 2 . In the case of an expert that predicts the final state using a model of the system dynamics, it might be LEk (f, Ek ) = f (x, c) − Ek (x∗ , x, c; θEk ) 2 . B.2 C RITIC The critic, L̂P , is an approximate model of the performance loss, LP , (1), which is used to backpropagate gradients to the controller and memory. This means the critic can either be an action-value function, which approximates L̂P = E0 ≈ LP directly, or a model of the system dynamics composed with a known loss function between the goal and future states, L̂P = L◦E0 ≈ L◦f . We train the critic, E0 : X × X × C → R, using the same procedure as the experts are trained (Figure 6d). A good expert may even be used as the critic. B.3 C ONTROLLER AND M EMORY As shown in Figure 6a, we trained the controller and memory using backpropagation through time (BPTT) with an actor-critic architecture. Specifically, rather than assuming f is known and differentiable, we use a critic and backpropagate through it (Heess et al., 2015): ∂L ∂L ∂E∗ ∂πnC ∂ + µn ∂L ∂L ∂E∗ ∂πnC ∂ + µn ∂π0C ∂µ0 = , = (7) ··· C ··· µ C C C µ ∂θ ∂E∗ ∂πn ∂µn ∂πn−1 ∂θ ∂θ ∂E∗ ∂πnC ∂µn ∂µn−1 ∂θ where E∗ is the critic, n is the maximum number of iterations the controller can use, and: ∂ + µn ∂µn ∂Ekn−1 ∂µn ∂ + µn ∂µn ∂ + µn = + , + = (8) C C C C ∂E ∂µ ∂µ ∂πn−1 ∂πn−1 ∂πn−1 kn−1 ∂πn−1 n−1 n−1 where we are using the ∂ + notation to indicate summed gradients, following Pascanu et al. (2013). Since kn has already been produced by the manager it can be treated as a constant and will produce an unbiased estimate of the gradient. This is convenient because it allows for training the controller and manager separately, or testing the controller’s behavior with arbitrary actions post-training. B.4 M ANAGER As discussed in the main text, we used the R EINFORCE algorithm Williams (1992) to train the manager (Figure 6b). One potential issue, however, is that when training the controller and manager simultaneously, the controller will result in high cost early on in training and thus the manager will learn to always choose the execute action. To discourage the manager from learning what is an essentially deterministic policy, we included a regularization term based on the entropy, LH (Williams & Peng, 1991; Mnih et al., 2016): LH ( · ; θM ) = λ EπM [log π M ( · ; θM )]  ∂ ∂EπM [r] = r − LH ( · ; θM ) log π M ( · ; θM ), ∂θM ∂θM r is the full return given by (3) and λ is the strength of the regularization term. C C.1 S PACESHIP TASK DATASETS We generated five datasets, each containing scenes with a different number of planets (ranging from a single planet to five planets). Each dataset consisted of 100,000 training scenes and 1,000 testing scenes. The target in each scene was always located at the origin, and each scene always had a sun with a mass of 100 units. The sun was located between 100 and 200 distance units away from the target, with this distance sampled uniformly at random. The other planets had a mass between 20 and 50 units, and were located 100 to 250 distance units away from the target, sampled uniformly at random. The spaceship had a mass between 1 and 9 units, and was located 150 to 250 distance units away from the target. The planets were always fixed (i.e., they could not move), and the spaceship always started at the beginning of each episode with zero velocity. 15 Published as a conference paper at ICLR 2017 C.2 E NVIRONMENT We simulated our scenes using a physical simulation of gravitational dynamics. The planets were always stationary (i.e., they were not acted upon by any of the objects in the scene) but acted upon the spaceship with a force of: mp ms Fp = G 3 (xp − xs ), (9) r where Fp is the force vector of the planet on the spaceship, G = 1000000 is a gravitational constant, mp is the mass of the planet, ms is the mass of the spaceship, r is the distance between the centers of masses of the planet and the spaceship, xp is the location of the planet, and xs is the location of the spaceship. We simulated this environment using the Euler method, i.e.: P ( p Fp ) − dvs + c as = x0s = xs + vs vs0 = vs + as (10) ms where as , vs , and xs are the acceleration, velocity, and position of the spaceship, respectively; d = 0.1 is a damping constant; c is the control force applied to the spaceship; and  is the step size. Note that we set c to zero for all timesteps except the first. D I MPLEMENTATION D ETAILS We used TensorFlow (Abadi et al., 2015) to implement and train all versions of the model. D.1 A RCHITECTURE In our implementation of the controller, we used a two-layer MLP each with 100 units. The first layer used ReLU activations and the second layer used a multiplicative interaction similar to van den Oord et al. (2016), which we found to work better in practice. In our implementation of the memory, we used a single LSTM layer of size 100. In our implementation of the manager, we used a MLP of two fully connected layers of 100 units each, with ReLU nonlinearities. We constructed three different experts to test the various controllers. The true simulation expert was the same as the world model, and consisted of a simulation for 11 timesteps with  = 0.05 (see Appendix C). The IN expert was an interaction network (Battaglia et al., 2016), which has previously been shown to be able to learn to predict n-body dynamics accurately for simple systems. The IN consists of a relational module and an object module. In our case, the relational module was composed of 4 hidden layers of 150 nodes each, outputting “effects” encodings of size 100. These effects, together with the relational model input are then used as input to the object model, which contained a single hidden layer of 100 nodes. The object model outputs the velocity of the spaceship and we trained it to predict the velocity on every timestep of the spaceship’s trajectory. The MLP expert was a MLP that predicted the final location of the spaceship and had the same architecture as the controller. As discussed in Appendix B, we used a critic to train the controller and memory. We always used the IN expert as the critic, except in the case when the true simulation expert was used, in which case we also used the true simulation as the critic. D.2 T RAINING P ROCEDURE All weights were initialized uniformly at random between 0 and 0.01. An iteration of training consisted of gradient updates over a minibatch of size 1000; in total, we ran training for 100,000 iterations. We additionally used a waterfall schedule for each of the learning rates during training, such that after 1000 iterations, if the loss was not decreasing, we would decay the step size by 5%. We trained the controller and memory together using the Adam optimizer (Kingma & Ba, 2014) with gradients clipped to a maximum global norm of 10 (Pascanu et al., 2013). The manager was trained simultaneously, but using a different learning rate than the controller and memory. The IN and MLP experts were also trained simultaneously, but again with different learning rates. Learning rates were determined using a grid search over a small number of values, and are given in Table 1 for the iterative agent, in Table 2 for the metacontroller with one expert, and in Table 3 for the metacontroller with two experts. 16 Published as a conference paper at ICLR 2017 The iterative agent was trained to take a fixed number of ponder steps, ranging from 0 (i.e., the reactive agent) to 10. The metacontrollers were allowed to take a variable number of ponder steps up to a maximum of 10. For the metacontroller with a single expert, we trained the manager using τ = 0 and 20 additional values of τ spaced logarithmically between 0.00004 and 0.4 (inclusive). For the metacontroller with multiple experts, we trained the manager on a grid of pairs of τ values, where each expert could have τ = 0 or one of 6 values spaced logarithmically between 0.00004 and 0.2 (inclusive). In all cases, the entropy penalty for the metacontroller was λ = 0.2. D.3 C ONVERGENCE Reactive agent. datasets. Training for the reactive agents was straightforward and converged reliably on all Iterative agent. For the iterative agent with the interaction network or true simulation experts, convergence was also reliable for small numbers of ponder steps. Convergence was somewhat less reliable for larger numbers of ponder steps. We believe this is because for some scenes, a larger number of ponder steps was more than necessary to solve the task (as is evidenced by the plateauing performance in Figure 2). So, the iterative agent had to effectively “remember” what the best control was while it took the last few ponder steps, which is a more complicated and difficult task to perform. For the iterative agent with the MLP expert, convergence was more variable especially when the task was harder, as can be seen in the variable performance on the five planets dataset in Figure 2 (left). We believe this is because the MLP agent was so poor, and that convergence would have been more reliable with a better agent. Metacontroller with a single expert. The metacontroller agent with a single expert converged more reliably than the corresponding iterative agent (see the bottom row of Figure 3). As mentioned in the previous paragraph, the iterative agent had to take more steps than actually necessary, causing it to perform less well for larger numbers of ponder steps, whereas the metacontroller agent had the flexibility of stopping when it had found a good control. On the other hand, we found that the metacontroller agent sometimes performed too many ponder steps for large values of τ (see Figures 3 and 7). We believe this is due to the entropy term (λ) added to the R EINFORCE loss. This is because when then ponder cost is very high, the optimal thing to do is to behave deterministically and always execute (never ponder); however, the entropy term encouraged the policy to be nondeterministic. We plan to explore different training regimes in future work to alleviate this problem, for example by annealing the entropy term to zero over the course of training. Metacontroller with multiple experts. The metacontroller agent with multiple experts was somewhat more difficult to train, especially for high ponder cost of the interaction network expert. For example, note how the proportion of steps using the MLP expert does not decrease monotonically in Figure 5 (right) with increasing cost for the MLP expert. We believe this is also an unexpected result of using the entropy term: in all of these cases, the optimal thing to do actually is to rely on the MLP expert 100% of the time, yet the entropy term encourages the policy to be non-deterministic. Future work will explore these difficulties further by using experts that complement each other better (i.e., so there is not one that is wholly better than the other). Experts. The experts themselves always converged quickly and reliably, and trained much faster than the rest of the network. R EFERENCES Martı́n Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, et al. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. URL http://tensorflow.org/. Software available from tensorflow.org. Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, and Koray Kavukcuoglu. Interaction networks for learning about objects, relations and physics. Advances in Neural Information Processing Systems, 2016. Alex Graves. Adaptive computation time for recurrent neural networks. arXiv:1603.08983, 2016. 17 Published as a conference paper at ICLR 2017 True sim. MLP IN Dataset # Ponder Steps αc αc αEIN αEMLP αc αEIN one planet one planet one planet one planet one planet one planet one planet one planet one planet one planet one planet 0 1 2 3 4 5 6 7 8 9 10 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 5e-04 1e-03 5e-04 5e-04 1e-03 5e-04 5e-04 5e-04 1e-03 5e-04 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 5e-04 1e-03 1e-03 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 two planets two planets two planets two planets two planets two planets two planets two planets two planets two planets two planets 0 1 2 3 4 5 6 7 8 9 10 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 5e-04 1e-03 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 3e-03 3e-03 3e-03 3e-03 3e-03 1e-03 3e-03 3e-03 3e-03 3e-03 3e-03 1e-03 5e-04 5e-04 5e-04 1e-03 1e-03 1e-03 5e-04 5e-04 5e-04 1e-03 3e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 5e-04 5e-04 3e-03 5e-04 3e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 3e-03 1e-03 three planets three planets three planets three planets three planets three planets three planets three planets three planets three planets three planets 0 1 2 3 4 5 6 7 8 9 10 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 5e-04 1e-03 1e-03 1e-03 5e-04 1e-03 1e-03 1e-03 5e-04 3e-03 3e-03 3e-03 1e-03 3e-03 1e-03 3e-03 3e-03 3e-03 3e-03 3e-03 1e-03 1e-03 1e-03 5e-04 5e-04 5e-04 5e-04 1e-03 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 5e-04 1e-03 1e-03 5e-04 1e-03 1e-03 3e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 four planets four planets four planets four planets four planets four planets four planets four planets four planets four planets four planets 0 1 2 3 4 5 6 7 8 9 10 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 5e-04 5e-04 1e-03 1e-03 5e-04 5e-04 5e-04 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 1e-03 3e-03 3e-03 3e-03 5e-04 1e-03 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 5e-04 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 five planets five planets five planets five planets five planets five planets five planets five planets five planets five planets five planets 0 1 2 3 4 5 6 7 8 9 10 1e-03 1e-03 5e-04 1e-03 5e-04 1e-03 1e-03 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 5e-04 5e-04 5e-04 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 3e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 3e-03 3e-03 1e-03 1e-03 Table 1: Hyperparameter values for the iterative controller. αc refers to the learning rate for the controller and memory, while αEIN refers to the learning rate for the IN expert, and αEMLP refers to the learning rate for the MLP expert. 18 Published as a conference paper at ICLR 2017 0.8 Cost of managed controller 0.7 True simulation expert Interaction network expert 0.6 0.5 0.4 0.3 0.2 0.1 0.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 Cost of best iterative controller 0.7 0.8 Figure 7: Cost of the best iterative controller compared to the managed controller. Each point represents the total cost of the best iterative agent under a particular value of τ (x-axis) versus the total cost achieved by the metacontroller trained with the same value of τ (y-axis). The best iterative agent was chosen by computing the cost for all the different number of ponder steps, and then choosing the whichever number of ponder stpes yielded the lowest cost (i.e., finding the minimum of the curves in Figure 3, top row). In almost all cases, the managed controller achieves a lower loss than the iterative controller: for the metacontroller with the IN expert, the cost is 11% lower than the iterative controller on average, and for the metacontroller with the true simulation expert, it is 15% lower on average. Nicholas Hay, Stuart J. Russell, David Tolpin, and Solomon Eyal Shimony. Selecting computations: Theory and applications. Proceedings of the 28th Conference on Uncertainty in Artificial Intelligence, 2012. Nicolas Heess, Gregory Wayne, David Silver, Tim Lillicrap, Tom Erez, and Yuval Tassa. Learning continuous control policies by stochastic value gradients. Advances in Neural Information Processing Systems, 2015. Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv:1412.6980, 2014. Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timothy P. Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. Proceedings of the 33rd International Conference on Machine Learning, 2016. Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural networks. Proceedings of the 27st International Conference on Machine Learning, pp. 1310–1318, 2013. Stuart Russell and Eric Wefald. Principles of metareasoning. Artificial Intelligence, 49(1):361 – 395, 1991. Aäron van den Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, and Koray Kavukcuoglu. Conditional image generation with PixelCNN decoders. arXiv:1606.05328, 2016. Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8(3-4):229–256, 1992. Ronald J. Williams and Jing Peng. Function optimization using connectionist reinforcement learning algorithms. Connection Science, 3(3):241–268, 1991. 19 Published as a conference paper at ICLR 2017 True sim. MLP IN τ αc αm αc αm αEIN αEMLP αc αm αEIN 0.00000 0.00004 0.00006 0.00011 0.00017 0.00028 0.00045 0.00073 0.00119 0.00193 0.00314 0.00510 0.00828 0.01344 0.02182 0.03543 0.05754 0.09343 0.15171 0.24634 0.40000 5e-04 1e-03 5e-04 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 5e-04 1e-04 5e-05 1e-04 1e-04 1e-03 1e-03 1e-04 5e-05 5e-05 1e-04 5e-05 5e-04 5e-05 1e-04 1e-04 5e-04 5e-05 1e-04 5e-05 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 5e-05 5e-04 1e-04 1e-03 1e-03 1e-04 1e-04 1e-04 5e-05 1e-04 5e-05 5e-04 5e-05 1e-04 1e-04 5e-04 5e-05 5e-04 1e-03 1e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 5e-04 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 3e-03 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 5e-04 5e-04 1e-03 5e-04 5e-04 5e-04 1e-03 5e-04 1e-03 5e-04 1e-03 5e-04 5e-04 1e-03 5e-04 5e-04 1e-03 5e-04 1e-03 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-04 1e-03 5e-05 5e-04 5e-05 5e-05 5e-05 5e-05 5e-04 5e-05 1e-04 5e-05 1e-03 5e-05 1e-04 1e-04 1e-04 1e-04 1e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 Table 2: Hyperparameter values for the metacontroller with a single expert. τ refers to the ponder cost, αc refers to the learning rate for the controller and memory, αm refers to the learning rate for the manager, αEIN refers to the learning rate for the IN expert, and αEMLP refers to the learning rate for the MLP expert. 20 Published as a conference paper at ICLR 2017 IN + MLP τIN τMLP αc αm αEIN αEMLP 0.00000 0.00000 0.00000 0.00000 0.00000 0.00000 0.00000 0.00000 0.00121 0.00663 0.03641 0.20000 0.30000 0.40000 1e-03 1e-03 1e-03 1e-03 1e-03 5e-04 5e-04 5e-05 5e-04 1e-03 5e-05 5e-05 1e-04 5e-05 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 0.00121 0.00121 0.00121 0.00121 0.00121 0.00121 0.00121 0.00000 0.00121 0.00663 0.03641 0.20000 0.30000 0.40000 1e-03 1e-03 1e-03 1e-03 1e-03 5e-04 1e-03 1e-04 5e-05 1e-03 1e-04 5e-04 5e-05 1e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 0.00663 0.00663 0.00663 0.00663 0.00663 0.00663 0.00663 0.00000 0.00121 0.00663 0.03641 0.20000 0.30000 0.40000 1e-03 5e-04 5e-04 1e-03 5e-04 5e-04 5e-04 1e-03 5e-05 1e-04 1e-04 5e-04 1e-03 1e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 0.03641 0.03641 0.03641 0.03641 0.03641 0.03641 0.03641 0.00000 0.00121 0.00663 0.03641 0.20000 0.30000 0.40000 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 5e-04 5e-04 1e-03 5e-04 1e-04 5e-05 1e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 0.20000 0.20000 0.20000 0.20000 0.20000 0.20000 0.20000 0.00000 0.00121 0.00663 0.03641 0.20000 0.30000 0.40000 1e-03 1e-03 1e-03 1e-03 5e-04 1e-03 1e-03 5e-04 5e-04 5e-04 1e-04 1e-03 5e-05 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 0.30000 0.30000 0.30000 0.30000 0.30000 0.30000 0.30000 0.00000 0.00121 0.00663 0.03641 0.20000 0.30000 0.40000 5e-04 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-04 1e-03 1e-03 5e-04 1e-03 1e-04 5e-05 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 0.40000 0.40000 0.40000 0.40000 0.40000 0.40000 0.40000 0.00000 0.00121 0.00663 0.03641 0.20000 0.30000 0.40000 1e-03 5e-04 1e-03 5e-04 1e-03 5e-04 5e-04 1e-03 1e-03 5e-04 1e-04 1e-03 1e-03 5e-04 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 1e-03 Table 3: Hyperparameter values for the metacontroller with two experts. τIN refers to the ponder cost for the interaction network expert, τMLP refers to the ponder cost for the MLP expert, αc refers to the learning rate for the controller and memory, αm refers to the learning rate for the manager, αEIN refers to the learning rate for the IN expert, and αEMLP refers to the learning rate for the MLP expert. 21
2cs.AI
Proceedings of the 2009 INFORMS Simulation Society Research Workshop L.H. Lee, M.E. Kuhl, J.W. Fowler, and S. Robinson, eds. INVESTIGATING OUTPUT ACCURACY FOR A DISCRETE EVENT SIMULATION MODEL AND AN AGENT BASED SIMULATION MODEL Mazlina Abdul Majid Uwe Aickelin School of Computer Science Nottingham University Nottingham, NG8 1BB, U.K. School of Computer Science Nottingham University Nottingham, NG8 1BB, U.K. ABSTRACT In this paper, we investigate output accuracy for a Discrete Event Simulation (DES) model and Agent Based Simulation (ABS) model. The purpose of this investigation is to find out which of these simulation techniques is the best one for modelling human reactive behaviour in the retail sector. In order to study the output accuracy in both models, we have carried out a validation experiment in which we compared the results from our simulation models to the performance of a real system. Our experiment was carried out using a large UK department store as a case study. We had to determine an efficient implementation of management policy in the store’s fitting room using DES and ABS. Overall, we have found that both simulation models were a good representation of the real system when modelling human reactive behaviour. 1 INTRODUCTION Simulation has become a preferred tool in Operation Research (OR) for modelling complex systems. Studies in human behaviour modelling have received increased focus and attention from simulation research in the UK (Robinson 2004). The research in human behaviour modelling has been applied to various application areas such as manufacturing, health care, military and many more. As found in the literature, researchers choose either Discrete Event Simulation (DES) or Agent Based Simulation (ABS) as tools to investigate human behaviour problems. The choice of which simulation technique to be used relies on the individual judgment of the simulation model characteristics and their experience with the model. The representation of human behaviour contains complexity and variability; therefore when investigating such systems it is very important to choose a suitable modelling and simulation technique. In this research, we aim to provide an empirical study in order to find out which simulation modelling technique Peer-Olaf Siebers School of Computer Science Nottingham University Nottingham, NG8 1BB, U.K. is a good representation of a real system in our validation experiment. In the validation experiment we have compared the results from traditional DES and ABS models to the performance of the real system. The main difference between traditional DES and ABS is that in the first one the modelling focuses on the process flow, while in the ABS the modelling focus is on the individual entities in the system and their interactions. Human reactive behaviour means “how a certain individual responds to a certain request. For example, sales staff provides help when needed. In this work, we investigate the output accuracy of DES and ABS models when modelling human reactive behaviour in a department store. Statistical tests were used to compare the models. The content of the report is as follows: a background section which gives a taxonomy of simulation techniques and a discussion of previous related work. Background in Section 2 explores the theory and characteristics of three major OR simulation methods, i.e. DES, ABS and System Dynamics (SD). In Section 3, we define our case study and the model design. Validation experimentation is presented in Section 4, where we also compare our simulation models’ output against the real world output by using quantitative methods. In addition results from the experiment are also discussed. Finally in Section 5 we draw our conclusions and summarize the current progress of our research. 2 BACKGROUND There are several tools and techniques that can be used to model a system. Modelling is a process of abstracting a real world problem into modelling tools. Over the last three decades, simulation has become a frequently used modelling tool in OR (Kelton 2007). A simulation can be defined as a process of executing a model over time. Its ability to model complex systems has made simulation the preferred user choice when compared to mathematical models. Simulation can be classified Majid, Aickelin and Siebers into three types; Discrete Event Simulation (DES), System Dynamics (SD) and Agent Based Simulation (ABS). DES models represent a system based on a series of chronological sequencing of events where each event changes state in discrete time. Meanwhile, SD models represent real world phenomenon using stock and flow diagrams, causal loop diagrams (to represent a number of interacting feedback loops) and differential equations. In contrast to the DES and SD models, ABS models comprise of a number of autonomous, responsive and interactive agents that will interact with each other in order to achieve their objectives. We can summarise the three simulation techniques as follows: the DES and ABS models are suitable to work with discrete event. They both model changes in discrete time from one event to another. On the other hand, SD models are suitable for modelling a system with continuous state changes. We have found literature performing comparisons between the DES and ABS models together with SD models in terms of model characteristics; however, none of them are currently focusing on modelling human behaviour. One of the relevant papers comparing simulation technique with regards to model characteristics is (Wakeland et al. 2005), where they have compared SD and ABS in field of Biomedical. The authors found that the understanding of the aggregate behaviour in SD models and state changes in individual entities in ABS models is relevant in the study. SD and DES comparisons in the field of Fisheries were described by (Morecroft and Robinson 2006). They found that SD and DES are different approaches but both are suitable for modelling systems over time. Existing comparisons between DES and ABS were described by (Becker et al. 2006) in the field of transportation where they found that DES is less flexible than ABS, i.e. it is difficult to model different behaviours of shippers in DES. In addition to the existing comparison between DES and ABS, (Yu et al. 2007) performed a quantitative comparison on the DES and ABS models outputs in the field of transportation. We found just one work that describes the three techniques discussed in this study. (Owen et al 2008) tried to establishing a framework for comparing the different modelling techniques. We have also found a disparity in the quantity of work comparing SD and ABS or SD and DES contrasted to the amount of work comparing DES and ABS. Specifically, this disparity can be outstanded when referring to modelling human behavioural and where the focus is output accuracy. Therefore, in this research we choose to study the differences between DES and ABS in this regard. To study the differences in both models, we choose to focus on management practices in retail with regards to worker behaviour. Research in retail previously focused on consumer behaviour (i.e. Schenk et al. 2007). However, research in management practices has started to evolve as described by (Siebers et al. 2007, 2008). As discussed in the existing comparison of simulation techniques above, a lot of work has been done to compare simulation techniques in transportation and supply chain management. These studies have focused on modelling characteristics. On the other hand, we have decided to compare the accuracy of output in the DES and ABS models on management practices, as it is currently a developing area of study in the retail domain. 3 CASE STUDY FIELDWORK In order to achieve our aim, we have used a case study approach. The research has focused on the operation of the main fitting room in the womenswear department of one of the top ten retailers in the UK (see Figure 1). We wanted to identify the potential impact for fitting room performance when having different numbers of sales staff permanently present. We have investigated the staff behaviour on human reactive behaviour, which relates to staff responding to the customer when being available and requested. From the case study exploration, we have produced a flow chart diagram for DES conceptual models (see Figure 2) as in DES we focus on process flow. Our ABS conceptual models (see Figure 3 for an example) are state chart diagrams for the different types of people we had to represent (in our case customers and staff) as in ABS we focus on the individual ‘actors’ and their interactions. In the fitting room operation, the staff reactive behaviours can be seen in three jobs. Job 1 is about counting the number of garments and giving the green card, job 2 is about providing help and lastly job 3 is about receiving the green card and unwanted garments from the customers. The case study data has been transformed into the simulation inputs for the experimentation. 4 EXPERIMENTATION Two similar simulation models of DES and ABS were developed using multi-paradigm software known as Anylogictm. A conventional M/M/1 queuing system is constructed in both techniques. The model consists of an arrival process (customers), three single queues (customer entry queue, customer return queue, customer help queue) and resource (sales staff). The run length for both simulation models is one day from 9.00 am to 5.00 pm and was replicated to 100 runs. Both simulation models used the same inputs. The simulation model we developed has one member of staff that does all three jobs mentioned above, with a workload of 45% for job 1 (counting garments on entry), a workload of 10% for job 2 (providing help) and a workload of 45% for job 3 (counting garments on exit). Along with developing the simulation models for DES and ABS, the verification and validation process is performed simultaneously with the models. In the next sub-section, Proceedings of the 2009 INFORMS Simulation Society Research Workshop L.H. Lee, M.E. Kuhl, J.W. Fowler, and S. Robinson, eds. Figure 1: An illustration of the main fitting room operation Figure 2: Flow chart for DES model we describe how we have performed our validation experiment. 4.1 Validation Experiment We have used black box validation to compare the simulation outputs from DES and ABS with the real system output using quantitative method. By using statistical methods for this comparison, we are able to find out which simulation model is a good representation of a real system. We compared output data observed in the department store to the distribution of the predicted output generated by each model. By assuming the alternative hypothesis is the opposite of null hypothesis in all tests, thus in the paper we will only state the null hypothesis. The main hypothesis for the validation test was constructed as following: Ho = DES model is a good representation of a real system Ho = ABS model is a good representation of a real system Figure 3: State chart for ABS model (customer) We have used mean waiting time from the three queues as our performance measurement in the experiment. This was the only performance data we were able to collect from the real system. Two tests have been setup in answering the hypothesis. These tests are described in the following sub-section. 4.1.1 Test 1: Comparing Medians Using a Non Parametric Test In Normal distributed data, the observed mean, median and mode will have similar values because they are identical in the population. Since waiting time data are not normally distributed, we have chosen to compare medians as a more robust measure of central tendency rather than means and have chosen a Non Parametric statistical test known as Mann-Whitney to avoid assuming Normality. We have constructed the following hypothesis for the test: Majid, Aickelin and Siebers Ho = DES models are not significantly different with real system in mean customer waiting time simulation models and real system. To execute the test, we have produced the following hypothesis: Ho = ABS models are not significantly different with real system in mean customer waiting time Ho = DES model shows similar variability compare to the real system In performing the Mann-Whitney test, we have used the open source statistical software, R. The median of waiting time from DES and ABS models and real system were used for the comparison purpose. We have chosen 0.05 as the level of significance. If the probability of seeing data as or more different than expected under the model is smaller than 0.05, we reject the null hypothesis. If it is not then we fail to reject the null hypothesis. Put some simply we can conclude that the data we observed are consistent with the model’s predictions (without assuming that the model must itself therefore be correct). In the statistic test for DES model against real system, p-value is 0.3269.Since the pvalue is not less than the chosen sigma value of 0.05; there is insufficient evidence to reject our hypothesis. The statistic test for ABS model against real system showed similar results when it had a p-value of 0.2958, and it was larger than the chosen level of 0.05.Therefore, we fail to reject our hypothesis for both simulation models. We find the distribution of median waiting times is consistent with both models. Next we look at the variability of waiting time to see if the one we get from our simulation models matches the variability we can observe in the real system. Ho = ABS model shows similar variability compare to the real system 4.1.2 Test 2: Measuring Variability In finding out which DES and ABS models are having more variation against real system we have plotted a frequency distribution of customer waiting time for a single day (represents one replication) from both simulation models and real system (Figure 4). Frequency Waiting Time Outputs from Real System, DES and ABS in One Day 160 140 120 100 80 60 40 20 0 REAL DES ABS 0 1 2 3 4 5 6 7 8 9 Custimer Waiting Time Figure 4: Waiting time outputs for one day in real system, DES and ABS models. Next, we calculate the variance (a measure of dispersion) of customer waiting times. This allows us to compare the variability of the simulation outputs data against real system on a statistical basis by study the spread of frequency distribution of customer waiting time from both The spread of frequency distribution for DES and ABS models (as shown in Figure 4) seem slightly different and to look into their differences, we have compared the variances of the outputs. The variance shows how close the simulation outputs i.e. waiting time, in the distribution are to the middle of the distribution. Table 1: Waiting time outputs for DES, ABS and real system in one day with standard deviation and variance. Models Real system DES ABS Mean waiting time (minutes) 1.68 1.69 1.61 Standard deviation Variance 1.73 1.59 1.70 3.01 1.96 2.89 By looking at the result in Table 1 above, the variance between DES model and the real system is significantly different with 35% difference and ABS model is similar with real system with 4% difference. The reason for the differences in variance for both simulation models may be due to the different operation structure in DES and ABS models. The entities in DES model are having a strong order and they are depending to each other to change their state. Meanwhile agents in ABS model are in decentralised order and they change their state independently. For this test we fail to reject our hypothesis for ABS model where the model has produced similar variability with the real system. However, we have to reject the hypothesis for DES model where the model has shown dissimilarity in variability when compared to the real system. 4.1.3 Validation Experiment Conclusion In conclusion we find on statistically significant differences between the DES and ABS models when compared to the real system in Test 1. However in Test 2 we found DES model produced different variability then the real system and ABS produced the similar variability with the real system. Therefore based on Test 2 result, we suggest that the ABS model is more suitable in representing the behaviour of the real system where the operation involving human is the main focus of the system. Nevertheless based on the validation experiment both simulation models are good in representing the real system and ABS shows better representation of the real system when passed the Test 2. Majid, Aickelin and Siebers 5 CONCLUSION AND FUTURE WORK In the validation experiments between DES and ABS models, we are able to demonstrate which simulation models is a good representation of real system when modelling human reactive behaviour. To achieve that, we compared the output accuracy of the simulation outputs with the real system in mean customer waiting time. Statistical methods to analyse the outputs have helped us to identify the statistical significance of the similarity and difference of the simulation models to one another. The testing suggested that even though both DES and ABS models produce similar outputs in Test 1, when we compared their medians to the real system, they showed different variation in their model outputs in Test 2. ABS models reflect the real system behaviour much better than DES model in terms of their predicted variability in waiting times. The system that we modelled in DES and ABS is a typical queuing system with no extra features of complex human behaviour. Moreover, based on the validation results we have concluded that both DES and ABS models are a good representation of the real system that contains human reactive behaviour. Investigating the outputs’ behaviour using different sample data and different scenarios will be carried out as part of our future work. We also want to evolve our experiment by modelling human proactive behaviour in ABS only and look into the benefit of modelling such behaviour in a queuing system for the OR simulation study. Siebers, P.O. et al. 2007. “Using Intelligent Agents to Understand Management Practices and Retail Productivity.” In Proceedings of the Winter Simulation Conference, (Dec 9-12 2007).Washington DC, USA, 22122220 Siebers, P.O. et al, 2008. “An Agent-Based Simulation of In-Store Customer Experiences.” for the Proceedings of the 2008 Operational Research Society Simulation Workshop, (April 1-2). Worcestershire, UK. Yu, T. T, J. P. Scanlan., and G. B. Wills. 2007. AgentBased and Discrete-Event Modelling: A quantitative approach. 7th AIAA Aviation Technology, Integration and Operation Conference (ATIO). Wakeland, et al. 2005. “A Tale of Two Methods—AgentBased Simulation and System Dynamics—Applied in a Biomedical Context: Acute Inflammatory Response,” 6th European Congress of Systems Science. AUTHORS BIOGRAPHIES MAZLINA ABDUL MAJID is a PhD student in the School of Computer Science, University of Nottingham. Her interest is in discrete event simulation and agent based simulation. Her email is<[email protected]>. Prof. UWE AICKELIN is a Professor in the School of Computer Science, University of Nottingham. His interests include agent-based simulation, heuristic optimization and artificial immune system His email is<[email protected]>. REFERENCES Becker, M. 2006. Agent-based and Discrete Event Simulation of Autonomous Logistic Processes for the Proceedings of: Borutzky, W.; Orsoni, A.; Zobel, R. (eds.): 20th European Conference on Modelling and Simulation. pp. 566-571. Kelton, W.D. et al. 2007. Simulation with ARENA: Fourth Edition. New York: McGraw-Hill. ISBN 0-07110685-5. Morecroft. J., and S. Robinson. 2006. Comparing DiscreteEvent Simulation And System Dynamics: Modelling A Fishery for the Proceedings of the 2006 OR Society Simulation Workshop S. Robinson, S. Taylor, S. Brailsford and J.Garnett, eds. Owen, C. , D. D. Love, and P. Albores.2008.Selection of Simulation Tools for Improving Supply Chain Performance for the Proceedings of the 2008 OR Society Simulation Workshop. Robinson, S. 2004.Discrete-event simulation: from the pioneers to the present, what next? Journal of the Operational Research Society. 56, 619–629. Schenk, et al. 2007. Agent-based simulation of consumer behavior in grocery shopping on a regional level. Journal of Business Research 60. 894–903 Dr. PEER –OLAF SIEBERS is a Research Fellow in the School of Computer Science, University of Nottingham. His main interest includes agent-based simulation and human oriented complex adaptive system. His email is <pos@ cs.nott.ac.uk>.
5cs.CE
arXiv:1802.04986v1 [cs.SE] 14 Feb 2018 Convolutional Neural Networks over Control Flow Graphs for Software Defect Prediction Anh Viet Phan, Minh Le Nguyen Lam Thu Bui Japan Advanced Institute of Information Technology Nomi, Japan 923-1211 Email: {anhphanviet,nguyenml}@jaist.ac.jp Le Quy Don Technical University Ha Noi, Vietnam Email: [email protected] Abstract—Existing defects in software components is unavoidable and leads to not only a waste of time and money but also many serious consequences. To build predictive models, previous studies focus on manually extracting features or using tree representations of programs, and exploiting different machine learning algorithms. However, the performance of the models is not high since the existing features and tree structures often fail to capture the semantics of programs. To explore deeply programs’ semantics, this paper proposes to leverage precise graphs representing program execution flows, and deep neural networks for automatically learning defect features. Firstly, control flow graphs are constructed from the assembly instructions obtained by compiling source code; we thereafter apply multi-view multi-layer directed graph-based convolutional neural networks (DGCNNs) to learn semantic features. The experiments on four real-world datasets show that our method significantly outperforms the baselines including several other deep learning approaches. Index Terms—Software Defect Prediction, Control Flow Graphs, Convolutional Neural Networks (a) File 1.c (b) File 2.c Fig. 1: A motivating example Recently, several software engineering problems have been successfully solved by exploiting tree representations of Software defect prediction has been one of the most attractive programs - the Abstract Syntax Trees (ASTs) [12]. In the research topics in the field of software engineering. According field of machine learning, the quality of input data directly to regular reports, semantic bugs result in an increase of affects the performance of learners. Regarding this, due to application costs, trouble to users, and even serious conse- containing rich information of programs, tree-based approaches quences during deployment. Thus, localizing and fixing defects have shown significant improvements in comparison with in early stages of software development are urgent requirements. previous research, especially software metrics-based. Mou Various approaches have been proposed to construct models et al. proposed a tree-based convolutional neural network to which are able to predict whether a source code contains extract structural information of ASTs for classifying programs defects. These studies can be divided into two directions: one by functionalities [15]. Wang et al. employed a deep belief is applying machine learning techniques on data of software network to automatically learn semantic features from AST metrics that are manually designed to extract features from tokens for defect prediction [23]. Kikuchi et. al measured the source code; the other is using programs’ tree representations similarities between tree structures for source code plagiarism detection [10]. and deep learning to automatically learn defect features. The traditional methods focus on designing and combining However, defect characteristics are deeply hidden in profeatures of programs. For product metrics, most of them are grams’ semantics and they only cause unexpected output in based on statistics on source code. For example, Halstead met- specific conditions [24]. Meanwhile, ASTs do not show the rics are computed from numbers of operators and operands [8]; execution process of programs; instead, they simply represent CK metrics are measured by function and inheritance counts the abstract syntactic structure of source code. Therefore, [6]; McCabe’s metric estimates the complexity of a program both software metrics and AST features may not reveal many by analyzing its control flow graph [13]. However, according types of defects in programs. For example, we consider the to many studies and surveys, the existing metrics often fail in procedures with the same name sumtoN in two C files File capturing the semantics of programs [9], [14]. As a result, 1.c and File 2.c (Fig. 1). Two procedures have a tiny although many efforts have been made such as adopting difference at line 7 File 1.c and line 15 in File 2.c. As robust learning algorithms and refining the data, the classifier can be seen a bug from File 2.c, the statement i=1 causes performance is not so high [4]. an infinite loop of the for statement in case N>=1. Whereas I. I NTRODUCTION using RSM tool1 to extract the traditional metrics, their feature vectors are exactly matching, since two procedures have the same lines of code, programming tokens, etc. Similarly, parsing the procedures into ASTs using Pycparser2 , their ASTs are identical. In other words, both metrics-based and tree-based approaches are not able to distinguish these programs. To explore more deeply programs’ semantics, this paper proposes combining precise graphs which represent execution flows of programs called Control Flow Graphs (CFGs), and a powerful graphical neural work. Regarding this, each source code is converted into an execution flow graph by two stages: 1) compiling the source file to the assembly code, 2) generating the CFG from the compiled code. Applying CFGs of assembly code is more beneficial than those of ASTs. Firstly, assembly code contains atomic instructions and their CFGs indicate step-by-step execution process of programs. After compiling two source files (Fig. 1) we observed that i+=1 and i=1 are translated into different instructions. Secondly, assembly code is refined because they are the products after AST processing of the compiler. The compiler applies many techniques to analyze and optimize the ASTs. For example, two statements int n; n=4; in the main procedure of File 2.c are treated in a similar way with the statement int n = 4; and, the if statement can be removed since the value of n is identified. Meanwhile, ASTs just describe the syntactic structures of programs, and they may contain many redundant branches. Assume that if the statement i=1 in procedure sumtoN of File 2.c is changed to i+=1 then both programs are the same. However, their AST structures have many differences including the positions of subtrees of two functions main and sumtoN, the separation of declaration and assignment of the variable n, and the function prototype of sumtoN in File 2.c. We thereafter leverage a directed graph-based convolutional neural network (DGCNN) on CFGs to automatically learn defect features. The advantage of the DGCNN is that it can treat large-scale graphs and process the complex information of vertices like CFGs. The experimental results on four real-world datasets show that applying DGCNN on CFGs significantly outperforms baselines in terms of the different measures3 . The main contributions of the paper can be summarized as follows: • Proposing an application of a graphical data structure namely Control Flow Graph (CFG) to software defect prediction and experimentally proving that leveraging CFGs is successful in building high-performance classifiers. • Presenting an algorithm for constructing Control Flow Graphs of assembly code. • Formulating an end-to-end model for software defect prediction, in which a multi-view multi-layer convolutional neural networks is adopted to automatically learn the defect features from CFGs. 1 http://msquaredtechnologies.com/m2rsm/ 2 https://pypi.python.org/pypi/pycparser 3 The source code and collected datasets are publicly available at https: //github.com/nguyenlab/DGCNN The model implementation is released to motivate related studies. The remainder of the paper is organized as follows: Section II explains step-by-step our approach for solving software defect prediction problem from processing data to adapting the learning algorithm. The settings for conducting the experiments such as the datasets, the algorithm hyper-parameters, and evaluation measures are indicated in Section III. We analyze experimental results in Section IV, and conclude in Section VI. • II. T HE P ROPOSED A PPROACH This section formulates a new approach to software defect prediction by applying graph-based convolutional neural networks over control flow graphs of binary codes. Our proposed method is illustrated in Fig. 3 includes two steps: 1) generating CFGs which reveal the behavior of programs, and 2) applying a graphical model on CFG datasets. In the first step, to obtain the graph representation of a program, the source code is compiled into an assembly code using g++ on Linux. The CFG thereafter is constructed to describe the execution flows of the assembly instructions. In the second step, we leverage a powerful deep neural network for directed labeled graphs, called the multiview multi-layer convolutional neural network, to automatically build predictive models based on CFG data. A. Control Flow Graphs A control flow graph (CFG) is a directed graph, G = (V,E) where V is the set of vertices {v1 , v2 , ..., vn } and E is the set of directed edges {(vi , vj ), (vk , vl ), ...} [1]. In CFGs, each vertex represents a basic block that is a linear sequence of program instructions having one entry point (the first instruction executed) and one exit point (the last instruction executed); and the directed edges show control flow paths. This paper aims to formulate a method for detecting faulty source code written in C language. Regarding this, CFGs of assembly code, the final products after compiling the source code, server as the input to learn faulty features using machine algorithms. Based on recent research, it has been proved that CFGs are successfully applied to various problems including malware analysis [2], [3], software plagiarism [5], [22]. Since semantic errors are revealed while programs are running, analyzing the execution flows of the assembly instructions may be helpful for distinguishing faulty patterns from nonfaulty ones. Fig. 2 illustrates an example of the control flow graph constructed from an assembly code snippet, in which each vertex corresponds to an instruction and a directed edge shows the execution path from an instruction to the other. The pseudocode to generate CFGs is shown in Algorithm 1. The algorithm takes an assembly file as the input, and outputs the CFG. Building the CFG from an assembly code includes two major steps. In the first step, the code is partitioned into blocks of instructions based on the labels (e.g. L1, L2, L3 in Fig. 2). The second step is creating the edges to represent the control flow transfers in the program. Specifically, the first line invokes procedure initialize_Blocks to read the file contents (a) (b) Fig. 2: An example of a Control Flow Graph (CFG).( 2a) a fragment of assembly code;( 2b) the CFG of the code fragment (each node is viewed by the line number and the name of the instruction). and return all the blocks. In line 2, the set of edges is initially set to empty. From line 3 to 24, the graph edges are created by traversing all instructions of each block and considering possible execution paths from the current instruction to others. For a block, because the instructions are executed in sequence, every node has an outgoing edge to the next one (line 5-9). Additionally, we consider two types of instructions which may have several targets. For jump instructions, an edge is added from the current instruction to the first one of the target block. We use two edges to model function calls, in which one is from the current node to the first instruction of the function and the other is from the final instruction of the function to the next instruction of the current node (line 10-24). Finally, the graphs are formed from the instruction and edge sets (line 25-26). B. Multi-view Multi-layer Graph-based Convolutional Neural Networks Algorithm 1: The algorithm for constructing Control Flow Graphs from assembly code Input : asm f ile - A file of assembly code Output : The graph representation of the code 1 blocks ← initialize Blocks(asm f ile); 2 edges ← {}; 3 for i ← 0 to |blocks| do 4 for j ← 0 to |blocks[i].instructions| do 5 if j > 0 then 6 inst 1 ← blocks[i].instructions[j − 1]; 7 inst 2 ← blocks[i].instructions[j]; 8 edges.add(new Edge(inst 1, inst 2)); 9 end 10 if inst 1.type=“jump” or inst 1.type=“call” then 11 label ← inst 1.params[0]; 12 to block ←find Block by Label(label); 13 if to block 6= NULL then 14 inst 2 ← to block.f irst instruction; 15 edges.add(new Edge(inst 1, inst 2)); 16 if inst 1.type=“call” then 17 inst 2 ← to block.last instruction; 18 inst 1 ← inst 1.next; 19 edges.add(new Edge(inst 2, inst 1)); 20 21 22 23 24 25 26 end end end end end instructions ← get All Instructions(blocks); return construct Graph(instructions, edges); into a vector. Finally, the feature vector is fed into a fullyconnected layer and an output layer to compute the categorical distributions for possible outcomes. Specifically, during the forward pass of convolutional layers, each filter slides through all vertices of the graph and computes dot products between entries of the filter and the input. Suppose that the subgraph in the sliding window includes d + 1 vertices (the current vertex and its neighbors) with vector representations of x0 , x1 , ..., xd ∈ Rvf ×nf , then the output of the filters is computed as follows: The directed graph-based convolutional neural network (DGCNN) is a dynamic graphical model that is designed to treat large-scale graphs with complex information of vertex labels. For instance, a CFG vertex is not simply a token but represents an instruction, which may contain many components including the instruction name and several operands. In addition, each instruction can be viewed in other perspectives (multiple vf d X X views), e.g. instruction types or functions. y = tanh( Wconv,i,j · xi,j + bconv ) (1) Fig. 3 depicts the overview architecture of DGCNN. In the i=0 j=1 DGCNN model, the first layer is called vector representations or the embedding layer, whereby a vertex is represented as a set where y, bconv ∈ Rvc ×nc , Wconv,i ∈ Rvc ×nc ×vf ×nf . tanh is of real-valued vectors corresponding with the number of views. the activation function. nf and vf are the vector size and the Next, we design a set of fixed-size circular windows sliding number of views of the input layer. nc and vc are the numbers over the entire graphs to extract local features of substructures. of filters and views of the convolutional layer. The input graphs are explored at different levels by stacking Due to arbitrary structures of graphs, the numbers of vertices several convolutional layers on the embedding layer. In our in subgraphs are different. As can be seen in Fig.3, the current work, we apply DGCNN with two layers in the convolution receptive field at the red node includes 5 vertices while only stage. After convolution, a dynamic pooling layer is applied 3 vertices are considered if the window moves right down. to gather extracted features from all the parts of the graphs Consequently, determining the number of weight matrices for Fig. 3: The overview of our approaches for software defect prediction using convolutional neural networks on Control Flow Graphs of assembly code. filters is unfeasible. To deal with this obstacle, we divide vertices into groups and treat items in each group in a similar way. Regarding the way, the parameters for convolution have only three weight matrices including W cur , W in , and W out for current, outgoing, and incoming nodes, respectively. In the pooling layer, we face the similar problem of dynamic graphs. The numbers of nodes are varying between programs’ CFGs. Meanwhile, convolutions preserve the structures of input graphs during feature extraction. This means that it is impossible to determine the number of items which will be pooled into the final feature vector. An efficient solution to this problem is applying dynamic pooling [18] to normalize the features such that they have the same dimension. Regarding this, one-way max pooling is adopted to gather the information from all parts of the graph to one fixed size vector regardless of graph shape and size. III. E XPERIMENTS A. Datasets The datasets for conducting experiments are obtained from a popular programming contest site CodeChef 4 . We created four benchmark datasets which each one involves source code submissions (written in C, C++, Python, etc.) for solving one of the problems as follows: • SUMTRIAN (Sums in a Triangle): Given a lower triangular matrix of n rows, find the longest path among all paths starting from the top towards the base, in which each movement on a part is either directly below or diagonally below to the right. The length of a path is the sums of numbers that appear on that path. • FLOW016 (GCD and LCM): Find the greatest common divisor (GCD) and the least common multiple (LCM) of each pair of input integers A and B. • MNMX (Minimum Maximum): Given an array A consisting of N distinct integers, find the minimum sum of cost to convert the array into a single element by following 4 https://www.codechef.com/problems/hproblem-namei TABLE I: Statistics on the datasets Dataset FLOW016 MNMX SUBINC SUMTRIAN Total Class 0 Class 1 Class 2 Class 3 Class 4 10648 3472 4165 231 2368 412 8745 5157 3073 189 113 213 6484 3263 2685 206 98 232 21187 9132 6948 419 2701 1987 operations: select a pair of adjacent integers and remove the larger one of these two. For each operation, the size of the array is decreased by 1. The cost of this operation will be equal to their smaller. • SUBINC (Count Subarrays): Given an array A of N elements, count the number of non-decreasing subarrays of array A. The target label of an instance is one of the possibilities of source code assessment. Regarding this, a program can be assigned to one of the groups as follows: 0) accepted - the program ran successfully and gave a correct answer; 1) time limit exceeded - the program was compiled successfully, but it did not stop before the time limit; 2) wrong answer: the program compiled and ran successfully but the output did not match the expected output; 3) runtime error: the code compiled and ran but encountered an error due to reasons such as using too much memory or dividing by zero; 4) syntax error - the code was unable to compile. We collected all submissions written in C or C++ until March 14th, 2017 of four problems. The data are preprocessed by removing source files which are empty code, and unable to compile. Table I presents statistical figures of instances in each class of the datasets. All of the datasets are imbalanced. Taking MNMX dataset as an example, the ratios of classes 2, 3, 4 to class 0 are 1 to 27, 46, and 24. To conduct experiments, each dataset is randomly split into three folds for training, validation, and testing by ratio 3:1:1. B. Experimental Setup We compare our model with tree-based approaches which have been successfully applied to programming language TABLE II: Structures and numbers of hyperparameters of the neural networks. Each layer is presented in form of the name followed by the number of neurons. Emb is a embedding layer. Rv, TC, GC, and FC stand for recursive, tree-based convolutional, graph-based convolutional, and fully-connected, respectively. C. Evaluation Measures The approaches are evaluated based on two widely used measures including accuracy and the area under the receiver operating characteristic (ROC) curve, known as the AUC. Predictive accuracy has been considered the most important criterion to evaluate the effectiveness of classification algoNetwork Architecture weights biases rithms. For multi-class classification, the accuracy is estimated RvNN Coding30-Emb30-Rv600-FC600-Soft5 1104600 1235 by the average hit rate. TBCNN Coding30-Emb30-TC600-FC600-Soft5 1140600 1235 SibStCNN Coding30-Emb30-TC600-FC600-Soft5 1140600 1235 In machine learning, the AUC that estimates the discriminaDGCNN-1V GC100-GC600-FC600-Soft5 552000 1305 tion ability between classes is an important measure to judge DGCNN-2V GC100-GC600-FC600-Soft5 561000 1305 the effectiveness of algorithms. It is equivalent to the nonparametric Wilcoxon test in ranking classifiers [7]. According processing tasks. For the tree-based methods, a source code is to previous research, AUC has been proved as a better and represented as an abstract syntax trees (AST) using a parser, more statistically consistent criterion than the accuracy [11], and then different machine learning techniques are employed to especially for imbalanced data. In the cases of imbalanced build predictive models. The settings for baselines are presented datasets that some classes have much more samples than others, most of the standard algorithms are biased towards as follows: The neural networks For neural networks includ- the major classes and ignore the minor classes. Consequently, ing DGCNN, tree based convolutional neural networks the hit rates on minor classes are very low, although the overall (TBCNN) [15], Sibling-subtree convolutional neural networks accuracy may be high. Meanwhile, in practical applications, (SibStCNN - an extension of TBCNN in which feature accurately predicting minority samples may be more important. detectors are redesigned to cover subtrees including a node, Taking account of software defect prediction, the essential task its descendants and siblings), and recursive neural networks is detecting faulty modules. However, many software defect (RvNN), we use some common hyper-parameters: initial datasets are highly imbalanced and the faulty instances belong learning rate is 0.1, vector size of tokens is 30. The structures to minority classes [17]. Because all experimental datasets are imbalanced, we adopt both measures to evaluate the classifiers. of the networks are shown in Table II. ROC curves which depict the tradeoffs between hit rates When adapting DGCNN for CFGs, we use two views of CFG nodes including instructions and their type. For instance, jne, and false alarm rates are commonly used for analyzing binary jle, and jge are assigned into the same group because they classifiers. To extend the use of ROC curves to multi-class are conditional jump instructions. For instructions, it should problems, the average results are computed based on two be noted that they may have many operands. In this case, we ways: 1) macro-averaging gives equal weight to each class, replace all operands such as block names, processor register and 2) micro-averaging gives equal weight to the decision of names, and literal values with symbols “name”, “reg”, and each sample [21]. The AUC measure for ranking classifiers is “val”, respectively. Following this replacement, the instruction estimated by the area under the macro-averaged ROC curves. addq $32, %rsp is converted into addq value, reg. IV. R ESULTS AND D ISCUSSION To generate inputs for DGCNN, firstly the symbol vectors are randomly initialized and then the corresponding vector of Table III shows the accuracies of classifiers on the four each view is computed based on the component vectors as datasets. As can be seen, CFG-based approaches significantly follows: outperform others. Specifically, in comparison with the second C 1 X best, they improve the accuracies by 12.39% on FLOW016, xj (2) x vi = 1.2% on MNMX, 7.71% on SUBINC, and 1.98% on SUMC j=1 TRIAN. As mentioned before (Section I), software defect where C is the number of symbols in the view, and xj is the prediction is a complicated task because semantic errors are vector representation of the j th symbol. Taking instruction hidden deeply in source code. Even if a defect exists in a proaddq $32, %rsp as an example, its vector is the linear gram, it is only revealed during running the application under combination of vectors of symbols addq, value, and reg specific conditions. Therefore, it is impractical to manually k-nearest neighbors (kNN) We apply kNN algorithm with design a set of good features which are able to distinguish tree edit distance (TED) and Levenshtein distance (LD) [16]. faulty and non-faulty samples. Similarly, ASTs just represent The number of neighbors k is set to 3. the structures of source code. Although tree-based approaches Support Vector Machines (SVMs) The SVM classifiers (SibStCNN, TBCNN, and RvNN) are successfully applied to are built based on hand-crafted features, namely bag-of-words other software engineering tasks like classifying programs by (BoW). By this way, the feature vector of each program is functionalities, they have not shown good performance on determined by counting the numbers of times symbols appear the software defect prediction. In contrast, CFGs of assembly in the AST. The SVM with RBF kernel has two parameters C code is precise graphical structures which show behaviors of and γ; their values are 1 and 0, respectively. programs. As a result, applying DGCNN on CFGs achieves the TABLE III: Comparison of classifiers according to accuracy. 1V and 2V following DGCNN means that CFG nodes are viewed by one and two perspectives. Op and NoOp are using instruction with or without operands. NoOp NoOp Op Op FLOW016 60.00 60.75 61.69 61.03 63.10∗ 62.25 73.80 74.32 75.49 75.12 MNMX 77.53 79.13 80.73 82.56 82.45 82.85∗ 83.19 83.82 84.05 83.71 SUBINC 67.23 66.62 68.31∗ 64.53 63.99 67.69 70.93 74.02 72.40 76.02 SUMTRIAN 64.87 65.81 66.97∗ 58.82 65.05 65.10 68.83 68.12 68.19 68.95 0.8 True Positive Rate Approach SVM-BoW LD TED RvNN TBCNN SibStCNN DGCNN 1V DGCNN 2V DGCNN 1V DGCNN 2V 1.0 0.6 0.2 0.0 0.0 TABLE IV: The performance of classifiers in terms of AUC. NoOp NoOp Op Op FLOW016 0.74 0.75 0.76 0.76 0.82 0.80 0.81 0.82 MNMX 0.76 0.79 0.77 0.79 0.82 0.81 0.80 0.79 SUBINC 0.73 0.69 0.72 0.71 0.74 0.72 0.75 0.74 SUMTRIAN 0.79 0.73 0.78 0.80 0.82 0.81 0.81 0.81 0.2 0.4 0.6 False Positive Rate 0.8 1.0 (a) 1.0 0.8 True Positive Rate Approach SVM-BoW RvNN TBCNN SibStCNN DGCNN 1V DGCNN 2V DGCNN 1V DGCNN 2V micro-average ROC curve (area = 0.89) macro-average ROC curve (area = 0.76) ROC curve of class 0 (area = 0.80) ROC curve of class 1 (area = 0.77) ROC curve of class 2 (area = 0.52) ROC curve of class 3 (area = 0.94) ROC curve of class 4 (area = 0.78) 0.4 0.6 micro-average ROC curve (area = 0.92) 0.4 macro-average ROC curve (area = 0.82) highest accuracies on the experimental datasets about software ROC curve of class 0 (area = 0.89) defects. ROC curve of class 1 (area = 0.87) 0.2 From the last four rows of Table III, the more the information ROC curve of class 2 (area = 0.59) ROC curve of class 3 (area = 0.96) is provided, the more efficient the learner is. In general, viewing ROC curve of class 4 (area = 0.78) graph nodes by two perspectives including instructions and 0.0 0.0 0.2 0.4 0.6 0.8 1.0 instruction groups helps boost DGCNN classifiers in both cases: False Positive Rate with and without the use of operands. Similarly, taking into (b) account of all components in instructions (Eq. 2) is beneficial. In this case, the DGCNN models achieve highest accuracies on Fig. 4: The illustration of the discrimination ability between the experimental datasets. Specifically, DGCNN with one view classes of classifiers on imbalanced datasets. Fig. 4a and Fig. 4b reaches the accuracies of 75.49% on FLOW016, and 84.05% are the ROC curves of TBCNN and DGCNN 1V NoOp on on MNMX; DGCNN with two views obtains the accuracies MNMX dataset, respectively. of 76.02% on SUBINC, and 68.95% on SUMTRIAN. We also assess the effectiveness of the models in terms of the discrimination measure (AUC) which is equivalent to belonging degrees of an instance to classes. There are two Wilcoxon test in ranking classifiers. For imbalanced datasets, groups including graph-based and tree-based approaches, in many learning algorithms have a trend to bias the majority which the approaches in each group has the similar AUC class due to the objective of error minimization. As a result, scores; and graph-based approaches show better performance the models mostly predict an unseen sample as an instance of than those of tree-based. It is worth noticing that, along with the majority classes, and ignore the minority classes. Fig. 4 the efforts of accuracy maximization, the approach based on plots the ROC curves of TBCNN and DGCNN 1V NoOp DGCNN and CFGs also enhance the distinguishing ability classifiers on MNMX dataset, an imbalanced data with the between categories even on imbalanced data. The DGCNN minority classes of 2 and 4. Both two classifiers have a notable classifier improves the second best an average of 0.03 on AUC lower ability in detecting minority instances from the others. scores. From above analysis, we can conclude that leveraging For predicting class 4, the TBCNN is even equivalent to a precise control flow graphs of binary codes is suitable for random classifier. After observing the other ROC curves we software defect prediction, one of the most difficult tasks in found the similar problem for all of the approaches on the the field of software engineering. experimental datasets. Thus, AUC is an essential measure for V. E RROR A NALYSIS evaluating classification algorithms, especially in the case of imbalanced data. We analyze cases of source code variations which methods Table IV presents the AUCs of probabilistic classifiers, are able to handle or not based on observations on classifiers’ which produce the probabilities or the scores to indicate the outputs, training and test data. We found that RvNN’s per- (a) File 3.c (a training sample) (b) File 4.c (G+, T-) (d) File 6.c (G-, T-) (c) File 5.c (G-, T+) Fig. 5: Some source code examples in FLOW016 dataset which may cause mistakes of tree-based (T) and CFG-based (G) approaches. Fig. 5a is a sample in the training set. Figs. 5b, 5c, and 5d are samples in the test set. Symbols “+” and “-” denote the sample is correctly and incorrectly classified by the approaches. formance is degraded when tree sizes increase. This problem new branches of ASTs (File 3.c and File 4.c). Because is also pointed out from other research on tasks of natural of the weight matrices for each node being determined based language processing [19], [20] and programming language on the position, SibStCNN and TBCNN are easily affected by processing [15]. From Tables I, III, and IV the larger trees, changes regarding tree shape and size. the lower accuracies and AUCs, RvNN obtains in comparison Meanwhile, graph-based approaches are able to handle these with other approaches, especially on SUMTRIAN dataset. changes. We observed that although loop statements like For, SibStCNN and TBCNN obtain higher performance than other While, and DoWhile have different tree representations, their baselines due to learning features from subtrees. For analyzing assembly instructions are similar by using a jump instruction tree-based methods in this section, we only take into account to control the loop. Similarly, moving a statement to possible SibStCNN and TBCNN. positions may not result in notable changes in assembly code. Effect of code structures: the tree-based approaches suffer Moreover, grouping a set of statements to form a procedure is from varying structures of ASTs. For example, given a also captured in CFGs by using edges to simulate the procedure program, we have many ways to reorganize the source code invocation (Section II-A). such as changing positions of some statements, constructing Effect of changing statements: CFG-based approaches may procedures and replacing statements by equivalent ones. These be affected by replacements of statements. Considering source modifications lead to reordering the branches and producing code in File 3.c, File 5.c, they have similar ASTs, but the assembly codes are different. In C language, statements are translated into different sets of assembly instructions. For example, with the same operator, the sets of instructions for manipulating data types of int and long int are dissimilar. Moreover, statements are possible replaced by others without any changes of program outcomes. Indeed, to show values, we can select either printf or cout. Since contents of CFG nodes are changed significantly, DGCNN may fail in predicting these types of variations. Effect of using library procedures: when writing a source code, the programmer can use procedures from other libraries. In Fig. 5, File 6.c applies the procedure __gcd in the library algorithm, while the others use ordinary C statements for computing the greatest common divisor of each integer pair. Both ASTs and CFGs do not contain the contents of external procedures because they are not embedded to generate assembly code from source code. As a result, tree-based and graph-based approaches are not successful in capturing program semantics in these cases. VI. C ONCLUSION This paper presents an end-to-end model for solving software defect prediction, one of the most difficult tasks in the field of software engineering. By applying precise representations (CFGs) and a graphical deep neural network, the model explores deeply the behavior of programs to detect faulty source code from others. Specifically, the CFG of a program is constructed from the assembly code after compiling its source code. Then DGCNN is leveraged to learn from various information of CFGs data to build predictive models. Our evaluation of four real-world datasets indicates learning on graphs could significantly improve the performance of feature-based and tree-based approaches according to both accuracy and discrimination measures. Our method improves the accuracies from 4.08% to 15.49% in comparison with the feature-based approach, and from 1.2% to 12.39% in comparison with the tree-based approaches. ACKNOWLEDGEMENTS This work was supported partly by JSPS KAKENHI Grant number 15K16048 and the first author would like to thank the scholarship from Ministry of Training and Education (MOET), Vietnam under the project 911. R EFERENCES [1] Frances E Allen. Control flow analysis. In ACM Sigplan Notices, volume 5, pages 1–19. ACM, 1970. [2] Blake Anderson, Daniel Quist, Joshua Neil, Curtis Storlie, and Terran Lane. Graph-based malware detection using dynamic analysis. Journal in computer virology, 7(4):247–258, 2011. [3] Danilo Bruschi, Lorenzo Martignoni, and Mattia Monga. Detecting selfmutating malware using control-flow graph matching. In International Conference on Detection of Intrusions and Malware, and Vulnerability Assessment, pages 129–143. Springer, 2006. [4] Cagatay Catal. Software fault prediction: A literature review and current trends. Expert systems with applications, 38(4):4626–4636, 2011. [5] Dong-Kyu Chae, Jiwoon Ha, Sang-Wook Kim, BooJoong Kang, and Eul Gyu Im. Software plagiarism detection: a graph-based approach. In Proceedings of the 22nd ACM international conference on Conference on information & knowledge management, pages 1577–1580. ACM, 2013. [6] Shyam R Chidamber and Chris F Kemerer. A metrics suite for object oriented design. IEEE Transactions on software engineering, 20(6):476– 493, 1994. [7] Tom Fawcett. An introduction to roc analysis. Pattern recognition letters, 27(8):861–874, 2006. [8] Maurice Howard Halstead. Elements of software science, volume 7. Elsevier New York, 1977. [9] C Jones. Strengths and weaknesses of software metrics. AMERICAN PROGRAMMER, 10:44–49, 1997. [10] Hiroshi Kikuchi, Takaaki Goto, Mitsuo Wakatsuki, and Tetsuro Nishino. A source code plagiarism detecting method using alignment with abstract syntax tree elements. In Software Engineering, Artificial Intelligence, Networking and Parallel/Distributed Computing (SNPD), 2014 15th IEEE/ACIS International Conference on, pages 1–6. IEEE, 2014. [11] Charles X Ling, Jin Huang, and Harry Zhang. Auc: a statistically consistent and more discriminating measure than accuracy. In IJCAI, volume 3, pages 519–524, 2003. [12] Kenneth C Louden et al. Programming languages: principles and practices. Cengage Learning, 2011. [13] Thomas J McCabe. A complexity measure. IEEE Transactions on software Engineering, (4):308–320, 1976. [14] Tim Menzies, Jeremy Greenwald, and Art Frank. Data mining static code attributes to learn defect predictors. IEEE transactions on software engineering, 33(1), 2007. [15] Lili Mou, Ge Li, Lu Zhang, Tao Wang, and Zhi Jin. Convolutional neural networks over tree structures for programming language processing. In Thirtieth AAAI Conference on Artificial Intelligence, 2016. [16] Viet Anh Phan, Ngoc Phuong Chau, and Minh Le Nguyen. Exploiting tree structures for classifying programs by functionalities. In Knowledge and Systems Engineering (KSE), 2016 Eighth International Conference on, pages 85–90. IEEE, 2016. [17] Daniel Rodriguez, Israel Herraiz, Rachel Harrison, Javier Dolado, and José C Riquelme. Preliminary comparison of techniques for dealing with imbalance in software defect prediction. In Proceedings of the 18th International Conference on Evaluation and Assessment in Software Engineering, page 43. ACM, 2014. [18] Richard Socher, Eric H Huang, Jeffrey Pennington, Andrew Y Ng, and Christopher D Manning. Dynamic pooling and unfolding recursive autoencoders for paraphrase detection. In NIPS, volume 24, pages 801– 809, 2011. [19] Richard Socher, Jeffrey Pennington, Eric H Huang, Andrew Y Ng, and Christopher D Manning. Semi-supervised recursive autoencoders for predicting sentiment distributions. In Proceedings of the conference on empirical methods in natural language processing, pages 151–161. Association for Computational Linguistics, 2011. [20] Richard Socher, Alex Perelygin, Jean Y Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, Christopher Potts, et al. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the conference on empirical methods in natural language processing (EMNLP), volume 1631, page 1642, 2013. [21] Marina Sokolova and Guy Lapalme. A systematic analysis of performance measures for classification tasks. Information Processing & Management, 45(4):427–437, 2009. [22] Xin Sun, Yibing Zhongyang, Zhi Xin, Bing Mao, and Li Xie. Detecting code reuse in android applications using component-based control flow graph. In IFIP International Information Security Conference, pages 142–155. Springer, 2014. [23] Song Wang, Taiyue Liu, and Lin Tan. Automatically learning semantic features for defect prediction. In Proceedings of the 38th International Conference on Software Engineering, pages 297–308. ACM, 2016. [24] Martin White, Christopher Vendome, Mario Linares-Vásquez, and Denys Poshyvanyk. Toward deep learning software repositories. In Mining Software Repositories (MSR), 2015 IEEE/ACM 12th Working Conference on, pages 334–345. IEEE, 2015.
9cs.NE
arXiv:1803.00829v1 [cs.DS] 2 Mar 2018 Independence number and the number of maximum independent sets in pseudofractal scale-free web and Sierpiński gasket Liren Shana,b , Huan Lia,b , Zhongzhi Zhanga,b b a School of Computer Science, Fudan University, Shanghai 200433, China Shanghai Key Laboratory of Intelligent Information Processing, Fudan University, Shanghai 200433, China Abstract As a fundamental subject of theoretical computer science, the maximum independent set (MIS) problem not only is of purely theoretical interest, but also has found wide applications in various fields. However, for a general graph determining the size of a MIS is NP-hard, and exact computation of the number of all MISs is even more difficult. It is thus of significant interest to seek special graphs for which the MIS problem can be exactly solved. In this paper, we address the MIS problem in the pseudofractal scale-free web and the Sierpiński gasket, which have the same number of vertices and edges. For both graphs, we determine exactly the independence number and the number of all possible MISs. The independence number of the pseudofractal scale-free web is as twice as the one of the Sierpiński gasket. Moreover, the pseudofractal scale-free web has a unique MIS, while the number of MISs in the Sierpiński gasket grows exponentially with the number of vertices. Keywords: Maximum independent set, Independence number, Minimum vertex cover, Scale-free network, Sierpiński gasket, Complex network 1. Introduction An independent set of a graph G with vertex set V is a subset I of V, such that each pair of vertices in I is not adjacent in G. A maximal independent set is an independent set that is not a subset of any other independent set. A largest maximal independent set is called a maximum independent set Email address: [email protected] (Zhongzhi Zhang) Preprint submitted to Theoretical Computer Science March 5, 2018 (MIS). In other words, a MIS is an independent set that has the largest size or cardinality. The cardinality of a MIS is referred to as the independence number of graph G. A graph G is called a unique independence graph if it has a unique MIS [1]. The MIS problem has a close connection with many other fundamental graph problems [2, 3, 4]. For instance, the MIS problem in a graph is equivalent to the minimum vertex cover problem [5] in the same graph, as well as the maximum clique problem in its complement graph [6]. In addition, the MIS problem is also closely related to graph coloring, maximum common induced subgraphs, and maximum common edge subgraphs [7]. In addition to its intrinsic theoretical interest, the MIS problem has found important applications in a large variety of areas, such as coding theory [8], collusion detection in voting pools [9], scheduling in wireless networks [10]. For example, it was shown in [8] that the problem of finding the largest error correcting codes can be reduced to the MIS problem on a graph. In [9] the problem of collusion detection was framed as identifying maximum independent sets. Moreover, finding a maximal weighted independent set in a wireless network is connected with the problem of organizing the vertices of the network in a hierarchical way [11]. Finally, the MIS problem also has numerous applications in mining of graph data [7, 12]. In view of the theoretical and practical relevance, in the past decades the MIS problem has received much attention from different disciplines, e.g., theoretical computer science [13, 14, 15, 16, 17, 18] and discrete mathematics [19, 20, 21]. It is well-known that solving the MIS problem of a generic graph is computationally difficult. Finding a MIS of a graph is a classic NP-hard problem [2, 4], while enumerating all MISs in a graph is even #Pcomplete [22, 23]. Due to the hardness of the MIS problem, exact algorithms for finding a MIS in a general graph take exponential time [24, 25, 26], which is infeasible for moderately sized graphs. For practical applications, many local or heuristic algorithms were proposed to solve the MIS problem for those massive and intractable graphs [27, 28, 29]. Comprehensive empirical study [30] has unveiled that large real networks are typically scale-free [31], with their vertex degree following a power-law distribution P (k) ∼ k−γ . This nontrivial heterogeneous structure has a strong effect on various topological and combinatorial aspects of a graph, such as average distances [32], maximum matchings [33, 34], and dominating sets [35, 36, 37]. Although there have been concerted efforts to understanding the MIS problem in general, there has been significantly less work focused on the MIS problem for power-law graphs [38]. In particular, exact result about the independence number and the number of all MISs in a power-law 2 graph is still lacking, despite the fact that exact result is helpful for testing heuristic algorithms. Moreover, the influence of scale-free behavior on the MIS problem is not well understood, although it is suggested to play an important role in the MIS problem. The ubiquity of power-law phenomenon makes it interesting to uncover the dependence of MISs on the scale-free feature, which is helpful for understanding the applications of MIS problem. In this paper, we study the independence number and the number of maximum independent sets in a scalefree graph, called pseudofractal scale-free web [39, 40], and the Sierpiński gasket. Both networks are deterministic and have the same number of vertices and edges. Note that since determining the independence number and counting all maximum independent sets in a general graph are formidable, we choose these two exactly tractable graphs. This is a fundamental route of research for NP-hard and #P-complete problems. For example, Lovász [41] pointed out that it is of great interest to find specific graphs for which the matching problem can be exactly solved, since the problem in general graphs is NP-hard. By using an analytic technique based on a decimation procedure [42], we find the exact independence number and the number of all possible maximum independent sets for both studied graphs. The independence number of the pseudofractal scale-free web is as twice as the one associated with the Sierpiński gasket. In addition to this difference, there is a unique maximum independent set in the pseudofractal scale-free web, while the number of all maximum independent sets in the Sierpiński gasket increases as an exponential function of the number of vertices. 2. Independence number and the number of maximum independent sets in pseudofractal scale-free web In this section, we study the independence number in the pseudofractal scale-free web, and demonstrate that its maximum independent set is unique. 2.1. Network construction and properties The pseudofractal scale-free web [39, 40] is constructed in an iterative way. Let Gn , n ≥ 1, denote the network after n iterations. When n = 1, G1 is a triangle. For n > 1, Gn is obtained by adding, for every edge (u, v) in Gn−1 a new vertex connected to u and v. Figure 1 illustrates the networks for the first several iterations. By construction, the total number of edges in Gn is En = 3n . 3 G1 G2 G3 Figure 1: The first three iterations of the scale-free graph. (2) (1) (2) C n Bn (1) Bn (1) Cn Gn (1) An (2) An (3) Bn Bn+1 (2) Gn (3) Gn (3) Cn ⇒ (3) (1) Gn (2) Gn Cn+1 An+1 (3) Gn Gn+1 An Figure 2: Alternative construction of the scale-free network. The network displays the striking properties observed in most real-life networks. First, it is scale-free, since the degree of its vertices obeys a power law distribution P (k) ∼ k− ln 3/ ln 2 [39], implying that the probability of a vertex chosen randomly having degree k is approximately k− ln 3/ ln 2 . Moreover, it is small-world, with its average distance growing logarithmically with the number of vertices [39, 43]. Finally, it is highly clustered, with its average clustering coefficient converging to 45 . Of particular interest is the self-similarity of network Gn , which is another ubiquitous property of real networks [44]. For Gn , the three vertices generated at n = 1 have the highest degree, which are called hub vertices, and are denoted by An , Bn , and Cn , respectively. The self-similar feature of the network can be seen from another construction approach [43]. Given the network Gn , Gn+1 can be obtained by joining three copies of Gn at their (θ) hub vertices, see Fig. 2. Let Gn , θ = 1, 2, 3, be three replicas of Gn , and (θ) (θ) (θ) (θ) denote the three hub vertices of Gn by An , Bn , and Cn , respectively. (θ) (1) Then, Gn+1 can be obtained by merging Gn , θ = 1, 2, 3, with An (resp. (1) (2) (3) (2) (3) Cn , An ) and Bn (resp. Bn , Cn ) being identified as the hub vertex An+1 (resp. Bn+1 , Cn+1 ) in Gn+1 . Let Nn stand for the total number of vertices in Gn . By the second 4 construction of the network, Nn satisfies relation Nn+1 = 3Nn − 3, which together with the initial value N1 = 3, is solved to give Nn = (3n + 3)/2. 2.2. Independence number and the number of maximum independent sets Let αn denote the independence number of network Gn . To determine αn , we introduce some intermediate quantities. Since the three hub vertices in Gn are connected to each other, any independent set of Gn contains at most one hub vertex. We classify all independent sets of Gn into two subsets Ω0n and Ω1n . Ω0n represents those independent sets with no hub vertex, while Ω1n denotes the remaining independent sets, with each having exactly one hub vertex. Let Θkn , k = 0, 1, be the subset of Ωkn , where each independent set has the largest cardinality (number of vertices), denoted by αkn . By definition, the independence number of network Gn , n ≥ 1, is αn = max{α0n , α1n }. The two quantities α0n and α1n can be evaluated by using the self-similar structure of the network. Lemma 2.1. For two successive generation networks Gn and Gn+1 , n ≥ 1, α0n+1 = max{3α0n , 2α0n + α1n , α0n + 2α1n , 3α1n }, (1) α1n+1 = max{2α1n + α0n − 1, 3α1n − 1}. (2) Proof. By definition, αkn+1 , k = 0, 1, is the cardinality of an independent set in Θkn+1 . Below, we will show that both Θ0n+1 and Θ1n+1 can be constructed iteratively from Θ0n and Θ1n . Then, α0n+1 and α1n+1 can be obtained from α0n and α1n . We now establish the recursive relations for α0n and α1n . We first prove graphically Eq. (1) . (θ) Notice that Gn+1 consists of three copies of Gn , Gn , θ = 1, 2, 3. By definition, for any independent set χ in Θ0n+1 , the three hub vertices of Gn+1 do not belong to χ, implying that the corresponding six identified hub (θ) vertices of Gn , θ = 1, 2, 3, are not in χ, see Fig. 2. Therefore, we can construct set χ from Θ0n and Θ1n by considering whether the hub vertices of (θ) Gn , θ = 1, 2, 3, are in χ or not. Fig. 3 illustrates all possible configurations of independent sets Ω0n+1 that include Θ0n+1 as subsets. From Fig. 3, we obtain α0n+1 = max{3α0n , 2α0n + α1n , α0n + 2α1n , 3α1n }. Similarly we can prove Eq. (2), the graphical representation of which is shown in Fig. 4. ✷ Lemma 2.2. For network Gn , n ≥ 2, α1n < α0n . 5 (2) (1) (2) C n Bn (1) Bn (1) Gn Cn (1) (1) Gn (2) An (3) Gn (2) Cn (1) Gn (2) An (3) An (3) Cn Bn (3) (3) Gn (2) (1) (2) C n Bn (1) Bn (2) Gn (1) An (3) Bn (2) (1) C n Bn (1) Bn (2) Gn (1) ×3 An (3) (3) (2) An An (3) Cn Gn (2) Cn (2) Gn (1) ×3 (3) (3) Bn Bn (1) Gn (2) An (3) Cn (2) (1) Cn Bn (1) Cn (2) Gn Bn (3) (3) Gn (3) Cn (3) An An An An 3αn0 2αn0 + αn1 αn0 + 2αn1 3αn1 Figure 3: Illustration of all possible configurations of independent sets Ω0n+1 of Gn+1 , (θ) which contain Θ0n+1 . Only the hub vertices of Gn , θ = 1, 2, 3, are shown. Filled vertices are in the independent sets, while open vertices are not. (2) (1) (2) C n Bn (1) Bn (1) Gn (1) (1) Gn (2) An (3) Bn Bn (2) Gn An (3) Gn (2) (1) (2) An (3) Bn (3) Cn (2) Gn (1) ×3 (3) Cn (2) C n Bn (1) Cn An (3) Gn ×3 (3) Cn (3) An An αn0 + 2αn1 − 1 3αn1 − 1 Figure 4: Illustration of all possible configurations of independent sets Ω1n+1 of Gn+1 , which contain Θ1n+1 . Proof. We prove this lemma by mathematical induction on n. For n = 2, we obtain α12 = 2, α02 = 3 by hand. Thus, the basis step holds immediately. Assume that the statement holds for t (t ≥ 2). Then, according to Eq. (1), α0t+1 = max{3α0t , 2α0t +α1t , α0t +2α1t , 3α1t }. By induction hypothesis, we have α0t+1 = 3α0t . (3) In an analogous way, we obtain relation α1t+1 = α0t + 2α1t − 1. (4) By comparing Eqs. (3) and (4) and using the induction hypothesis α1t < α0t , we have α1t+1 < α0t+1 . Thus, the lemma is true for t + 1. ✷ Theorem 2.3. The independence number of network Gn , n ≥ 1, is αn = 3n−1 . (5) Proof. Lemma 2.2 indicates that any maximum independent vertex set of Gn contains no hub vertices. By Eq. (3), we obtain αn+1 = α0n+1 = 3α0n = 3αn . (6) Considering the initial condition α1 = 1, the above equation is solved to give the result. ✷ 6 Corollary 2.4. The largest number of vertices in an independent vertex set of Gn , n ≥ 2, which contains exactly 1 hub vertex, is α1n = 3n−1 − 2n−1 + 1. (7) Proof. By Eqs. (5) and (6), we derive α0n = αn = 3n−1 . Using Eq. (4), we obtain the following recursive equation for α1n : α1n+1 = 2α1n + 3n−1 − 1, (8) which together with the boundary condition α12 = 2 is solved to yield Eq. (7). ✷ Theorem 2.5. For network Gn , n ≥ 2, there is a unique maximum independent set. Proof. Eq. (6) and Fig. 3 mean that for n ≥ 2 any maximum independent set of Gn+1 is actually the union of maximum independent sets, (3) (1) (2) Θ0n , of the three copies of Gn (i.e. Gn , Gn , and Gn ) constituting Gn+1 . (1) Thus, any maximum independent set of Gn+1 is determined by those of Gn , (3) (2) Gn , and Gn . Because the maximum independent set of G2 is unique, there is a unique maximum independent set for Gn for all n ≥ 2. Furthermore, the unique maximum independent set of Gn , n ≥ 2, is in fact the set of all vertices that are generated at the (n − 1)-th iteration. ✷ Theorem 2.5 indicates that the pseudofractal scale-free web is a unique independence graph. 3. Independence number and the number of maximum independent sets in Sierpiński gasket In this section, we consider the independence number and the number of maximum independent sets in the Sierpiński gasket, and compare the results with those of the pseudofractal scale-free web, with an aim to unveil the effect of network structure, in particular the scale-free property, on the independence number and the number of maximum independent sets. 3.1. Construction of Sierpiński gasket The Sierpiński gasket is also constructed iteratively. Let Sn , n ≥ 1, represent the n-generation graph. For n = 1, S1 is an equilateral triangle with three vertices and three edges. For n = 2, perform a bisection of the three edges of S1 forming four smaller replicas of the original equilateral 7 S1 S2 S3 Figure 5: The first three generations of the Sierpiński gasket. (1) An An+1 (1) Bn (1) Sn (1) Cn (2) (3) An An (2) Sn ⇒ Sn+1 (3) Sn Bn+1 (2) Bn (2) (3) C n Bn Cn+1 (3) Cn Figure 6: Alternative construction of the Sierpiński gasket. triangle, and remove the central downward pointing equilateral triangle to get S2 . For n > 2, Sn is obtained from Sn−1 by performing the above two operations for each triangle in Sn−1 . Fig. 5 illustrates the first several iterations of the Sierpiński gaskets Sn for n = 1, 2, 3. Both the number of vertices and the number of edges in the Sierpiński gasket Sn are the same as those for the scale-free network Gn , which are equal to Nn = (3n + 3)/2 and En = 3n , respectively. In contrast to the inhomogeneity of Gn , the Sierpiński gasket is homogeneous. The degree of vertices in Sn is equal to 4, except the topmost vertex An , the leftmost vertex Bn , and the rightmost vertex Cn , the degree of which is 2. These three vertices with degree 2 are called outmost vertices hereafter. Analogously to the scale-free network Gn , the Sierpiński gasket also exhibits the self-similar property, which suggests an alternative construction way of the graph. Given the nth generation graph Sn , the (n + 1)th generation graph Sn+1 can be obtained by amalgamating three copies of Sn at their (θ) outmost vertices, see Fig. 6. Let Sn , θ = 1, 2, 3, represent three copies of (θ) (θ) (θ) (θ) Sn . And denote the three outmost vertices of Sn by An , Bn , and Cn , (θ) respectively. Then, Sn+1 can be obtained by coalescing Sn , θ = 1, 2, 3, (1) (2) (3) with An , Bn , and Cn being the outmost vertices An+1 , Bn+1 , and Cn+1 8 of Sn+1 . 3.2. Independence number In this case without causing confusion, we employ the same notation as those for Gn to study related quantities for the Sierpiński gasket Sn . Let αn be the independence number of Sn . Note that all independent sets of Sn can be sorted into four types: Ω0n , Ω1n , Ω2n , and Ω3n , where Ωkn , k = 0, 1, 2, 3, stands for those independent sets, each of which includes exactly k outmost vertices of Sn . Let Θkn , k = 0, 1, 2, 3, denote the subsets of Ωkn , each independent set in which has the largest cardinality, denoted by αkn . Then, the independence number of the Sierpiński gasket Sn , n ≥ 1, is αn = max{α0n , α1n , α2n , α3n }. Therefore, to determine αn for Sn , one can alternatively determine αkn , k = 0, 1, 2, 3, which can be solved by establishing some relations between them, based on the self-similar architecture of the Sierpiński gasket. Lemma 3.1. For any integer n ≥ 3, the following relations hold. α0n+1 = max{3α0n , α0n + 2α1n − 1, 2α1n + α2n − 2, 3α2n − 3}, (9) α1n+1 = max{2α0n +α1n , α0n +α1n +α2n −1, 3α1n −1, 2α1n +α3n −2, α1n +2α2n −2, 2α2n +α3n −3}, (10) 2 0 1 0 2 1 2 2 1 2 3 αn+1 = max{αn +2αn , αn +2αn −1, 2αn +αn −1, 3αn −2, αn +αn +αn −2, α2n +2α3n −3}, (11) 3 1 1 2 2 3 3 αn+1 = max{3αn , αn + 2αn − 1, 2αn + αn − 2, 3αn − 3}. (12) Proof. This lemma can be proved graphically. Figs. 7-10 illustrate the graphical representations from Eq. (9) to Eq. (12). ✷ (2) (2) Sn (1) (3) Sn Sn 3αn0 (2) Sn (1) Sn (3) (1) Sn Sn αn0 + 2αn1 − 1 (2) Sn ×3 Sn ×3 (3) Sn 2αn1 + αn2 − 2 (1) Sn (3) Sn 3αn2 − 3 Figure 7: Illustration of all possible configurations of independent sets Ω0n+1 of Sn+1 , (θ) which contain Θ0n+1 . Only the outmost vertices of Sn , θ = 1, 2, 3, are shown. Filled vertices are in the independent sets, while open vertices are not. Lemma 3.2. For arbitrary n ≥ 2, α0n + 1 = α1n = α2n = α3n − 1. 9 (2) (2) Sn (1) (3) Sn (2) Sn (1) Sn 2αn0 + αn1 (3) Sn Sn (2) Sn ×2 (1) (3) Sn αn0 + αn1 + αn2 − 1 (2) Sn (1) Sn (3) Sn 3αn1 − 1 (2) Sn (1) Sn (3) Sn 2αn1 + αn3 − 2 Sn ×2 (1) Sn (3) Sn αn1 + 2αn2 − 2 Sn 2αn2 + αn3 − 3 Figure 8: Illustration of all possible configurations of independent sets Ω1n+1 of Sn+1 , which contain Θ1n+1 . Note that here we only illustrate the independent sets, each of which only includes An+1 but excludes Bn+1 , and Cn+1 . Since An+1 , Bn+1 , and Cn+1 are equivalent to each other, we omit other independent sets, including Bn+1 (resp. Cn+1 ) but excluding An+1 and Cn+1 (resp. An+1 and Bn+1 ) . (2) (2) Sn (1) Sn (2) Sn (3) (1) Sn αn0 + 2αn1 (3) Sn (2) Sn Sn αn0 + 2αn2 − 1 (1) (3) Sn (1) Sn (2) Sn (3) Sn 2αn1 + αn2 − 1 (2) Sn ×2 (1) Sn (3) Sn 3αn2 − 2 Sn ×2 (1) Sn Sn αn1 + αn2 + αn3 − 2 (3) Sn αn2 + 2αn3 − 3 Figure 9: Illustration of all possible configurations of independent sets Ω2n+1 of Sn+1 , which contain Θ2n+1 . Note that here we only illustrate the independent sets, each of which includes two outmost vertices Bn+1 and Cn+1 , but excludes the outmost vertex An+1 . Similarly, we can illustrate those independent sets, each including An+1 and Cn+1 (resp. An+1 and Bn+1 ), but excluding Bn+1 (resp. Cn+1 ) . (2) (2) Sn (1) (3) Sn Sn 3αn1 (2) Sn (1) Sn αn1 + (3) Sn (1) −1 2αn2 Sn 2αn2 (2) Sn ×3 (3) Sn + αn3 Sn ×3 −2 (1) Sn 3αn3 (3) Sn −3 Figure 10: Illustration of all possible configurations of independent sets Ω3n+1 of Sn+1 , which contain Θ3n+1 . 10 Proof. We prove this lemma by induction. For n = 2, it is easy to check that α02 = 1, α12 = 2, α22 = 2, and α32 = 3. Thus, the result holds for n = 2. Let us suppose that the statement is true for t, t ≥ 2. For t + 1, by induction assumption and Lemma 3.1, it is not difficult to check that the relation α0t+1 + 1 = α1t+1 = α2t+1 = α3t+1 − 1 is true. ✷ Theorem 3.3. The independence number of the Sierpiński gasket Sn , n ≥ n−1 2, is αn = 3 2 +3 . Proof. According to Lemmas 3.1 and 3.2, we obtain αn+1 = α3n+1 = 3α3n − n−1 3 = 3αn − 3. Considering α2 = 3, it is obvious that αn = 3 2 +3 holds for all n ≥ 2. ✷ Theorems 2.3 and Theorem 3.3 show that the independence number of the Sierpiński gasket Sn is larger than the one corresponding to the pseudofracal scale-free web Gn , with the former being as half as the latter for large n. Corollary 3.4. The largest possible number of vertices in an independent vertex set of Sn , n ≥ 2, which contains exactly 0, 1 and 2 outmost vertices, n−1 n−1 n−1 is α0n = 3 2 −1 , α1n = 3 2 +1 , and α2n = 3 2 +1 , respectively. n−1 Proof. Theorem 3.3 shows α3n = αn = 3 2 +3 . From Lemma 3.2, we obtain α0n = α3n − 2 and α1n = α2n = α3n − 1. Then, the results are obtained immediately. ✷ 3.3. The number of maximum independent sets In comparison with the scale-free network Gn with a unique maximum independent set, the number of maximum independent sets of Sn increases exponentially with the number of vertices. Theorem 3.5. For n ≥ 2, the number of maximum independent sets of the Sierpiński gasket Sn is 2 3n−2 −1 2 . Proof. Let xn denote the number of maximum independent sets of the Sierpiński gasket Sn . Let yn be the number of independent sets of Sn with maximum number of vertices, including only An but excluding Bn , and Cn . For the initial condition n = 2, we have x2 = 1, y2 = 1. For n ≥ 2, we can prove that the two quantities xn and yn obey the following relations: xn+1 = yn3 + x3n , 11 (13) yn+1 = yn3 + xn yn2 . (14) We first prove Eq. (13). By definition, xn is the number of different maximum independent sets for Sn , each of which contains all the three outmost vertices of Sn . According to Lemma 3.2 and Fig. 10, the two configurations in Fig. 10 which maximize α3n+1 (and thus αn+1 ) are when (θ) Sn , θ = 1, 2, 3, contains exactly one outmost vertex and when it contains the three outmost vertices. Then, we can establish Eq. (13) by using the rotational symmetry of the Sierpiński gasket. Eq. (14) can be proved analogously by using Lemma 3.2 and Fig. 8. Since x2 = 1 and y2 = 1, Eqs. (13) and (14) show that xn = yn for all n ≥ 2. Then, we obtain a recursion relation for xn as xn+1 = 2x3n , which together with the initial value x2 = 1 is solved to yield xn = 2 3n−2 −1 2 . ✷ Acknowledgements This work is supported by the National Natural Science Foundation of China under Grant No. 11275049. References [1] G. Hopkins, W. Staton, Graphs with unique maximum independent sets, Discrete Math. 57 (1985) 245–251. [2] J. M. Robson, Algorithms for maximum independent sets, J. Algorithms 7 (1986) 425–440. [3] P. Berman, M. Fürer, Approximating maximum independent set in bounded degree graphs, in: Proceedings of the 1994 Annual ACMSIAM Symposium on Discrete Algorithms, pp. 365–371. [4] M. M. Halldórsson, J. Radhakrishnan, Greed is good: Approximating independent sets in sparse and bounded-degree graphs, Algorithmica 18 (1997) 145–163. [5] R. M. Karp, Reducibility among combinatorial problems, in: Complexity of Computer Computations, Springer, 1972, pp. 85–103. [6] P. M. Pardalos, J. Xue, The maximum clique problem, J. Global Optim. 4 (1994) 301–328. 12 [7] Y. Liu, J. Lu, H. Yang, X. Xiao, Z. Wei, Towards maximum independent sets on massive graphs, Proceedings of the 2015 International Conference on Very Large Data Base 8 (2015) 2122–2133. [8] S. Butenko, P. Pardalos, I. Sergienko, V. Shylo, P. Stetsyuk, Finding maximum independent sets in graphs arising from coding theory, in: Proceedings of the 2002 ACM symposium on Applied computing, ACM, pp. 542–546. [9] F. Araujo, J. Farinha, P. Domingues, G. C. Silaghi, D. Kondo, A maximum independent set approach for collusion detection in voting pools, J. Parallel Distrib. Comput. 71 (2011) 1356–1366. [10] C. Joo, X. Lin, J. Ryu, N. B. Shroff, Distributed greedy approximation to maximum weighted independent set for scheduling with fading channels, IEEE/ACM Trans. Netw. 24 (2016) 1476–1488. [11] S. Basagni, Finding a maximal weighted independent set in wireless networks, Telecom. Syst. 18 (2001) 155–168. [12] L. Chang, W. Li, W. Zhang, Computing a near-maximum independent set in linear time by reducing-peeling, in: Proceedings of the 2017 ACM International Conference on Management of Data, ACM, pp. 1181– 1196. [13] C. Murat, V. T. Paschos, A priori optimization for the probabilistic maximum independent set problem, Theoret. Comput. Sci. 270 (2002) 561–590. [14] M. Xiao, H. Nagamochi, Confining sets and avoiding bottleneck cases: A simple maximum independent set algorithm in degree-3 graphs, Theoret. Comput. Sci. 469 (2013) 92–104. [15] G. Agnarsson, M. M. Halldórsson, E. Losievskaja, SDP-based algorithms for maximum independent set problems on hypergraphs, Theoret. Comput. Sci. 470 (2013) 1–9. [16] W.-K. Hon, T. Kloks, C.-H. Liu, H.-H. Liu, S.-H. Poon, Y.-L. Wang, On maximum independent set of categorical product and ultimate categorical ratios of graphs, Theoret. Comput. Sci. 588 (2015) 81–95. [17] V. Lozin, J. Monnot, B. Ries, On the maximum independent set problem in subclasses of subcubic graphs, J. Discrete Algorithms 31 (2015) 104–112. 13 [18] J. Chuzhoy, A. Ene, On approximating maximum independent set of rectangles, in: Proceedings of IEEE 2016 Annual Symposium on Foundations of Computer Science, IEEE, pp. 820–829. [19] E. Čenek, L. Stewart, Maximum independent set and maximum clique algorithms for overlap graphs, Discrete Appl. Math. 131 (2003) 77–91. [20] M. Xiao, H. Nagamochi, An exact algorithm for maximum independent set in degree-5 graphs, Discrete Appl. Math. 199 (2016) 137–155. [21] R. Mosca, A sufficient condition to extend polynomial results for the maximum independent set problem, Discrete Appl. Math. 216 (2017) 281–289. [22] L. Valiant, The complexity of computing the permanent, Theor. Comput. Sci. 8 (1979) 189–201. [23] L. Valiant, The complexity of enumeration and reliability problems, SIAM J. Comput. 8 (1979) 410–421. [24] F. V. Fomin, D. Kratsch, Exact exponential algorithms, Springer Science & Business Media, 2010. [25] R. E. Tarjan, A. E. Trojanowski, Finding a maximum independent set, SIAM J. Compu. 6 (1977) 537–546. [26] P. Beame, R. Impagliazzo, A. Sabharwal, The resolution complexity of independent sets and vertex covers in random graphs, Comput. Complex. 16 (2007) 245–297. [27] D. V. Andrade, M. G. Resende, R. F. Werneck, Fast local search for the maximum independent set problem, J. Heuristics 18 (2012) 525–547. [28] J. Dahlum, S. Lamm, P. Sanders, C. Schulz, D. Strash, R. F. Werneck, Accelerating local search for the maximum independent set problem, in: Proceedings of the 2016 International Symposium on Experimental Algorithms, Springer, pp. 118–133. [29] S. Lamm, P. Sanders, C. Schulz, D. Strash, R. F. Werneck, Finding near-optimal independent sets at scale, J. Heuristics 23 (2017) 207–229. [30] M. E. J. Newman, The structure and function of complex networks, SIAM Rev. 45 (2003) 167–256. 14 [31] A. Barabási, R. Albert, Emergence of scaling in random networks, Science 286 (1999) 509–512. [32] F. Chung, L. Lu, The average distances in random graphs with given expected degrees, Proc. Natl. Acad. Sci. 99 (2002) 15879–15882. [33] Y.-Y. Liu, J.-J. Slotine, A.-L. Barabási, Controllability of complex networks, Nature 473 (2011) 167–173. [34] Z. Zhang, B. Wu, Pfaffian orientations and perfect matchings of scalefree networks, Theoret. Comput. Sci. 570 (2015) 55–69. [35] J. C. Nacher, T. Akutsu, Dominating scale-free networks with variable scaling exponent: heterogeneous networks are not difficult to control, New J. Phys. 14 (2012) 073005. [36] M. Gast, M. Hauptmann, M. Karpinski, Inapproximability of dominating set on power law graphs, Theoret. Comput. Sci. 562 (2015) 436–452. [37] L. Shan, H. Li, Z. Zhang, Domination number and minimum dominating sets in pseudofractal scale-free web and sierpiński graph, Theoret. Comput. Sci. 677 (2017) 12–30. [38] A. Ferrante, G. Pandurangan, K. Park, On the hardness of optimization in power-law graphs, Theoret. Comput. Sci. 393 (2008) 220–230. [39] S. N. Dorogovtsev, A. V. Goltsev, J. F. F. Mendes, Pseudofractal scalefree web, Phys. Rev. E 65 (2002) 066122. [40] Z. Z. Zhang, Y. Qi, S. G. Zhou, W. L. Xie, J. H. Guan, Exact solution for mean first-passage time on a pseudofractal scale-free web, Phys. Rev. E 79 (2009) 021127. [41] L. Lovász, M. D. Plummer, Matching Theory, volume 29 of Annals of Discrete Mathematics, North Holland, New York, 1986. [42] M. Kneževic, J. Vannimenus, Large-scale properties and collapse transition of branched polymers: Exact results on fractal lattices, Phys. Rev. Lett. 56 (1986) 1591. [43] Z. Z. Zhang, S. G. Zhou, L. C. Chen, Evolving pseudofractal networks, Eur. Phys. J. B 58 (2007) 337–344. [44] C. Song, S. Havlin, H. Makse, Self-similarity of complex networks, Nature 433 (2005) 392–395. 15
8cs.DS
A note on sub-orthogonal lattices João Eloir Strapasson arXiv:1608.01172v2 [cs.IT] 8 Aug 2017 School of Applied Sciences Abstract It is shown that, given any k-dimensional lattice Λ, there is a lattice sequence Λw , w ∈ Z, with a sub-orthogonal lattice Λo ⊂ Λ, converging to Λ (up less equivalence), also we discuss the conditions for the faster convergence. Keywords: Subortoghonal Lattice, Dense Packing, Spherical Code 2010 MSC: 11H31, 11H06, 94A15 1. Introduction A large class of the problems in the coding theory are related to the properties of the lattices, in special, with sublattices generated by orthogonal basis (suborthogonal lattices). Several authors investigated the relationship of the suborthogonal with a spherical codes, and with a q-ary codes, see [1, 8, 13, 16, 17, 18]), but, of course, that does not restrict to these problems ([2, 4, 5, 12]. In general, the problems in the lattices are concentrated in obtaining certain parameters: as the shortest vector, the packing radius and the packing density, the radius coverage radius and the coverage density. The points of the lattices are interpreted as elements of a code. Thus finding efficient coding and decoding schemes is essential. There are several schemes in the literature that establish the relationship of the linear codes with the lattices, a good reference is [7]. Preprint submitted to Elsevier February 14, 2018 This article is organized as follows. In the Section 2, we present the notations, definitions and small properties. In the Section 3, we present a new scheme of obtaining lattices with a sub-orthogonal. In the Section 4, we present a case study for special lattices: Dn and En (n = 7, 8) and Leech lattice Λ24 . 1.1. Background Definitions and Results A lattice in Rn is an the discrete additive subgroup of Rn , Λ, which has a generator matrix with full rank, k × n, B, e.g, v ∈ Λ ↔ v = ut B (u ∈ Zk , k is said rank of Λ. The determinant of a lattice is det(Λ) = det(G), there G = p BBt is a Gram matrix of the lattice Λ and the volume of the lattice is det(Λ) (volume of the parallelotope generate for rows of B). The minimum norm of the Lattice Λ, ρ(Λ), is min{kvk; v ∈ Λ and v 6= 0} and center density ρ(Λ)n packing of Λ is δΛ = n . Two lattices Λ1 and Λ2 , with generator 2 vol(Λ) matrices B1 and B2 are equivalence if, only if B1 = c UB2 O, there c ∈ R, U is uni-modular matrix (integer, k × k matrix with det(U) = ±1) and O is the orthogonal, n × n matrix (OOt = In , In identity matrix n × n). Dual lattice of Λ is a lattice, Λ∗ , obtained for all vectors u ∈ span(B) (there span(B) is a vector space generated by the rows of B) with that u · v ∈ Z, ∀v ∈ Λ, the generator matrix of Λ∗ is B∗ = (BBt )−1 B, in particular, B∗ = B−t if n = k. A sub-lattice, Λ′ , is a subset of Λ which is also lattice, if Λ′ has generator matrix is formed by the orthogonal row vectors we will say that it is a sub-orthogonal lattice. Since the lattice is a group, remember that the quotiente of a lattice Λ by a sublattice Λ′ , Λ , Λ′ is as a finite abelian group with M elements, where M is the ratio of the volume of Λ′ by the volume of Λ, e.g., M = vol(Λ′ ) . vol(Λ) The M elements of the lattice Λ, can be seen as an orbit of vector in the k-dimensional torus Λ . Λ′ This essentially establishes the relationship with a central spherical class codes, as well as a class of the linear codes track construction “A” and similar constructions, see more details in [7]. 2 2. Suborthogonal sequences Consider the lattice Λ ⊂ Rn , of rank n, contain an orthogonal sub-lattice, Λo ⊂ Λ, such that Λo is equivalence to Zn , e.g., the generator matrix of the Λo is c O, with OOt = In . Let B and B∗ = B−t the generator matrices of the Λ and Λ∗ (respectively). Assuming that B∗ has integer entries, Then the lattice, Λ, with generator matrix B = adj(B∗ ) = det(B∗ )B∗ −t has a sub-orthogonal lattice, Λo , with generator matrix B∗ B = det(B∗ )In . The ratio of the volume measured quantities points and in this case it is vol(Λo ) vol(Λ) = det(B∗ B) det(B) = det(B∗ ). We want to build codes with a large number of the points. And we observe that as we want to increase the number of points we must increase the determinant of the matrix of the dual lattice. Proposition 1. Let Λ be the n-dimensional lattice and Λ∗ be its dual, with generator matrices B and B∗ (respectively). Assuming that B∗ has integer entries. Define Λ∗w with generator matrix Bw∗ = wB∗ + P (w is integer and P is an integer matrix any). Then the lattices Λ∗w and Λw with generator matrices Bw∗ and Bw = adj(Bw∗ ) (respectively) to satisfy w1 Λ∗w −→ Λ∗ (w → ∞) and by continuity of the matrix inversion process det( 11 B∗ ) Λw −→ Λ (w → ∞). w w Proof. The proof is trivial by the fact that the convergence of each entry of the generator matrix, and the convergence of the generating matrix defines convergence groups. Recalling that the cardinality of the points in the quotient, Λ∗w is a polynomial, specifically M(w) = det(Λ∗w ). Corollary 1. Let Λ be the n-dimensional lattice and Λ∗ be its dual, with a generator matrices B and B∗ (respectively). Define Λ∗w with generator matrix Bw∗ = wB∗ + P (P = ⌊wB∗ ⌉ − wB∗ ), in other words, Bw∗ = ⌊wB∗ ⌉ (rounded entries). Then the lattices Λ∗w and Λw with generator matrices Bw and Bw = adj(Bw∗ ) (respectively) to satisfy w1 Λ∗w −→ Λ∗ (w → ∞) and by continuity of the matrix inversion process det( 11 B∗ ) Λw −→ Λ (w → ∞). w w The corollary allows to extend the use of the proposition for whose lattices dual have not, up less equivalence, integer generator matrix. In the following propositions we establish the speed of the convergence: 3 Proposition 2 (faster dual convergence). Let Λ∗w and Λw as in Proposition 1. Then faster convergence is obtained minimizing inputs PP t , considering P = S B (S is antisymmetric matrix n × n, whose parameters will be minimized). Of course, if P is identically zero is the best convergence, because there is no error. Proof. A better sequence is that in which the vectors is closest sizes and angles of the desired lattice. And so, we must analyse the sequence formed by Gram matrix 1/w 2 Gw∗ = (1/wBw )(1/wBw )t . Therefore 1/w 2 Gw∗ = = = = (1/wBw∗ )(1/wBw∗ )t (B∗ + 1/wP)(B∗ + 1/wP)t (B∗ + 1/wP)(B∗t + 1/wP t ) B∗ B∗ t + 1/w(PB∗ t + B∗ P t ) + 1/w 2 PP t From now on we will say that convergence is linear if PB∗ t + B∗ P t = αB∗ B∗ t and quadratic if PB∗ t + B∗ P t = αB∗ B∗ t and constant if P = 0 (P identically zero), unless a change in w variable, we can assume α = 0, in these conditions: PB∗ t + B∗ P t = 0 ⇒ PB∗ t = −B∗ P t = −(PB∗ t )t (skew-symmetric). Then P = SB∗ −t = SB, there any antisymmetric S with P = SB integer matrix. In the case, of convergence is quadratic, the convergence coefficient also depends on the inputs PP t , which must be minimized. Proposition 3 (faster convergence). Let Λ∗w and Λw as in Proposition 1. Then faster convergence is obtained minimizing inputs BP t GPBt , considering P = B∗ S (S is antisymmetric matrix n × n, whose parameters will be minimized). Of course, if P is identically zero is the best convergence, because there is no error. Proof. We recall that the inverse of a sum of matrix with matrix P∞ identity n −1 can be calculated by Neumann series ([11]) (A + I) = n=0 (−A) ), so the dual generator matrix of lattice sequence is: Bw∗ = wB∗ + P = wB∗ (In B∗ −1 P) = wB∗ (In + 1/wBt P), 4 and the inverse transpose is: Bw = = = = = = = ≈ adj(Bw∗ ) det(Bw∗ )Bw∗ −t (β := det(Bw∗ )) β(wB∗ (In + 1/wBt P))−t β/w n B∗ −t (In + 1/wBt P)−t β/w n B(In + 1/wP t B)−1 β/w n B(In + 1/wP t B)−1 β/w n B(In − 1/wP t B + (1/wP t B)2 − 1/wP t B)3 + · · · ) β/w n B(In − 1/wP t B). Therefore, the gram matrix approximated is: Gw ≈ = = = (β/w)2B(In − 1/wP t B)(In − 1/wP t B)t Bt (β/w)2B(In − 1/wP t B)(In − 1/wBt P)Bt (β/w)2B(In − 1/w(P t B + Bt P) + 1/w 2P t GP)Bt (β/w)2(G − 1/wB(P t B + Bt P)Bt + 1/w 2BP t GPBt ). It is desirable that: P t B + Bt P = 0 ⇒ Bt P = −P t B = −(Bt P)t = S(skew-symmetric), from which it follows that P = B∗ S, there any antisymmetric S with P = B∗ S integer matrix. In the case, of convergence is quadratic, the convergence coefficient also depends on the inputs BP t GPBt , which must be minimized. The structure of the group obtained by the quotient of the lattice sequence, Λw , by their respective an orthogonal sublattice can be determined and extended, applying the Theorem 2.4.13 in [6]. In particular, B∗ is lower triangular matrix and P = Cn = (ci,j ) (cyclic perturbation), where ci,j = 1 if j = i + 1 and ci,j = 0 otherwise, the quotient is cyclic group although convergence is not nearly quadratic. Lattices of rank n that, up less equivalence, are sublattices the integer lattice Zn , play an interesting role with regard to the convergence as discussed 5 below with case study, next section. 3. Case study In this section, we present a construction applied to the special cases: Dn , En (n = 7, 8) and Λ24 and we show the best perturbations found. Although the null perturbation is optimal, it is never associated with cyclic quotient groups, it does not offer the best solution in terms of the spherical codes, as we shall see below. Being cyclical and optimal is unlikely. In addition, the results presented here complement results obtained in [13] (not considering the initial vector problem) and extend and simplify the results obtained in [1]. 3.1. The root lattice Dn (n ≥ 3) We consider the generate matrix of D∗n as Dn∗ and good perturbation is Pn :  0 1 0 ···  −1 0 1 ···     0 −1 0 . . . 2 0 · · · 0    .. .. . . . . .. . . .. ..  and P =  .  . . . . . . . n    .. 0 · · · 2 0 . 0 0 0    1 ··· 1 1 0 0 ··· 0 2 0 ···  0  . Dn∗ =  ..  0  1  0 0  −1 0 0 ···  0 1  0 0 0   0 0 0  . . .. ..  . . .  . (1)   0 1 0  −1 0 1  0 0 1 0 The good perturbation is Pn in (1), this case the quotient is cyclic case odd n, the performance is illustrated in Tables 1 and 2. 3.2. The root lattice En Up less equivalence, assuming that E∗7 , E∗8,1 and E∗8,2 are generated by ∗ ∗ matrices E7∗, E8,1 and E8,2 and the good perturbation P7 , P8,1 and P8,2 . 6 M(w) δ(Λw ) Group M(w) δ(Λw ) Group M(w) δ(Λw ) Group 4 0.176777 Z2 ⊕ Z2 7 0.133631 Z7 3 0.0721688 Z3 32 0.176777 Z2 ⊕ Z4 ⊕ Z4 38 0.162221 Z38 26 0.0969021 Z26 108 0.176777 Z3 ⊕ Z6 ⊕ Z6 117 0.169842 Z117 93 0.116923 Z93 256 0.176777 Z4 ⊕ Z8 ⊕ Z8 268 0.172774 Z268 228 0.129349 Z228 500 0.176777 Z5 ⊕ Z10 ⊕ Z10 515 0.174183 Z515 455 0.137602 Z455 864 0.176777 Z6 ⊕ Z12 ⊕ Z12 882 0.174964 Z882 798 0.143442 Z798 1372 0.176777 Z7 ⊕ Z14 ⊕ Z14 1393 0.175439 Z1393 1281 0.147780 Z1281 2048 0.176777 Z8 ⊕ Z16 ⊕ Z16 2072 0.175750 Z2072 1928 0.151126 Z1928 2916 0.176777 Z9 ⊕ Z18 ⊕ Z18 2943 0.175964 Z2943 2763 0.153783 Z2763 4000 0.176777 Z10 ⊕ Z20 ⊕ Z20 4030 0.176117 Z4030 3810 0.155943 Z3810 Table 1: Show performance in 3-dimensional case, for perturbations 0n , Pn and Cn respectively. δ(Λw ) δ(D3 ) δ(Λw ) w) w) Group δ(Λ Group Group δ(Λ Group δ(D4 ) δ(D5 ) δ(D6 ) 0.7559 Z7 1. Z3 ⊕ Z6 0.6718 Z41 0.675 Z10 ⊕ Z10 0.9177 Z38 1. Z9 ⊕ Z18 0.8732 Z682 0.8576 Z17 ⊕ Z170 0.9608 Z117 1. Z19 ⊕ Z38 0.9371 Z4443 0.9269 Z74 ⊕ Z370 0.9774 Z268 1. Z33 ⊕ Z66 0.9631 Z17684 0.9565 Z65 ⊕ Z2210 0.9853 Z515 1. Z51 ⊕ Z102 0.9759 Z52525 0.9714 Z202 ⊕ Z2626 0.9897 Z882 1. Z73 ⊕ Z146 0.9831 Z128766 0.9799 Z145 ⊕ Z10730 0.9924 Z1393 1. Z99 ⊕ Z198 0.9875 Z275807 0.9851 Z394 ⊕ Z9850 0.9942 Z2072 1. Z129 ⊕ Z258 0.9904 Z534568 0.9885 Z257 ⊕ Z33410 0.9954 Z2943 1. Z163 ⊕ Z326 0.9924 Z959409 0.9909 Z650 ⊕ Z26650 0.9963 Z4030 1. Z201 ⊕ Z402 0.9938 Z1620050 0.9926 Z401 ⊕ Z81002 Table 2: Show performance in 3 to 6-dimensional case, for perturbations Pn .  1  0   0  ∗ E7 =  0  0  0  0 0 0 0 −1 0 1   0   −1 1 0 0 −1 −1 1      0 1 0 −1 −1 0  0    0 0 1 0 −1 −1 , P7 =  −1   0 0 0 0 2 0 −2    0 0 0 0 0 2 0  7 1 0 0 0 0 0 2 1 0 0 0 0 0 1 0 0 0 −1 0 0 0 0 −1 −1 0 −1 0 0 −1 1 −1 0 0 1 1 1 0 0 0 −1 1 −1 0   0   0  0   −1  1  0  ∗ E8,1 1  0   0  0  = 0   0  0  0 0 0 0 −1 0 0 1 0 −1 −1 0 0 0 1 0 −1 −1 0 0 0 2 0 −2 0 0 0 0 2 0 0 0 0 0 0 2 0 0 0 0 0 0 1  −1   0  0  = 0   0  0  0 0 1 1     0 −1 0 0   0 1 0 1   0 0 0  −1   , P8,1 =   0 −1 −1 −2     0 0 0 1   −1 −1 0 0   1 1 0 2 1 1 0 0 −1 −1 1  ∗ E8,2 1 1 1 1 1 1 1 −7 1 1 1 1 1 1 0 2 2 2 2 2 0 0 2 2 2 2 0 0 0 2 2 2 0 0 0 0 2 2 0 0 0 0 0 2 0 0 0 0 0 0     −5    −10  −8    , P8,2 −6    −4   −2   4  −1  −1   −1  −1  = −1   −1  −1  0 0 0 0 0 0 0 0 0 0 −1 0 0 −1 0 0 0 0 0 0 0 1 1 1 0 1 0 0 −1 0 −1 0 0 −1 0 0 0 0 1 −1 0 0 0 0 0 −1 −1 0 1 0 0 0 −1 −1 0 1 0 0 0 −1 0 0 0 −1 1 −1 0 −1 1 −1 0 0 0 0 0 1 −1 0 0 0 0 0   0   0  0   −1   0  1  0  0  0   0  0  . 0   0  0  0 The performance is illustrated in Table 3 (note that the density ratio is deployed close and the amount of the associated points are: 1.664.641.200 points for dual lattice 10 E∗8,1 +P8,1 and 11.430.630.576 for dual lattice 9 E∗8,2 + P8,2 (very more points in the second case). The Table 4, illustrates the performance applied in spherical codes, details in [13], the non-null perturbation is better in the case of E8,1 representation, moreover, point out that the performance is similar to the second representation with null perturbation. 8 δ(Λw ) δ(E7 ) δ(Λw ) δ(Λw ) Group Group Group δ(E8,1 ) δ(E8,2 ) 0.2346 Z2 ⊕ Z68 0.1204 Z2 ⊕ Z78 0.2706 Z2 ⊕ Z2 ⊕ Z364 0.4161 Z2 ⊕ Z1552 0.4022 Z4 ⊕ Z2316 0.5065 Z2 ⊕ Z4 ⊕ Z15128 0.5966 Z2 ⊕ Z15468 0.622 Z6 ⊕ Z26154 0.6918 Z2 ⊕ Z6 ⊕ Z189252 0.7208 Z2 ⊕ Z92192 0.7521 Z8 ⊕ Z165912 0.7993 Z2 ⊕ Z8 ⊕ Z1251376 0.8005 Z2 ⊕ Z391540 0.8284 Z10 ⊕ Z729030 0.8616 Z2 ⊕ Z10 ⊕ Z5612060 0.8522 Z2 ⊕ Z1313328 0.8754 Z12 ⊕ Z2495268 0.8997 Z2 ⊕ Z12 ⊕ Z19429704 0.8869 Z2 ⊕ Z3708572 0.9059 Z14 ⊕ Z7137186 0.9243 Z2 ⊕ Z14 ⊕ Z55966708 0.911 Z2 ⊕ Z9191488 0.9266 Z16 ⊕ Z17842224 0.9411 Z2 ⊕ Z16 ⊕ Z140558432 0.9284 Z2 ⊕ Z20572452 0.9412 Z18 ⊕ Z40176702 0.9529 Z2 ⊕ Z18 ⊕ Z317517516 0.9412 Z2 ⊕ Z42432080 0.952 Z20 ⊕ Z83232060 0.9615 Z2 ⊕ Z20 ⊕ Z659296120 Table 3: Show performance in 7 to 8-dimensional case, for representations E∗7 , E∗8,1 E∗8,2 and perturbations P7 , P8,1 and P8,2 . 3.3. The Leech lattices Λ24 The laminate lattice is generally dense in their respective dimensions in special dimensions, n = 9, 15, 16, 19, 20, 21, 24 admit integer representation, un less equivalence, and in these cases can analyse the fast convergence, consider n = 24 the matrix generator of Leech Lattice, unless equivalence is: 4 0  00  2  0  0  0  0  0  0 0  −1   2  1  1  −1  −1  2  1  1 1 1 2  L24,1 = 0 4 0 0 2 2 0 2 0 0 0 0 1 1 1 2 1 −1 −1 1 1 −1 0 2 0 0 4 0 2 2 2 0 0 0 0 0 1 −1 −2 1 2 −1 1 2 1 2 1 0 0 0 0 4 0 2 2 2 0 0 0 0 2 1 −1 1 −1 −1 0 1 0 2 0 −1 0 0 0 0 2 0 2 2 0 0 0 0 1 0 0 0 0 1 0 0 1 0 0 2 0 0 0 0 0 2 0 0 0 0 0 0 0 1 0 0 0 1 0 0 0 1 −1 1 0 0 0 0 0 0 2 0 0 0 0 0 0 0 1 0 1 1 1 1 0 0 −1 −1 0 0 0 0 0 0 0 2 0 0 0 0 0 0 0 1 0 1 1 0 0 1 0 −1 0 0 0 0 0 0 0 0 4 0 0 0 −1 2 1 1 2 0 0 0 2 2 0 0 0 0 0 0 0 0 0 0 0 4 0 0 1 1 1 2 0 2 0 0 0 0 −2 2 9 0 0 0 0 0 0 0 0 0 0 4 0 1 −1 2 1 0 0 2 0 0 2 2 −2 0 0 0 0 0 0 0 0 0 0 0 4 2 1 −1 1 0 0 0 2 2 0 0 2 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 1 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0, 0 0 0 0 0 0 0 0 0 0 1 or 0 P8,i E8,1 Distance M − Points 0.707107 4096 0.707107 104976 0.500000 1048576 0.415627 6250000 0.366025 26873856 0.306802 92236816 0.270598 268435456 0.241845 688747536 0.218508 1600000000 Distance M − Points 0.839849 9264 0.641669 156924 0.509472 1327296 0.419589 7290300 0.355527 29943216 0.307914 99920604 0.271283 285475584 0.242296 723180636 0.218821 1664641200 E8,2 Distance M − Points 0.707107 65536 0.500000 1679616 0.382683 16777216 0.309017 100000000 0.258819 429981696 0.222521 1475789056 0.195090 4294967296 0.173648 11019960576 0.156434 25600000000 Distance M − Points 0.639702 121024 0.468092 2271024 0.366403 20022016 0.299852 112241200 0.253223 466312896 0.218878 1567067824 0.192596 4497869824 0.171869 11430630576 0.155124 26371844800 Table 4: Show spherical code performance 8-dimensional case, for different representations.  L24,2 = 4 −4  44  4  4  4  2  4  4  4  2  4  2  2  2  4  2  2  2  00 0 −3 4 4 0 0 0 0 0 2 0 0 0 2 0 2 0 0 0 0 0 2 2 0 0 1 0 0 4 0 0 0 0 2 0 0 0 2 0 0 2 0 0 2 0 0 2 0 0 1 0 0 0 4 0 0 0 2 0 0 0 2 0 0 0 2 0 0 2 0 2 0 0 1 0 0 0 0 4 0 0 2 0 0 0 0 0 2 2 2 0 2 2 2 2 0 0 1 0 0 0 0 0 4 0 2 0 0 0 0 0 2 0 0 0 0 2 0 0 0 0 1 0 0 0 0 0 0 4 2 0 0 0 0 0 0 2 0 0 0 0 2 0 0 0 1 0 0 0 0 0 0 0 2 0 0 0 0 0 0 0 2 0 2 0 0 0 0 0 1 0 0 0 0 0 0 0 0 4 0 0 2 0 2 2 2 0 2 2 2 2 2 2 1 0 0 0 0 0 0 0 0 0 4 0 2 0 2 0 0 0 2 0 0 0 2 0 1 0 0 0 0 0 0 0 0 0 0 4 2 0 0 2 0 0 0 2 0 0 0 2 1 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 2 0 0 0 2 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 4 2 2 2 0 0 0 0 2 2 2 1 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 0 0 0 0 2 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 0 0 0 0 2 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 4 2 2 2 2 2 2 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 2 0 1 −t ∗ Consider here L∗24,1 = 4L−t 24,1 and L24,2 = 8L24,2 . 10 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 2 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 2 2 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 1  0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 log10 M 10.1917 15.5128 19.1994 21.9813 24.2006 26.0413 27.6113 28.9791 30.1901 31.2763 32.2609 33.161 33.99 distance 0.633946 0.484887 0.370468 0.294144 0.24225 0.205264 0.177774 0.156625 0.13989 0.126336 0.115147 0.10576 0.097775 log10 M 10.8371 18.0618 22.288 25.2865 27.6124 29.5127 31.1194 32.5112 33.7389 34.8371 35.8305 36.7374 37.5717 distance 0.57735 0.408248 0.288675 0.220942 0.178411 0.149429 0.128473 0.112635 0.100256 0.0903175 0.0821655 0.0753593 0.0695919 Table 5: Show spherical code performance 24-dimensional case, for two different representations and respectives perturbations. We know from the literature that the Leech lattice can be regarded as a sub-lattice of the lattice E8 × E8 × E8 , as in the 8-dimensional   case: we P8,1 0 0 use the non-null perturbation the first case (P24,1 = 0 P8,1 0 ) and the 0 0 P8,1 second case to null perturbation (P24,2 = O) and we analyse the performance point of view of the spherical codes, vide Table 5 (the first two columns refer to the first case). 3.4. The lattice E6 and An In the case of n-dimensional lattices which do not have full size representation in n, for example lattices An and E6 . The construction will be done by Matrix cholesky decomposition Gram its dual lattice in according Corollary 1. We exemplify this through the lattice E6 , note that in this case different perturbations can induce the same number of points but distinct distances, see Table 6. 11 δ(Λw ) w M δ(E6 ) 1 1 0.2165 2 16 0.3059 3 216 0.3761 4 2160 0.4011 5 11520 0.4538 6 27440 0.5469 7 76800 0.7639 8 183708 0.6381 9 252000 0.7376 10 569184 0.7247 11 1078272 0.6326 12 1514240 0.7356 13 2806650 0.7436 14 4224000 0.7454 15 6714048 0.7095 16 9173736 0.7707 17 14555520 0.8081 18 21294000 0.7831 δ(Λw ) w M δ(E6 ) 9. 252000 0.7376 9.1 277200 0.6141 9.2 431200 0.77 9.35 431200 0.7247 9.4 474320 0.6437 9.55 474320 0.6706 9.6 521752 0.6987 9.7 521752 0.6643 10. 569184 0.7247 10.05 569184 0.6856 10.1 569184 0.6624 10.3 620928 0.7601 10.45 698544 0.6643 10.5 762048 0.7247 10.65 995328 0.792 10.7 995328 0.7917 10.85 1078272 0.6711 11. 1078272 0.6326 Table 6: Show density rate for 6-dimensional case, for w integer and no integer. 4. Conclusions We conclude that all n-dimensional lattice, up less scale, can be approximated by a sequence of lattices that have orthogonal sub-lattice. Furthermore, there is a degree of freedom (n(n − 1)/2) for quadratic convergence, this freedom induces quotient group with different number of generators and can make convergency more fast in certain applications, for example in the case of spherical codes are reticulated target has some multiple minimum vectors of some canonical vector, we find a non-null pertubation as it will be more efficient. We present here a method for finding lattices with suborthogonal, our method is simpler, more general and more efficient than the one presented in [1]. 12 References [1] C. Alves, S.I.R. Costa, Commutative group codes in R4 , R6 , R8 and R16 –approaching the bound. Discrete Math. 313 (2013), no. 16, 16771687. [2] M. Berstein, N.J.A. Sloane, P.E. Wright, On sublattices of the hexagonal lattice, Discrete Mathematics 170 (1997) 29-39. [3] E. Biglieri, M. Elia, Cyclic-group codes for the Gaussian channel, IEEE Transactions on Information Theory 22 (1976) 624-629. [4] A. Campello, J.E. Strapasson, On sequences of projections of the cubic lattice. Comput. Appl. Math. 32 (2013), no. 1, 57-69. [5] A. Campello, J.E. Strapasson, S.I.R. Costa, On projections of arbitrary lattices. Linear Algebra Appl. 439 (2013), no. 9, 2577-2583. [6] H.C. Cohen, A Course in Computational Algebraic Number Theory, Springer-Verlag, New York, 1993. [7] J.H. Conway, N.J.A. Sloane, Sphere Packings, Lattices and Groups, Springer-Verlag, New York, 1999. [8] S.I.R. Costa, M. Muniz, E. Augustini, R. Palazzo, Graphs, tessellations and perfect codes on flat tori, IEEE Transactions on Information Theory 50 (2004) 2363-2377. [9] I. Ingemarsson, Group codes for the Gaussian channel, in: Lecture Notes in Control and Information Sciences, vol. 128, Springer Verlag, 1989, pp. 73-108. [10] H. Loeliger, Signals sets matched to groups, IEEE Transactions on Information Theory 37 (1991) 1675-1682. 13 [11] C. D. Meyer, Matrix Analysis and Applied Linear Algebra, Society for Industrial Mathematics (SIAM), Philadelphia PA, USA, 2000. [12] D. Micciancio, S. Goldwasser, Complexity of Lattice Problems - A Cryptographic Perspective, Kluwer Academic Publishers, Norwell, Massachusetts, USA, 2002. [13] R.M. Siqueira, S.I.R. Costa, Flat tori, lattices and bounds for commutative group codes, Designs, Codes and Cryptography 49 (2008) 307-321. [14] D. Slepian, Group codes for the Gaussian channel, Bell System Technical Journal 47 (1968) 575-602. [15] N.J.A. Sloane, V.A. Vaishampayan, S.I.R. Costa, A note on projecting the cubic lattice. Discrete Comput. Geom. 46 (2011), no. 3, 472-478. [16] C. Torezzan, S.I.R. Costa, V.A. Vaishampayan, Spherical codes on torus layers, in: Proceedings of the IEEE International Symposium on Information Theory, June-July 2009, pp. 2033-2037. [17] C. Torezzan, J.E. Strapasson, S.I.R. Costa, R.M. Siqueira, Optimum commutative group codes, Des. Codes Cryptogr. 74 (2015), no. 2, 379394. [18] V.A. Vaishampayan, S.I.R. Costa, Curves on a sphere, shift-map dynamics and error control for continuous alphabet sources, IEEE Transactions on Information Theory 49 (2003) 1658-1672 14
7cs.IT
X-TORSION AND UNIVERSAL GROUPS arXiv:1610.00313v1 [math.GR] 2 Oct 2016 MAURICE CHIODO, ZACHIRI MCKENZIE Abstract. For a set X ⊆ N, we define the X-torsion of a group G to be all elements g ∈ G with g n = e for some n ∈ X. With X recursively enumerable, we give two independent proofs (group-theoretic, and model-theoretic) that there exists a universal finitely presented X-torsion-free group; one which contains all finitely presented X-torsion-free groups. We also show that, if X is recursively enumerable, then the set of finite presentations of Xtorsion-free groups is Π02 -complete in Kleene’s arithmetic hierarchy. 1. Introduction Torsion is a well-studied object in group theory. We let o(g) denote the order of a group element g. Recalling that g ∈ G is torsion if 1 ≤ o(g) < ω, we write Tor(G) ∶= {g ∈ G ∣ g is torsion}. So, what if we were to ‘restrict’ the type of torsion we are looking at? Perhaps we are only concerned with 2-torsion, or torsion elements of prime order. So, for any set X ⊆ N, we define the X-torsion of a group G, written TorX (G), as TorX (G) ∶= {g ∈ G ∣ ∃n ∈ X with g n = e} It is clear that, for the case X = N, we have TorN (G) = Tor(G) in the usual sense. For any set X ⊆ N, we define the factor completion of X to be X fc ∶= {n ∈ N ∣ ∃m ≥ 1, nm ∈ X}, and we say X is factor complete if X fc = X. We say that a group G is X-torsion-free if TorX (G) = {e}; equivalently, if Tord(G) ∩ X fc = ∅. We use the following notation for the set of orders of torsion elements in a group: Tord(G) ∶= {n ∈ N ∣ ∃g ∈ Tor(G) with o(g) = n ≥ 2} One famous consequence of the Higman Embedding Theorem [6] is the fact that there is a universal finitely presented (f.p.) group; that is, a finitely presented group into which all finitely presented groups embed. Recently, Belegradek in [1] and Chiodo in [4] independently showed that there exists a universal f.p. torsion-free group; that is, an f.p. torsion-free group into which all f.p. torsion-free groups embed. In this paper we generalise this result further, in the context of X-torsion-freeness, as follows: 2010 AMS Classification: 20F05, 03D55, 03C98. Keywords: Groups, torsion, torsion quotients, embeddings, universal Horn theory. This work is part of a project that has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No. 659102. Date: March 13, 2018. 1 2 MAURICE CHIODO, ZACHIRI MCKENZIE Theorem 2.12. Let X be a recursively enumerable set of integers. Then there is a universal finitely presented X-torsion-free group G. That is, G is Xtorsion-free, and for any finitely presented group H we have that H ↪ G if (and only if ) H is X-torsion-free. There are two key steps in proving this theorem. The first is to construct a free product of all such groups, in an algorithmic way: Theorem 2.9. Let A be an r.e. set of integers. Then there is a countably generated recursive presentation Q of a group Q which is A-torsion-free, and contains an embedded copy of every countably generated recursively presentable A-torsion-free group. From this, we can construct a finitely presented example, using the following theorem of Higman: Theorem 2.10. There is a uniform algorithm that, on input of a countably generated recursive presentation P = ⟨X∣R⟩, constructs a finite presentation T(P ) such that P ↪ T(P ) and Tord(P ) = Tord(T(P )), along with an explicit embedding φ ∶ P ↪ T(P ). We prove Theorem 2.9 in the following two independent ways, for reasons which we outline shortly. Firstly, in Section 2, we generalise the construction in [4, Theorem 3.10] of a universal f.p. torsion-free group, using arguments in group theory. Then, in Section 3, we generalise the construction in [1, Theorem A.1] of a universal f.p. torsion-free group, using arguments in model theory. The results in Section 3 provide the framework for proving that there might be other group properties for which there exist universal f.p. examples. Our main result in Section 3, which we apply directly to the theory of X-torsion-free groups to prove Theorem 2.9, is the following: Theorem 3.8. If T ⊇ TGrp is an r.e. universal Horn LGrp -theory then there exists a recursively presented group G ⊧ T such that every recursively presented group H ⊧ T embeds into G. The reason for providing both proofs of Theorem 2.9 is twofold. On the one hand, the group-theoretic proof in Section 2 is more direct, and gives a clear picture of why Theorem 2.9 holds. It is explicit, and algorithmic. On the other hand, the model-theoretic proof in Section 3 is more general, and might lead to showing that other group properties possess universal f.p. examples. Indeed, it was only by looking at the model-theoretic arguments in [1, Theorem A.1] that we realised we could consider X-torsion as an object, and that the results of [4] could be generalised to this; we would never have made the connection otherwise. X-TORSION AND UNIVERSAL GROUPS 3 On this, it would be interesting further work to see what other grouptheoretic properties ρ possess universal f.p. examples. A potential proof technique would be as follows: (1) Show that the property ρ satisfies the conditions of Theorem 3.7 (that is, closed under free products, identity, and subgroups). (2) Show that ρ has an r.e. universal Horn LGrp -theory, thus allowing us to apply Theorem 3.8. (3) Show that ρ is possessed by finitely generated free groups, and preserved by free products and HNN extensions (which is what is needed to show that it is preserved under the Higman embedding of Theorem 2.10). Of course, satisfying all of the above is quite difficult. One can easily find many group properties that satisfy (1), and some of these also satisfy (2) quite trivially. But there are very few properties of groups that satisfy (3), and therein lies the problem. We finish our paper with a generalisation a result from [4] on the complexity of recognising torsion-freeness, and show the following: Theorem 2.13. Let X ⊆ N≥2 be a non-empty r.e. set of integers. Then the set of finite presentations of X-torsion-free groups is Π02 -complete. This is remarkable; even if X is a finite set, or stronger still, a single prime (say X = {2}), the set of finite presentations of groups with no X-torsion will still form a Π02 -complete set. 2. X-torsion and universality This section is a generalisation of the definitions and results in [4, Section 3] on universal torsion-free quotients and universal finitely presented torsion-free groups. 2.1. Notation. If P is a group presentation, we denote by P the group presented by P , and w by the group element represented by the word w. A presentation P = ⟨X∣R⟩ is said to be a recursive presentation if X is a finite set and R is a recursive enumeration of relations; P is said to be a countably generated recursive presentation if instead X is a recursive enumeration of generators. A group G is said to be finitely (respectively, recursively) presentable if G ≅ P for some finite (respectively, recursive) presentation P . If P, Q are group presentations then we denote their free product presentation by P ∗ Q: this is given by taking the disjoint union of their generators and relations. If g1 , . . . , gn are elements of a group G, we write ⟨g1 , . . . , gn ⟩ for the subgroup in G generated by these elements and ⟪g1 , . . . , gn ⟫G for the normal closure of these elements in G. Let ω denote the smallest infinite ordinal. Let ∣X∣ denote the cardinality of a set X. If X is a set, let X −1 be a set of the same cardinality as and disjoint from X along with a fixed bijection ∗−1 ∶ X → X −1 . Write X ∗ for the set of finite words on X ∪ X −1 . We will make use of Σ0n sets and Π0n sets; see [8] for an introduction to these. 4 MAURICE CHIODO, ZACHIRI MCKENZIE 2.2. X-torsion. If G, H are groups, and X a set of integers with H X-torsion-free, a surjective homomorphism h ∶ G ↠ H is universal if, for any X-torsion-free K and any homomorphism f ∶ G → K, there is a homomorphism φ ∶ H → K such that f = φ ○ h ∶ G → K, i.e., the following diagram commutes: h / H ❆❆ ❆❆ ❆ φ f ❆  G❆ K Note that if φ exists then it will be unique. Indeed, if φ′ also satisfies f = φ′ ○ h, then φ ○ h = φ′ ○ h, and hence φ = φ′ as h is a surjection and thus is rightcancellative. Moreover, any such H is unique, up to isomorphism. Such an H is called the universal X-torsion-free quotient for G, denoted GX−tf . Observe that if G is itself X-torsion-free, then GX−tf exists and GX−tf ≅ G, as the identity map idG ∶ G → G has the universal property above. A standard construction, showing that GX−tf exists for every group G, is done via taking the quotient of G by its X-torsion-free radical ρX (G), where ρX (G) is the intersection of all normal subgroups N ⊲ G with G/N X-torsion-free (a generalisation of the torsion-free radical, ρ(G), in [2]). It follows immediately that G/ρX (G) has all the properties of an X-torsion-free universal quotient for G. We present here an alternative construction for GX−tf which, though isomorphic to G/ρX (G), lends itself more easily to an effective procedure for finitely (or recursively) presented groups Definition 2.1. Given a group G, and a set of integers X ⊆ N, we inductively define TorX n (G) as follows: TorX 0 (G) ∶= {e}, X X X G TorX n+1 (G) ∶= ⟪ {g ∈ G ∣ g Torn (G) ∈ Tor (G/ Torn (G))} ⟫ , X TorX ω (G) ∶= ⋃ Torn (G). n∈N TorX i (G) Thus, is the set of elements of G which are annihilated upon taking i successive quotients of G by the normal closure of all X-torsion elements, and TorX ω (G) is the union of all these. Lemma 2.2. If G is a group, then G/ TorX ω (G) is X-torsion-free. X X X n Proof. Suppose g TorX ω (G) ∈ Tor (G/ Torω (G)). Then g Torω (G) = e in X n G/ TorX ω (G) for some 1 ≤ n ∈ X, so g ∈ Torω (G). Thus there is some i ∈ N X X X such that g n ∈ TorX i (G), and hence g Tori (G) ∈ Tor (G/ Tori (G)). Thus X X X X g ∈ Tori+1 (G) ⊆ Torω (G), and so g Torω (G) = e in G/ Torω (G).  Proposition 2.3. If G is a group, then ρX (G) = TorX ω (G). X Proof. Clearly ρX (G) ⊆ TorX ω (G), by definition of ρ (G) and the fact that X G/ Torω (G) is torsion-free (Lemma 2.2). It remains to show that TorX ω (G) ⊆ X ρX (G). We proceed by contradiction, so assume TorX (G) ⊈ ρ (G). Then ω X-TORSION AND UNIVERSAL GROUPS 5 there is some N ⊲ G with G/N X-torsion-free, along with some minimal i such X that TorX i (G) ⊈ N (clearly, i > 0, as Tor0 (G) = {e}). Then, by definition of X Tori (G) and the fact that N is normal, there exists e ≠ g ∈ TorX i (G) such X X X X that g Tori−1 (G) ∈ Tor (G/ Tori−1 (G)) and g ∉ N (or else Tori (G) ⊆ N ). But then gn ∈ Tori−1 (G) for some 1 < n ∈ X. Since TorX i−1 (G) ⊆ N by minimality of i, we have that gN is a (non-trivial) X-torsion element of G/N , contradicting X X-torsion-freeness of G/N . Hence TorX  ω (G) ⊆ ρ (G). X−tf Corollary 2.4. If G is a group, then G/ TorX . ω (G) ≅ G What follows is a standard result, which we state without proof. Lemma 2.5. Let P = ⟨X∣R⟩ be a countably generated recursive presentation. Then the set of words {w ∈ X ∗ ∣ w = e in P } is r.e. Lemma 2.6. Let P = ⟨X∣R⟩ be a countably generated recursive presentation, and A an r.e. set of integers. Then the set of words {w ∈ X ∗ ∣ w ∈ TorA (P )} is r.e. Proof. Take any recursive enumeration {w1 , w2 , . . .} of X ∗ . Using Lemma 2.5, start checking if w ni = e in P for each wi ∈ X ∗ and each n ∈ A (by proceeding along finite diagonals). For each wi we come across which is A-torsion, add it to our enumeration. This procedure will enumerate all words in TorA (P ), and only words in TorA (P ). Thus the set of words in X ∗ representing elements in TorA (P ) is r.e.  We use this to show the following: Lemma 2.7. Given a countably generated recursive presentation P = ⟨X∣R⟩, and r.e. set of integers A, the set TiA ∶= {w ∈ X ∗ ∣ w ∈ TorA i (P )} is r.e., uniformly over all i and all such presentations P . Moreover, the union TωA ∶= ⋃ TiA is r.e., and is precisely the set {w ∈ X ∗ ∣ w ∈ TorA ω (P )}. Proof. We proceed by induction. Clearly TorA 1 (P ) is r.e., as it is the normal closure of TorA (P ), which is r.e. by Lemma 2.6. So assume that TorA i (P ) is A A r.e. for all i ≤ n. Then Torn+1 (P ) is the normal closure of Tor (P / TorA n (P )), which again is r.e. by the induction hypothesis and Lemma 2.6. The rest of the lemma then follows immediately.  Proposition 2.8. There is a uniform algorithm that, on input of a countably generated recursive presentation P = ⟨X∣R⟩ of a group P , and an r.e. set of integers A, outputs a countably generated recursive presentation P A−tf = ⟨X∣R′ ⟩ A−tf (on the same generating set X, and with R ⊆ R′ as sets) such that P A−tf ≅ P , with associated surjection given by extending idX ∶ X → X. A−tf Proof. By Corollary 2.4, P is the group P / TorA ω (P ). Then, with the notation of Lemma 2.7, it can be seen that P A−tf ∶= ⟨X∣R ∪ Tω ⟩ is a countably A−tf , uniformly constructed from P .  generated recursive presentation for P 6 MAURICE CHIODO, ZACHIRI MCKENZIE 2.3. Universality and complexity of X-torsion. With the above machinery, we can now prove the main technical result of this section; Theorem 2.9. We will re-prove this result again in Section 3, using tools from model theory. Theorem 2.9. Let A be an r.e. set of integers. Then there is a countably generated recursive presentation Q of a group Q which is A-torsion-free, and contains an embedded copy of every countably generated recursively presentable A-torsion-free group. Proof. Take an enumeration P1 , P2 , . . . of all countably generated recursive presentations of groups, and construct the countably generated recursive presentation Q ∶= P1A−tf ∗P2A−tf ∗. . .; this is the countably infinite free product of the universal A-torsion-free quotient of all countably generated recursively presentable groups (with some repetition). As each PiA−tf is uniformly constructible from Pi (by Proposition 2.8), we have that our construction of Q is indeed effective, and hence Q is a countably generated recursive presentation. Also, Proposition 2.8 shows that Q is an A-torsion-free group, as we have successfully annihilated all the A-torsion in the free product factors, and the free product of A-torsionfree groups is again A-torsion-free. Moreover, Q contains an embedded copy of every A-torsion-free countably generated recursively presentable group, as the universal A-torsion-free quotient of an A-torsion-free group is itself.  As detailed in [3, Lemma 6.9 and Theorem 6.10], the following is implicit in Rotman’s proof [9, Theorem 12.18] of the Higman Embedding Theorem. Theorem 2.10. There is a uniform algorithm that, on input of a countably generated recursive presentation P = ⟨X∣R⟩, constructs a finite presentation T(P ) such that P ↪ T(P ) and Tord(P ) = Tord(T(P )), along with an explicit embedding φ ∶ P ↪ T(P ). We can now prove our main result: Theorem 2.11. Let A be an r.e. set of integers. Then there is a finitely presentable group G which is A-torsion-free, and contains an embedded copy of every countably generated recursively presentable A-torsion-free group. Proof. We construct Q as in Theorem 2.9, and then use Theorem 2.10 to embed Q into a finitely presentable group T(Q). By construction, Tord(Q) = Tord(T(Q)), so T(Q) is A-torsion-free. Finally, T(Q) has an embedded copy of every countably generated recursively presentable A-torsion-free group, since Q did. Taking G to be T(Q) completes the proof.  As all f.p. groups are recursively presentable, we have the following corollary: Theorem 2.12. Let A be an r.e. set of integers. Then there is a universal finitely presented A-torsion-free group G. That is, G is A-torsion-free, and for any finitely presented group H we have that H ↪ G if (and only if ) H is A-torsion-free. The following is an unexpected and very strong generalisation of [4, Theorem 4.2], classifying the computational complexity of recognising f.p. A-torsion-free groups. X-TORSION AND UNIVERSAL GROUPS 7 Theorem 2.13. Let A ⊆ N≥2 be a non-empty set of integers. Then the set of finite presentations of A-torsion-free groups is Π02 -hard. Moreover, if A is r.e., then this set of presentations is Π02 -complete. Proof. This follows the proofs of [3, Lemma 6.11] and [4, Theorem 4.2]. First, there must be some element 1 < a ∈ A. Now, given n ∈ N, form the recursive presentation Pn ∶= ⟨x1 , x2 , . . . ∣ xai = e ∀i ∈ N, xj = e ∀j ∈ Wn ⟩. Then form the finite presentation Qn using Higman’s Embedding Theorem (Theorem 2.10) so that Tord(Qn ) = Tord P n . Now note that Qn is A-torsion-free ⇔ P n is Atorsion free ⇔ xi = e for all i ∈ N ⇔ Wn = N; the latter being a Π02 -complete set ([4, Lemma 4.1]). So the set of finite presentations of A-torsion-free groups is Π02 -hard, for any non-empty A ⊆ N≥2 . Moreover, when A is also r.e., this set has the following Π02 description: G is A-torsion-free ⇔ (∀w ∈ G)(∀n ∈ A)(wn ≠G e or w =G e) and is thus Π02 -complete.  Note that the recursive presentation Pn can be constructed uniformly from n whenever A is r.e. and non-empty, as to find our 1 < a ∈ A we simply begin an enumeration of A and take the first output. 2.4. X-torsion-length. We finish this section by generalising the notion of torsion length, which was first introduced in [5, Definition 2.5]. Definition 2.14. Given X ⊆ N, we define the X-Torsion Length of G, TorLenX (G), X by the smallest ordinal n such that TorX n (G) = Torω (G). Rather than go and re-work all the theory developed in [5], we simply state here the main results [5, Theorem 3.3 and Theorem 3.10], generalised to Xtorsion. Going through the work in [5], it is straightforward to see that all results there generalise to X-torsion, and thus we refrain from doing so here. Theorem 2.15. Given any ∅ ≠ X ⊆ N≥2 , there is a family of finite presentations {Pn }n∈N of groups satisfying TorLenX (P n ) = n and P n / TorX 1 (P n ) ≅ P n−1 . Theorem 2.16. Given any ∅ ≠ X ⊆ N≥2 , there exists a 2-generator recursive presentation Q for which TorLenX (Q) = ω. If X is r.e., then we can algorithmically construct such a finite presentation Q from X. 3. Presentations The purpose of this section is to re-prove Theorem 2.9 using model-theoretic arguments. We follow the idea in [1, Theorem A.1]. 3.1. Notation. Throughout this section LGrp will be used to denote the language of group theory, that is the language of first-order logic supplemented with a binary function symbol ⋅ whose intended interpretation is the group operation, a constant symbol e whose intended interpretation is the identity element, and a unary function symbol ✷−1 whose intended interpretation is the function that sends 8 MAURICE CHIODO, ZACHIRI MCKENZIE elements to their inverses. We use the standard abbreviation of writing xn instead of x ⋅ ⋯ ⋅ x. If X is a set of constant symbols that are not in L endowed ´¹¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¶ n with a well-ordering of its elements then we write LX for the language obtained by adding the constant symbols in X to L. If L is a language, X is a set of new constant symbols that are not in L endowed with an implicit well-ordering, M is an L-structure, and A ⊆ M with a canonical bijection witnessing that ∣A∣ = ∣X∣, then we write ⟨M, A⟩ for the LX -structure obtained by interpreting the constant symbols in X with the elements of A. If M is an L-structure and A ⊆ M, then we say that M is generated by A if all of M is obtained by closing A and the constants of M under applications of the interpretation of functions from L in M. 3.2. Model theory preliminaries. We begin by recalling some definitions and results from Chapter 9 of [7]. Definition 3.1. Let L be a language. We say that an L∞∞ -formula φ is basic Horn if φ is in the form ⋀ Φ ⇒ ψ, where Φ is a, possibly infinite, set of atomic L-formulae, and ψ is either an atomic L-formula or –. We say that an L∞∞ -formula φ is universal Horn, and write ∀1 Horn, if φ is in the form ∀⃗ xθ(⃗ x) where θ is basic Horn. We say that an L-theory T is universal Horn if T has an axiomatisation that only consists of ∀1 Horn sentences. Let TGrp be the obvious LGrp -theory that axiomatises the class of groups. It is clear that that TGrp can be written as a finite set of finitary ∀1 Horn sentences. Definition 3.2. Let X ⊆ N. We write TX−tf for the LGrp -theory with axioms: TGrp ∪ {∀x(xn = e ⇒ x = e) ∣ n ∈ X}. It is clear that for all X ⊆ N, TX−tf is a finitary universal Horn theory that axiomatises the class of X-torsion-free groups. Definition 3.3. Let L be a language and let K be a class of L-structures. A K-presentation is a tuple ⟨X, Φ⟩K such that X is a set of new constant symbols endowed with an implicit well-ordering, called generators, that are not in L, and Φ is a set of atomic LX -sentences. We will write ⟨X, Φ⟩ instead of ⟨X, Φ⟩K when K is clear from the context. If both Φ and X are finite then we say that ⟨X, Φ⟩K is a finitely presented K-presentation. If both Φ and X are r.e. then we say that ⟨X, Φ⟩K is a recursively presented K-presentation. Definition 3.4. Let L be a language and let K be a class of L-structures. Let ⟨X, Φ⟩K be a K-presentation. We say that an LX -structure ⟨M, A⟩ is a model of ⟨X, Φ⟩K if M ∈ K and ⟨M, A⟩ ⊧ ⋀ Φ. Definition 3.5. Let L be a language and let K be a class of L-structures. Let ⟨X, Φ⟩K be a K-presentation. We say that ⟨X, Φ⟩K presents an LX -structure ⟨M, A⟩ if X-TORSION AND UNIVERSAL GROUPS 9 (i) ⟨M, A⟩ is a model of ⟨X, Φ⟩K , (ii) M is generated by A, (iii) for every model ⟨N , B⟩ of ⟨X, Φ⟩K , there exists a homomorphism f ∶ M Ð→ N such that f (cM ) = cN for all c ∈ X. We say that K admits presentations if every K-presentation presents a structure ⟨M, A⟩ with M ∈ K. Note that if ⟨X, Φ⟩ presents ⟨M, A⟩ and ⟨N , B⟩ is a model of ⟨X, Φ⟩, then, since M is generated by A, the homomorphism whose existence is guaranteed by Definition 3.5(iii) is unique. The following is Lemma 9.2.1 of [7]: Lemma 3.6. Let Φ be a set of atomic LX -sentences where X is a set of constants not in L. Let ⟨X, Φ⟩ be a K-presentation and let ⟨M, A⟩ be an LX structure with M ∈ K. The following are equivalent: (i) ⟨X, Φ⟩ presents ⟨M, A⟩, (ii) A generates M; and for every atomic formula ψ(⃗ x) of L and for every a ⃗ ∈ A, M ⊧ ψ(⃗ a) if and only if every structure in K is a model of ∀⃗ x (⋀ Φ ⇒ ψ) . And this is Lemma 9.2.2 of [7]: Lemma 3.7. Let L be a language and let K be a class of L-structures which is closed under isomorphic copies. The following are equivalent: (i) K is closed under products, 1 and substructures, (ii) K admits presentations, (iii) K is axiomatised by a universal Horn theory in the language L∞∞ . 3.3. A proof of Theorem 2.9 using model theory. The following is an adaptation of the proof of Theorem A.1 in [1]. Theorem 3.8. If T ⊇ TGrp is an r.e. universal Horn LGrp -theory then there exists a recursively presented group G ⊧ T such that every recursively presented group H ⊧ T embeds into G. Remark 3.9. A group property ρ having an r.e. universal Horn theory does not imply that the finite presentations of groups with ρ are r.e. Indeed, for X a non-empty r.e. set, TX−tf is an r.e. universal Horn theory (Definition 3.2), but by Theorem 2.13 the set of finite presentations of such groups is Π02 -complete. Proof. Let T ⊇ TGrp be a r.e. universal Horn LGrp -theory. Let K be the class of LGrp -structures that satisfy T . Let K′ be the class of LGrp -structures that satisfy TGrp . By Lemma 3.7, both K and K′ admit presentations. If τ is a presentation, then we will write Gτ for the element of K presented by τ , and Gτ for the element of K′ presented by τ . Let ⟨πn ∣ n ∈ N⟩ be effective enumeration of recursive presentations. Let π be the disjoint union of all the πn ’s; π ∶= π1 ∗ π2 ∗ ⋯. Therefore, π is a recursive presentation. We claim that Gπ is the desired universal group satisfying T . It is immediate that Gπ ⊧ T . If τ is a recursive presentation such that Gτ ⊧ T then Gτ = Gτ , and so Gτ embeds into Gπ . This shows that Gπ is universal. It remains to show that Gπ 10 MAURICE CHIODO, ZACHIRI MCKENZIE is recursively presented in K′ . Let π = ⟨Y, Φ⟩ and let Y G be the interpretations of the constant symbols Y in Gπ . Let Φ′ be the set of atomic LGrp,Y -sentences (the language obtained by adding new constants for the generators in Y ) that π π hold in ⟨Gπ , Y G ⟩. Lemma 3.6 implies that ⟨Y, Φ′ ⟩K′ presents ⟨Gπ , Y G ⟩. We need to show is that Φ′ is r.e. Let S = T ∪ Φ. We claim that for all atomic LGrp,Y -sentences σ, π (1) S ⊢ σ if and only if σ ∈ Φ′ . Since Gπ ⊧ S, it follows that for all atomic LGrp,Y -sentences σ, if S ⊢ σ then σ ∈ Φ′ . We need to show the converse. Let σ ∈ Φ′ , and suppose that S ⊬ σ. Note that, since σ is atomic, it is of the form w1 = w2 where w1 and w2 are words in the generators Y . Let H be a group such that Y H is the interpretation of the constant symbols Y in H, ⟨H, Y H ⟩ ⊧ S, and H ⊧ (w1 ≠ w2 ). Therefore ⟨H, Y H ⟩ is a model of ⟨Y, Φ⟩K . It follows from (iii) of Definition 3.5 that the π map g ∶ Y G Ð→ Y H defined by the implicit ordering on Y must lift to a homomorphism f ∶ Gπ Ð→ H. But this map is not well-defined since w1 = w2 in Gπ and f (w1 ) ≠ f (w2 ) in H, which is a contradiction. Therefore (1) holds. Since S is r.e., (1) yields a recursive enumeration of the elements of Φ′ .  If X ⊆ N is r.e. then Definition 3.2 shows that TX−tf is an r.e. universal Horn theory that extends TGrp . This immediately re-proves Theorem 2.9. It is then an application of Theorem 2.10 again, to re-prove Theorem 2.12. This time, we have done most of the work using model-theoretic techniques and in a way which could be applied to other group properties (not just X-torsion-freeness), as discussed in the introduction. References [1] I. Belegradek, A. Szczepański, Endomorphisms of relatively hyperbolic groups with an appendix by O. Belegradek, Internat. J. Algebra Comput., 18, No. 1, 97–110 (2008). [2] S. D. Brodsky, J. Howie, The universal torsion-free image of a group, Israel J. Math. 98, 209–228 (1997). [3] M. Chiodo, Finding non-trivial elements and splittings in groups, J. Algebra. 331, 271– 284 (2011). [4] M. Chiodo, On torsion in finitely presented groups, Groups Complex. Cryptol. (2) 6, 1–8 (2014). [5] M. Chiodo, R. Vyas, A note on torsion length, Comm. Algebra, 43, No. 11, 4825–4835 (2015). [6] G. Higman, Subgroups of finitely presented groups, Proc. Roy. Soc. Ser. A 262, 455–475 (1961). [7] W. Hodges, Model Theory. Encyclopedia of mathematics and its applications. No. 42. Cambridge University Press, 1993. [8] H. Rogers Jr, Theory of recursive functions and effective computability, MIT Press, (1987). [9] J. Rotman, An introduction to the theory of groups, Springer-Verlag, New York, (1995). Department of Pure Mathematics and Mathematical Statistics, University of Cambridge, Wilberforce Road, Cambridge, CB3 0WB, UK. [email protected] Department of Philosophy, Linguistics and Theory of Science, Gothenburg University, Olof Wijksgatan 6, 405 30 Gothenburg, Sweden. [email protected]
4math.GR
arXiv:1802.08215v1 [cs.RO] 22 Feb 2018 ArduSoar: an Open-Source Thermalling Controller for Resource-Constrained Autopilots Samuel Tabor Iain Guilliard Andrey Kolobov [email protected] Australian National University Canberra, Australia [email protected] Microsoft Research Redmond, WA-98052 [email protected] Abstract—Autonomous soaring capability has the potential to significantly increase time aloft for fixed-wing UAVs. In this paper, we introduce ArduSoar, the first soaring controller integrated into a major autopilot software suite for small UAVs. We describe ArduSoar from the algorithmic standpoint, outline its integration with the ArduPlane autopilot, and discuss tuning procedures and parameter choices for it. I. I NTRODUCTION Autonomous soaring has long been considered for its promise of extending the range and time in the air for Uninhabited Aerial Vehicles (UAVs). These two UAV characteristics are crucial in a range of scenarios from aerial mapping to crop monitoring to package delivery, and are currently limited by a combination of the UAV’s powerplant and the limited energy available onboard. Exploiting the energy of the sun [25] and of the moving air masses, the latter of which is at the focus of our work, have the potential to increase fixed-wing drone range by many times [6]. While there are many atmospheric phenomena that can help an aircraft gain altitude without the help of a motor, the most ubiquitous of them are thermals, rising plumes of air generated by certain areas on the ground giving up heat into the atmosphere. Including the first feasibility study of autonomous thermalling in 1998 [31], researchers have proposed at least 5 approaches to modeling thermals computationally and soaring in them, evaluating these techniques in simulation [30, 17, 26] and in live tests [7, 14]. Each of them has a significant number of intricacies related to implementation and parameter tuning. Since none of the implementations, which were written in different languages and for different platforms and setups, are publicly available, none of these techniques have been properly compared against each other. This paper introduces and describes the implementation of a soaring controller called ArduSoar as part of a popular opensource UAV autopilot system, ArduPilot [1]. ArduPilot can be built to run on many different autopilot hardware platforms, including Pixhawk [4] and Rasberry Pi 3 [5], as well as as a software-in-the-loop simulation. Being part of an open-source project, ArduSoar’s software components can be reused or inspected to serve as an inspiration, basis or benchmark for other soaring controller designs. Fig. 1: Rayleigh-Bénard convection simulation showing two rising thermal plumes, with red color indicating hotter air. The lower boundary of the simulation heats the adjacent air hotter than the ambient air temperature. The hotter air gains buoyancy and rises in turbulent plumes to mix with the cooler air above. The small black vertical bars indicate vertical air speed at an altitude and show the typical bell shaped distribution of uplift through the center of a thermal. II. S OARING , T HERMALS , T HERMAL E STIMATION Soaring usually refers to process of using a rising air mass to gain altitude by birds and aircraft. Technically, while in dynamic soaring the source of lift is a wind gradient [23, 24, 9], not rising air, here we focus on static soaring, which always relies on air being pushed upwards in some way. The most commonly encountered type of rising air regions is a thermal. Thermals start immediately above parts of Earth’s surface that accumulate more heat than surrounding areas and then emit the heat back into the atmosphere, warming up the air above them. Thermals can form over darker terrain, e.g. asphalt or plowed fields, but may also arise above forested and other areas. They are invisible, but are frequently pictured as air columns, possibly slanted due to wind. Vertical air velocity in the center of such a column is higher, becomes lower towards the column fringes, and eventually turns into sink. In reality and even under realistic simulation models, thermals do not always conform to this idealized concept. Figure 1 is a screenshot from a Rayleigh-Bénard simulation (see, e.g., Kadanoff [18]), a standard high-fidelity convection model that gives a more accurate idea of thermals’ irregular and chaotic nature. In all existing work on autonomous soaring in thermals, autopilots/soaring controllers use thermal models that can be estimated by observing the aircraft’s vertical air velocity and combining this data with information about the aircraft’s 3D position changes. Soaring controllers try to fit such a thermal model to this data and, if the fit is good enough, synthesize a trajectory that exploits the thermal’s lift distribution, as predicted by the model, to gain altitude. The model at the heart of Rayleigh-Bénard simulations has been used in this way to learn sailplane trajectories [26] for simulated aircraft, but is expensive computationally and nontrivial to tune. Therefore, for computational and ease-of-use reasons, ArduSoar relies on the thermal model proposed by Wharington and Herszberg [31] and modified by Allen and Lin [7]. It makes a number of implicit assumptions, which we make explicit here, all of which apply to the thermal’s horizontal cross-section at a given altitude: • A thermal is stationary: its position and distribution of lift within it, i.e., the air’s vertical velocity, does not change with time. Note, however, that this assumption is for a fixed altitude only, and the model allows for a thermal’s properties to vary with altitude due to air cooling down or due to the thermal getting affected by wind. • A thermal has a single “center”, called core, at position xth , y th . • The vertical air velocity within a thermal’s horizontal cross-section has a bell-shaped distribution w.r.t. the thermal core. It is strongest at (xth , y th ), and has speed W th m/s there. Thus, the lift strength at position x, y is given by w(x, y) = W th  exp − (x − xth )2 + (y − y th )2 Rth 2  , (1) where Rth is a parameter describing the thermal radius. Parameters W th , Rth , xth and y th can be continually estimated using a Kalman filter [20] if we assume a Gaussian distribution over them and keep updating this distribution with sensor observations. Hazard [17] has used an Unscented Kalman filter (UKF) for this purpose. In ArduSoar, due to the computational constraints of the hardware platforms on which it typically runs, Pixhawk and APM, we use an Extended Kalman Filter (EKF) instead. An EKF is a type of Kalman filter that allows maintaining a Gaussian estimate of the system state given a sequence of observations if the system transition function f (X, u) and observation function h(X) are nonlinear. Specifically, suppose we start with a state distribution N (Xk , Pk ), and want to compute a new estimate N (Xk+1 , Pk+1 ) after a system makes a transition over one time step and we receive an observation o about it. To begin with, we linearize of the transition function f about the current state estimate by computing the Jacobian F = ∂f ∂X (2) Xk Then we update the state estimate using the state transition X̂k+1 = f (Xk , uk ) | P̂k+1 = F Pk F + Q (3) (4) Then we linearize h around this new state estimate by computing the Jacobian H= ∂h ∂X (5) X̂k+1 and calculate the approximate Kalman gain K = P̂k+1 H | (H P̂k+1 H | + R)−1 (6) As the final step, the state distribution so far needs to be corrected for the observation process: Xk+1 = X̂k+1 + K(o − h(X̂k+1 )) (7) Pk+1 = (I − KH)P̂k+1 (8) Matrices Q and R are process and observation noise covariances, respectively, and are tunable parameters in our model. III. A LGORITHMIC A SPECTS OF A RDU S OAR Gaining altitude with a thermal’s help can be broken down into four stages — thermal detection, identification, exploitation, and exit. At an abstract level, an aircraft traveling through the air with its motor shut down will tend to lose altitude at a certain rate that depends on the aircraft’s airspeed via a function called the polar curve that we discuss shortly. The autopilot knows both the polar curve and the current airspeed, and therefore at any point can easily determine how fast the aircraft would be descending if the air around it was still. If the aircraft is flying through a thermal, the air around it is not still; it is rising and lifting the aircraft with it. The autopilot can detect this by comparing the aircraft’s actual altitude change rate to the one predicted by the polar curve, and enter the thermalling mode. In the thermalling mode, the autopilot continually keeps refitting the altitude change rate data it is receiving to a thermal model, e.g., the model in Equation 1 using an EKF, and trying to keep the aircraft within the thermal to gain altitude. If the altitude change rate becomes close to the “natural” sink rate predicted by the polar curve, it is an indication that the thermal is weak and it is time to exit it. While this reasoning is straightforward, operationalizing it is significantly complicated not only by the inherent uncertainty in thermal parameters but also by the subtleties in attributing the aircraft’s altitude gain to thermal lift as opposed to other factors. ArduSoar, our soaring controller integrated into an opensource fixed-wing UAV autopilot called ArduPlane [2], overcomes these difficulties and implements the above intuition using the following data streams from ArduPlane’s Attitude and Heading Reference System (AHRS): • h - altitude • φ - roll (bank) angle • v - airspeed abs • x - GPS latitude abs • y - GPS longitude wind • v - wind velocity We now describe ArduSoar’s operation in each stage. A. Polar Curve and Variometer Data The key to ArduSoar’s operation is a stream of variometer data, synthesized from data streams described above. To a first approximation, a variometer aims to measure the vertical speed of the air around an aircraft, thereby helping the autopilot determine whether the aircraft is in a thermal. In practice, however, no onboard instrument has direct access to the true vertical speed of the air. A variometer tries to deduce it from the aircraft’s own vertical speed w.r.t. the ground, which, in the case of a sailplane with its motor turned off, can be imparted by the sailplane being in a thermal or by pitching up or down. Therefore, instead of measuring the vertical airspeed, we implement a variometer that measures the rate of change of a sailplane’s total specific energy. The total energy of a sailplane in a glide (i.e. with the motor off, if present), is 1 E = mgh + mv 2 2 (9) and the specific energy (in meters) results from dividing through by mg: e=h+ v2 2g where CD0 is the zero lift drag coefficient, AR is the aspect ratio and e is the span efficiency, all of which are constants of the airframe. The lift coefficient is defined: CL = L where L is the lift force generated by the wings, ρ is the air density, S is the wing area, and v is the airspeed. Since the pitch angle is typically very shallow in a glide, we can approximate L ≈ W , where W = mg, Which gives: CL ≈ mg and setting K = 2mg ρS , which we can assume is a constant for the airframe at least at low altitudes, gives the expression: K (15) v2 Further, we can correct CD for the effect of load factor in a steady turn at bank angle φ, by dividing CL by cos φ. Setting 1 B = πeAR as an airframe constant, gives a new expression for CD from Equation 12:  2 CL CD = CD0 + B (16) cos φ The glide ratio, the ratio of distance traveled to height lost, or equivalently the ratio of airspeed to sink rate, is given by “L over D”: CL = CL v L = = (17) D CD vz and rearranging gives us an expression for sink rate vz : (10) (11) CD0 BCL + CL cos2 φ and the netto variometer signal is given as: vz = v  vz (v, φ) = v Note that the measurement units of ė are m/s. Computing variometer data involves estimating ḣ and v̇, which can be approximated by ∆e ∆t computed using successive h and v data points from the altitude and airspeed data streams. To recover the true vertical speed of the air mass surrounding the sailplane, we need to further correct ė with the aircraft’s natural sink at its current airspeed v and bank angle φ. This corrected variometer is known as a netto variometer, and has a value of 0 m/s in a steady glide through still air. The sink rate can be derived from the sailplane’s aforementioned drag polar curve. From Anderson [8], the standard simplified formula for the drag polar curve is: CD = CD0 + CL2 πeAR (14) 1 2 2 ρv S CD (18) CL Finally, substituting Equation 16 into Equation 18, gives the sink rate vz in terms of v and φ: Its rate of change over time is v v̇ ė = ḣ + g (13) 1 2 2 ρv S (12)  , ėnet = ė + vz (v, φ) (19) (20) Estimation of the CD0 , K, and B airframe-specific constants is described in Section V. B. Thermal Detection ArduSoar constantly monitors a first order low-pass-filtered version of ėnet : ėf iltn = Tc ėnet + (1 − Tc ) ėf iltn−1 (21) Where Tc = 0.03. When ėf ilt exceeds a manually settable threshold (parameter SOAR VSPEED in ArduSoar’s implementation, see Table I), ArduSoar switches to the thermalling mode. Using this filtered version of the vario signal helps to remove turbulence which might otherwise trigger the algorithm. C. Thermal Identification and Exploitation ArduSoar naturally interleaves these two stages: it uses observation data — variometer readings associated with GPS locations — to update its distribution over possible thermal parameters, and at the same time chooses its trajectory — a circle centered at a location x0 , y0 of radius R̂th — based on the thermal parameter distribution. Recall from Equation 1 that ArduSoar’s thermal model has 4 parameters that need to be estimated: W th , Rth , xth and y th . To simplify the math, we make the following assumption: Given wind velocity vector ~v wind , aircraft velocity vector ~v a , and aircraft and thermal center locations p~a and p~th in the Earth reference frame, after time dt their locations in the 0 Earth reference frame are p~a = p~a + (~v wind + ~v a )dt and th0 th wind p~ = p~ + ~v dt respectively. In other words, we assume that the aircraft and the thermal location are affected by wind in the same way. This assumption isn’t completely accurate [29], because thermals tend to move slower than the surrounding air mass. Nonetheless, for small dt, i.e., high re-estimation frequency, it is a very good approximation of reality. Moreover, it allows us to easily carry out all calculations in the aircraft’s reference frame, since, under this assumption, the relative displacement of the the aircraft and the thermal is entirely due to the aircraft’s velocity vector. As in Hazard [17], we define the aircraft to be always at the origin of its own reference frame, so the thermal state vector to be tracked as follows: dx = ∆xabs − vxwind (24) abs (25) dy = ∆y − vywind , i.e., we need to perform wind correction to obtain thermal motion w.r.t. the aircraft. Since f (Xk ) effectively has no process dynamics, we get   1 0 0 0 0 1 0 0 ∂Xk+1 ∂f  (26) = = F = 0 0 1 0 = I ∂X ∂X 0 0 0 1 For the observation function, the only readings relevant to the thermal state vector that we get is vertical air velocity w at the aircraft’s position w.r.t. the thermal center, so   2 x + y2 (27) o = w = h(X) = W th exp − 2 Rth and  2 2  exp − xR+y th 2  th 2 2  2 2   2W (x +y )  | exp − xR+y   3 ∂h th th 2 R |     H = = 2 2 th  x +y 2W x ∂X exp −   Rth 2 Rth 2     2 2 th x +y 2W y − 2 exp 2 th th R R  (28) Note the two aspects that make ArduSoar’s EKF update extra efficient:   th   • Due to Equation 26, covariance update (Equation 4) Strength W   Rth   becomes simply P̂k+1 = Pk + Q. Radius    State vector X =   x  = Center distance north of aircraft • Since our observation vector has only a single component, Center distance east of aircraft y ∂h (22) is a 1 × 4 vector (Equation vertical air velocity, and ∂X 28), it follows that in Equation 6 (H P̂k+1 H | + R) is Recall that state tracking with an EKF involves updating an integer and the matrix inversion during Kalman gain the current state distribution’s mean with a linearization of the computation in that equation is just a division. transition function f (Equation 2) and observation function h (Equation 5). In our case, assuming that the aircraft has travelled distance dx to the north and dy to the east during a ArduSoar’s EKF update runs at 5Hz, giving ArduSoar a given time step, where dx and dy are both known, stream of thermal parameter estimates. In order to use this information to gain height, ArduSoar enters the autopilot into   the LOITER mode (see the next section), in which ArduPlane 0  0  ensures that the aircraft orbits a specified location in a circle of  Xk+1 = f (Xk ) = Xk +  (23) −dx . a specified radius. ArduSoar sets (x, y) from the mean thermal estimate to be the center of this circle, but setting the radius −dy R̂ is less straightforward. In theory, using the mean Rth estimate, ArduSoar could Despite its apparent simplicity, the computation of f (Xk ) involves an important subtlety. As stated above, we aim to choose R̂ to strike the optimal balance between estimated calculate the motion of the thermal relative to the aircraft. lift R̂ meters away from the thermal center and the lift loss However, we have direct access only to the aircraft’s GPS due to bank angle required to stay in a turn of radius R̂. In positions xabs , y abs in the Earth reference frame. Since the practice, however, we made R̂ a manually settable parameter aircraft can be affected by wind, dx 6= ∆xabs and dy 6= ∆y abs for ArduSoar, for reasons described at the end of Section V. for two successive xabs (resp. y abs ) measurements. Rather, LOITER LOITER Throttle active Throttle suppressed Manual cmd Entry condition h > SOAR ALT MAX or Exit condition h < SOAR ALT MIN AUTO Fig. 2: ArduPlane’s typical flight profile with ArduSoar enabled. Image reproduced with permission. Source: http://ardupilot.org/plane/docs/soaring.html h > SOAR ALT CUTOFF Throttle active AUTO Throttle suppressed h < SOAR ALT MIN Fig. 3: State transition diagram. Normal transitions in thick lines. D. Thermal Exit ArduSoar constantly monitors the strength of the thermal to decide whether to exit and search for a better thermal. It is assumed that the sailplane is currently circling the thermal center at distance R̂, and a lift estimate at this distance is calculated using Equation 1. The estimate is then adjusted using a constant correction factor, Ksink , representing sink from the  drag polarand bank angle. Thus, ArduSoar computes 2 2 th − Ksink , and compares it against the W exp − x R̂+y 2 SOAR VSPEED threshold (Table I) to determine if the thermal has become too weak to continue with. Note that this exit condition doesn’t rely on the actual lift measurements, only on those predicted by the thermal model. This approach has the advantage over other methods [7, 14] of not being overly sensitive to the speed of the controller’s response to fluctuations in thermal size and position. IV. A RDU P LANE I NTEGRATION Although a soaring controller can enable an aircraft to stay aloft by extracting energy from the atmosphere, per se it lacks a lot of the useful functionality of a full-blown autopilot, including an AHRS system for estimating aircraft state, waypoint following capability, safety mechanisms such as geofencing, etc. Integrating ArduSoar into ArduPlane resulted in a flight controller for fixed-wing UAV’s that has the rich feature set of a major open-source autopilot in addition to thermalling functionality. To integrate ArduSoar into ArduPlane, we have augmented ArduPlane with a number of parameters configurable via ground control station software such as Mission Planner [3]. These parameters are listed in Table I. Some of them, such as SOAR VSPEED and SOAR POLAR ∗, pertain to ArduSoar itself. The rest, particularly SOAR ALT ∗, control ArduSoar’s interoperation with ArduPlane. In particular, ArduPlane has two modes relevant to ArduSoar, AUTO and LOITER: • In AUTO mode, given a set of waypoints, a target airspeed, and several other parameters, ArduPlane’s default behavior is to fly the aircraft fully autonomously, following the specified waypoint course. • In LOITER mode, ArduPlane’s default behavior is to orbit a point with specified GPS coordinates at a specified radius. We modified ArduPlane’s AUTO mode to opportunistically use thermalling functionality as shown in Figure 2. If the aircraft just took off or if it descended to the altitude given by the new SOAR ALT MIN parameter (stage 1 of the flight in Figure 2), it climbs under motor to SOAR ALT CUTOFF altitude (stage 2 in Figure 2) while following pre-set waypoints. Upon reaching SOAR ALT CUTOFF, it turns off the motor and glides down at ArduPlane’s target airspeed (stage 3). During this stage, ArduSoar keeps monitoring the aircraft state. If it detects a thermal, it takes over navigational control and starts thermalling by putting the aircraft into LOITER mode (stage 4 in Figure 2). The point ArduSoar commands the aircraft to orbit is the current mean of the thermal center estimate, (xth , y th ). The radius is given by a special ArduPlane parameter for LOITER mode called WP LOITER RAD. While thermalling, the aircraft does not follow its waypoint course and may deviate from it. It may also reach an altitude given by the SOAR ALT MAX parameter. This safety parameter terminates thermalling to prevent the aircraft from climbing too high. At that point, ArduSoar stops thermalling, gives navigational control back to ArduPlane, and the aircraft starts another motorless glide (stage 5). This process can repeat many times without any human intervention. Several of the newly introduced parameters are meant to give the human operator the ability to experiment with ArduSoar’s behavior. Specifically, SOAR VSPEED determines how strong a thermal needs to be for ArduSoar to catch it; SOAR MIN THML S allows forcing ArduSoar to stay thermalling for a minimum time period, even if ArduSoar would otherwise decide to leave the thermal. SOAR MIN CRSE S prevents ArduSoar from starting to thermal too soon after it exited a thermal previously, e.g., to make sure it doesn’t keep entering and exiting the same thermal over and over again. The state transition diagram of the resulting autopilot is shown in Figure 3. In AUTO mode, ArduPlane’s default behavior is to use the motor and follow waypoints, so implementing ArduSoar involved overriding ArduPlane’s throttle setting and navigation waypoints. Code was inserted into the main ArduPlane vehicle update functions to override the relevant variables when the soaring controller is active. Parameter name SOAR ENABLE SOAR VSPEED SOAR Q1 SOAR Q2 SOAR R SOAR DIST AHEAD SOAR MIN THML S SOAR MIN CRSE S SOAR SOAR SOAR SOAR POLAR CD0 POLAR B POLAR K ALT MIN SOAR ALT CUTOFF SOAR ALT MAX Description Enables or disables ArduSoar functionality A threshold for vertical air velocity that determines whether the aircraft that observes air moving upwards at SOAR VSPEED should enter/exit the thermalling mode. Standard deviation of the process noise for thermal strength Standard deviation of the process noise for thermal position and radius Standard deviation of the observation noise Initial guess of the distance to the thermal center along the aircraft heading Minimum time the aircraft should keep trying to soar once it has entered the thermalling mode (LOITER), in seconds Minimum time the aircraft should avoid entering the thermalling mode (LOITER) after it last exited it, in seconds Polar curve parameter (see eq. (12)) Polar curve parameter (see eq. (16)) Polar curve parameter (see eq. (15)) The minimum altitude at which soaring or off-motor gliding is allowed, in meters. If the aircraft descends to SOAR ALT MIN, it turns on the motor and starts climbing in AUTO mode. The altitude, in meters, at which the aircraft turns off the motor after a climb from SOAR ALT MIN (or from the ground) and starts gliding in AUTO mode, trying to detect thermals. The maximum altitude, in meters, that the aircraft is allowed to reach. Upon reaching it, the aircraft exits the thermalling mode if necessary and starts a glide in AUTO mode. and methods that trigger flight mode changes. 3) Variometer provides the total energy rate estimate from height, airspeed and roll measurements. 4) ExtendedKalmanFilter provides the thermal parameter estimate from the total energy rate estimate and the aircraft position. At any instant the system has two primary state variables that determine its behaviour. Flight mode is an ArduPilot feature that selectively enables control features. Throttle suppressed is a flag within the soaring controller that determines whether the throttle signal is forced to zero. The transitions between the states are triggered by update soaring() based on altitude and thermal parameter estimates. The transition diagram is illustrated in Figure 3. A number of other interactions were required between ArduSoar and ArduPlane. For example, the rapid changes of operating regime when switching between throttle suppressed and throttle active states require large changes in trim pitch angle. This mandated resetting the pitch integrator on state transitions to avoid overshoots. V. F LIGHT T ESTING Flight testing was conducted in two phases: airframe characterization, in which the polar curve parameters of a Parkzone Radian Pro, a popular remote-controlled sailplane, were estimated; and controller tuning, in which key parameters affecting the performance of ArduSoar were adjusted. The Radian Pro was equipped with a 3DR Pixhawk flight controller hardware (32-bit ARM processor, 256KB RAM, 2MB flash memory, 168MHz clock speed), a 3DR GPS and compass, a 3DR pitot/static tube and a sensor, a 3DR 915MHz telemetry radio, and a FrSky X8R receiver. The remote pilot could intervene at any point using a FrSky X9D+ 2.4GHz trasmitter. Power was provided by a single 3-cell 1300 mAh LiPo battery. The equipment diagram is shown in Figure 4. TABLE I: Parameters of the ArduSoar controller settable in ArduPlane. The main logic of the controller is divided into modules: 1) soaring.cpp provides a single function, update soaring(), that is called by the task scheduler at 50Hz. Depending on the current flight mode this function updates the soaring controller and checks for required changes to flight mode and navigation target. It exists in the main ArduPlane namespace, allowing it to set the required variables directly. 2) AP SoaringController.cpp provides the main controller logic, including methods to update the thermal estimator and the variometer, calculation of the navigation target Fig. 4: Radian Pro with its electronic equipment (the canopy is temporarily removed)  SOAR Q12  0 Q=  0 0  0  0   0 SOAR Q22 (29) (see Table I). These were tuned largely by analysis of filter performance during thermal encounters. The underlying model assumes a static updraft distribution, so SOAR R captures variations in vario readings due to actual measurement noise, unmodelled aircraft dynamics and turbulence. SOAR R was therefore taken as the variance in reading during circling flight in turbulent air. In the initial seconds of thermal encounter it is essential to maintain uncertainty in position estimates xth and y th . This allows movement of the estimated thermal center and is achieved by large initial variances of xth and y th . The aircraft will follow this estimated center, yielding additional information that rapidly localizes the thermal center and reduces the state variances. The initial variance of W th and Rth also affect the ability of xth and y th to change, and so are selected relatively smaller. After the initial filter transients, the balance between SOAR R and the elements of Q determines the steady-state magnitude of the elements of the state covariance matrix P . For a given SOAR R, small values of Q lead to low state covariances. These small values in P give a small response to innovations; the states become almost constant over time. Conversely, large values of Q increase state uncertainty so P is larger. This gives the filter the ability to refine the thermal position estimate. The elements of Q are selected to give good filter response characteristics in steady-state thermalling. The strength and radius estimates should change slowly (on the order of the period of thermal orbit) relative to the position estimates. As mentioned previously, despite the EKF providing ArduSoar with estimates of the thermal radius, we opted to set the radius R̂ that ArduSoar uses in LOITER mode to a manually specified parameter, WP LOITER RAD. We did this because for thermals at low altitudes where we conducted our test flights, we observed a large variance in the Rth estimates, indicating that they were unreliable. Moreover, we observed large covariances between W th and Rth estimates, indicating that, 0 SOAR Q22 0 0 0 0 SOAR Q22 0 Fixed Loiter Radius Performance 1.4 1.2 Vertical speed (m/s) The polar curve was estimated by conducting test glides in calm air. ArduPlane’s AUTO mode was used to track a target airspeed around a triangular course, for a range of target airspeeds. The descent rate was measured for each, and parameters SOAR POLAR B and SOAR POLAR CD0 were fitted to the resulting data using a least squares fit of Equations 12 and 16. SOAR POLAR K was determined directly from the wing dimensions and weight of the Radian Pro, and ρ = 1.225 kg m−3 was assumed since all flight testing was conducted near sea level. Controller tuning required deriving estimates of variometer measurement noise R = [SOAR R2 ], initial state values and variances xinit and Pinit and state process noise 1.0 0.8 0.6 10m 15m 20m optimal loiter radius 0.4 0.2 0.0 0 20 40 60 Thermal radius, Rth (m) 80 100 Fig. 5: Simulation results showing the trade-offs in soaring performance from using three different fixed loiter radii over a wide range of thermal sizes (Rth ) with W th = 2.5 m s−1 , compared to using the optimal loiter radius for that thermal size. at least for the irregular thermals at low altitudes, ArduSoar had difficulty distinguishing between strong narrow thermals and weak large ones. This is not surprising given ArduSoar’s lack of exploration of the thermal — once it is confident of the center location, it orbits it and doesn’t attempt to reduce the variance of its W th and Rth estimates. An exploration strategy such as deliberately varying the orbit radius could alleviate this problem. At the same time, we observed that the smallest turn radius that a Radian Pro can reliably maintain in LOITER mode, 15 m, works well across a wide range of thermals. Figure 5, based on simulated thermals, shows the trade-off in soaring performance with the choice of loiter radius. Using a small loiter radius extends performance to smaller thermals but at the expense of performance in large thermals. Additionally, trying to orbit tightly at speeds close to the stall speed, where soaring performance is best, can be challenging for the autopilot. For the Radian Pro glider a radius of 15 m at an airspeed of 9 m/s provides a good compromise. We found that setting the loiter radius too tightly resulted in occasional tip stalls leading to dangerous flat spins, which required manual intervention to recover. VI. R ELATED W ORK Besides autonomous dynamic soaring and soaring in thermals, researchers have considered many other ways of extracting energy from the atmosphere to maintain flight, such as exploiting gusts [22], wind fields [24], and orographic lift [21, 16]. Of these, ArduSoar can and successfully has been used for soaring in orographic lift, generated by wind blowing into a vertical obstacle such a tree line or a hill and forcing the air to move upwards. Although the true shape of a rising air mass in this case can be almost arbitrary and is certainly far from a column, ArduSoar’s strategy of circling in order to stay within the airmass is a valid appoach to exploiting this kind of lift. Techniques for flight path planning to increased endurance due to natural lift sources has also received attention in the literature [15, 24, 11, 12, 13], but is far from the topic of our work. Among previously proposed autonomous soaring approaches, those by Wharington et al [30, 31], Allen and Lin [7], and Edwards [14] are the closest to ours. All three use a Gaussian thermal model first proposed in [31] and later modified by Allen and Lin [7]. The difference lies in the set of hypotheses each approach maintains, and its thermal exploitation strategy. In their simulation scenarios, Warrington et al [30, 31] maintain a fixed grid of candidate thermal center locations over the flight area, and identify the node where the thermal parameter fit is best, as the center. A control strategy developed with the help of reinforcement learning (RL) [28], was trained using a simple simulator. Convergence of the solver was slow prohibiting any practical implementation. Allen and Lin [7] place the thermal center at the centroid of the recent history of ė observation locations, each weighted by the square of its measurement. Thermal radius is fitted to the observations using gradient descent. The thermal is then exploited using a controller based on a form of Reichmann rules [7, 27]. Edwards [14] mitigated the problem of biased thermal center estimation in [7] to the inside of a turn, by augmenting Allen’s centroid locater with a surrounding grid of candidate thermal locations (similar to [30]), but makes grid scale and placement adaptive. An evolutionary search recursively refines the thermal parameters by choosing the grid point with the best fit to the observation history at each step, using least squares. Edwards and Silberberg [15] also presents a weighted least squares method to simultaneously fit the all parameters of generalized ellipsoidal thermal model, without the need of a search grid. However in flight testing the method was found to be unreliable and sensitive to observation noise. Bird and Langelaan [10], presents a method to map the vertical wind field around the sailplane by fitting the weights of a parametric spline surface to observations using an extended kalman filter. To exploit the arbitrary uplift map, a contour solver is used to extract the optimal lift isotherm along which the sailplane is then guided, augmented with a sinusoidal dither signal to promote exploration and refinement of the map. Simulation results show no significant performance improvement over the previously described controllers even for irregular shaped thermals. Reddy et al. [26], like Wharington [30], employ an RL method to learn a policy, however they utilize the much more detailed thermal simulation based on Rayleigh-Bénard convection. Like Wharington [30]’s and other RL based approaches, theirs requires a random RL trial restart capability, which is far too impractical to conduct in the real world, limiting training to simulation only. Without flight testing, it is is not clear how well the learnt control strategies would carry over to the real world. On the other hand, Edwards [14] conducted a live flight test of his controller and kept a sailplane airborne for 5.28 hours along a closed course. During the tests, thermal centering ran on a laptop on the ground; due to its use of adaptive grid for thermal hypotheses, Edwards’s approach is still too expensive for a computationally constrained device as a Pixhawk. An efficient alternative means of thermal identification for simulated autonomous thermalling is by Hazard [17]. It derives a UKF-based thermal tracking approach and evaluates its performance at determining a thermal’s parameters assuming various open-loop environment exploration policies, such as flying a horizontal circular or sinusoidal pattern. Our work uses several elements of Hazard’s work, particularly the thermal model and insights regarding tracking it using a Kalman filter variant. In another implementation [19], an EKF estimator is used to identify thermal dynamics. Its state variables are thermal strength, radius, distance from aircraft and bearing to aircraft. State dynamics are introduced to cause a decay in the thermal strength state. The estimation is separated into two 2-variable EKFs, because the dynamics of the thermal strength and radius are decoupled from the position states. The formulation can be executed iteratively on an embedded autopilot. This differs from our approach in that the more complex state dynamics increase computational complexity and the decoupled estimation approach does not capture the coupling between the uncertainties in the states. These coupling terms are captured as off-diagonal terms in the covariance matrix of our 4-variable estimator. VII. C ONCLUSIONS This paper has presented ArduSoar, an embedded thermal soaring controller that has been implemented and tested in the codebase of a popular open-source autopilot for fixed-wing UAVs called ArduPlane. ArduSoar is based on a 4-variable estimator that uses an extended Kalman filter with trivial state dynamics to achieve a computational cost low enough for ArduSoar to be executable on highly resource-constrained drone autopilot hardware such as Pixhawk or even APM. The implementation demonstrates good performance and has been used by several research groups as well as many individuals for leisure, research and commercial purposes. Acknowledgements. We would like to thank Peter Braswell for making available his soaring controller code that ArduSoar’s implementation is derived from. R EFERENCES [1] Ardupilot project website. http://www.ardupilot.org, . Accessed: 2018-02-10. [2] ArduPlane, . URL http://ardupilot.org/plane/index.html. [3] Mission planner. URL http://ardupilot.org/planner/. [4] Pixhawk project website. https://pixhawk.org/. Accessed: 2018-02-10. [5] Raspberry pi project website. http://raspberrypi.org/. Accessed: 2018-02-10. [6] M. Allen. Updraft model for development of autonomous soaring uninhabited air vehicles. In Proc. of the 44th AIAA Aerospace Sciences Meeting and Exhibit, 2006. doi: https://doi.org/10.2514/6. [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] 2006-1510. URL https://ntrs.nasa.gov/archive/nasa/casi. ntrs.nasa.gov/20060004052.pdf. M.J. Allen and V. Lin. Guidance and control of an autonomous soaring UAV. Technical report, 2007. URL https://ntrs.nasa.gov/archive/nasa/casi.ntrs. nasa.gov/20070022339.pdf. J. D. Anderson. Introduction to flight, 1989. J. J. Bird, J. W. Langelaan, C. Montella, J. Spletzer, and J. Grenestedt. Closing the loop in dynamic soaring. In Proc. of the 54th AIAA Guidance, Navigation and Control Conference and Exhibit, 2014. John Bird and Jack Langelaan. Spline mapping to maximize energy exploitation of non-uniform thermals. Technical Soaring, 37(3):38–44, 2013. A. Chakrabarty and Jack W. Langelaan. Energy-based long-range path planning for soaring-capable uavs. Journal of Guidance, Control and Dynamics, 34(4):1002– 1015, 2011. J. J. Chung, N. R. J. Lawrance, and S. Sukkarieh. Learning to soar: Resource-constrained exploration in reinforcement learning. The International Journal of Robotics Research, 34(2):158–172, 2015. N. T. Depenbusch, J. J. Bird, and Jack W. Langelaan. The autosoar autonomous soaring aircraft part 2: Hardware implementation and flight results. Journal of Field Robotics. doi: 10.1002/rob.21747. URL http: //onlinelibrary.wiley.com/doi/10.1002/rob.21747/full. D. Edwards. Implementation details and flight test results of an autonomous soaring controller. In Proc. of the 46th AIAA Guidance, Navigation and Control Conference and Exhibit, 2008. doi: https://doi.org/10.2514/6.2008-7244. URL http://citeseerx.ist.psu.edu/viewdoc/download?doi= 10.1.1.559.4025&rep=rep1&type=pdf. D. J. Edwards and L. M. Silberberg. Autonomous soaring: The montague cross-country challenge. Journal of Aircraft, 47(5):1763–1769, 2010. doi: https://doi.org/ 10.2514/1.C000287. URL https://arc.aiaa.org/doi/abs/10. 2514/1.C000287. A. Fisher, M. Marino, R. Clothier, S. Watkins, L. Peters, and J. L. Palmer. Emulating avian orographic soaring with a small autonomous glider. Bioinspiration & Biomimetics, 11(1), 2016. URL http://stacks.iop.org/ 1748-3190/11/i=1/a=016002. M. W. Hazard. Unscented kalman filter for thermal parameter identification. In Proc. of the 48th AIAA Aerospace Sciences Meeting and Exhibit, 2010. L. P. Kadanoff. Turbulent heat flow: structures and scaling. Physics Today, 54(8):34–39, 2001. A. M. Kahn. Atmospheric thermal location estimation. Journal of Guidance, Control and Dynamics, 40, 2017. R.E. Kalman. A new approach to linear filtering and prediction problems. Transactions of the ASME–Journal of Basic Engineering, 82(Series D):35–45, 1960. J. W. Langelaan. Long distance/duration trajectory optimization for small UAVs. In Proc. of the 47th AIAA Guidance, Navigation and Control Conference and Exhibit, 2007. [22] J. W. Langelaan. Gust energy extraction for mini- and micro- uninhabited aerial vehicles. In Proc. of the 46th AIAA Aerospace Sciences Meeting and Exhibit, 2008. [23] N. R. J. Lawrance and S. Sukkarieh. A guidance and control strategy for dynamic soaring with a gliding uav. In Proc. of IEEE International Conference on Robotics and Automation (ICRA), pages 3632–3637, 2009. URL http://ieeexplore.ieee.org/abstract/document/5152441/. [24] N. R. J. Lawrance and S. Sukkarieh. Path planning for autonomous soaring flight in dynamic wind fields. In Proc. of IEEE International Conference on Robotics and Automation (ICRA), pages 2499–2505, 2011. URL http: //ieeexplore.ieee.org/abstract/document/5979966/. [25] P. Oettershagen, A. Melzer, T. Mantel, K. Rudin, T. Stastny, B. Wawrzacz, T. Hinzmann, S. Leutenegger, K. Alexis, and R. Siegwart. Design of small handlaunched solar-powered uavs: From concept study to a multi-day world endurance record flight. volume 34, pages 1352–1377, 2017. doi: 10.1002/rob.21717. URL http://dx.doi.org/10.1002/rob.21717. [26] G. Reddy, A.Celani, T. J. Sejnowski, and M. Vergassola. Learning to soar in turbulent environments. In Proc. of the National Academy of Science (PNAS), 2016. [27] H. Reichmann. Cross-Country Soaring. 7th ed. edition, 1993. [28] R. S. Sutton and A. G. Barto. Reinforcement Learning: An Introduction. MIT Press, 1998. [29] J. W. Telford. Convective plumes in a convective field. Journal of the Atmospheric Sciences, 27(3):347–358, 1970. [30] J. Wharington. Autonomous Control of Soaring Aircraft by Reinforcement Learning. Doctorate thesis, Royal Melbourne Institute of Technology, Melbourne, Australia, 1998. [31] J. Wharington and I. Herszberg. Control of a high endurance unmanned air vehicle. In Proc. of the 21st ICAS Congress, 1998.
3cs.SY
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 1 A Short-Term Voltage Stability Index and case studies Wenlu Zhao  Abstract—The short-term voltage stability (SVS) problem in large-scale receiving-end power systems is serious due to the increasing load demand, the increasing use of electronically controlled loads and so on. Some serious blackouts are considered to be related to short-term voltage instability. In China, the East China Grid (ECG) is especially vulnerable to short-term voltage instability due the its increasing dependence on power injection from external grids through HVDC links. However, the SVS criteria used in practice are all qualitative and the SVS indices proposed in previous researches are mostly based on the qualitative SVS criteria. So a Short-Term Voltage Stability Index (SVSI), which is continuous, quantitative and multi-dimensional, is proposed in this paper. The SVSI consists of three components, which reflects the transient voltage restoration, the transient voltage oscillation and the steady-state recovery ability of the voltage signal respectively after the contingency has been cleared. The theoretical backgrounds and affected factors of these three components of SVSI are analyzed, together with some feasible applications. The verification of the validity of SVSI are tested through more 10,000 cases based on ECG. Additionally, a simple case of selecting candidate locations to install dynamic var using SVSI is presented to show its feasibility to solve the optimization problem for dynamic var allocation. Index Terms—Short-Term Voltage Stability, Short-Term Voltage Stability Index (SVSI), case studies I. INTRODUCTION A CCORDING to the definition of short-term voltage stability (SVS) [1], it involves dynamics of fast acting load components such as induction motors, electronically controlled loads, and HVDC converters. Modern power systems are operating under more stressed situations due to the increasing load demands and electricity exchange between regions. Additionally, their dynamic responses are more complicated with the increasing use of induction motor loads (such as air-conditioners), electronically controlled loads, HVDCs and renewable power generations. As a result, the power systems are more vulnerable to short-term voltage instability. It is considered that part of the reason for the North American blackout in August 2003 may be related to SVS [2]. In China, most energy resources (such as coal, hydro and natural gas) are distributed in the west, while most loads are distributed in the east. So long-distance transmission is needed to deliver electric power from the west of China to the east of China. For instance, East China Grid (ECG) is a typical Recently Modified on Feb 22, 2017. receiving-end system and receives power through HVDC links. According to the requirements for Chinese environmental protection, the increase of fossil-fuel generation is strictly controlled unless Combined Heat and Power (CHP) units. With the dependence on power injection from external grid increases year by year, the SVS in Yangtze River Delta is getting more serious. Therefore, the SVS in ECG, especially in Yangtze River-Delta, should be taken seriously. Indices for describing voltage stability are required to evaluate the SVS. Currently the indices for describing transient stability can be divided into three categories: indices based on Transient Energy Function [3-6], Critical Clearing Time (CCT) [7] and indices based on voltage signals. The front two categories are most used in the study of rotor angle stability, while the latter one category is most used in the study of SVS. Some SVS criteria which used in practice are listed below [8-13]: Western Electricity Coordinating Council (WECC):  Voltage dip should not to exceed 25% at load buses or 30% at non-load buses and not to exceed 20% for more than 20 cycles at load buses under N-1 contingencies.  Voltage dip should not to exceed 30% at any bus and not to exceed 20% for more than 40 cycles at load buses under N-2 contingencies. Tennessee Valley Authority (TVA):  Transmission system voltage should recover to 90% of nominal system voltage within 0.5 seconds of fault clearing under N-1 contingencies.  Transmission system voltage should recover to 90% of nominal system voltage within 0.5 seconds of fault clearing or unit tripping under N-2 contingencies. State Grid Corporation of China (SGCC): Voltage should restore to 0.8p.u. within 10s after the fault has been cleared: China Southern Power Grid (CSG) Voltage should not drop to under a certain level such as 0.75p.u. lasting for longer than 1 s Therefore, the SVS criteria used in practice can only determine the system is stability/instability, but they cannot reflect the degree of stability/instability. Furthermore, these criteria are merely dependent on the operating experiences and are lack of theoretical backgrounds. Therefore, the circumstance in which the criteria of WECC and TVA are in the same form but have different threshold occurs. Due to the increasing complexity of modern power systems, the current SVS criteria can hardly satisfy the demand for the > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < evaluation and optimization of SVS. Some indices to evaluate the SVS have been introduced in the previous researches [14-19], but they are mostly dependent on the above-mentioned SVS criteria and their theoretical backgrounds have not been discussed. Therefore, a Short-Term Voltage Stability Index (SVSI), which is continuous, quantitative and multi-dimensional, is proposed in this paper to evaluate the SVS, and it is tested in the SVS evaluation of ECG to verify its effectiveness. The SVSI consists of three components, and each of them has theoretical backgrounds and affected factors. Moreover, the proposed SVSI can be additionally used in the researches on grid-partition, evaluation and optimization of dynamic vars and SVS evaluation using data-mining. II. SHORT-TERM VOLTAGE STABILITY INDEX The Short-Term Voltage Stability Index (SVSI) proposed in this paper consists of three components: SVSIr, SVSIo, SVSIs. These three components of SVSI reflects transient voltage restoration, transient voltage oscillation and ability to reach steady-state respectively. The bus voltage signals considered in this paper are all normalized. Namely, if v0(t) is the voltage signal in transient process corresponding to a bus and v0(0) is the voltage of a bus before contingency, then the normalized voltage signal is: v(t) = v0(t)/v0(0) (1) After normalization, all of the voltage signals starts at 1. The advantages of normalization will be stated below. Based on the SVS criteria listed in Section I, a new definition of transient voltage restoration is proposed as follows: if the average of v(t) is above 0.8pu in the time period which is from 9s to 10s after the contingency has been cleared, then the bus is considered as transient voltage recovered; otherwise is considered as transient voltage unrecovered. The definition and algorithm of SVSI is different between the transient voltage recovered scenario and transient voltage unrecovered scenario. So SVSIr, SVSIo, SVSIs will be presented separately under these two scenarios. A. SVSIr under transient voltage recovered scenario Two important arguments: VS and TSVSIr are needed to be defined before the calculation of SVSIr. The argument VS is defined to represent steady-state voltage after the transient process. Since the time for simulation is limited, the true value of VS is hard to obtain. Therefore, some logical judgments for approximation of VS is proposed. The most common circumstance of the voltage signals after the contingency has been cleared is: the voltage firstly surges to a relative high status, then go through an oscillation damping period and finally comes to a steady-state. Therefore, the computational expression for VS under common circumstances is: VS = (max(v(t))end + min(v(t))end)/2 (2) max(v(t))end is the last peak of v(t) and min(v(t))end is the last valley of v(t). In addition, there are two special circumstances need to be considered while the fluctuation cycle and fluctuation 2 amplitude of v(t) cannot be recognized: Firstly, if the slope of v(t) tends to be gentle since the time when v(t) reached the last extreme value, then: VS = v(TEnd) (3) Tend is the end time of simulation (i.e. end time of the SVS considered in this paper). Secondly, if the slope of v(t) tends to be steep since the time when v(t) reached the last extreme value, then: VS = (max(v(t))end + min(v(t))end)/2 (4) The definition of TSVSIr is: the moment when v(t) firstly reaches VS during ramping periods after the contingency has been cleared. Under the most common circumstances, the computational expression for TSVSIr is: min TSVSIr  (TClear , TEnd ) s.t. v(t ) t T SVSIr   0,  VS dv(t ) 0 dt t(TSVSIr  ,TSVSIr  ) t0  (TSVSIr   , TSVSIr   ), (5) dv(t ) 0 dt t t0 TClear is the moment in simulation when contingency is cleared. In addition, there is a special circumstance need to be considered in which the inequality v(t) > VS holds from TClear to TEnd, thus: TSVSIr = TClear (6) Based on the arguments defined above and simulation result, the computational expression of tSVSIr is: SVSI r =  TSVSIr TFlt v  t   VS dt (7) TFlt is the moment when contingency starts. Fig 1 The schematic of SVSIr under transient voltage recovered scenario In Fig 1, the area of yellow region is SVSIr. v(t) maintains nearly post-contingency steady-state (VS) after tSVSIr, but experiences a continuous low-voltage stage before tSVSIr. In this paper, the time period from tFlt to tSVSIr is defined as transient voltage restoration stage and the time period from tSVSIr to tEnd is defined as after transient voltage restoration stage. SVSIr is the transient voltage restoration component of SVSI. It reflects the dynamic reactive power balance and the transient voltage restoration in the short time after the contingency has been cleared. It is closely related to the dynamic response of induction motors (such as AC air-conditioners), over-excitation limits of generators, dynamic vars and HVDCs. The SVSIr reflects the key factors of SVS, and the greater the value is, the > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < greater is the possibility of short-term voltage instability. During the process of transient voltage restoration, the induction motors and HVDCs consumes a massive of reactive power and thus threaten the voltage stability. On the contrary, the over-excitation of generators and the dynamic vars such as STATCOMs and SVCs can rapidly send out reactive power to support the SVS. After the transient voltage restoration process, the variation of operating states of the induction motors and the HVDCs are not threatening because their consumption of reactive power returns to the normal level and thus the threaten to SVS is reduced. The induction motors consume a massive of reactive power when they are operating under low voltage or stalled status. The induction motors threaten the SVS during transient voltage restoration stage because the bus remains under-voltage. The operating states of the induction motors depend on the balance between the electromagnetic torque (Te) and the mechanical torque (Tm). Te is proportional to the square of bus voltage and Tm usually increases with the speed of the motor. The bus voltage reaches the post-contingency steady-state level at the end of transient voltage restoration stage under transient voltage recovered scenario, so Te recovers to the post-contingency steady-state level as well. So Te is sufficient to maintain the balance with Tm after transient voltage restoration stage and thus the variation of operating states of the induction motors are not threatening. Coupled with the effect of damping, the induction motors will return to steady-state eventually. The HVDCs consume a massive of reactive power in the restoration process after commutation failures, which are harmful to the SVS. The operating states of the HVDCs are closely related to the voltage at the coupling point with AC system (VAC) and the extinction angles (γ). When γ decreases, which may cause by a voltage dip, to a threshold value, the commutation failure of the HVDC will occurs. The bus remains under-voltage during transient voltage restoration stage, so the commutation failure of the HVDC may occur and thus threaten the SVS. Regardless of whether the commutation failure occurred in transient voltage restoration stage, the HVDC will operate normally after transient voltage restoration stage under transient voltage recovered scenario. Because the HVDC operates normally at post-contingency steady-state and its control strategy of commutation failure prevention after it has been restored from commutation failure allows it to have a relative strong ability against voltage disturbance. To reduce SVSIr and enhance SVS, we can focus on: modification of induction motors, adjustment of HVDC controllers, optimize AVR parameters and increase dynamic var reservation. B. SVSIo under transient voltage recovered scenario The definition of VSVSIo(t) is needed before the calculation of SVSIo. VSVSIo(t) reflects the mid-line of oscillation in v(t) as shown in Fig 2. The oscillation of v(t) can be considered in the area of power system small signal analysis, as the transient process can be studied in the area of small signal stability analysis when the voltage recovers to a certain level and the 3 system tends to a steady-state after the contingency has been cleared. So, the aperiodic or low frequency fluctuations (usually lower than 0.1 Hz [20]) component of v(t) is retained in VSVSIo(t), while the high frequency fluctuations (usually lower than 2 Hz [20]) of v (t ) are considered as noises as shown in Fig 5. The computational expression for SVSIo is: SVSI o   tEnd tClear vTVSIo  t   v  t  dt (8) Fig 2 The schematic of SVSIo under transient voltage recovered scenario In Fig 2, the area of yellow region is SVSIo. It is the transient voltage oscillation component of SVSI, which reflects the balance of active power between generators and loads and the damping ability of the system. It is closely related to dynamic responses of high-gain fast-excitation devices [21], generator speed control systems [22] and PSSs. The greater the SVSIo is, the slower the oscillation attenuates after the contingency has been cleared, and thus the greater is the possibility of short-term voltage instability. To reduce SVSIo and enhance SVS, we can focus on optimizing of the dynamic control of generators and performance of PSSs. C. SVSIs. under transient voltage recovered scenario The definition of TStable is need to be defined before the calculation of SVSIs. TStable is the moment when v(t) reaches a post-contingency steady-state. Since the simulation time is limited, the true value of TStable is hard to obtain. Therefore, some logical judgments for approximation of TStable is proposed. The computational expression for TStable under the most common circumstances is: min TStable (9) s.t.t  TStable , TEnd  , VS  v (t )  VwTh VwTh is the amplitude threshold of voltage fluctuation/wave for determining steady-state. Refers to the requirements for voltage steady-state judgement in [21], VwTh is set to be 0.01p.u. In addition, there are two special circumstances to be considered. Firstly, if a continuous oscillation with increasing amplitude is observed in v(t), then: TStable = TEnd (10) Secondly, if t  TClear , TEnd  , VS  v(t )  VwTh holds, then: TStable = TClear (11) Based on the arguments defined above and the voltage signal, the computational expression for SVSIs is: SVSI s  (v(0)  VS )  (tStable  tClear ) (12) > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 4 Fig 3 The schematic of SVSIs under transient voltage recovered scenario Fig 5 The schematic of SVSIo under transient voltage unrecovered scenarios In Fig 3, the area of yellow region is SVSIs. It is the steady-state recovery ability component of SVSI, which reflects the speed to reach post-contingency steady-state and voltage level of bus at post-contingency steady-state. It is closely related to the topology and operating mode (such as: tripping transmission lines, load-shedding, switching shunt compensations) of the system. The greater the SVSIs is, the slower the post-contingency steady-state reaches. and the lower the voltage of bus is at post-contingency steady-state, and thus the greater is the possibility of short-term voltage instability. To reduce SVSIs and enhance SVS, we can focus on the modification of post-contingency operating mode of the system. Fig 6 The schematic of SVSIs under transient voltage unrecovered scenarios D. SVSI under transient voltage unrecovered scenarios The main difference of SVSIr, SVSIo and SVSIs under transient voltage unrecovered scenario is the different definitions of VS, TSVSIr, TStable and SVSIr. The definitions of rest of the arguments and indices are the same as the transient voltage recovered scenario. The new definitions of VS, TSVSIr, TStable and SVSIr are: VS = vSVSIo(TEnd) (13) TSVSIr = TEnd (14) TStable = TEnd (15) SVSI r =  TTVSIr TFlt v(t )  v(0) dt (16) The new schematics for SVSIr, SVSIo and SVSIs are shown in Fig 4-6. Fig 4 The schematic of SVSIr under transient voltage unrecovered scenarios E. Some Discussions 1) Motivation of Normalization The major factor of most short-term voltage instability cases is the continuous under-voltage at buses. So when suffering from a same voltage dip, the SVS of the buses whose voltage is originally in a lower level is worse, while by contrast the buses whose voltage is originally in a higher level is better. Normalization increases the SVSI corresponding to the buses whose voltage is originally lower than base level, and decreases the SVSI corresponding to the buses whose voltage is originally higher than base level, which makes the result of SVS evaluation more reasonable. 2) Probable application of SVSI a) Grid-Partition According to SVSI Utilizing SVSI to describe the response of voltage after the contingency has been cleared, thus the grid-partition can be made according to the SVSI corresponding to buses under the set of credible contingencies. The grid-partition results are mainly affected by the topology and the set of credible contingencies of the system, which usually change little for a long time. Therefore, the grid-partition result can be used in SVS evaluation until the next change of topology or set of credible contingencies of the system, thus reducing the computational amount. b) Evaluation and Optimization of Dynamic VAR Reserve SVSI can quantitatively evaluate the SVS, so the sensitivity function between dynamic var and SVSI can be used for the evaluation and optimization of dynamic var reserve. c) Real-Time Evaluation of SVS based on Data-Mining The theoretical backgrounds and the affected factors of the > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < SVSI are stated above. Utilizing these affected factors as input and the SVSI as output, the supervised learning artificial neural network (ANN) can be constructed. The constructed ANN model can evaluate the SVS much faster than Time Domain Simulation, so it can be used in on-line evaluation of SVS. In summary, the SVSI consists of three components, which can reflect the key features of voltage signals after the contingency has been cleared. Moreover, each of the components has theoretical backgrounds and affected factors. Therefore, SVSI is a comprehensive criterion to evaluate the SVS and it can be applied to various researches. III. CASE STUDIES The validity of SVSI are verified based on ECG system, which contains 3,435 buses and 5,838 transmission lines. The typical operating mode in summer is chosen for analysis, because the load level is summer is high and thus the SVS of the system is vulnerable. More than 10,000 cases have been carried out for the verification, but only four typical and one practical instance are shown due to limited space. A. Case I: Modify the proportion of induction motor loads Fig 7 The voltage signals and SVSI under the scenarios with different proportion of induction motor loads Fig 7 shows the voltage signals and the SVSI as the proportion of induction motors in loads changes. From the voltage signals on the left, the restoration ability of voltage signals becomes weaker as the proportion of induction motor increases. This conforms to the statement about the impact of induction motors on SVS as above-mentioned. From the SVSI on the right, SVSIr increases significantly as the proportion of induction motor increases. As a contrast, SVSIo and SVSIs change little when the proportion of induction motor is less than 50%. From the voltage signals, when the proportion of induction motor increases to 55%, the voltage signal is nearly unstable; when the proportion increase to 60%, the short-term voltage instability occurs. From the SVSI, the SVSIr, SVSIo and SVSIs are all increased significantly when the proportion of induction motor increased to more than 55%. So if all of the three components of SVSI increases significantly, the voltage signal may nearly unstable or already unstable. This conclusion remains correctness in the following cases. 5 B. Case II: Modify the installations of dynamic var Fig 8 The voltage signals and SVSI under the scenarios with different configurations of dynamic var Fig 8 show the voltage signals and the SVSI as the installation of dynamic var changes. From the voltage signals on the left, the restoration ability of voltage signals becomes better when the dynamic var is installed. Moreover, the voltage restoration ability of voltage signals is better when the dynamic var locates near the measurement bus in contrast to the dynamic var locates remote from the measurement bus. From the SVSI on the right, all of the three components increases as the dynamic var arrangement changes in the sequence of no dynamic var, dynamic var is remote and dynamic var is nearby. C. Case III: Modify the configuration of PSS Fig 9 The voltage signals and SVSI under the scenarios with different configuration of PSS Fig 9 shows the voltage signals and the SVSI as the configuration of PSS changes. From the voltage signals on the left, the oscillation of voltage signals is more obvious when the system is configured with proper PSS in contrast to when the system is configured with improper PSS. From the SVSI on the right, the SVSIo is smaller when the PSS is proper. The reason for the SVSIr is larger under the ‘Proper PSS’ case may be that the PSS restricts the response of fast-excitation to prevent the sustained oscillation after voltage restoration. D. Case IV: Modify the strategy of under-voltage load-shedding Fig 10 The voltage signals and SVSI under the scenarios with different strategies of under-voltage load-shedding > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < Fig 10 shows the voltage signals and SVSI as the strategy of under-voltage load-shedding changes. From the voltage signals on the left, the ability of voltage signals to reach steady-state and the voltage level at post-contingency steady-state is better as the proportion of load-shedding increases. From the SVSI on the right, the SVSIs decreases as the proportion of load-shedding increases. E. Case V: select installation locations for dynamic var 6 contrast to when the dynamic var is installed at Location B. However, under the case that contingency occurred at location 4, only part of the SVSI in this part is smaller when the dynamic var is installed at Location B (closer to contingency location) in contrast to when the dynamic var is installed at Location A. Therefore, Location A is a better choice for dynamic var installation in contrast to Location B. Furthermore, the optimal locations to install dynamic vars can be solved based on the above-mentioned procedure. IV. CONCLUSION Fig 11 Diagram of part of ECG Fig 11 is the diagram of a part of ECG. The buses are anonymous due to confidential considerations. This part locates at the end of East China Gird, which connects to the rest part only through Bus1. The red numbers represents the candidate locations of contingencies and the blue letters represents the candidate location to install dynamic vars. A continuous, quantitative and multidimensional index (SVSI) is presented in this paper for evaluation of SVS. SVSIr, SVSIo and SVSIs are the three components of SVSI, which reflects the transient voltage restoration, the transient voltage oscillation and the steady-state recovery ability of the voltage signal respectively after the contingency has been cleared. The theoretical backgrounds and affected factors of the three components are analyzed in detail to reveal the wide applicability of SVSI. More than 10,000 cases based on ECG have been carried out to verify the effectiveness of SVSI and five of them are presented in this paper. The presented cases shown that: 1) The SVSI can reflect the key characteristics of the voltage signal after the contingency has been cleared. 2) The three components of SVSI can be adjusted through the modifications of their affected factors respectively. 3) The SVSI can be applied to determining the optimal locations to install dynamic vars. Moreover, the SVSI can also be applied to grid-partitioning, evaluation and optimization of dynamic var reserve, Real-Time evaluation of SVS based on Data-Mining and so on. These are also the following research contents of this study. REFERENCES [1] Fig 12 Effect of dynamic var when contingency occurred at Location 2 Fig 13 Effect of dynamic var when contingency occurred at Location 4 Fig 12 and Fig 13 show the composite value of SVSI when contingency occurred at two of the candidate locations. Under the case that the contingency was occurred at location 2, all of the SVSI in this part are much smaller when the dynamic var is installed at Location A (closer to contingency location) in Termsdefinitions I J F O S. Definition and Classification of Power System Stability[J]. IEEE Transactions on Power Systems Pwrs, 2004. [2] Potamianakis E G, Vournas C D. Short-term voltage instability: effects on synchronous and induction machines[J]. IEEE Transactions on Power Systems, 2006, 21(2):791-798. [3] Fouad A A, Vittal V. Power System Response to a Large Disturbance: Energy Associated with System Separation[J]. IEEE Transactions on Power Apparatus & Systems, 1983, 102(11):3534-3540. [4] Llamas A, De L R L J, Mili L, et al. Clarifications of the BCU approach for transient stability analysis[J]. Power Systems IEEE Transactions on, 1995, 10(1):210-219. [5] Xue Y, Van Custem T, Ribbens-Pavella M. Extended equal area criterion justifications, generalizations, applications[J]. IEEE Transactions on Power Systems, 1989, 4(1):44-52. [6] Yu Yixin. A Study on Dynamic Security Regions of Power Systems[J]. Proceedings of Electric Power System and Automation, 1990. [7] Ardyono Priyadi †, Yorino N, Tanaka M, et al. A direct method for obtaining critical clearing time for transient stability using critical generator conditions[J]. IEEE Transactions on Power Systems, 2010, 25(3):1620-1626. [8] NERC Planning Standards, http://www.nerc.com [9] NERC/WECC, “Planning Standards”, 2003, https://www.wecc.biz [10] TVA, “TVA 2003 Transient Stability Planning Criteria”, 2003, https://www.tva.gov/ [11] National Energy Board ． Technique specification of power system security and stability calculation[S]．Beijing：China Electric Power Press， 2013 > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < [12] SGCC. Q / GDW 404-2010, Security and stability of the national grid computing specification [M]. China Electric Power Press, 2010 [13] CSG, Q / CSG 11004-2009, Guide on security and stability analysis for CSG [S], 2009 [14] Xu Y, Dong Z Y, Meng K, et al. Multi-Objective Dynamic VAR Planning Against Short-Term Voltage Instability Using a Decomposition-Based Evolutionary Algorithm[J]. IEEE Transactions on Power Systems, 2014, 29(6):2813-2822. [15] Paramasivam M, Salloum A, Ajjarapu V, et al. Dynamic optimization based reactive power planning to mitigate slow voltage recovery and short term voltage instability[C]// Pes General Meeting | Conference & Exposition. IEEE, 2014:1-1. [16] Xu Y, Zhang R, Zhao J, et al. Assessing SVS of Electric Power Systems by a Hierarchical Intelligent System.[J]. IEEE Transactions on Neural Networks & Learning Systems, 2015, 27(8):1-1. [17] Dong Y, Xie X, Zhou B, et al. An Integrated High Side Var-Voltage Control Strategy to Improve SVS of Receiving-End Power Systems[J]. IEEE Transactions on Power Systems, 2015, 31(3):1-11. [18] Tiwari A, Ajjarapu V. Optimal Allocation of Dynamic VAR Support Using Mixed Integer Dynamic Optimization[J]. IEEE Transactions on Power Systems, 2011, 26(1):305-314. [19] Wildenhues S, Rueda J L, Erlich I. Optimal Allocation and Sizing of Dynamic Var Sources Using Heuristic Optimization[J]. IEEE Transactions on Power Systems, 2014, 30(5):1-9. [20] Kundur P. Power system stability and control[M]. New York: McGraw-hill, 1994 [21] Brown P G, Demello F P, Lenfest E H, et al. Effects of excitation, turbine energy control, and transmission on transient stability[J]. IEEE Transactions on Power Apparatus and Systems, 1970 (6): 1247-1252. [22] Vournas C D. Unstable frequency oscillations in a slow-response reheat-turbine generator[C]//Power Engineering Society 1999 Winter Meeting, IEEE. IEEE, 1999, 1: 140-144. [23] China Machinery Industry Standard (CMIS) GB/T 12326-2008,Power quality – Voltage fluctuation and flicker[S], 2008 7
3cs.SY
Set membership with non-adaptive bit probes Mohit Garg∗1 and Jaikumar Radhakrishnan2 1 arXiv:1612.09388v1 [cs.DS] 30 Dec 2016 2 Tokyo Institute of Technology, Tokyo, [email protected] Tata Institute of Fundamental Research, Mumbai, [email protected] Abstract We consider the non-adaptive bit-probe complexity of the set membership problem, where a set S of size at most n from a universe of size m is to be represented as a short bit vector in order to answer membership queries of the form “Is x in S?” by non-adaptively probing the bit vector at t places. Let sN (m, n, t) be the minimum number of bits of storage needed for such a scheme. Buhrman, Miltersen, Radhakrishnan, and Srinivasan [4] and Alon and Feige [1] investigated sN (m, n, t) for various ranges of the parameter t. We show the following. 2 2 General upper bound (t ≥ 5 and odd): For odd t ≥ 5, sN (m, n, t) = O(tm t−1 n1− t−1 lg 2m n ). This improves on a result of Buhrman et al. that states for odd t ≥ 5, sN (m, n, t) = 4 1 lg lg m) and n ≤ m1−ǫ (ǫ > 0), we O(m t+1 n). For small values of t (odd t ≥ 3 and t ≤ 10 2 2 obtain adaptive schemes that use a little less space: O(exp(e2t )m t+1 n1− t+1 √ lg m). Three probes (t = 3) lower bound: We show that sN (m, n, 3) = Ω( mn) for n ≥ n0 for some constant n0 . This improves on a result of Alon and Feige that states that for q mn n ≥ 16 lg m, sN (m, n, 3) = Ω( lg m ). The complexity of the non-adaptive scheme might, in principle, depend on the function that is used to determine the answer based on the three bits read (one may assume that all queries use the same function). Let sfN (m, n, 3) be the minimum number of bits of storage required in a three-probe non-adaptive scheme where the function f : {0, 1}3 → {0, 1} is used to answer the queries. We show that for large class of functions f (including the majority function on three bits), we in fact have 1 sN (m, n, 3) = Ω(m1− cn ) for n ≥ 4 and some c > 0. In particular, three-probe non-adaptive schemes that use such query functions f do not give any asymptotic savings over the trivial characteristic vector when n ≥ log m. 1 Introduction The set membership problem is a fundamental problem in the area of data structures and information compression and retrieval. In its abstract form we are given a subset S of size at most n from a universe of size m and required to represent it as a bit string so that membership queries of the form “Is x in S?” can be answered using a small number of probes into the bit string. The characteristic function representation provides a solution to this problem: just one bit-probe is needed to answer queries, but all sets are represented using m-bit strings (which is very wasteful when n is promised to be small). The trade-off between the number of bits in the representation and the number of probes is the subject of several previous works: it was studied by Minsky and Papert in their 1969 book Perceptrons [10]; more recently, Buhrman, Miltersen, Radhakrishnan and Venkatesh [4] showed the existence of randomized schemes that answer queries with just one bit probe and use near optimal space. In contrast, they showed that deterministic schemes that answer queries ∗ A part of this work was done when the first author was at the Tata Institute of Fundamental Research. 1 by making a constant number of probes cannot use optimal space. The deterministic worstcase trade-off for this problem was also considered in the same paper and in several subsequent works (e.g., Radhakrishnan, Raman and Rao [11], Alon and Feige [1], Radhakrishnan, Shah and Shannigrahi [13], Viola [14], Lewenstein, Munro, Nicholson and Raman [7], Garg and Radhakrishnan [5]). For sets where each element is included with probability p, Makhdoumi, Huang, Médard and Polyanskiy [9] showed, in particular, that no savings over the characteristic vector can be obtained in this case for non-adaptive schemes with t = 2. In this work, we focus on deterministic schemes with non-adaptive probes, where the probes are made in parallel (or equivalently the location of probes do not depend on the value read in previous probes). Such schemes have been studied in several previous works. Let sN (m, n, t) be the minimum number of bits of storage required in order to answer membership queries with t non-adaptive probes. Definition 1. A non-adaptive (m, n, s, t)-scheme consists of a storage function and a query  s scheme. The storage function has the form φ : [m] ≤n → {0, 1} that takes a set of size at most n and returns its s-bit representation. The query scheme associates with each element x the t probe locations (i1 (x), . . . , it (x)) ∈ [s]t and a function fx : {0, 1}t → {0, 1}. We require that  for all S ∈ [m] ≤n and all x ∈ [m]: x ∈ S iff fx (φ(S)[i1 (x)], φ(S)[i2 (x)], . . . , φ(S)[it (x)]) = 1. Let sN (m, n, t) denote the minimum s such that there is an (m, n, s, t)-scheme. In our discussion, we use s(m, n, t) (without the subscript N ) to denote the minimum space required for adaptive schemes. Using the above notation, we now describe our results and their relation to what was known before. All asymptotic claims below hold for large m. 1.1 General non-adaptive schemes Theorem 2 (Result 1, non-adaptive schemes). For odd t ≥ 5, we have 2 2 sN (m, n, t) = O(tm t−1 n1− t−1 lg 2m ). n 4 In comparison, for odd t ≥ 5, Buhrman et al. showed that sN (m, n, t) = O(m t+1 n). The exponent of m in their upper bound result is roughly four times the exponent of m appearing in their lower bound result. Their schemes are non-adaptive and use the MAJORITY function to answer membership queries. We exhibit schemes that still use MAJORITY but need less space. 1 1 Buhrman et al. also show a lower bound of s(m, n, t) = Ω(tm t n1− t ) valid (even for adaptive schemes) when n ≤ m1−ǫ (for ǫ > 0 and t ≪ lg m). Note that the exponent of m in our result is twice the exponent of m appearing in the lower bound result. These schemes, as well as the non-adaptive scheme for t = 4 due to Alon and Feige [1], have implications for the problem studied by Makhdoumi et al. [9]; unlike in the case of t = 2, significant savings are possible if t ≥ 4, even with non-adaptive schemes1 . Using a similar proof idea, we obtain slightly better upper bound with adaptive schemes when t is small and n is at most m1−ǫ . Theorem 3 (Result 2, adaptive schemes). For odd t ≥ 3 and t ≤ (ǫ > 0), we have s(m, n, t) = O(exp(e2t )m 2 t+1 n 2 1− t+1 lg m). 1 10 lg lg m and for n ≤ m1−ǫ Technique. To justify our claim, we need to describe the query scheme, that is, (i1 (x), i2 (x), . . . , it (x)) for each x ∈ [m] and the query function fx : {0, 1}t → {0, 1}. For fx we use the MAJORITY on t bits (t is odd). The locations to be probed for each element will be obtained using a probabilistic argument. Once a query scheme is fixed, we need to show how the assignment to the memory is obtained. For this, we describe a sequential algorithm. We show that the random assignment of locations ensures sufficient expansion allowing us to start with a greedy argument 1 We are grateful to Tom Courtade and Ashwin Pananjady for this observation. 2 arrange that most queries are answered correctly, and then use Hall’s bipartite graph matching theorem to find the required assignment for the remaining elements. Versions of this argument have been used in previous works [8, 4, 6, 1, 5]. 1.2 Three non-adaptive probes For one probe and m ≥ 2, it is easy to show that no space can be saved over the characteristic vector representation. For two non-adaptive probes, only for the special case n = 1, some non√ trivial savings over the characteristic vector representation are possible: sN (m, 1, 2) = θ( m)). For n ≥ 2, Buhrman et al. [4] showed sN (m, n, 2) = m. The smallest number of probes for which the complexity of problem with non-adaptive probes is not settled is three. Observe that any scheme with two adaptive probes can be converted to a scheme with three non-adaptive probes; the two probe decision tree has at most three nodes. Thus, using the two adaptive probes upper 1 bound result of Garg and Radhakrishnan [5], we have sN (m, n, 3) ≤ s(m, n, 2) = O(m1− 4n+1 ). Thus, non-trivial savings in space over the characteristic vector representation is possible when n = o(lg m). Consequently, the question is whether more space can be saved or is this upper bound tight? We are not aware of any three-probe non-adaptive scheme that manages with o(m) space q for sets of size ω(lg m). Alon and Feige [1] show the following lower bound: sN (m, n, 3) = Ω( lgmn m ) for n ≥ 16 lg m. In order to obtain better lower bounds for three-probe non-adaptive schemes, we proceed as follows. In any three-probe non-adaptive scheme, the query scheme specifies, for each element, the three locations to probe and a three variable boolean function to be applied on three values read. In principle, for different elements, the query scheme can specify different boolean functions. But since there are only a finite number (256) of boolean functions on three variables, some set of at least m/256 elements of the universe use a common function. We may thus restrict attention to this part of the universe, and assume that the function being employed to answer queries is always the same. Furthermore, we may place functions obtained from one another by negating and permuting variables in a common equivalence class, and restrict our attention to one representative in each class. For three variable boolean functions, Pólya counting yields that there are twenty-two equivalence classes. This classification of the 256 functions into twenty-two classes is already available in the literature [15]. We show the following. Theorem 4 (Result 3). (a) If the query function f : {0, 1}3 → {0, 1} is not equivalent to 1 (x, y, z) 7→ (x ∧ y) ⊕ z or (x, y, z) 7→ 1 iff x + y + z = 1, then sN (m, n, 3) = Ω(m1− cn ) for n ≥ 4 and some c > 0. (b) If the query function f : {0, 1}3 → {0, 1} is equivalent to (x, y, z) 7→ (x ∧ y) ⊕ z or √ (x, y, z) 7→ 1 iff x + y + z = 1, then sN (m, n, 3) = Ω( mn). (c) If the query function f : {0, 1}3 → {0, to (x, y, z) 7→ (x ∧ y) ⊕ z and q1} is equivalent lg m m n lg m ≤ n ≤ lg m , then sN (m, n, 3) = Ω( mn lg lg m ). The best upper bounds for non-adaptive schemes with four or more probes use the MAJORITY function to answer membership queries. Our result implies that for three non-adaptive probes, when queries are answered by computing MAJORITY, the space required is at least 1 Ω(m1− cn ) for some constant c. In fact, similar lower bound holds if membership queries are answered using most boolean functions. Our results do not yield a similar lower bound for (x, y, z) 7→ (x ∧ y) ⊕ z and (x, y, z) 7→ 1 iff x + y + z = 1 types. For these two types of query functions, we get a slightly better lower bound than what is implied by [1]. Thus, further investigations on three probes non-adaptive schemes need to focus on just (x, y, z) 7→ 1 iff x + y + z = 1 and (x, y, z) 7→ (x ∧ y) ⊕ z as the query functions. Technique. As mentioned above, there are twenty-two types of functions for which we need to prove a lower bound. Seven of the twenty-two classes contain functions that can be represented 3 by a decision tree of height at most two. Thus, for these functions, the two probe adaptive lower bound in [5] implies the result. These functions are: constant 0, constant 1, the DICTATOR function (x, y, z) 7→ x, the function (x, y, z) 7→ x∧ y, its complement (x, y, z) 7→ x̄∨ ȳ, (x, y, z) 7→ (x ∧ y) ∨ (x̄ ∧ z), and (x, y, z) 7→ (x ∧ y) ∨ (x̄ ∧ ȳ). After this, fifteen classes remain. Functions in some eleven of the remaining fifteen classes admit a density argument, similar in spirit to the adaptive two-probes lower-bound proof in [5]. To streamline the argument, we classify these eleven classes into two parts. The first part contains the MAJORITY function. The second part contains the AND function, the ALLEQUAL function, the functions (x, y, z) 7→ (x ⊕ y) ∧ z, (x, y, z) 7→ (x ∨ y) ∧ z, (x, y, z) 7→ (x ∧ y ∧ z) ∨ (ȳ ∧ z̄), and their complements. For functions in the second part we deal with two functions—a function and its complement—with a single proof. In these proofs, we produce sets S and T of size at most n such that storing S and not storing T leads to a contradiction. The proof for the complement function works with a small twist: storing T and not storing S leads to the contradiction. Thus, these eleven cases are handled by six proofs. In each of these proofs we roughly argue (sometimes probabilistically) that if the scheme is valid, it must conceal a certain dense graph that avoids small cycles. Standard graph theoretic results (the Moore bound) that relate density and girth then gives us the lower bound. For the remaining four classes, we employ linear-algebraic arguments. Representatives chosen from these classes are PARITY, (x, y, z) 7→ 1 iff x + y + z 6= 1, (x, y, z) 7→ (x ∧ y) ⊕ z, and (x, y, z) 7→ 1 iff x + y + z = 1. For PARITY and (x, y, z) 7→ 1 iff x + y + z 6= 1, we show using standard dimension argument, that if the space used is smaller than the universe size m, then there is some element u ∈ [m] that is (linearly) dependent on the other elements. Not storing the other elements, leaves the scheme with no choice for u, thus leading to a contradiction. For (x, y, z) 7→ (x ∧ y) ⊕ z and (x, y, z) 7→ 1 iff x + y + z = 1 a modification of an algebraic argument √ of Radhakrishnan, Sen and Venkatesh [12] implies a lower bound of mn. (Interestingly, we need to choose an appropriate characteristic of the field (2 or 3) based on which function we deal with.) For (x, y, z) 7→ (x ∧ y) ⊕ z, we further improve on this argument by employing random restrictions. These results together improve the previous best lower bound (due to Alon and Feige [1]) irrespective of the query function used. 2 General non-adaptive upper bound In this section, we prove the general non-adaptive upper bound result: Theorem 2. Definition 5. A non-adaptive (m, s, t)-graph is a bipartite graph G with vertex sets U = [m] and V (|V | = ts). V is partitioned into t disjoint sets: V1 , . . . , Vt ; each Vi has s vertices. Every u ∈ U has a unique neighbour in each Vi . A non-adaptive (m, s, t)-graph naturally gives rise to a non-adaptive (m, ts, t)-query scheme TG as follows. We view the memory (an array L of ts bits) to be indexed by vertices in V . On receiving the query “Is u in S?”, we answer “Yes” iff the MAJORITY of the locations in the neighbourhood of u contain a 1. We say that the query scheme TG is satisfiable for a set S ⊆ [m], if there is an assignment to the memory locations (L[v] : v ∈ V ), such that TG correctly answers all queries of the form “Is x in S?”. We now restrict attention to odd t ≥ 5. First, we identify an appropriate property of the underlying non-adaptive (m, s, t)-graph G that guarantees that TG is satisfiable for all sets S of 2 2 size at most n. We then show that such a graph exists for some s = O(m t−1 n1− t−1 lg 2m n ). Definition 6 (Non-adaptive admissible graph). We say that a non-adaptive (m, s, t)-graph G is admissible for sets of size at most n if the following two properties hold:   t+1 (P1) ∀R ⊆ [m] (|R| ≤ n + 2n lg 2m n ): |ΓG (R)| ≥ 2 |R|, where ΓG (R) is the set of neighbors of R in G. 4   (P2) ∀S ⊆ [m] (|S| = n): |TS | ≤ 2n lg 2m n , where TS = {y ∈ [m] \ S : |ΓG (y) ∩ ΓG (S)| ≥ t+1 2 }. Our theorem will follow from the following claims. Lemma 7. If a non-adaptive (m, s, t)-graph G is admissible for sets of size at most n, then the non-adaptive (m, ts, t)-query scheme TG is satisfiable for every set S of size at most n. 2 2 Lemma 8. There is a non-adaptive (m, s, t)-graph, with s = O(m t−1 n1− t−1 lg 2m n ), that is admissible for every set S ⊆ [m] of size at most n. Proof of Lemma 7. Fix an admissible graph G. Thus, G satisfies (P1) and (P2) above. Fix a set S ⊆ [m] of size at most n. We will show that there is a 0-1 assignment to the memory such that all queries are answered correctly by TG .   ′ and |S ′ | = n. From (P2), we know |T ′ | ≤ 2n lg 2m . Let S ′ ⊆ [m] be such that S ⊆ S S n   Hence, |S ′ ∪TS ′ | ≤ n+ 2n lg 2m n . From (P1) and Hall’s theorem, we may assign to each element u ∈ S ′ ∪ TS ′ a set Au ⊆ V such that (i) |Au | = t+1 2 and (ii) the Au ’s are disjoint. For each u ∈ S ⊆ S ′ , we assign the value 1 to all locations in Au . For each u ∈ (S ′ ∪ TS ′ )\S, we assign the t ′ ′ value 0 to all locations in Au . Since t+1 2 > 2 , all queries for u ∈ S ∪ TS are answered correctly. ′ ′ Assign 0 to all locations in ΓG ([m]\(S ∪TS ′ )). For y ∈ [m]\(S ∪TS ′ ), |ΓG (y)∩ΓG (S)| ≤ t−1 2 . ′ As a result, queries for elements in [m]\(S ∪TS ′ ) are answered correctly, as the majority evaluates to 0 for each one of them. Proof of Lemma 8. In the following, set   2 2 2m 1− s = 60m t−1 n t−1 lg . n We show that a suitable random non-adaptive (m, s, t)-graph G is admissible for sets of size S at most n with positive probability. The graph G is constructed as follows. Recall that V = i Vi . For each u ∈ U , one neighbor is chosen uniformly and independently in each Vi .   (P1) holds. If (P1) fails, then for some non-empty W ⊆ U , (|W | ≤ n + 2n lg 2m n ), we have |ΓG (W )| ≤ t+1 Fix a set W of size r ≥ 1 and L ⊆ V of size t+1 2 |W | − 1. P 2 r − 1. Let L have ℓi elements in Vi ; thus, i ℓi = t+1 r − 1. Then, 2  t  Y ℓi r ≤ Pr[ΓG (W ) ⊆ L] ≤ |Vi | ( t+1 2 )r − 1 ts i=1 !tr , where the last inequality is a consequence of GM ≤ AM. We conclude, using the union bound over choices of W and L, that (P1) fails with probability at most n+⌈2n lg 2m n X r=1 ≤ ≤ n+⌈2n lg X ⌉    m ts t+1 r 2 r−1 2m n r=1 n+⌈2n lg X r=1 2m n ⌉ ⌉" em r r (e t+1 2 r tes t+1 2 r−1 t+3 − 1r 2 )mr 1 r (s )s −1 ts ! t+1 r−1 t−1 −1+ 1r 2 t−1 2 !tr 2 #r (1) t+1 2 r −1 ts !tr 1 ≤ , 3 where the last inequality holds because we have chosen s large enough. 5 (2)   (P2) holds. For (P2) to fail, there must exist a set S ⊆ [m] of size n such that |TS | > 2n lg 2m n . Fix a set S of size n. Fix a y ∈ [m] \ S.     t+1 t n 2 n Pr[y ∈ TS ] ≤ t+1 ≤ , s 10m 2 n where the last inequality holds because of choice of s and m is large. Thus, E[|TS |] ≤ 10 . To conclude that |TS | is bounded with high probability, we will use the following version of P Chernoff bound: if X = N i=1 Xi , where each random variable Xi ∈ {0, 1} independently, then if γ > 2eE[X], then Pr[X > γ] ≤ 2−γ . Then, for all large m, Pr[|TS | > 2n lg 2m 2m ] ≤ 2−2n lg n . n Using the union bound, we conclude that Pr[(P2) fails] ≤ Thus, with probability at least 3 1 3  em n n 2−2n lg 2m n 1 ≤ . 3 the random graph G is admissible. General upper bound: adaptive In this section, we prove the general adaptive upper bound result: Theorem 3. We will use the following definitions from previous works. Definition 9 (storing scheme, query scheme, scheme, systematic). An (m, n, s)-storing scheme is a method for representing a subset of size at most n of auniverse of size m as an s-bit string. s Formally, an (m, n, s)-storing scheme is a map φ from [m] ≤n to {0, 1} . A deterministic (m, s, t)-query scheme is a family {Tu }u∈[m] of m Boolean decision trees of depth at most t. Each internal node in a decision tree is marked with an index between 1 and s, indicating the address of a bit in an s-bit data structure. For each internal node, there is one outgoing edge labeled “0” and one labeled “1”. The leaf nodes of every tree are marked ‘Yes’ or ‘No’. Such a tree Tu induces a map from {0, 1}s to {Yes, No}; this map will also be referred to as Tu . An (m, n, s)-storing scheme φ and an (m, s, t)-query scheme {Tu }u∈[m] together form an  (m, n, s, t)-scheme if ∀S ∈ [m] ≤n , ∀u ∈ [m] : Tu (φ(S)) =Yes if and only if u ∈ S. Let s(m, n, t) be the minimum s such that there is an (m, n, s, t)-scheme. We say that an (m, n, s, t)-scheme is systematic if the value returned by each of its trees Tu is equal to the last bit it reads (interpreting 0 as No/False and 1 as Yes/True). In order to show that s(m, n, t) is small, we will exhibit efficient adaptive schemes to store sets of size exactly n. This will imply our bound (where we allow sets of size at most n) because we may pad the universe with n additional elements, and extend S (|S| ≤ n)by adding n − |S| additional elements, to get a subset is of size exactly n in a universe of size m + n ≤ 2m. Definition 10. An adaptive (m, s, t)-graph is a bipartite graph G with vertex sets U = [m] and V (|V | = (2t − 1)s). V is partitioned into 2t − 1 disjoint sets: A, A0 , A1 , A00 ,. . . , that is, one Aσ for each σ ∈ {0, 1}≤(t−1) ; each Aσ has s vertices. Between each u ∈ U and each Aσ there is exactly one edge. Let Vi := ∪σ:|σ|=i−1 Aσ . An (m, s, t)-graph naturally gives rise to a systematic (m, (2t − 1)s, t)-query scheme TG as follows. We view the memory (an array L of (2t − 1)s bits) to be indexed by vertices in V . For query element u ∈ U , if the first i − 1 probes resulted in values σ ∈ {0, 1}i−1 , then the i-th probe is made to the location indexed by the unique neighbor 6 of u in Aσ . In particular, the i-th probe is made at a location in Vi . We answer “Yes” iff the last bit read is 1. In addition, we use following notation. We refer to Vt as the leaves of G and for y ∈ [m], let leaves(y) := Vt ∩ ΓG (y). For R ⊆ [m], let leaves(R) := Vt ∩ ΓG (R). We say that the query scheme TG is satisfiable for a set S ⊆ [m], if there is an assignment to the memory locations (L[v] : v ∈ V ), such that TG correctly answers all queries of the form “Is x in S?”. We assume that t ≥ 3 is odd and show that ∀ǫ > 0 ∀n ≤ m1−ǫ ∀t ≤ 2 2 1 10 lg lg m s(m, n, t) = O(exp(e2t )m t+1 n1− t+1 lg m). Our t-probe scheme will have two parts: a t1 -probe non-adaptive part and a t2 -probe adaptive part, such that t1 + t2 = t. The respective parts will be based on appropriate non-adaptive (m, s, t1 )-graph G1 and adaptive (m, s, t2 )-graph G2 respectively. To decide set membership, we check set membership in the two parts separately and take the AND, that is, we answer “Yes” iff all bits read in TG1 are 1 and the last bit read in TG2 is 1. We refer to this scheme as TG1 ∧ TG2 . First, we identify appropriate properties of the underlying graphs G1 and G2 that guarantee that all queries are answered correctly for sets of size n. We then show that such graphs exist 2 2 with s = O(exp(e2t − t)m t+1 n1− t+1 lg m). We will use the following constants in our calculations: α := 2t2 − 1 and β := 2t2 − t2 . Note that α is the total number of nodes in a t2 -probe adaptive decision tree. In any such decision tree, for every choice of β nodes and every choice b ∈ {0, 1} of the answer, it is possible to assign values to those β nodes so that the decision tree returns the answer b. Definition 11 (admissible-pair). We say that a non-adaptive (m, s, t1 )-graph G1 and an adaptive (m, s, t2 )-graph G2 form an admissible pair (G1 , G2 ) for sets of size n if the following conditions hold. t / S : ΓG1 (y) ⊆ (P1) ∀S ⊆ [m] (|S| = n): |survivors(S)| ≤ 10m ns 1 , where survivors(S) = {y ∈ ΓG1 (S)}. (P2) For S ⊆ [m] (|S| = n), let survivors+ (S) = {y ∈ survivors(S) : leavesG2 (S) ∩ leavesG2 (y) 6= ∅}. Then, ∀S ⊆ [m] (|S| = n) ∀T ⊆ S ∪ survivors+ (S): ΓG2 (T ) ≥ β|T |. Lemma 12. If a non-adaptive (m, s, t1 )-graph G1 and an adaptive (m, s, t2 )-graph G2 form an admissible pair for sets of size n, then the query scheme TG1 ∧ TG2 is satisfiable for every set S ⊆ [m] of size n. t+3 Lemma 13. Let t ≥ 3 be an odd number; let t1 = t−3 2 and t2 = 2 . Then, there exist an admissible pair of graphs consisting of a non-adaptive (m, s, t1 )-graph G1 and an adaptive 2 2 (m, s, t2 )-graph G2 with s = O(exp(e2t − t)m t+1 n1− t+1 lg m). Proof of Lemma 12. Fix an admissible pair (G1 , G2 ). Thus, G1 satisfies (P1) and G2 satisfies (P2) above. Fix a set S ⊆ [m] of size n. We will show that there is an assignment such that TG1 ∧ TG2 answers all questions of the form “Is x in S?” correctly. The assignment is constructed as follows. Assign 1 to all locations in ΓG1 (S) and 0 to the remaining locations in ΓG1 ([m]). Thus, TG1 answers “Yes” for all query elements in S and answers “No” for all query elements outside S ∪ survivors(S). However, it (incorrectly) answers “Yes” for elements in survivors(S). We will now argue that these false positives can be eliminated using the scheme TG2 . Using (P2) and Hall’s theorem, we may assign to each element u ∈ S ∪ survivors+ (S) a set Lu ⊆ V (G2 ) such that (i)|Lu | = β and (ii) the Lu ’s are disjoint. Set bu = 1 for u ∈ S and bu = 0 for u ∈ survivors+ (S) (some of the false positives). As observed above for each u ∈ S ∪ survivors+ (S) we may set the values in the locations in Lu such that the value returned on the query element u is precisely bu . Since the Lu ’s are disjoint we may take such an action independently for each u. After this partial assignment, it remains to ensure that queries for 7 elements y ∈ survivors(S) \ survivors+ (S) (the remaining false positives) return a “No”. Consider any such y. By the definition of survivors+ (S), no location in leavesG2 (y) has been assigned a value in the above partial assignment. Now, assign 0 to all unassigned locations in V (G2 ). Thus TG2 returns the answer “No” for queries from survivors(S) \ survivors+ (S). Proof of Lemma 13. In the following, let l m 2 2 s = exp(e2t − t)m t+1 n1− t+1 lg m . We will construct the non-adaptive (m, s, t1 )-graph G1 and the (m, s, t2 )-graph G2 randomly, and show that with positive probability the pair (G1 , G2 ) is admissible. The graphSG1 is constructed as in the proof of Lemma 8, and the analysis is similar. Recall that V (G1 ) = i∈[t1 ] Vi (G1 ). For each u ∈ U , one neighbor is chosen uniformly and independently from each Vi (G1 ). (P1) holds. Fix a set S of size n. Then, E[|survivors(S)|] ≤ (m − n) before, using the Chernoff bound, we conclude that  n t1 n t1 ] ≤ 2−10m( s ) . Pr[|survivors(S)| > 10m s  n t1 s ≤ m  n t1 . s As Then, by the union bound,   m −10m( n )t1 s Pr[P1 fails] ≤ 2 n 1 , ≤ 10 where the last inequality follows from our choice of s. Fix a graph G1 suchSthat (P1) holds. The random graph G2 is constructed as follows. Recall that V (G2 ) = z∈{0,1}≤t2 −1 Az . For each u ∈ [m], one neighbor is chosen uniformly and independently from each Az . (P2) holds To establish (P2), we need to show that all sets of the form S ′ ∪ R, where S ′ ⊆ S and R ⊆ survivors+ (S) expand. To restrict the choices for R, we first show in Claim 14 (a) that with high probability survivors+ (S) is small. Then, using direct calculations, we show that whp the required expansion is available in the random graph G2 . t +1 Claim 14. (a) Let Ea ≡ ∀S ⊆ [m](|S| = n) : |survivors+ (S)| ≤ 100 · 2t2 m ns 1 ; then, 9 . Pr[Ea ] ≥ 10 (b) Let Eb ≡ ∀R ⊆ [m] (|R| ≤ n + ⌈n lg m⌉) : |ΓG2 (R)| ≥ β|R|; then, Pr[Eb ] ≥ 9 10 . (c) Let Ec = ∀S ⊆ [m](|S| = n), ∀S ′ ⊆ S, ∀R ⊆ survivors+ (S)(⌈n lg m⌉ ≤ |R| ≤ 100 · t +1 9 2t2 m ns 1 ) : |ΓG2 (S ′ ∪ R)| ≥ β|S ′ ∪ R|; then, Pr[Ec ] ≥ 10 . Proof of claim 14. Part (a) follows by a routine application of Chernoff bound, as in several previous proofs. For a set S of size n, we have E[survivors+ (S)] ≤ |survivors(S)|2t2 ( ns ) t +1 ≤ 2t2 10m ns 1 . Then,   m −2t2 10m( n )t1 +1 s Pr[¬Ea ] ≤ 2 n 1 ≤ , 10 where the last inequality holds because of our choice of s. 8 Next consider part (b). If Eb does not hold, then for some non-empty W ⊆ [m], (|W | ≤ n + ⌈n lg m⌉), we have |ΓG2 (W )| ≤ β|W | − 1. Fix a set W of size r ≥ 1 and L ⊆ V (G2 ) of size βr − 1. Let L have ℓz elements in Az . Then, Y  ℓz r  βr − 1 αr ≤ . Pr[ΓG2 (W ) ⊆ L] ≤ |Az | αs z We conclude, using the union bound over choices of W and L, that the probability that Eb does not hold is at most n+⌈n lg m⌉     αs βr − 1 αr βr − 1 αs X m r X em r r ≤ n+⌈n lg m⌉  X βr αes " em β e r ≤ n+⌈n lg m⌉  X βr αes "  α−β  α−β−1 #r β mr eβ+1 α sα−β r=1 ≤ ≤ n+⌈n lg m⌉  r=1 r=1 r=1 1 , 10  αes βr − 1 βr−1   βr αs βr − 1 αs #r αr α−β where the last inequality holds because of our choice of s. Finally, we justify part (c). To bound the probability that Ec fails, we consider a set S ⊆ [m] of size n, a subset S ′ ⊆ S of size i (say), a subset R ⊆ survivors+ (S) of size r t +1 (where ⌈n lg m⌉ ≤ r ≤ 100 · 2t2 m ns 1 ) and L ⊆ V (G2 ) of size ℓ = β(i + r) and define the event E(S, S ′ , R, L) ≡ (∀y ∈ R : leavesG2 (S) ∩ leavesG2 (y) 6= ∅) ∧ ΓG2 (S ′ ∪ R) ⊆ L. Then, r  (α−1)r   ℓ αi αs  t2 r  (α−1)(i+r)   2 n β(i + r) β(i + r) i ≤ , s (α − 1)s αs Pr[E(S, S ′ , R, L)] ≤  2t2 n s  2t2 n s ℓ (α − 1)s (3) (4) r is justified because of the requirement that every y ∈ R has  (α−1)r ℓ at least one neighbour in leavesG2 (S); the factor (α−1)s is justified because all  ℓ αi the remaining neighbours must lie in L (we use AM ≥ GM); the last factor αs is justified because all neighbors of elements in S ′ lie in L (again we use AM ≥ GM). To complete the argument we apply the union bound over the choices of (S, S ′ , R, L). Note that we may restrict attention to ℓ = β(i + r) (because for our choice of s, we have β(i + r) ≤ |V (G2 )| = αs). Thus, the probability that Ec fails to hold is at most X Pr[Ec (S, S ′ , R, L)], where the factor S,S ′ ,R,L where S ranges over sets of size n, S ′ ⊆ S of size i, R ⊆ survivors(S) of size r such that t +1 ⌊n lg m⌋ ≤ r ≤ 1002t2 m ns 1 , L is a subset of V (G2 ) of size β(i + r). We evaluate this 9 sum as follows. j  k X X m 10m n t1 n αs   2t2 n r  β(i + r) (α−1)(i+r)  β(i + r) i s s (α − 1)s αs r i β(i + r) n r i (5) ≤  n X X  em  i+r  n r i t ! r   1   10em ns 1 i+r n 1+r β(i + r) α−1 r (α − 1)s i r  β  t2  i+r  i #i+r eαs 2 n β(i + r) i+r β(i + r) s (α − 1)s  t ! r   1   n X X  em  i+r 10em ns 1 i+r n 1+r β(i + r) α−β−1  ≤ n r (α − 1)s i r  (6) i  eα α−1 β  β(i + r) t 2 2 n(α − 1)  i i+r  2t2 n s #i+r (7) . We will show that the quantity inside the square brackets is at most r ≥ n lg m and i ≥ 0 ! ! X X 1 −i −r ≤ . 2 2 Pr[¬Ec ] ≤ 10 r 1 2. Then, since i The quantity in the brackets can be decomposed as a product of two factors, which we will bound separately. Factor 1: Consider the following contributions    i   1   em  n i+r r eα β β(i + r) n i+r i+r i+r . (10e) n α−1 2t2 n(α − 1) i Since r ≥ n lg m and i ≤ n, we have enough m, this quantity is at most i i+r ≤ n n+r ≤ 1 lg m ≤ 1 lge m . Thus, for all large e2 · 10e · e2 · (2e)β · e ≤ exp(e2t − t). Factor 2: We next bound the contribution for the remaining factors. r   n t1  i+r    m( s ) β(i + r) α−β−1 2t2 n r (α − 1)s s  n t1   α−β−1  t2  m( s ) 2r 2 n ≤ r s s (8) (9) mnt1 +1 2α−β+t2 −1 r α−β−2 . (10) sα−β+t1 t +1 To justify (9), recall that r ≤ 100 · 2t2 m ns 1 and l m 2 2 m( ns )t1 ≥ 1. Then, the above quantity is s = exp(e2t − t)m t+1 n1− t+1 lg m ; thus r bounded by α−β−2 mnt1 +1 22(t2 −1) 100 · 2t2 mnt1 +1 (11) s(t1 +1)(α−β−2) sα−β+t1  α−β−1 100 · 22t2 mnt1 +1 ≤ . (12) st1 +2 = 10 m l 2 2 Thus, since s = exp(e2t − t)m t+1 n1− t+1 lg m , then the product of the factors is at most 1 10 , 4 as required. Three non-adaptive probes lower bound In this section, we prove the three probe lower bound result: Theorem 4. Definition 15 (Equivalent). Two boolean functions are called equivalent if one can be obtained from the other by negating and permuting the variables. Proposition 16. Let f, g : {0, 1}t → {0, 1} be equivalent. If s1 and s2 are the minimum bits of space required for non-adaptive (m, n, s1 , t) and (m, n, s2 , t)-schemes with query functions f and g respectively, then s1 = s2 . For three variable boolean functions, there are twenty-two equivalence classes (see [15]). To prove Theorem 4, we provide proofs for these twenty-two query functions, each from a different class. In many proofs below we assume that the memory consists of three arrays of size s each, and the three probes are made on different arrays. Given any scheme that uses space s, we can always modify it to meet our assumption, by expanding the space by factor 3. 4.1 Decision trees of height two Seven of the twenty-two classes contain functions that can be represented by a decision tree of height at most two. Thus, for these functions, the two probe adaptive lower bound [5] implies the result. These functions are: constant 0, constant 1, the DICTATOR function (x, y, z) 7→ x, the function (x, y, z) 7→ x ∧ y, its complement (x, y, z) 7→ x̄ ∨ ȳ, (x, y, z) 7→ (x ∧ y) ∨ (x̄ ∧ z), and (x, y, z) 7→ (x ∧ y) ∨ (x̄ ∧ ȳ). 4.2 MAJORITY Let Φ be a non-adaptive (m, n, s, 3)-scheme with MAJORITY as the query function. The memory is a bit array A[1, · · · , s] of length s. For each element u ∈ [m], x(u), y(u), z(u) ∈ [s] are the three distinct locations in A that are probed to determine whether u is in the set or not. For each set S ⊆ [m] of size at most n, the assignment σ(S) ∈ {0, 1}s to A is such that for all elements u ∈ [m], Maj(A[x(u)], A[y(u)], A[z(u)]) is 1 iff u ∈ S, where Maj is the MAJORITY of 3 bits. Definition 17. (model-graph for Φ, third vertex, meet) Let Φ be a (m, n, s, 3)-scheme with MAJORITY as the query function. Fix a graph G such that V (G) = [s], |E(G)| = m and edge labels: {lab(e)|e ∈ E(G)} = [m] (there is a unique edge for each label in [m]). G is called a model-graph for Φ if for each u ∈ [m] the edge labelled u has the set of endpoints in u {{x(u), y(u)}, {y(u), z(u)}, {z(u), x(u)}}. For example, the graph G = ([s], {x(u) ←→ y(u)|u ∈ [m]}) is a model graph for Φ. In a model-graph for Φ, let e be the set of endpoints of the edge with label u. The element in the singleton ({x(u), y(u), z(u)} \ e) is defined to be the third vertex of u. Two edge-disjoint cycles C1 and C2 are said to meet in a model-graph for Φ if there exist elements u, v ∈ [m] such that the third vertices of u and v are the same vertex and the edges labelled u and v are in the different cycles C1 and C2 respectively. Definition 18. A model-graph G for an (m, n, s, 3)-scheme with MAJORITY as the query function is said to be forced if at least one of the following three conditions hold. 11 (P1) ∃ edge-disjoint odd cycles C1 , C2 in G with lengths at most n each that intersect at a vertex. (P2) ∃ edge disjoint even cycles C1 , C2 in G with lengths at most n each and C1 and C2 meet. (P3) ∃ an even cycle C of length at most n, such that some two edges in C, labelled e and f say, have an even number of edges between them (while traversing the edges of the cycle in order) and the third vertices of e and f are the same vertex. Lemma 19. A model-graph for a scheme with MAJORITY as the query function cannot be forced.   1 1 1− ⌊ n +1 ⌋ 2 , 3)-scheme with MAJORITY as the query function has Lemma 20. Any (m, n, 6 m a forced model-graph. From lemmas 19 and 20, it follows that when MAJORITY is used as the query function, sN (m, n, 3) > 16 m 1− 1 ⌊n 2 ⌋+1 . Proof of Lemma 19. Fix a (m, n, s, 3)-scheme Φ with MAJORITY as the query function. Fix a model-graph G for Φ. Assume G is forced, that is, it satisfies (P1) or (P2) or (P3) above. Case: (P1) holds. (P1) implies that there are edge-disjoint cycles C1 and C2 in G such that, e e e2k+1 f2 f2l+1 1 2 C1 : u0 −→ u1 −→ · · · −→ u2k+1 = u0 , f1 C2 : u0 −→ v1 −→ · · · −→ v2l+1 = u0 , and 2k+1, 2l+1 ≤ n. Let S0 = {e1 , e3 , · · · , e2k+1 }∪{f2 , f4 , · · · , f2l } and S1 = {e2 , e4 , · · · , e2k }∪ {f1 , f3 , · · · , f2l+1 }. Note that |S0 | = |S1 | ≤ n. We claim that Φ cannot represent any set S such that S0 ⊆ S ⊆ S¯1 . In particular, Φ cannot represent the set S0 . Assume Φ represents such an S. We claim that u0 cannot be assigned a 0. If u0 is assigned a 0, then since e1 ∈ S, u1 must be assigned a 1. Otherwise, Maj(A[x(e1 )], A[y(e1 )], A[z(e1 )]) = Maj(0, 0, b) = 0, where b is the bit assigned to the location in {x(e1 ), y(e1 ), z(e1 )} \ {u0 , u1 }. Since, u1 is assigned a 1 and e2 ∈ / S, u2 must be assigned a 0. Similarly, since e3 ∈ S, u3 must be assigned a 1 and so on. Finally, u2k+1 = u0 must be assigned a 1. A contradiction. Hence u0 cannot be assigned a 0. Again, we claim u0 cannot be assigned a 1. For if u0 is assigned a 1, since f1 ∈ / S, v1 must be assigned a 0. Again, since f2 ∈ S, v2 must be assigned a 1 and so on. Finally, v2k+1 = u0 must be assigned a 0. A contradiction. Since u0 can neither be assigned a 0 or a 1, we get a contradiction. Remark. In the proofs below, we will often encounter similar arguments, where we will have a cycle of dependencies: assigning a particular bit to a location will force the assignment to the next location along the cycle. Case: (P2) holds. (P2) implies that there are edge-disjoint cycles C1 and C2 in G such that, e e e f1 f2 f 1 2 2k C1 : u0 −→ u1 −→ · · · −→ u2k = u0 , 2l C2 : v0 −→ v1 −→ · · · −→ v2l = v0 , 12 2k, 2l ≤ n, and the third vertices of e1 and f1 are the same vertex w. Let S0 = {e1 , e3 , · · · , e2k−1 } ∪ {f2 , f4 , · · · , f2l } and S1 = {e2 , e4 , · · · , e2k } ∪ {f1 , f3 , · · · , f2l−1 }. Note that |S0 | = |S1 | ≤ n. We claim that Φ cannot represent any set S such that S0 ⊆ S ⊆ S¯1 . In particular, Φ cannot represent the set S0 . Assume Φ represents such an S. Since e1 ∈ S, either the location u0 or the location u1 of the memory A must be assigned a 1, otherwise Maj(A[x(e1 )], A[y(e1 )], A[z(e1 )]) = Maj(A[u0 ], A[u1 ], A[w]) = Maj(0, 0, A[w]) = 0. Assume u1 is assigned a 1. Then, since e2 is not in the set, using a similar argument, u2 must be assigned a 0. Similarly, u3 must be assigned a 1 and so on. Finally, u2k = u0 must be assigned a 0. Similarly, if u0 was assigned a 1, then u1 must be assigned a 0. Thus, Maj(x(e1 ), y(e1 ), z(e1 )) = Maj(0, 1, A[w]) = A[w]. Hence w must be assigned a 1. Again, since f1 is not in S, either v0 or v1 is assigned a 0. If v1 is assigned 0, v2 must be assigned a 1, v3 a 0, and so on. Finally, v2l = v0 must be assigned a 1. Similarly, if v0 is assigned 1, then v1 is assigned a 0. Therefore, Maj(x(f1 ), y(f1 ), z(f1 )) = Maj(A[v0 ], A[v1 ], A[w]) = Maj(0, 1, A[w]) = A[w]. Since f1 ∈ / S, w must be assigned a 0. A contradiction. Case: (P3) holds. (P3) implies that there is a cycle C: e e e 1 2k 2l v0 −→ v1 · · · −→ v2k · · · −→ v2l = v0 , 2k ≤ 2l ≤ n and the third vertices of e1 and e2k are the same vertex w. Let S0 = {e1 , e3 , · · · , e2l−1 } and S1 = {e2 , e4 , · · · , e2k , · · · , e2l }. Note that |S0 | = |S1 | ≤ n. We claim that Φ cannot represent any set S such that S0 ⊆ S ⊆ S¯1 . In particular, Φ cannot represent the set S0 . Assume Φ represents such an S. Since e1 ∈ S, either v0 or v1 must be assigned a 1. Assume that v1 is assigned a 1. Then, since e2 ∈ / S, v2 must be assigned a 0. Again, since e3 ∈ S, v3 must be assigned a 1 and so on. All locations in R := {v2r |0 ≤ r ≤ l} must be assigned a 0 and all locations in Q := {v2r+1 |0 ≤ r ≤ l − 1} must be assigned a 1. Else if v0 is assigned a 1, then all locations in R must be assigned a 1 and all locations in Q must be assigned a 0. Now, Maj(x(e1 ), y(e1 ), z(e1 )) = Maj(A[v0 ], A[v1 ], A[w]) = Maj(0, 1, A[w]) = A[w]. Since e1 ∈ S, w must be assigned a 0. Similarly, Maj(x(e2k ), y(e2k ), z(e2k )) = Maj(A[v2k−1 ], A[v2k ], A[w]) = Maj(0, 1, A[w]) = A[w]. Since e2k ∈ / S, w must be assigned a 1. A contradiction. In order to prove Lemma 20 we will make use of the following proposition, which is a consequence of a theorem of Alon, Hoory and Linial [2] (see also Ajesh Babu and Radhakrishnan [3]). Proposition 21. Fix a graph G such that the average degree d ≥ 2. Then, (d − 1)k > |V (G)| =⇒ ∃ a cycle C ⊆ E(G), |C| ≤ 2k.   1 1 1− ⌊ n +1 ⌋ 2 Proof of Lemma 20. Fix an (m, n, 6 m , 3)-scheme Φ that uses MAJORITY as the   1− n 1 query function. Note that s := 16 m ⌊ 2 ⌋+1 implies m≥s 1+ 1 ⌊n 2⌋ 13 + 4s + 1. (⋆) For Φ we will come up with a model-graph which is forced, that is, one of (P1), (P2) or (P3) holds. We will start with an initial model-graph G for Φ. We will observe that the average degree of G is high and invoke Proposition 21 to find a small cycle C. If |C| is odd, we will bin it in ODD, delete C and repeat. If |C| is even and all the third vertices of the labels of edges in C are distinct, we will bin C in EVEN, delete the edges of C and repeat; otherwise, we will either discover that property (P3) holds or we will modify our model-graph and find an odd cycle in it and bin it in ODD, delete it and repeat. The moment the sum of the lengths of the deleted cycles exceeds 2s, we know either the sum of the lengths of odd or even cycles exceeds s and two odd cycles intersect or even cycles with distinct third vertices meet, which means either (P1) or (P2) holds. Formally, the procedure can be described as below. We will maintain the following invariant. EVEN will contain edge-disjoint cycles of length even and at most n each and the third vertices of the labels in such a cycle will be all distinct. ODD will contain edge-disjoint cycles of length odd and at most n. Furthermore ([s], E(G) ∪ EVEN ∪ ODD) will always be a model-graph for Φ. u Step 0: Initialization. EVEN = ∅. ODD = ∅. G = ([s], {x(u) ←→ y(u)|u ∈ [m]}). Observe G is a model-graph for Φ. P Step 1. If C∈EVEN∪ODD |C| > 2s , END (this ensures that either (P1) or (P2) holds). Else, using Proposition 21 fix a cycle C ⊆ E(G) such that |C| ≤ n. Step 2. If |C| is odd, ODD ← ODD ∪ {C} and E(G) ← E(G) \ C and GOTO Step 1. Step 3. If |C| is even and all the third vertices of the labels of edges in C are distinct, EVEN ← EVEN ∪ {C} and E(G) ← E(G) \ C and GOTO Step 1. Step 4. If |C| is even and the third vertices of the labels of two edges in C which have an even number of edges between them while traversing the edges of C in order, then END (Note this means that (P3) holds). Step 5. If |C| is even and the third vertices of the labels of two edges in C have an odd number of edges between them (while traversing the edges of C in order), then represent C as e2k+1 e e 1 2l C : v0 −→ v2l = v0 , v1 · · · v2k −→ v2k+1 · · · −→ such that the third vertices of e1 and e2k+1 are the same vertex w. We modify the modelgraph G by changing the endpoints of the edges appearing with labels e1 , e2k+1 from {v0 , v1 }, {v2k , v2k+1 } to {v1 , w}, {v2k , w} respectively, thus obtaining a shorter odd cycle C ′ in G: e2k+1 e e e2k+1 1 1 E(G) ← (E(G) \ {v0 ←→ v1 , v2k ←→ v2k+1 }) ∪ {v1 ←→ w, v2k ←→ w} (Observe: G with E(G) ∪ {e|e ∈ ODD ∪ EVEN} continues to be a model-graph for Φ). e e e2k+1 1 2 C ′ ⊆ E(G) : w −→ v1 −→ v2 · · · v2k −→ w is an odd length cycle of length at most n in G. ODD ← ODD ∪ {C ′ }. E(G) ← E(G) \ C ′ . GOTO Step 1. 1 n > s ⌊ 2 ⌋ + 2 (from ⋆) In Step 1, if |E(G)| ≤ 2s, then the average degree d is at least m−2s s n and (d − 1)⌊ 2 ⌋ > s which implies from Proposition 21 that there is a cycle of length at most n. We claim that the procedure terminates only by encountering an END statement in Step 1 or in Step 4. Observe that once the procedure finds a cycle in Step 1, then exactly one of the four if conditions in Steps 2-5 holds. If the procedure does not encounter an END statement 14 in Step 4, then the procedure moves to Step 1 again as each of the Steps 2, 3 and 5 end in a ‘GOTO Step 1’ statement. If the procedure encounters the END statement in Step 4, then (P3) holds. If the procedure encounters the P P END statement in Step 1, then from the pigeonhole principle, either C∈ODD |C| > s or C∈EVEN |C| > s. In the first case, (P1) holds. In the second case, since each edge in a cycle in EVEN has a distinct third vertex, two cycles in EVEN meet. Finally, we observe that the procedure terminates. If the procedure does not terminate in Step 4, then the procedure repeatedly finds edge disjoint cycles and deletes them. If the number of edges in the deleted cycle exceeds 2s, then the procedure will terminate when it encounters the END statement in Step 1. 4.3 Density argument The query functions considered in this section all admit a density argument. For such a query function, a valid scheme that supports sets from a large universe using only small space, must conceal a certain dense graph that avoids certain forbidden configurations. On the other hand, standard graph theoretic results (e.g., the Moore bound) would imply that dense graphs must have at least one of those forbidden configurations, which would contradict the existence of any such scheme. In this section we provide lower bounds for the following ten query functions: AND function, (x, y, z) 7→ (x ⊕ y) ∧ z, (x, y, z) 7→ (x ∨ y) ∧ z, the ALL-EQUAL function, (x, y, z) 7→ (x ∧ y ∧ z) ∨ (ȳ ∧ z̄), and their complements. For these functions, we deal with two functions—a function and its complement—with a single proof. In these proofs, we produce sets S and T of size at most n such that storing S and not storing T leads to a contradiction. The proof for the complement function works with a small twist: storing T and not storing S leads to the contradiction. We now develop a common framework to prove lower bounds for the above mentioned query functions. We assume the query function is f , where f could be any of these ten functions. Fix a scheme Φf for the query function f , where the memory consists of three distinct bit arrays: A[1, · · · , s], B[1, · · · , s] and C[1, · · · , s]. For any element u ∈ [m], the scheme Φf probes three distinct locations x(u) ∈ A, y(u) ∈ B, z(u) ∈ C to determine if u is in the set or not. Given any set S ⊆ [m] of size at most n, the assignment σf (S) ∈ {0, 1}3s to the memory is such that for all elements u ∈ [m], f (A[x(u)], B[y(u)], C[z(u)]) is 1 iff u ∈ S. We will need the following definitions. Definition 22 (GA,B (E), GB,C (E), GA,C (E)). For a scheme Φf , and any subset of elements E ⊆ [m], we define the bipartite graph GA,B (E) as follows. The vertex sets are A = [s] and B = [s]. For each element u ∈ E, we have an edge labelled u with end points x(u) ∈ A and y(u) ∈ B. u Similarly, we define the bipartite graphs GB,C (E) = (B := [s], C := [s], {y(u) ←→ z(u) : u ∈ u E}) and GA,C (E) = (A := [s], C := [s], {x(u) ←→ z(u) : u ∈ E}). Definition 23 (private vertex, private z vertex). For an element u ∈ [m], we say u does not have a private vertex if each of its three probe locations is also the probe location of another element, that is, there exists elements v1 , v2 , v3 ∈ [m] such that u ∈ / {v1 , v2 , v3 }, x(u) = x(v1 ), y(u) = y(v2 ) and z(u) = z(v3 ). If it is not the case that u does not have a private vertex, we say that u has a private vertex. For an element u ∈ [m], we say u has a private z vertex if its probe location in C is not shared by any other element, that is for all v ∈ [m] \ {u}, z(v) 6= z(u). Proposition 24. If every u ∈ [m] has a private vertex, then 3s ≥ m. Proposition 25. If at least m 2 u’s in [m] have a private z vertex, then s ≥ 15 m 2. 4.3.1 AND Let the query function f be the AND function. From Proposition 24, if every u ∈ [m] has a private vertex, we are done. Otherwise, fix an element u ∈ [m] such that u does not have a private vertex. Let elements v1 , v2 , v3 ∈ [m] be such that x(u) = x(v1 ), y(u) = y(v2 ) and z(u) = z(v3 ). Let S = {v1 , v2 , v3 } and T = {u}. Clearly, |S|, |T | ≤ n. We now show that Φf cannot represent any set S ′ such that S ⊆ S ′ ⊆ [m] \ T. In particular Φf cannot represent the set S. To show this, consider the assignment for the set S ′ : σf (S ′ ) to the memory A, B, C. Since v1 , v2 , v3 ∈ S ′ , and f is the AND function, A[x(v1 )] = B[y(v2 )] = C[z(v3 )] = 1. This implies A[x(u)] = B[y(u)] = C[z(u)] = 1. Thus, f (A[x(u)], B[y(u)], C[z(u)]) = 1. But u ∈ / S ′ . A contradiction. The same argument works for f¯ when we try to represent any set S ′ such that T ⊆ S ′ ⊆ [m] \ S. 4.3.2 (x, y, z) 7→ (x ⊕ y) ∧ z Let the query function f be (x, y, z) 7→ (x ⊕ y) ∧ z. From Proposition 25, if at least m 2 elements have a z private vertex, we are done. Otherwise, fix a set E ⊆ [m] of size at least m that has no element with a private z vertex. 2 Now, we assume s ≤ 14 m GA,B (E) is at least 2m 1 ⌊n 2⌋ 1 1− ⌊ n ⌋ 2 and prove a contradiction. The average degree of vertices in ≥ 2, for large m. Since 1 n 1 n n n (2m ⌊ 2 ⌋ − 1)⌊ 2 ⌋ ≥ (m ⌊ 2 ⌋ )⌊ 2 ⌋ ≥ m > 2s, from Proposition 21, there exists a cycle of length 2k ≤ 2⌊ n2 ⌋ ≤ n in GA,B (E): u u u 1 2 2k v0 −→ v1 −→ · · · −→ v2k = v0 . Since u1 ∈ E does not have a private z vertex, fix an element v such that z(v) = z(u1 ). Let S = {v, u2 , u3 , · · · , u2k } and T = {u1 }. Clearly, |S|, |T | ≤ n. We now show that Φf cannot represent any set S ′ such that S ⊆ S ′ ⊆ [m] \ T. In particular Φf cannot represent the set S. Under the assignment σS ′ , since v ∈ S ′ , the location z(v) = z(u1 ) is assigned 1; otherwise if C[z(v)] = 0, then (A[x(v)] ⊕ B[y(v)]) ∧ C[z(v)] = 0. Again, under the assignment σS ′ , exactly one of v0 and v1 is assigned 0. To see this, let us assume v1 is assigned the bit b, then since u2 ∈ S ′ , v2 must be assigned the bit b̄; otherwise if both v1 and v2 are assigned b, then (A[x(u2 )] ⊕ B[y(u2 )]) ∧ C[z(u2 )] = (b ⊕ b) ∧ C[z(u2 )] = 0. Similarly, v3 must be assigned the bit b, v4 must be assigned the bit b̄ and so on. Thus, each location in {v1 , v3 , · · · , v2k−1 } must be assigned the bit b and each location in {v2 , v4 , · · · , v2k = v0 } must be assigned the bit b̄. Now, (A[x(u1 )] ⊕ B[y(u1 )]) ∧ C[z(u1 )] = (b ⊕ b̄) ∧ 1 = 1. But u1 ∈ / S′. A 1− 1 n contradiction. Thus, s > 14 m ⌊ 2 ⌋ . The same argument works for f¯ when we try to represent any set S ′ such that T ⊆ S ′ ⊆ [m] \ S. 16 4.3.3 (x, y, z) 7→ (x ∨ y) ∧ z Let the query function f be (x, y, z) 7→ (x ∨ y) ∧ z. From Proposition 25, if at least m 2 elements have a private z vertex, we are done. Otherwise, fix a set E ⊆ [m] of size at least m 2 that has no element with a private z vertex. Now make disjoint pairs of distinct elements (u, v), u, v ∈ E, that share the same z location, that is, z(u) = z(v). Make as many pairs as possible. At most s many, one per location in C, elements m can remain unpaired. Delete these unpaired elements from E. Thus, m′ := |E| ≥ 21 ( m 2 − s) ≥ 8 if s ≤ m 4 (which if not true, we are immediately done). For each such pair (u, v), define v to be the reserved partner of u. Let E ′ ⊆ E be the set of all unreserved elements. Clearly, m′ m E ′ = |E| 2 = 2 ≥ 16 . Find elements u, v, w ∈ E ′ such that u ∈ / {v, w}, x(u) = x(v) and y(u) = y(w). If no such triple of elements exists then each element has at least one of its x or y probe locations not m and we are done. shared with any other element in E ′ , and thus 2s ≥ |E ′ | ≥ 16 Let S = {u} ∪ {reserved partner of t : t ∈ {v, w}}, T = {v, w}. Clearly, |S|, |T | ≤ n. We now show that Φf cannot represent any set S ′ such that S ⊆ S ′ ⊆ [m] \ T. In particular Φf cannot represent the set S. Any assignment σS ′ must assign 1 to each location z(t), for each t ∈ {v, w}; since, the reserved partner of each such t is in S ′ and t shares the same z vertex with it. The assignment σS ′ must assign 0 to each of x(v), y(v), x(w), y(w); since, if any of them is assigned a 1, together with the fact that z(v), z(w) are assigned 1, either (A[x(v)] ∨ B[y(v)]) ∧ C[z(v)] or (A[x(w)] ∨ B[y(w)]) ∧ C[z(w)] will evaluate to 1 which cannot happen as v, w ∈ / S′. Now, since x(u) = x(v) and y(u) = y(w), (A[x(u)]∨B[y(u)])∧C[z(u)] = (0∨0)∧C[z(u)] = 0. But u ∈ S ′ . A contradiction. The same argument works for f¯ when we try to represent any set S ′ such that T ⊆ S ′ ⊆ [m] \ S. 4.3.4 ALL-EQUAL Let the query function f be the ALL-EQUAL function. 1 We assume s < 1− ⌊ n ⌋+1 1 4 600 m and derive a contradiction. Lemma 26. If GA,B and GB,C contain cycles C1 and C2 respectively of size at most n2 each such that for some e1 ∈ C1 and e2 ∈ C2 , labels of e1 and e2 are the same, then the scheme Φf cannot represent all sets of size n. 1− n 1 1 Lemma 27. If s < 600 m ⌊ 4 ⌋+1 , then GA,B and GB,C will contain cycles C1 and C2 respectively each of length at most n2 such that for some e1 ∈ C1 and e2 ∈ C2 , labels of e1 and e2 are the same. Our claim follows immediately from the above two lemmas. Proof of Lemma 26. Let the two cycles C1 and C2 as promised by Lemma 26 be u u u u v v 1 2 2k C1 : r0 −→ r1 −→ · · · −→ r2k = r0 , 1 2 2l C2 : r0 −→ t1 −→ · · · −→ t2l = r0 , 17 where r0 = y(u1 ) ∈ B, r1 = x(u1 ) ∈ A, t1 = z(u1 ) ∈ C and 2k, 2l ≤ n2 . Let S = {u2 , u3 , · · · , u2k } ∪ {v2 , v3 , · · · , v2l } and T = {u1 }. Clearly, |S|, |T | ≤ n. We now show that Φf cannot represent any set S ′ such that S ⊆ S ′ ⊆ [m] \ T. In particular Φf cannot represent the set S. Let the assignment σS ′ assign the bit b to the location r0 . Then, all locations in {r1 , r2 , · · · , r2k } ∪ {t1 , t2 , · · · , t2l } must be assigned the bit b: since, r2k = r0 is assigned the bit b and since u2k ∈ S ′ , r2k−1 must be assigned the bit b, similarly since r2k−1 is assigned the bit b and since u2k−1 ∈ S ′ , r2k−2 must be assigned the bit b and so on and thus all locations in {r1 , r2 , · · · , r2k } must be assigned the bit b. Arguing similarly, since r0 = t2l is assigned the bit b and since v2l ∈ S ′ , t2l−1 must be assigned the bit b, and so on and thus all locations in {t1 , t2 , · · · , t2l } must be assigned the bit b. Now, since r0 , r1 , t1 are all assigned the bit b and they are the three probe locations for the element u1 , the ALL-EQUAL function for u1 evaluates to 1. But u1 ∈ / S ′ . A contradiction. n Proof of Lemma 27. In the graph GA,B ([m]), as long as the number of edges is at least 6s1+⌊ 4 ⌋ we can find a cycle of length at most n2 from Proposition 21; since, the average degree then would be at least n 1 6s1+⌊ 4 ⌋ n ≥ 3s ⌊ 4 ⌋ ≥ 2, 2s and since 1 n n (3s ⌊ 4 ⌋ − 1)⌊ 4 ⌋ > 2s. We repeatedly remove cycles: C1 , C2 , · · · from the graph GA,B each of length at most n2 using Proposition 21 (that is, we delete the edges appearing in the picked cycle), till no more cycle n 1 remains; then, the number of remaining edges in the graph will be at most 6s1+⌊ 4 ⌋ ≤ 100 m; since, 1 1+ n 1 1− ⌊ n 1⌋+1 4 m =⇒ m > 600s ⌊ 4 ⌋ . s< 600 Similarly, following the same argument, we can remove cycles: D1 , D2 , · · · from the graph GB,C ([m]), where the length of each cycle is at most n2 , till no more cycle remains. The number 1 m. of remaining edges in the graph will be at most 100 Now, pick a random element u ∈ [m]. The probability that the edge {x(u), y(u)} with label u appears in some cycle Ci and the edge {y(u), z(u)} with label u appears in some cycle Dj is 1 1 98 at least 1 − ( 100 + 100 ) = 100 (using the union bound). Thus, there exists an element u and cycles Ci and Dj each of length at most n2 in GA,B and GB,C respectively, each containing an edge labelled u. The same argument works for f¯ when in Lemma 26, we try to represent any set S ′ such that T ⊆ S ′ ⊆ [m] \ S. 4.3.5 (x, y, z) 7→ (x ∧ y ∧ z) ∨ (ȳ ∧ z̄) Let the query function f be (x, y, z) 7→ (x ∧ y ∧ z) ∨ (ȳ ∧ z̄). Definition 28 (forced). We say that the scheme Φf is forced if in the graph GB,C ([m]) at least one of the following two conditions hold. (P1) There exists a cycle C of length at most n2 such that there are two elements u1 , u2 ∈ [m] which appear as labels of some two edges in C and x(u1 ) = x(u2 ). (P2) There exist cycles C1 and C2 of lengths at most n2 each and some two elements u, v ∈ [m] that appear as the labels of an edge in C1 and an edge in C2 respectively have the same x probe location, that is, x(u) = x(v). 18 Lemma 29. If the scheme Φf is forced, then it cannot represent all sets of size at most n. Lemma 30. If s < 17 m 1− ⌊ n 1⌋+1 4 , then Φf is forced. Our claim follows immediately from these two lemmas. Proof of Lemma 29. We assume either (P1) or (P2) above holds and derive a contradiction. Case: (P1) holds. Let C be a cycle: v v v v 1 2 k l C : r0 −→ r1 −→ · · · rk−1 −→ rk · · · −→ rl = r0 in GB,C of length at most n2 , that contains two edges labelled v1 and vk such that x(v1 ) = x(vk ). Let S = {v2 , · · · , vl } and T = {v1 }. Clearly |S|, |T | ≤ n We now show that Φf cannot represent any set S ′ such that S ⊆ S ′ ⊆ [m] \ T. In particular Φf cannot represent the set S. The assignment σS ′ has to assign the same bit to all locations in {r1 , · · · , rl = r0 }. To see this, first observe that for any element u ∈ S ′ , for the query function f to evaluate to 1, y(u) and z(u) must be assigned the same bit. Now, if r1 is assigned the bit b, then, since v2 ∈ S, r2 must be assigned the bit b, and again r2 is assigned the bit b and v3 ∈ S, r3 must be assigned the bit b and so on and thus all locations in {r1 , · · · , rl = r0 } is assigned the same bit b. Now, since the locations x(v1 ), x(vk ) are the same and the locations y(v1 ), z(v1 ), y(vk ), z(vk ) are assigned the same bit the query function will evaluate to the same value for both v1 and vk . But vk ∈ S ′ and v1 ∈ / S ′ . A contradiction. Case: (P2) holds. Let C1 and C2 be cycles: u u v v u u 1 2 k l C1 : r0 −→ r1 −→ · · · rk−1 −→ rk · · · −→ rl = r0 vp 1 2 C2 : t0 −→ t1 −→ · · · −→ tp = t0 each of length at most n2 and x(uk ) = x(v1 ). Let S = {u2 , u3 · · · ul } ∪ {v2 , v3 , · · · , vp } and T = {u1 , v1 }. Clearly, |S|, |T | ≤ n. We now show that Φf cannot represent any set S ′ such that S ⊆ S ′ ⊆ [m] \ T. In particular Φf cannot represent the set S. Following a similar reasoning as in the earlier case, we have the assignment σS ′ must assign all locations in {r1 , r2 , · · · , rl = r0 } the same bit b and all locations in {t1 , t2 , · · · , tp } the same bit b′ . Since u1 ∈ / S ′ , the locations in {y(u1 ), z(u1 )} = {r0 , r1 } cannot be assigned both 0’s and since both of them are assigned the same bit b, it must be that b = 1. Now, since uk ∈ S ′ and {y(uk ), z(uk )} = {rk−1 , rk } are both assigned b = 1, x(uk ) must be assigned 1. x(uk ) = x(v1 ) is assigned 1 and the locations y(v1 ), z(v1 ) are assigned the bit b′ implies f (A[x(v1 )], B[y(v1 )], C[z(v1 )]) = f (1, b′ , b′ ) = 1. But v1 ∈ / S ′ . A contradiction. Proof of Lemma 30. If the scheme uses space less and (P1) above does not hold then we show (P2) above must hold. n In the graph GB,C ([m]), as long as the number of edges is at least 6s1+⌊ 4 ⌋ we can find a cycle of length at most n2 from Proposition 21; since, the average degree then would be at least n 1 6s1+⌊ 4 ⌋ n ≥ 3s ⌊ 4 ⌋ ≥ 2, 2s 19 and since 1 n n (3s ⌊ 4 ⌋ − 1)⌊ 4 ⌋ > 2s. We repeatedly remove cycles: C1 , C2 , · · · from the graph GB,C each of length at most n2 using Proposition 21 till no more cycle remains; then, the number of remaining edges in the n graph will be at most 6s1+⌊ 4 ⌋ ; since, 1 1 1− n 1 1+ n s < m ⌊ 4 ⌋+1 =⇒ m > 7s ⌊ 4 ⌋ . 7 Thus, the sum of the lengths of the removed cycles exceeds s. Now, for any two edges labelled u, v in a removed cycle, z(u) 6= z(v); otherwise (P1) holds. Thus, the sum of the probe locations in C (z probe locations) of the labels of the edges appearing in the cycles exceeds s and hence some two cycles must have edge labels u, v respectively, such that z(u) = z(v). The same argument works for f¯ when in Lemma 29, we try to represent any set S ′ such that T ⊆ S ′ ⊆ [m] \ S. 4.4 Dimension argument In the proofs below we make use of algebraic arguments. The query functions considered in this section admit a dimension argument. The argument will go as follows. To each element in the universe we will associate a vector. If the space used by a scheme is small, then all these vectors will reside in a vector space of small dimension, which, in turn, will force a linear dependence between them. We will then argue that if we keep one element u aside and not store any other elements appearing in the linear dependence, then the scheme will be left with no choice for u leading to a contradiction. We now develop a common framework to prove lower bounds for the query functions- PARITY and (x, y, z) 7→ 1 iff x + y + z 6= 1 (over R). We assume the query function is f , where f could be either of these two functions. Fix a scheme Φf for the query function f , where the memory is a bit array A[1, · · · , s] consisting of s locations. For any element u ∈ [m], the scheme Φf probes three distinct locations x(u), y(u), z(u) ∈ A to determine if u is in the set or not. Given any set S ⊆ [m] of size at most n the assignment σf (S) ∈ {0, 1}s to A is such that for all elements u ∈ [m], f (x(u), y(u), z(u)) is 1 iff u ∈ S. Definition 31 (Fields F2 , R, vector spaces V , W , vector). Let F2 = {0, 1} be the field with modulo 2 arithmetic. Let R be the field of real numbers with the usual arithmetic. Let V = Fs2 be the vector space of all s dimensional vectors over the field F2 . Let W = Rs be the vector space of all s dimensional vectors over the field R. For an element u ∈ [m], define vector(u) to be an s-dimensional vector which contains three 1’s in the positions x(u), y(u) and z(u) and 0’s everywhere else. Clearly vector(u) ∈ V ⊆ W . 4.4.1 PARITY Now we fix f to be the parity function. Proposition 32. For a set S ⊆ [m] of size at most n, the assignment σf (S) is such that for all u ∈ [m], the dot product of the two vectors vector(u) and σf (S) in V is 1 iff u ∈ S. If m > s, then the set of m vectors {vector(u) : u ∈ [m]} ⊆ V is linearly dependent; V is s-dimensional. Thus, there exists distinct elements u, v1 , · · · , vt ∈ [m], such that vector(u) = t X i=1 20 vector(vi ). Taking dot product on both sides with the assignment vector σf ({u}) for the singleton {u}, we have vector(u).σf ({u}) = =⇒ 1 = t X i=1 t X vector(vi ).σf ({u}) 0 = 0, i=1 where the last step follows from Proposition 32 and the fact that u ∈ {u} and vi ∈ / {u} for all i ∈ [t]. A contradiction. Thus s ≥ m. 4.4.2 (x, y, z) 7→ 1 iff x + y + z 6= 1 Now we fix f to be the (x, y, z) 7→ 1 iff x + y + z 6= 1 function. Proposition 33. For a set S ⊆ [m] of size at most n, the assignment σf (S) is such that for all u ∈ [m], the dot product of the two vectors vector(u) and σf (S) in W is 0 iff u ∈ S. If m > s, then the set of m vectors {vector(u) : u ∈ [m]} ⊆ W is linearly dependent; W is s-dimensional. Thus, there exists distinct elements u, v1 , · · · , vt ∈ [m] and α1 , · · · , αt ∈ R, such that vector(u) = t X αi vector(vi ). (13) i=1 Taking dot product on both sides with the assignment vector σf ({u}) for the singleton {u}, we have vector(u).σf ({u}) = t X αi vector(vi ).σf ({u}) i=1 =⇒ 0 = t X αi i=1 where the last step follows from Proposition 33 and the fact that u ∈ {u} and vi ∈ / {u} for all i ∈ [t]. Again, taking dot product on both sides of Equation (13) with the assignment σf (∅) for the empty set, we have vector(u).σf ({u}) = =⇒ 1 = t X i=1 t X αi vector(vi ).σf ({u}) αi = 0 i=1 where the last step follows from Proposition 33 and the fact that u, vi ∈ / ∅ for all i ∈ [t]. A contradiction. Thus, s ≥ m. 4.5 Degree argument In this section we provide lower bound proofs for the query functions (x, y, z) 7→ (x ∧ y) ⊕ z and (x, y, z) 7→ 1 iff x + y + z = 1. 21 4.5.1 (x, y, z) 7→ (x ∧ y) ⊕ z Let Φ be a scheme with (x, y, z) 7→ (x ∧ y) ⊕ z as the query function. The memory consists of three distinct bit arrays: A[1, · · · , s], B[1, · · · , s] and C[1, · · · , s]. For any element u ∈ [m], the scheme probes three locations x(u) ∈ A, y(u) ∈ B and z(u) ∈ C to determine if u is in the set or not. We treat each location as a boolean variable. Given any set S ⊆ [m] of size at most n the assignment σ(S) ∈ {0, 1}3s to A, B and C is such that for all elements u ∈ [m], (x(u) ∧ y(u)) ⊕ z(u) is 1 iff u ∈ S. √ We first prove that s = Ω( mn) by specializing the lower bound proof in [12] to our case. Definition 34 (Field F2 , vector space V , polynomials PS ). Let F2 denote the field {0, 1} with mod 2 arithmetic. The query function (x, y, z) 7→ (x ∧ y) ⊕ z is same as (x, y, z) 7→ xy + z (over F2 ). Let V be the vector space over the field F2 of all multilinear polynomials of total degree at most 2n in the 3s variables: A[1], · · · , A[s], B[1], · · · , B[s], C[1], · · · , C[s] with coefficients coming from F2 . For each set S ⊆ [m], we define the polynomial PS in 3s variables and coefficients coming from the field F2 as follows: Y PS = (x(u)y(u) + z(u)). u∈S We make PS multilinear by reducing the exponents of each variable using the identity x2 = x for each variable x. This identity holds since we will be considering only 0-1 assignment to the variables. To prove the theorem for (x, y, z) 7→ (x ∧ y) ⊕ z, we use the following two lemmas.  Lemma 35. The set of m n multilinear polynomials {PS : |S| = n} is linearly independent in the vector space V .  . Lemma 36. V has a spanning set of size at most 3s+2n 2n Using these two lemmas, we first prove the theorem and provide the proofs of the lemmas later. Proof. Now, since the size of a linearly independent set is at most the size of a spanning set, using Lemmas 35 and 36, we have     m 3s + 2n ≤ n 2n  m n  e(3s + 2n) 2n ≤ =⇒ n 2n √ 2√ √ =⇒ 3s ≥ n( m − e n) e √ 18 √ m 1√ { =⇒ e n ≤ =⇒ 3s ≥ mn (when n ≤ m}). 10e 900 10 m , the fact that the assignments to the memory for storing different sets of size When n ≥ 900 √ m  m ⌈ 900 ⌉ are different implies that the space required is at least lg ⌈ m ⌉ ≥ Ω(m) ≥ Ω( mn). 900 Now, we prove the two lemmas. Proof of Lemma 35. First observe that any S of size n, the polynomial PS has n factors of degree 2 each. Hence, the degree of PS is at most 2n. 22 For sets S, S ′ ⊆ [m] of size n each, the evaluation of the polynomial PS on the assignment σ(S ′ ) is ( 0 if S 6= S ′ ′ PS (σ(S )) = 1 if S = S ′ . Since S 6= S ′ and |S| = |S ′ | = n ≥ 1, there exists u ∈ S such that u ∈ / S ′ and thus under the ′ ′ assignment σ(S ), the factor x(u)y(u) + z(u) in PS (σ(S )) evaluates to 0. While, when S = S ′ , for each u ∈ S the factor x(u)y(u) + z(u) in PS (σ(S ′ )) evaluates to 1.  In particular, this proves that {PS : |S| = n} has size m . Further we use this observation n below toPprove the lemma. Let S:|S|=n αS PS = 0 where each αS ∈ F2 . To show that the PS ’s are linearly independent, we need to show that each αS is 0. Consider an arbitrary set S ′ of size n, consider the assignment σ(S ′ ) to the variables in the above identity. X 0= αS PS (σ(S ′ )) S:|S|=n X = αS ′ PS ′ (σ(S ′ )) + αS PS (σ(S ′ )) S:S6=S ′ ,|S|=n = αS ′ PS ′ (σ(S ′ )) (since, PS (σ(S ′ )) = 0 for each S 6= S ′ ) = αS ′ (since, PS ′ (σ(S ′ )) = 1). Proof of Lemma 36. The monomials of total degree at most 2n form a spanning set; each polynomial in V can be written as a linear combination of these monomials. Thus, the size of this spanning set is 2n   X 3s k=0 k   3s + 2n ≤ , 2n where the last inequality follows from the fact that T 7→ T ∩ [3s] is an onto map from [3s]  to ≤2n . 4.5.2 [3s+2n] 2n (x, y, z) 7→ 1 iff x + y + z = 1 The lower bound proof for (x, y, z) 7→ 1 iff x + y + z = 1 is similar to the lower bound proof for (x, y, z) 7→ (x ∧ y) ⊕ z. The only difference here is that instead of looking at the query function over the field F2 , we consider the query function over the field F3 (the set of three elements {0, 1, 2} with mod 3 arithmetic). Over the field F3 , the query function (x, y, z) = 1 iff x + y + z = 1 is same as (x, y, z) 7→ x + y + z + xy + yz + zx (a degree 2 polynomial). Accordingly, the multilinear polynomial corresponding to a set S of size n is defined to be Y PS = (x(u) + y(u) + z(u) + x(u)y(u) + y(u)z(u) + z(u)x(u)). u∈S where we reduce the exponents using the identity x2 = x for each variable x (this identity holds as we consider only 0-1 assignments). Notice that PS has degree at most 2n and the rest of the proof is same as before. 23 4.6 Improved bound for (x, y, z) 7→ (x ∧ y) ⊕ z We now prove part (c) of Theorem 4. We extend the idea of the lower bound proof for the query function (x, y, z) 7→ (x ∧ y) ⊕ z for lg m ≤ n ≤ lgmm to get a better lower bound. We continue to use the framework of Section 4.5.1. For the scheme Φ, we first define a bipartite graph GA,B with vertex sets A, B ⊆ [s]. The sets A, B and the edges will be determined as follows. Initially let A = B = [s] and for each element u ∈ [m] add an edge labelled u between x(u) ∈ A and y(u) ∈ B. Our goal is to ensure m m′ that there are m′ ≥ 4m 5 edges in the graph and all vertices have degree at least 10s ≥ 10s . m , to get A and B. We will lose at most Repeatedly delete all vertices with degree less than 10s m m 2s × 10s = 5 edges. Also delete the corresponding elements of the universe [m] which appeared as labels of the deleted edges. Φ restricted to the remaining elements of the universe gives a we assume the remaining m′ elements scheme on m′ ≥ 4m 5 elements. Without loss of generality, r q ′ lg m lg m form the set [m′ ]. Proving a lower bound of Ω( m′ n lg lgnm′ ) = Ω( mn lg lgnm ) on the space for this restricted scheme will prove the theorem. Thus, GA,B has m′ edges labelled with distinct elements from the universe [m′ ] ⊆ [m] and m′ vertex sets A, B ⊆ [s] with minimum degree at least 10s . The edge labelled u ∈ [m′ ] has endpoints x(u) ∈ A and y(u) ∈ B.   m′ √ lg 16n ′ We assume s = cm n for c = − 1 and show a contradiction. m′ lg lg n Definition 37 (The parameters s, m, m′ , n, c, c′ , D, k and their relations). In the following, we assume ' & m′ lg 16n − 1, c= ′ lg lg mn m lg m ≤ n ≤ , lg m √ s = cm′ n. From before, we have 4 m ≥ m′ ≥ m. 5 Define r 1 m′ D := , 10 cn c′ := 2e · e2 · (64 · 1002 )2 c ln c,   1 n (1 − ′ ) . k := 2 c Clearly, using the above definitions we have minimum degree of a vertex in GA,B is at least D (by the assumption on s), 2 e2 m′ ≥ max{2.1002 cn, 64 · e2 n, 212 ′ n}, 2 n ≥ max{4c , 6c, 64 · 100 c}, min{D, c, c′ , n} ≥ 100 for all large m′ . These relations will be used in proving the various lemmas and the theorem below. 24 Definition 38 (Trap, gain, good gainer, restriction). In a bipartite graph (A, B, E), two edges e1 and e2 are said to trap an edge e = {a, b} if a ∈ e1 and b ∈ e2 . In a bipartite graph (A, B, E), a set S ⊆ E is said to gain if there exists two edges e1 , e2 ∈ S which either intersect at an end point, that is, e1 ∩ e2 6= ∅ or if they trap an edge e ∈ / S. In a bipartite graph (A, B, E), a set of edges S ⊆ E of size n is called a good gainer if every subset S ′ ⊆ S of size 2k gains (k is defined above). For a set T ⊆ [m], define restriction(T ) to be the set of all assignments to the memory specified by the storage scheme for sets which do not contain any element from T , that is, restriction(T ) = {σ(S) : |S| ≤ n and T ∩ S = ∅}. Lemma 39. Fix a bipartite graph G = (A, B, E) with |A|, |B| ≤ s, |E| = m′ and minimum degree at least D. A uniformly random set of edges S ⊆ E of size n is a good gainer with probability at least 12 for all large m′ . Lemma 40. If S ⊆ [m′ ] such that ES := {e ∈ E(GA,B ) : lab(e) ∈ S} is a good gainer in GA,B , n then there exist a polynomial PˆS of degree at most 2n − ⌊ 2c ⌋ and a set TS ⊆ [m′ ] such that n and PS (x̄) = PˆS (x̄) for every x̄ ∈ restriction(TS ). TS ∩ S = ∅, |TS | ≤ 2c Using these two lemmas, we first prove the theorem, and provide the proofs of the lemmas later. Proof of part (c). To each set S ⊆ [m′ ] of size n such that ES := {e ∈ E(GA,B ) : lab(e) ∈ S} is n ⌋ and set TS ⊆ [m′ ] such a good gainer, we associate a polynomial PˆS of degree at most 2n − ⌊ 2c n that |TS | ≤ 2c , TS ∩ S = ∅ and PS (x̄) = PˆS (x̄) for every x̄ ∈ restriction(TS ). This is possible due to Lemma 40. Let T be a random subset of [m′ ] where each element of [m′ ] is independently included in T with probability 21 . For each S above, let IS be the indicator random variable of the event S ∩ T = ∅ and TS ⊆ T . Then, the expected sum of these indicator random variables is     X 1 1 m′ 1 m′ 1 , · n ≥ 2n E[IS ] ≥ n 2 n 2n 2 c 2 S:ES is good gainer ′ where in the first inequality, the factor 12 mn appears because there are at least that many good gainers (from Lemma 39), the factor 21n · 1nc appears because for a given good gainer S, the n 2 elements in S must fall outside T and the elements in TS (at most nc of them) must fall inside T . The second inequality follows from the fact that cn ≥ n + c which holds for n, c ≥ 2. 1 m′ indicator random variables IS ’s are 1. Thus, there exists a T for which at least 22n n Fix such a T and let S = {S ⊆ [m′ ] : S is a good gainer, S ∩ T = ∅ , TS ⊆ T }. We have  1 m′ |S| ≥ 22n n . We first prove that the collection of polynomials {PˆS |S ∈ S} is linearly independent. First, observe that R ⊆ T implies that restriction(T ) ⊆ restriction(R). Thus, for any S ∈ S, PˆS (x̄) = PS (x̄) for all x̄ ∈ restriction(T ) ⊆ restriction(TS ). Since, each S ∈ S is disjoint from T , the assignment σ(S) ∈ restriction(T ). Consequently, for all S, S ′ ∈ S, we have the property that ( 0 if S 6= S ′ PˆS (σ(S ′ )) = PS (σ(S ′ )) = 1 if S = S ′ , which can be used to prove {PˆS |S ∈ S} is linearly independent as we did in Lemma 35. 25 Since the polynomials PˆS for S ∈ S are linearly independent and have degree at most n 1 n ⌋ ≤ 2n − 3c = 2n(1 − 6c ) (for n ≥ 6c), we have 2n − ⌊ 2c   1 m′ n 22n n  ′ n 1 m =⇒ 2n 2 n r ! 11   1 m′ 1− 6c 1 =⇒ 2n 1 − 4c 2 n r !1+ 6c1 1 m′ =⇒ n 2 n r !1+ 6c1 1 m′ =⇒ n 4 n   1 1 m′ 12c · =⇒ 16 n  n  ⌋ 3s + 2n − ⌊ 2c ≤ n 2n − ⌊ 2c ⌋ !2n(1− 1 ) 6c 3s + 2n ≤ e· 1 ) 2n(1 − 4c ≤ 3es + 2en ≤ 3es + 2en (14) √ ≤ 3e cm′ n (15) √ ≤ 12e c =⇒ c ≥ ′ m lg 16n ′ 12 lg lg mn , (16) 1 1 2 1 where (14) holds because c ≥ 12 =⇒ 1 − 4c ≥ 12 and 1 ≥ 1 − ( 6c ) =⇒ 1−1 1 ≥ 1 + 6c , 6c q q √ ′ ′ (15) holds for s = cm′ n and m′ ≥ 64e2 n =⇒ 21 mn − 2e ≥ 41 mn , and (16) holds because q  1  m′ ′ ′ lg 16n m′ 12c m′ 1 ′ ≥ 2122 e2 n =⇒ · and m =⇒ > lg c< lg mn ≥ 12e =⇒ lg mn ≥ ′ 16 n n 12 lg lg m n q √ ′ 12e lg mn ≥ 12e c. A contradiction. Proof of Lemma 40. For a S ⊆ [m′ ] such that ES isQ a good gainer, we get the desired TS and PˆS by running the following procedure. Recall PS = u∈S (x(u)y(u) + z(u)) consists of n factors each of degree 2. n ⌋ Initially we set Ŝ = S, EŜ = ES , PˆS = PS , and TS = ∅. We run the procedure for ⌊ 2c ˆ steps. In each step we maintain the following invariants: TS ∩ S = ∅, PS (x̄) = PS (x̄) for all assignments x̄ ∈ restriction(TS ). At the end of each step we delete two elements from each of Ŝ and EŜ . IF EŜ has two intersecting edges e1 and e2 (with labels v and w say), that is, x(v) = x(w) or y(v) = y(w), then we multiply out the two factors in PˆS corresponding to v and w to get a degree 3 factor: (x(v)y(v) + z(v))(x(w)y(w) + z(w)) =x(v)x(w)y(v)y(w) + x(v)y(v)z(w) + x(w)y(w)z(v) ( x(v)y(v)y(w) + x(v)y(v)z(w) + x(w)y(w)z(v) if x(v) = x(w) = x(v)x(w)y(v) + x(v)y(v)z(w) + x(w)y(w)z(v) if y(v) = y(w). n ⌋ steps and delete ELSE (no two edges in EŜ intersect) since we run the procedure for only ⌊ 2c n some 2 elements at the end of each step, |Ŝ| ≥ n − 2(⌊ 2c ⌋ − 1) ≥ 2k = 2⌈ n2 (1 − 1c )⌉. We invoke Lemma 39 with S ′ = EŜ , to find edges e1 and e2 (with labels v and w say) in 26 EŜ ⊆ ES that traps edge e ∈ / ES (with label t say). Note that e is not just not in EŜ as promised by Lemma 39 but also not in ES ; otherwise e and e1 would be both in ES and intersecting which will satisfy the IF part. We add the element t to the set TS . Observe that for any assignment in restriction(TS ), x(t)y(t) + z(t) = 0, that is, x(t)y(t) = z(t). We use this relation, to simplify the product of the two factors in PˆS corresponding to v and w to get a degree 3 factor: (x(v)y(v) + z(v))(x(w)y(w) + z(w)) =x(v)x(w)y(v)y(w) + x(v)y(v)z(w) + x(w)y(w)z(v) ( z(t)x(w)y(v) + x(v)y(v)z(w) + x(w)y(w)z(v) if trapped edge e = {x(v), y(w)} = z(t)x(v)y(w) + x(v)y(v)z(w) + x(w)y(w)z(v) if trapped edge e = {x(w), y(v)}. We delete v and w from Ŝ and delete e1 and e2 from EŜ and repeat. n n ⌋ steps, clearly |TS | ≤ ⌊ 2c ⌋, S ∩ TS = ∅, degree At the end of the procedure, that is, after ⌊ 2c n of PˆS ≤ 2n − ⌊ 2c ⌋, and PˆS (x̄) = PS (x̄) for all assignments x̄ ∈ restriction(TS ). To prove Lemma 39 we require the following lemma, which we prove later. Lemma 41. Fix any bipartite graph G = (A, B, E) with |A|, |B| ≤ s, |E| = m′ and minimum degree at least D. A uniformly random set of edges S ′ ⊆ E of size 2k gains with probability at n ′ least 1 − exp(− 32·100 2 c ) for all large m . Proof of Lemma 39. For a uniformly random set S of n edges from E, the probability that S is a good gainer can be lower bounded using the union bound and Lemma 41 as follows. Pr[S is a good gainer] = 1 − Pr[∃S ′ ⊆ S, |S ′ | = 2k, S ′ does not gain] X ≥1− Pr[S ′ does not gain] S ′ ⊆S,|S ′ |=2k     n n exp − 32 · 1002 c 2k     n n ≥1− exp − n − 2k 32 · 1002 c  n−2k   en n ≥1− . exp − n − 2k 32 · 1002 c ≥1− Now, since 2k ≥ n(1 − c1′ ) therefore n − 2k ≤ cn′ , and since 2k ≤ n(1 − c1′ ) + 2 therefore n − 2k ≥ cn′ − 2 ≥ 2cn′ for n ≥ 4c′ . Therefore, using these bounds we get,   n′   en c n Pr[S is a good gainer] ≥ 1 − n exp − 32 · 1002 c 2c′    ′ ln 2ec 1 = 1 − exp n − c′ 32 · 1002 c   n ≥ 1 − exp − 64 · 1002 c 1 ≥ , 2 where the last inequality holds for n ≥ 64 · 1002 c and the second last inequality holds for c′ = 2e · e2 · (64 · 1002 )2 c ln c, since 2e · e2 · 64 · 1002 ≥ ln(2e · e2 · 64 · 1002 )2 =⇒ 2e · e2 · 64 · 1002 − 2 ≥ ln(2e · 64 · 1002 )(2e · e2 · 64 · 1002 ), 27 and since ln c ≥ 1, we have (2e · e2 · 64 · 1002 − 2) ln c + 2 ln c ≥ ln(2e · 64 · 1002 )(2e · e2 · 64 · 1002 ) + ln c + ln ln c =⇒ (2e · e2 · 64 · 1002 ) ln c ≥ ln((2e · 64 · 1002 )(2e · e2 · 64 · 1002 )c ln c) = ln 2ec′ c′ ≥ 64 · 1002 c ln 2ec′ ln 2ec′ 1 ≥ 2 . =⇒ 32 · 1002 c c′ =⇒ Proof of Lemma 41. To aid our calculations, we pick a uniformly random set S ′ with 2k edges in two steps. In step I, we randomly pick k vertices: v1 , · · · , vk from A with replacement with each vertex being picked with probability proportional to its degree, that is, for all v ∈ A, Pr[vi = v] = deg(v) m′ . For each vi we pick an edge ei incident on vi uniformly at random, that is, 1 . Thus for each edge e ∈ E whose endpoint in A is v, with probability deg(v i) Pr[ei = e] = Pr[vi = v] Pr[ei = e|vi = v] = 1 1 deg(vi ) · = ′. ′ m deg(vi ) m This way, some t distinct edges are picked, where t ≤ k (note that t can be strictly less than k if for some i 6= j, ei = ej ). In step II, we delete the edges picked in step I, and pick a set of 2k − t edges uniformly at random without replacement from the remaining edges. Let BAD denote the event that S ′ does not gain. In step I, after sampling the vertices {v1 , · · · , vk } we get some fixed multiset of k vertices with t distinct vertices. It is enough to show that conditioning on each choice for v1 , · · · , vk , the  n conditional probability of BAD is at most exp − 32·1002 c for large m′ . After conditioning, the only source of randomness in the calculations below comes from sampling the 2k edges in steps I and II. To upper bound the probability that BAD occurs we have the following two cases. Let B0 ⊆ B be the neighbourhood of {v1 , · · · , vk }. Case: |B0 | ≥ KD 100 . For the event BAD to occur, none of the edges picked in step II should be incident on any vertex in B0 , the neighbourhood of {v1 , · · · , vk }. For if some edge f = {a, b} picked in step II is incident on a vertex b in the neighbourhood of some vi , then either f and ei trap an edge not in S with endpoints b and vi or f intersects some edge ej with endpoints b and vi . Since each vertex has degree at least D, the number of edges 2 incident on B0 is at least kD 100 and after deleting the edges {e1 , · · · , ek } picked in step I, at D2 least k( 100 − 1) edges remain. Since, at least k edges are picked in step II out of at most m′ edges, the probability that BAD occurs can be upper bounded as follows. 2 Pr[BAD] ≤ k( D − 1) 1 − 100 ′ m Now, using the fact that 2k ≥ n(1 − 1 c′ ), !k D≥ k2 (D 2 − 100) ≤ exp − 100m′ 1 10  q m′ cn ,  . we get ! ! n(1 − c1′ )2 n(1 − c1′ )2 1002 cn Pr[BAD] ≤ exp − , (1 − ) ≤ exp − 4 · 1002 c m′ 8 · 1002 c where the last inequality holds since m′ ≥ 2 · 1002 cn. 28 kD . For each i ∈ [k], define Γ′ (vi ) := Γ({v1 , · · · , vi−1 }) ∩ Γ(vi ) and deg′ (vi ) := Case: |B0 | < 100 |Γ′ (vi )| − 1, where Γ stands for the neighbourhood function. For the event BAD to occur, the following event BAD(vi ) must occur for each vi . When vi ∈ / {v1 , · · · , vi−1 }, BAD(vi ) ≡ ei is not incident on any vertex in Γ′ (vi ). When vi = vj for some j < i, BAD(vi ) ≡ ei = ej , where j is the smallest number such that vi = vj . For if some BAD(vi ) does not occur, then S ′ gains; in the case when vi ∈ / {v1 , · · · , vi−1 } and ei is incident on some vertex b in Γ′ ({vi }), then b must be the neighbour of some vertex vj where j < i and then either ei and ej trap an edge not in S with endpoints b, vj or ei intersects ej at b. Again, in the case when vi = vj for some j < i and ei 6= ej (where j is the smallest number for which vi = vj , then clearly ei and ej intersect at the vertex vi . Thus we bound the probability of BAD as follows. Pr[BAD] ≤ Pr[∀i ∈ [k] BAD(vi )] = k Y Pr[BAD(vi )], i=1 where the equality follows from the fact that, in the conditional space, the BAD(vi )’s are independent events. Now, if vi ∈ / {v1 , · · · , vi−1 }, Pr[BAD(vi )] ≤ 1 − deg′ (vi ) + 1 deg′ (vi ) ≤1− . deg(vi ) deg(vi ) Else, when vi = vj for some j < i, Pr[Bad(vi )] ≤ 1 − deg′ (vi ) . deg(vi ) Using this upper bound, we have Pr[BAD] ≤ k Y i=1 Pr[BAD(vi )] ≤ k  Y deg′ (vi ) 1− deg(vi )  ≤ exp − k X deg′ (vi ) deg(vi )  49 1 49 n(1 − ′ ) , ≤ exp(− k) ≤ exp − 50 100 c i=1  i=1 where the second last inequality can be shown as follows. Ṡk Since B0 = i=1 Γ(vi ) \ Γ′ (vi ), we have k X  kD deg(vi ) − (deg′ (vi ) + 1) ≤ 100 i=1   k X deg′ (vi ) kD deg(vi ) 1 − +k =⇒ ≤ deg(vi ) 100 i=1  k  X kD deg′ (vi ) =⇒ D +k ≤ 1− deg(vi ) 100 i=1  k  X deg′ (vi ) k k =⇒ 1− + ≤ deg(vi ) 100 D i=1 =⇒ k X i=1 − deg′ (vi ) k k ≥k−( + ) deg(vi ) 100 D 49 k, 50 where the last inequality follows from the fact that D ≥ 100 for large m′ . ≥ 29 ! Thus combining the two cases, for large m′ , the probability that a random set of 2k edges gains is at least !   n(1 − c1′ )2 49 1 , exp − n(1 − ′ ) } 1 − max{exp − 8 · 1002 c 100 c !   n(1 − c1′ )2 n ≥ 1 − exp − ≥ 1 − exp − , 8 · 1002 c 32 · 1002 c where the last two inequalities follow from the fact that c′ , c ≥ 2. References [1] Noga Alon and Uriel Feige. On the power of two, three and four probes. In Proceedings of the Twentieth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2009, New York, NY, USA, January 4-6, 2009, pages 346–354, 2009. URL: http://dl.acm.org/citation.cfm?id=1496770.1496809. [2] Noga Alon, Shlomo Hoory, and Nathan Linial. The Moore bound for irregular graphs. Graphs and Combinatorics, 18(1):53–57, 2002. URL: http://dx.doi.org/10.1007/s003730200002, doi:10.1007/s003730200002. [3] Ajesh Babu and Jaikumar Radhakrishnan. An entropy based proof of the Moore bound for irregular graphs. CoRR, abs/1011.1058, 2010. URL: http://arxiv.org/abs/1011.1058. [4] Harry Buhrman, Peter Bro Miltersen, Jaikumar Radhakrishnan, and Srinivasan Venkatesh. Are bitvectors optimal? SIAM J. Comput., 31(6):1723–1744, 2002. URL: http://dx.doi.org/10.1137/S0097539702405292, doi:10.1137/S0097539702405292. [5] Mohit Garg and Jaikumar Radhakrishnan. Set membership with a few bit probes. In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2015, San Diego, CA, USA, January 4-6, 2015, pages 776–784, 2015. URL: http://dx.doi.org/10.1137/1.9781611973730.53, doi:10.1137/1.9781611973730.53. [6] Michael T. Goodrich and Michael Mitzenmacher. Invertible Bloom lookup tables. In 49th Annual Allerton Conference on Communication, Control, and Computing, Allerton 2011, Allerton Park & Retreat Center, Monticello, IL, USA, 28-30 September, 2011, pages 792–799, 2011. URL: http://dx.doi.org/10.1109/Allerton.2011.6120248, doi:10.1109/Allerton.2011.6120248. [7] Moshe Lewenstein, J. Ian Munro, Patrick K. Nicholson, and Venkatesh Raman. Improved explicit data structures in the bitprobe model. In Algorithms - ESA 2014 22th Annual European Symposium, Wroclaw, Poland, September 8-10, 2014. Proceedings, pages 630–641, 2014. URL: http://dx.doi.org/10.1007/978-3-662-44777-2_52, doi:10.1007/978-3-662-44777-2_52. [8] Michael Luby, Michael Mitzenmacher, Mohammad Amin Shokrollahi, and Daniel A. Spielman. Efficient erasure correcting codes. IEEE Trans. Information Theory, 47(2):569–584, 2001. URL: http://dx.doi.org/10.1109/18.910575, doi:10.1109/18.910575. [9] Ali Makhdoumi, Shao-Lun Huang, Muriel Médard, and Yury Polyanskiy. On locally decodable source coding. In 2015 IEEE International Conference on Communications, ICC 2015, London, United Kingdom, June 8-12, 2015, pages 4394–4399, 2015. URL: http://dx.doi.org/10.1109/ICC.2015.7249014, doi:10.1109/ICC.2015.7249014. 30 [10] Marvin Minsky and Seymour Papert. Perceptrons. MIT press, Cambridge, MA, 1969. [11] Jaikumar Radhakrishnan, Venkatesh Raman, and S. Srinivasa Rao. Explicit deterministic constructions for membership in the bitprobe model. In Algorithms - ESA 2001, 9th Annual European Symposium, Aarhus, Denmark, August 28-31, 2001, Proceedings, pages 290–299, 2001. URL: http://dx.doi.org/10.1007/3-540-44676-1_24, doi:10.1007/3-540-44676-1_24. [12] Jaikumar Radhakrishnan, Pranab Sen, and Srinivasan Venkatesh. The quantum complexity of set membership. Algorithmica, 34(4):462–479, 2002. URL: http://dx.doi.org/10.1007/s00453-002-0979-0, doi:10.1007/s00453-002-0979-0. [13] Jaikumar Radhakrishnan, Smit Shah, and Saswata Shannigrahi. Data structures for storing small sets in the bitprobe model. In Mark de Berg and Ulrich Meyer, editors, Algorithms - ESA 2010, 18th Annual European Symposium, Liverpool, UK, September 6-8, 2010. Proceedings, Part II, volume 6347 of Lecture Notes in Computer Science, pages 159–170. Springer, 2010. URL: http://dx.doi.org/10.1007/978-3-642-15781-3_14, doi:10.1007/978-3-642-15781-3_14. [14] Emanuele Viola. Bit-probe lower bounds for succinct data structures. SIAM J. Comput., 41(6):1593–1604, 2012. URL: http://dx.doi.org/10.1137/090766619, doi:10.1137/090766619. [15] Wikiversity. The 22 becs, 3-ary boolean functions — wikiversity, 2016. [Online; accessed 7-August-2016]. URL: https://en.wikiversity.org/w/index.php?title=3-ary_Boolean_functions&oldid=1587287. 31
8cs.DS
arXiv:1302.0368v1 [math.AC] 2 Feb 2013 COHEN-MACAULAY-NESS IN CODIMENSION FOR BIPARTITE GRAPHS HASSAN HAGHIGHI, SIAMAK YASSEMI, AND RAHIM ZAARE NAHANDI Abstract. Let G be an unmixed bipartite graph of dimension d − 1. Assume that Kn,n , with n ≥ 2, is a maximal complete bipartite subgraph of G of minimum dimension. Then G is Cohen-Macaulay in codimension d − n + 1. This generalizes a characterization of Cohen-Macaulay bipartite graphs by Herzog and Hibi and a result of Cook and Nagel on unmixed Buchsbaum graphs. Furthermore, we show that any unmixed bipartite graph G which is Cohen-Macaulay in codimension t, is obtained from a Cohen-Macaulay graph by replacing certain edges of G with complete bipartite graphs. We provide some examples. 1. Introduction Cohen-Macaulay simplicial complexes are among central research topics in combinatorial commutative algebra. While characterization of such complexes is a far reaching problem, one appeals to study specific families of Cohen-Macaulay simplicial complexes. Flag complexes are among important families of complexes recommended to study [10, page 100]. However, it is known that a simplicial complex is Cohen-Macaulay if and only if its barycentric subdivision is a Cohen-Macaulay flag complex. Therefore, a characterization of Cohen-Macaulay flag complexes is equivalent to a characterization of Cohen-Macaulay simplicial complexes. Nevertheless, after all, the ideal of a flag complex is generated by quadratic square-free monomials, which are simpler compared with arbitrary square-free monomial ideals. Furthermore, it seems that, expressing many combinatorial properties in terms of graphs are more convenient. As some evidences, the characterization of unmixed bipartite graphs by Villarreal [11] and Cohen-Macaulay bipartite graphs by Herzog and Hibi [5] are well expressed in terms of graphs. On the other hand, in the hierarchy of families of graphs with respect to CohenMacaulay property, Buchsbaum complexes appear right after Cohen-Macaulay ones. Unmixed bipartite Buchsbaum graphs were characterized by Cook and Nagel [1] (also by the authors [3]). Natural families of graphs in this hierarchy are bipartite CM t graphs, i.e., graphs that their independence complexes are pure and CohenMacaulay in codimension t. The concept of CM t simplicial complexes were introduced in [4] which is the pure version of simplicial complexes Cohen-Macaulay in codimension t studied by Miller, Novik and Swartz [6]. In this note, we give characterizations of unmixed bipartite CM t graphs in terms of its dimension and the minimum dimension of its maximal nontrivial complete bipartite subgraphs. Cook and Nagel showed that the only non-Cohen-Macaulay unmixed bipartite graphs are 2000 Mathematics Subject Classification. 13H10, 05C75. Key words and phrases. Flag complex, Cohen-Macaulay complex in codimension. Emails: [email protected], [email protected], [email protected]. 1 2 HASSAN HAGHIGHI, SIAMAK YASSEMI, AND RAHIM ZAARE NAHANDI complete bipartite graphs [1, Theorem 4.10] and [3, Theorem 1.3]. Our results are generalizations of this fact to unmixed bipartite graphs which are Cohen-Macaulay in arbitrary codimension. In the next section we gather necessary definitions and known results to be used in the rest of the paper. In Section 3 we improve some results on joins of simplicial complexes and disjoint unions of graphs with respect to the CM t property. Section 4 is devoted to two characterizations of bipartite CM t graphs and some examples. 2. Preliminaries For basic definitions and general facts on simplicial complexes we refer to the book of Stanley [10]. By a complex we will always mean a simplicial complex. Let G = (V, E) be a simple graph with vertex set V and edge set E. The inclusive neighborhood of v ∈ V is the set N [v] consisting of v and vertices adjacent to v in G. The independence complex of G = (V, E) is the complex Ind(G) with vertex set V and with faces consisting of independent sets of vertices of G, i.e., sets of vertices of G where no two elements of them are adjacent. These complexes are called flag complexes, and their Stanley-Reisner ideal is generated by quadratic square-free monomials. By dimension of a graph G we mean the dimension of the complex Ind(G). A graph G is said to be unmixed if Ind(G) is pure. For an integer t ≥ 0, a complex ∆ is called CM t if it is pure and for every face F ∈ ∆ with #(F ) ≥ t, link ∆ (F ) is Cohen-Macaulay. This is the same as pure complexes which are Cohen-Macaulay in codimension t. Accordingly, CM 0 and CM 1 complexes are precisely Cohen-Macaulay and Buchsbaum complexes, respectively. Clearly, a CM t complex is CM r for all r ≥ t and a complex of dimension d − 1 is always CM d−1 . One uses the convention that for t < 0, CMt would mean CM0 . A graph G is called CM t if Ind(G) is CM t . A basic tool for checking CM t property of complexes is the following lemma. Lemma 2.1. ([4, Lemma 2.3]) Let t ≥ 1 and let ∆ be a nonempty complex. Then the following are equivalent: (i) ∆ is a CMt complex. (ii) ∆ is pure and link ∆ (v) is CMt−1 for every vertex v ∈ ∆. By the straightforward identity link Ind(G) (v) = Ind(G \ N [v]), the counter-part of this lemma for graphs would be the following: Lemma 2.2. Let t ≥ 1 and let G be a graph. Then the following are equivalent: (i) G is a CMt graph. (ii) G is unmixed and G \ N [v] is a CMt−1 graph for every vertex v ∈ G. We recall some basic relevant facts on bipartite graphs. A graph G = (V, E) is called bipartite if V is a disjoint union of a partition V1 and V2 and E ⊂ V1 × V2 . If #(V1 ) = m and #(V2 ) = n and E = V1 × V2 , then G is the complete bipartite graph Km,n . We will be interested in unmixed complete bipartite graphs Kn,n . Unmixed bipartite graphs are characterized by Villarreal in the following result. Theorem 2.3. [11, Theorem 1.1] Let G be a bipartite graph without isolated vertex. Then G is unmixed if and only if there is a partition V1 = {x1 , · · · , xn } and V2 = {y1 , · · · , yn } of vertices of G such that COHEN-MACAULAY-NESS IN CODIMENSION FOR BIPARTITE GRAPHS 3 (1) xi yi is an edge in G for 1 ≤ i ≤ n and (2) If xi yj and xj yk are edges in G, for some distinct i, j and k, then xi yk is an edge in G. In this case, such a partition and ordering is called a pure order of G. The edges xi yi , i = 1, · · · , n are called a perfect matching edges of G. A pure order is said to have a cross if, for some i 6= j, xi yj and xj yi are both edges in G. Otherwise, the order is called cross-free (see [1, § 4]). For unmixed bipartite graphs, being cross-free is independent of an ordering of vertices of G. More precisely, if G has a cross in some pure ordering, it has a cross in every pure ordering [1, Lemma 4.5]. An immediate consequence of Theorem 2.3 is the following useful lemma. Lemma 2.4. Let G be an unmixed bipartite graph with pure order of vertices ({x1 , · · · , xd }, {y1, · · · , yd }) and let Kn,n be a complete bipartite subgraph of G on ({xi1 , · · · , xin }, {yi1 , · · · , yin }). (i) If xj yik is an edge in G for some j and k, then xj yil is an edge in G for all l = 1, · · · , n. (ii) If xik yj is an edge in G for some k and j, then xil yj is an edge in G for all l = 1, · · · , n. Proof. The assertion (i) is immediate by Theorem 2.3 because xik yil is an edge in Kn,n ⊂ G for all l = 1, · · · , n. Also (ii) follows because xil yik is an edge in Kn,n ⊂ G for all l = 1, · · · , n.  There are also at least two nice characterization of Cohen-Macaulay bipartite graphs. Theorem 2.5. [5, Theorem 3.4] Let G be a bipartite graph without isolated vertices. Then G is Cohen-Macaulay if and only if there is a pure ordering V1 = {x1 , · · · , xn } and V2 = {y1 , · · · , yn } of vertices of G such that xi yj being in G implies i ≤ j. The ordering in Theorem 2.5 is called a Macaulay order of vertices of G. Proposition 2.6. [1, Proposition 4.8] Let G be a bipartite graph. Then G is Cohen-Macaulay if and only if G has a cross-free pure order. Bipartite Buchsbaum graphs are also classified. First recall that a complex is Buchsbaum if and only if it is pure and the link of each vertex is Cohen-Macaulay [9]. Thus, a graph is Buchsbaum if and only if G is unmixed and for each vertex v ∈ G, G\N [v] is Cohen-Macaulay. For bipartite graphs there is a sharper result. Complete bipartite graphs are well-known to be Buchsbaum (e.g., see [12, Proposition 2.3]). But indeed, the converse is also true. Theorem 2.7. (see [1, Theorem 4.10] or [3, Theorem1.3]) Let G be a bipartite graph. Then G is Buchsbaum if and only if G is a complete bipartite graph Kn,n for some n ≥ 2, or G is Cohen-Macaulay. 3. Joins of CM t complexes and disjoint unions of CM t graphs It is known that the join of two complexes is Cohen-Macaulay if and only if they are both Cohen-Macaulay (see [8] and [2]). If ∆ is a CM r complex of dimension d − 1 and ∆′ is a CM r′ complex of dimension d′ − 1, then their join ∆ ∗ ∆′ is a CM t complex where t = max{d + r′ , d′ + r} [4, Proposition 2.10]. However, if one 4 HASSAN HAGHIGHI, SIAMAK YASSEMI, AND RAHIM ZAARE NAHANDI of the complexes is Cohen-Macaulay, this result could be strengthened. Below we combine this with relevant known results. Theorem 3.1. Let ∆ and ∆′ be two complexes of dimensions d − 1 and d′ − 1, respectively. Then (i) The join complex ∆ ∗ ∆′ is Cohen-Macaulay if and only if both ∆ and ∆′ are so. (ii) If ∆ is Cohen-Macaulay and ∆′ is CM r′ for some r′ ≥ 1, then ∆ ∗ ∆′ is CM d+r′ (independent of d′ ). This is sharp, i.e., if ∆′ is not CM r′ −1 , then ∆ ∗ ∆′ is not CM d+r′ −1 . In particular, a cone on ∆′ is CM r′ +1 . (iii) If ∆ is CM r and ∆′ is CM r′ for some r, r′ ≥ 1, then ∆ ∗ ∆′ is CM t where t = max{d + r′ , d′ + r}. Conversely, if ∆ ∗ ∆′ is CM t , then ∆ is CM t−d′ and ∆′ is CM t−d . Proof. The statement in (i) is proved by Sava [8] and Fröberg [2]. The assertion (iii) is proved in [4, Theorem 2.10]. We prove (ii) using induction on d + r′ ≥ 2. Let d + r′ = 2, i.e., d = 1 and r′ = 1. Then ∆ = {v} is a singleton. Thus link ∆∗∆′ (v) = ∆′ , which is CM 1 . For v ∈ ∆′ , link ∆∗∆′ (v) = ∆ ∗ link ∆′ (v), which is Cohen-Macaulay by (i). Thus by Lemma 2.1, ∆ ∗ ∆′ is CM 2 . Now let d + r′ ≥ 2. Let v ∈ ∆. Then, link ∆∗∆′ (v) = link ∆ (v) ∗ ∆′ . But link ∆ (v) is Cohen-Macaulay of dimension less than d − 1, and ∆′ is CM r′ . Thus by induction hypothesis link ∆∗∆′ (v) is CM d−1+r′ . If v ∈ ∆′ , then link ∆∗∆′ (v) = ∆′ ∗ link ∆′ (v). But link ∆′ (v) is CM r′ −1 and hence link ∆∗∆′ (v) is again CM d+r′ −1 . Therefore, ∆ ∗ ∆′ is CM d+r′ . To prove that this result is sharp, proceed by induction on d ≥ 1. Indeed, in this case, for any v ∈ ∆, link ∆ (v) has dimension less than d − 1 and hence by induction hypothesis, link ∆∗∆′ (v) = link ∆ (v) ∗ ∆ is not CM d+r′ −2 . Therefore, ∆ ∗ ∆′ is not CM d+r′ −1  Let G ⊔ G′ denote the disjoint union of graphs G and G′ . By the fact that Ind(G ⊔ G′ ) = Ind(G)∗Ind(G′ ), the counter-part of Theorem 3.1 for graphs will be the following. Theorem 3.2. Let G and G′ be two graphs on disjoint sets of vertices and of dimensions d − 1 and d′ − 1, respectively. Then (i) The graph G ⊔ G′ is Cohen-Macaulay if and only if both G and G′ are so. (ii) If G is Cohen-Macaulay and G′ is CM r′ for some r′ ≥ 1, then G ⊔ G′ is CM d+r′ . If G′ is not CM r′ −1 , then G ⊔ G′ is not CM d+r′ −1 . (iii) If G is CM r and G′ is CM r′ for some r, r′ ≥ 1, then G ⊔ G′ is CM t where t = max{d + r′ , d′ + r}. Conversely, if G ⊔ G′ is CM t , then G is CM t−d′ and G′ is CM t−d . 4. Two characterizations of bipartite CM t graphs We now restrict to the case of bipartite graphs. Since Cohen-Macaulay bipartite graphs are characterized by Herzog and Hibi [5, Theorem 3.4], and also in a different version by Cook and Nagel [1, Proposition 4.8], we consider the nonCohen-Macaulay case. Theorem 4.1. Let G be an unmixed bipartite graph of dimensions d − 1. Let Kn,n , with n ≥ 2, be a maximal complete bipartite subgraph of G of minimum dimension. Then G is CM d−n+1 but it is not CM d−n . COHEN-MACAULAY-NESS IN CODIMENSION FOR BIPARTITE GRAPHS 5 Proof. We prove both assertions by induction on d ≥ 2. If d = 2 then G = K2,2 which is CM 1 but it is not Cohen-Macaulay. Assume that d > 2. We show that for every v ∈ G, G \ N [v] is CM n−d and for some v ∈ G it is not CM n−d−1 . Let ({x1 , · · · , xd }, {y1, · · · , yd }) be a pure order of G. Let xi be a vertex of some maximal bipartite subgraph Km,m with m ≥ n. Then G\N [xi ] is a disjoint union of c ≥ m−1 isolated vertices and an unmixed bipartite graph H of dimension d−c−2. The graph H is unmixed because Ind(G \ N [xi ]) = link xi (Ind(G)), and any link of a pure complex is pure. But G \ N [xi ] = {xi1 , · · · , xic } ⊔ H is unmixed if and only if H is so. Observe that if yj0 is a vertex of a maximal bipartite subgraph of G and yj0 ∈ N [xi ], then by Lemma 2.4, all yj vertices of this subgraph belong to N [xi ]. Thus if H has no crosses, by Proposition 2.6 it is Cohen-Macaulay. Otherwise, the minimum dimension of maximal complete bipartite subgraphs of H will not be less than the minimum dimension of such subgraphs in G. Hence by the induction hypothesis H is CM d−c−n and by Theorem 3.2(ii), G \ N [xi ] is CM n−d . If xi does not belong to any maximal bipartite subgraph of G of positive dimension, then G \ N [xi ] is a disjoint union of c ≥ 0 isolated vertices and an unmixed bipartite graph H of dimension d − c − 2. Hence H is CM d−c−n and by Theorem 3.2(ii), G \ N [xi ] is CM d−n . A similar argument reveals that for any yi ∈ G, the graph G\N [yi ] is CM d−n . Therefore, by Lemma 2.2, G is CM d−n+1 . We now proceed the induction step to show that this result is sharp. Let d > 2 and let Kn,n , n ≥ 2, be a maximal bipartite subgraph of G of minimum dimension. Take xi ∈ G \ Kn,n. First assume that xi is not adjacent to any vertex in Kn,n and consider G \ N [xi ]. Let G \ N [xi ] be the disjoint union of c ≥ 0 isolated vertices and an unmixed bipartite graph H of dimension d − c − 2. Then H contains Kn,n and hence by induction hypothesis H is sharp CM d−c−n and G \ N [xi ] is sharp CM d−n . Therefore, G can not be CM d−n . Now assume that xi yj ∈ G for some j with yj ∈ Kn,n . Then by purity of the order, all yk ∈ Kn,n is adjacent to xi . But then yi is not adjacent to any vertex of Kn,n , because otherwise, Kn,n will not be maximal. In this case, consider G \ N [yi ] and proceed similar to the previous case.  As a second characterization of bipartite CM t graphs, we show that any CM t graph is obtained from a Cohen-Macaulay graph H by replacing the perfect matching edges of H by complete bipartite graphs. This statement will be more precise in the next theorem. But first we provide a definition and a lemma. Definition 4.2. Let H be an unmixed bipartite graph with pure order ({x1 , · · · , xr }, {y1 , · · · , yr }). For a fixed i, by replacing the edge xi yi ∈ H with a complete bipartite graph Kni ,ni = {xi1 , · · · , xini } × {yi1 , · · · , yini } we mean a bipartite graph H ′ with vertex set ({x1 , · · · , xi−1 , xi1 , · · · , xini , xi+1 , · · · , xr }, {y1 , · · · , yi−1 , yi1 , · · · , yini , yi+1 , · · · , yr }), preserving all adjacencies, i.e., (i) xs yt ∈ H ′ for all s, t 6= i if and only if xt ys ∈ H, (ii) xik yj ∈ H ′ for all k if and only if xi yj ∈ H, (iii) xj yik ∈ H ′ for all k if and only if xj yi ∈ H. 6 HASSAN HAGHIGHI, SIAMAK YASSEMI, AND RAHIM ZAARE NAHANDI Lemma 4.3. Let G be an unmixed bipartite graph with pure order on the vertex set V (G) = V ∪ W where V = {x1 , · · · , xd } and W = {y1 , · · · , yd }. Let n1 , · · · , nd be any positive integers. Let G′ = G(n1 , · · · , nd ) be the graph obtained by replacing each edge xi yi with the complete bipartite graph Kni ,ni = {xi1 , · · · , xini } × {yi1 , · · · , yini } for all i = 1, . . . , d. Then G′ is also unmixed. Proof. Let Kni ,ni = {xi1 , · · · , xini } × {yi1 , · · · , yini }. Then V (G′ ) = ({x11 , · · · , x1n1 , · · · , xd1 , · · · , xdnd }, {y11 , · · · , y1n1 , · · · , yd1 , · · · , ydnd }) is a pure order of G′ . In fact, for all i, r, xir yir ∈ G′ . Also if xir yjs ∈ G′ and xjs ykt ∈ G′ , then xi yj ∈ G and xj yk ∈ G, and hence, xi yk ∈ G. Thus by the construction of G′ , xir ykt ∈ G′ .  Theorem 4.4. Let G be a Cohen-Macaulay bipartite graph with a Macaulay order on the vertex set V (G) = V ∪ W where V = {x1 , · · · , xd } and W = {y1 , · · · , yd }. Let n1 , · · · , nd be any positive integers with ni ≥ 2 for at list one i. Let G′ = G(n1 , · · · , nd ) be the graph obtained by replacing each edge xi yi with the complete bipartite graph Kni ,ni for all i = 1, . . . , d. Let ni0 = min{ni > 1 : i = 1, · · · , d}, n = Pd ′ i=1 ni . Then G is exclusively a CM n−ni0 +1 graph. Furthermore, any bipartite CM t graph is obtained by such a replacement of complete bipartite graphs in a unique bipartite Cohen-Macaulay graph. Proof. The first claim follows by Lemma 4.3 and Theorem 4.1. We settle the second claim. Let G be a bipartite CM t graph with a pure order of vertices. Let Kn1 ,n1 , · · · , Knd ,nd be the maximal bipartite subgraphs of G, where ni ≥ 1 for all i. Observe that, by maximality, these complete subgraphs of G are disjoint. Choose one edge xi1 yi1 from each subgraph Kni ,ni for all i = 1, · · · , d. Let H be the induced subgraph of G on the vertex set ({x11 , · · · , xd1 }, {y11 , · · · , yd1 }). By Lemma 2.4, H is independent of the choice of particular edge xi1 yi1 from Kni ,ni and hence H is unique. Since the ordering of vertices of G is a pure order, its restriction to H is also pure. Thus, H is an unmixed bipartite graph. But by the maximality of the complete bipartite subgraphs Kni ,ni , and the construction of H, it is cross-free. Therefore, by Proposition 2.6, H is Cohen-Macaulay. Now any edge xi1 yi1 replace in H with Kni ,ni for all i = 1, · · · , d, preserving all other adjacencies. Let H ′ be the resulting graph. Then by the construction, G = H ′ , as required.  Remark 4.5. Let H be a bipartite Cohen-Macaulay graph and let G = H ′ be a bipartite CM t graph obtained from H by the replacing process described above. Assume that G is is not CM t−1 and t ≥ 2. Using the the results of this section, the following observations are immediate. First of all, 1 ≤ dimH ≤ t − 1. Because if dimH ≥ t and we replace just one Kn,n with n ≥ 2, then G is strictly CM r with r ≥ t + 1. On the other hand, if dimH = 0, then G is CM 1 . If dimH = t − 1, then only one Kn,n with n ≥ 2 can be replaced. Because replacing at least two Kn,n with n ≥ 2, G is strictly CM r with r ≥ t + 1. If dimH = t − 1, for replacing just one Kn,n , n is arbitrary and hence G is of dimension n + t − 2. If dimH ≤ t − 2, the number of replacements should be at least 2. Again because if with one replacement of Kn,n , n ≥ 2, G would be CM r with r ≤ t − 1. When dimH ≤ t − 2, the maximum number of replacements of Kn,n , n ≥ 2, is at most t − dimH which may occur replacing K2,2 ’s. COHEN-MACAULAY-NESS IN CODIMENSION FOR BIPARTITE GRAPHS 7 For dimH ≤ t − 2, the maximum size of Kn,n to be replaced is also n = t − dimH which may occur when we have two replacements. Using these remarks we may easily distinguish all bipartite CM t graphs for t = 2, 3, 4. Example 4.6. Bipartite CM 2 graphs which are not Buchsbaum. Using the notation of Remark 4.5 we have dimH = 1. There are just two non-isomorphic bipartite Cohen-Macaulay graphs of dimension one. By replacing process, they produce two types of bipartite CM 2 graphs which are not Buchsbaum. They are of arbitrary dimensions. More precisely, one such graph is the disjoint union of an edge x1 y1 with Kn2 ,n2 = {x21 , · · · , x2n2 } × {y21 , · · · , y2n2 }, n2 ≥ 2, and the other one consists of the first graph together with the edges x1 y2i for all i = 1, · · · , n2 . The second graph with n2 = 3 could be depicted in Figure 1. xt1 xt21 xt22 xt23 PP ❍ ✟ ❍ ❅❍PP ❅❍ ❅ ✟ ❅❍❍PPP ❅✟ ❅❍❍ ✟ ❍ ✟ ❍ ❅ ❅P✟ ❍ P P❍❍❅ ❍✟ ❅ ❍❅ ❅ PP ❍✟ P ❍P P ❍ ❅✟ ❅t ❅t t t ✟ ❍❍ y1 y21 y22 y23 F igure 1 Example 4.7. Bipartite CM 3 graphs which are not CM 2 . For these graphs dimH = 1, 2. If dimH = 1, by Example 4.6, there are just two bipartite CM 3 graphs by replacing two edges of a perfect matching by K2,2 ’s. In this case, dimG = 3. (see Figure 2, and Figure 3). If dimH = 2, then there are 4 non-isomorphic bipartite Cohen-Macaulay graphs of dimension 2. By replacing one perfect matching edge with Kn,n of arbitrary size in each Cohen-Macaulay graph, they produce 7 types of bipartite CM 3 graphs which are not CM 2 . Note that depending on the choice of the edge to be replaced in each case, we may get non-isomorphic bipartite graphs. In this case dimG = n + 1. 8 HASSAN HAGHIGHI, SIAMAK YASSEMI, AND RAHIM ZAARE NAHANDI xt11 ❅ ❅ ❅ t y11 xt12 ❅ ❅t y12 xt21 xt22 ❅ ❅ ❅ ❅ t ❅t y21 y22 F igure 2 x11 x12 x21 x22 t t t t P ❍ ❍ PP ❅❍ ❍ ❅ ❅ P ❍ ❍ P ❅ ❍❍ PP ❅ ❍❍ ❅ ❅ ❍❍P ❅PP❍❍❅ ❅ ❅ ❍❍ ❅ PP❍ P❍ P❍ ❅t ❅ Pt ❍ t ❅t y11 y12 y21 y22 F igure 3 Example 4.8. Bipartite CM 4 graphs which are not CM 3 . For these graphs dimH = 1, 2, 3. If dimH = 1, there are two bipartite CM 4 graphs obtained by replacing two edges of a perfect matching by K3,3 ’s. In this case, dimG = 5. And, similarly, there are two others obtained by replacing one edge with K2,2 and another edge with K3,3 . In this case, dimG = 5. If dimH = 2, then while there are 4 non-isomorphic bipartite Cohen-Macaulay graphs of dimension 2, by replacing two perfect matching edges with K2,2 ’s in each Cohen-Macaulay graph, they produce 7 bipartite CM 4 graphs which are not CM 3 . They all have dimension 4. If dimH = 3, then there are 10 non-isomorphic bipartite Cohen-Macaulay graphs of dimension 3. Replacing one perfect matching edge with Kn,n , n ≥ 2, in each Cohen-Macaulay graph, they produce 25 bipartite CM 4 graphs which are not CM 3 . They all have dimension n + 2. Out of all 36 bipartite CM 4 graphs, 21 graphs are connected. Acknowledgments The authors would like to thank J. Herzog and U. Nagel for some fruitful discussions on the subject of this paper. This work has been supported by Center for International Studies & Collaboration (CISSC) and French Embassy in Tehran in COHEN-MACAULAY-NESS IN CODIMENSION FOR BIPARTITE GRAPHS 9 the framework of the Gundishapur project 27462PL on the Homological and Combinatorial Aspects of Commutative Algebra and Algebraic Geometry. The research of R. Zaare-Nahandi has been partially supported by research grant no. of University of Tehran. The research of S. Yassemi was in part supported by a grant from IPM (No. 91130214). References 1. D. Cook and U. Nagel, Cohen-Macaulay graphs and face vectors of flag complexes, SIAM J. Discrete Math. 26 (2012), no. 1, 89–101. 2. R. Fröberg, A note on the Stanley-Reisner ring of a join and of a suspension, Manuscripta Math. 60 (1988), no. 1, 89–91. 3. H. Haghighi, S. Yassemi and R. Zaare-Nahandi, Bipartite S2 graphs are Cohen-Macaulay, Bull. Math. Soc. Sci. Roumanie. 53 (2010), 125–132. 4. H. Haghighi, S. Yassemi and R. Zaare-Nahandi, A generalization of k-Cohen-Macaulay simplicial complexes, Ark. Mat. 50 (2012), 279–290. 5. J. Herzog and K. Hibi, Distributive lattices, bipartite graphs and Alexander duality, J. Algebraic Combin. 22 (2005), 289–302. 6. E. Miller, I. Novik and E. Swartz, Face rings of simplicial complexes with singularities, Math. Ann. (2011), 351:857-875. 7. H. Sabzrou, M. Tousi and S. Yassemi, Simplicial join via tensor products, Manuscripta Math. 126 (2008), 255–272. 8. C. Sava, On the Stanley-Reisner ring of a join, An. Sti. Univ. ”Al. l. Cuza”, lasi, Tom XXXI, Ser. Ia, (1985), 145–148. 9. P. Schenzel, On the number of faces of simplicial complexes and the purity of Frobenius, Math. Z. 178 (1981), 125–142. 10. R. Stanley, Combinatorics and Commutative Algebra, Second Edition, Birkhäuser, Boston, 1995. 11. R. H. Villarreal, Unmixed bipartite graphs, Rev. Colombiana Mat. 41 (2007), 393–395. 12. R. Zaare-Nahandi and R. Zaare-Nahandi, The minimal free resolution of a class of square-free monomial ideals, J. Pure Appl. Algebra 189 (2004) 263–278. Hassan Haghighi, Department of Mathematics, K. N. Toosi University of Technology, Tehran, Iran. Siamak Yassemi, School of Mathematics, Statistics & Computer Science, University of Tehran, and School of Mathematics, Institute for Research in Fundamental Sciences (IPM), Tehran, Iran. Rahim Zaare Nahandi, School of Mathematics, Statistics & Computer Science, University of Tehran, Tehran, Iran.
0math.AC
Variable-mixing parameter quantized kernel robust mixed-norm algorithms for combating impulsive interference✩ Lu Lua , Haiquan Zhaoa∗ , Badong Chenb a)School of Electrical Engineering, Southwest Jiaotong University, Chengdu, China b)School of Electronic and Information Engineering, Xi’an Jiaotong University, Xi’an,China. arXiv:1508.05232v5 [cs.SY] 9 Feb 2018 Abstract Although the kernel robust mixed-norm (KRMN) algorithm outperforms the kernel least mean square (KLMS) algorithm in impulsive noise, it still has two major problems as follows: (1) The choice of the mixing parameter in the KRMN is crucial to obtain satisfactory performance. (2) The structure of the KRMN algorithm grows linearly as the iteration goes on, thus it has high computational complexity and memory requirements. To solve the parameter selection problem, two variable-mixing parameter KRMN (VPKRMN) algorithms are developed in this paper. Moreover, a sparsification algorithm, quantized VPKRMN (QVPKRMN) algorithm is introduced for nonlinear system identification with impulsive interferences. The energy conservation relation (ECR) and convergence property of the QVPKRMN algorithm are analyzed. Simulation results in the context of nonlinear system identification under impulsive interference demonstrate the superior performance of the proposed VPKRMN and QVPKRMN algorithms as compared with the existing algorithms. Keywords: Kernel method, Impulsive noise, Variable mixing parameter, Quantization scheme, Convergence analysis 1. Introduction Kernel method has received increasing attention in machine learning and adaptive signal processing. The main idea of the kernel method is to transform the input data into a high-dimensional feature space via a reproducing kernel Hilbert space (RKHS). Some successful applications motivate to improve the robustness of the nonlinear adaptive filter (e.g., support vector machine (SVM) [1, 2] and the kernel principal component analysis (KPCA) [3, 4]). Recently, the kernel adaptive filters became popular due to their modeling capabilities in the feature space. By using Mercer kernels [5], many linear filters have been recast in high-dimensional reproducing kernel Hilbert spaces (RKHSs) to yield more powerful nonlinear extensions, such as the kernel recursive least squares (KRLS) algorithm [6, 7, 8, 9], the kernel least mean square algorithm (KLMS) [10, 11, 12] and the kernel affine projection algorithm (KAPA) [13, 14]. These algorithms have been successfully applied in nonlinear active ✩ * Corresponding author at: School of Electrical Engineering, Southwest Jiaotong University, Chengdu, Sichuan, 610031, PR China. E-mail addresses: [email protected](L. Lu), hqzhao [email protected](H. Zhao), [email protected](B. Chen). Preprint submitted to Elsevier February 12, 2018 noise cancellation [5, 10] and nonlinear acoustic echo cancellation (NLAEC) [15]. Although the above-mentioned kernel adaptive filters achieve improved performance, they are not suitable for online applications, as their structures grow linearly with the number of processed patterns. In the past years, some sparsification techniques that constrain the growth of the network size were proposed [5, 6, 16, 17, 18, 19, 20]. In 2012, the quantized KLMS (QKLMS) algorithm has been successfully applied to static function estimation and time series prediction [17]. It has a mechanism to utilize the redundant input data, which is helpful to achieve a better accuracy and a more compact network with fewer centers. On the other hand, in most signal processing applications, impulsive noise exists widely. It is well known that the impulsive noises have infinite variance, which makes the traditional l2 -norm algorithms diverge. Thus, a family of norm stochastic gradient adaptive filter algorithms was proposed, such as least mean absolute third (LMAT) algorithm [21, 22], least-mean-fourth (LMF) algorithm [23, 24], and least-mean mixed-norm (LMMN) algorithm [25, 26]. In [27], a robust mixed-norm (RMN) algorithm was developed based on a convex function of the error norms that underlie the least mean square (LMS) and least absolute difference (LAD) algorithms. Therefore, the RMN algorithm has robustness performance in the presence of impulsive noise. To achieve improved performance in impulsive noise, several variants of the kernel adaptive filter were proposed [28, 29, 30, 31, 32]. Particularly, in [33], the KRMN algorithm was proposed by deriving the RMN algorithm in RKHS. Unfortunately, the unsuitable selection of mixing parameter degrades the performance of KRMN algorithm. To overcome this problem, in this paper, we proposed two adaptation rules for the KRMN algorithm, called variable mixing parameter KRMN (VPKRMN). Based on the VPKRMN algorithms, we further proposed a quantized VPKRMN (QVPKRMN) algorithm to curb the growth of the networks. Furthermore, the energy conservation relation (ECR) and convergence property of QVPKRMN algorithm are analyzed. This paper is organized in the following manner. Section 2 introduces a brief description of the kernel method and the KRMN algorithm. In Section 3, two novel VPKRMN algorithms are proposed to adapt the mixing parameter. And, the QVPKRMN algorithm is proposed to control the growth of the kernel structure. In Section 4, the analysis of the convergence property is performed. Then, simulations in the context of nonlinear system identification are conducted in Section 5. Finally, conclusions are found in Section 6. 2. Kernel method and KRMN algorithm 2.1. Kernel method The kernel method is a useful nonparametric modeling tools to deal with the nonlinearity problem. The power of this idea is to transform input data (input space U) into a high-dimensional feature space F using a certain nonlinear mapping, which can be expressed as: ϕ: U →F 2 (1) where ϕ is the feature vector in the kernel method. Based on Mercer theorem, a shift-invariant Mercer kernel can be expressed as [5]: κ(u, u′ ) = ϕ(u)ϕT (u′ ) = ∞ X φi ϕi (u)ϕTi (u′ ) (2) i=1 where φ is the nonnegative eigenvalue, and ϕ is the corresponding eigenfunction. The eigenvalues and eigenfunctions constitute the feature vector ϕ: ϕ(u) = hp iT p φ1 ϕ1 (u), φ2 ϕ2 (u), ... . (3) It is well known that a Mercer kernel is a continuous, symmetric and positive-definite kernel. The commonly used Gaussian kernel can be expressed as   2 κ(u, u′) = exp −hku − u′ k (4) where h is the kernel bandwidth. By using (2) and (4), the feature space can be calculated by inner product. Consequently, the output of adaptive filter can be expressed by inner product of the transformed test data ϕ(u) and training data ϕ(uj ) f (u) = n X aj yj hϕ(u),ϕ(uj )i (5) j=1 where aj is the coefficient at discrete time n, and h·i is the inner product operation, respectively. 2.2. KRMN algorithm When the desired or the input signal is corrupted by impulsive noise, the performance of the KLMS algorithm degrades. To overcome this problem, the KRMN algorithm was proposed by using the kernel method [33]. The input data of RMN algorithm are transformed into RKHS as ϕ(n), and the weight vector in feature space is defined as Ω(n), Ω(1) = 0. The error signal is defined as: e(n) , d(n) − ΩT (n)ϕ(n). (6) The KRMN algorithm is based on minimization of the following mixed-norm error [27, 33]:  J(n) = λE e2 (n) + (1 − λ) E {|e(n)|} (7) where E {·} is the statistical expectation operator, λ is limited in the range of (0, 1). From (7), the cost function of KRMN algorithm is a linear combination of the KLMS and KLAD 1 algorithms. That is, the combination of l2 norm and l1 norm. The gradient vector of J(n) with respect to Ω(n) is ∇Ω(n) J(n) = − [2λe(n) + (1 − λ)sign {e(n)}] ϕ(n) (8) where sign{x} denotes the sign function, i.e., if x > 0, then sign{x} = 1, if x = 0, sign{x} returns to 0, otherwise sign{x} = −1. Hence, the adaptive rule of KRMN is solved iteratively on the new example sequence {ϕ(n), d(n)} Ω(n + 1) = Ω(n) + µ [2λe(n) + (1 − λ)sign {e(n)}] ϕ(n). 1 The (9) KLMAD algorithm can be easily derived by casting the LAD algorithm into RKHS. When the mixing parameter of VPKRMN is equal to zero, the VPKRMN becomes KLAD. 3 Reusing (9), we have Ω(n + 1) = µ n+1 X [2λe(j) + (1 − λ)sign {e(j)}] ϕ(j). (10) j=1 By using the Mercer kernel in (2), the filter output can be calculated through kernel evaluations y(n + 1) = µ n X [2λe(j) + (1 − λ)sign {e(j)}] κ(j, n + 1). (11) j=1 For the sake of simplicity, we define aj (n + 1) , µ [2λe(j) + (1 − λ)sign {e(j)}] , j = 1, ..., n + 1 (12) and codebook C(n) refer as a center set in time n C(n + 1) = [C(n), u(n + 1)] . (13) It can be observed that if the kernel function is replaced by a radial kernel, the KRMN algorithm produces a growing radial basis function (RBF) network by allocating a new kernel unit for every new example with input u(n + 1). The main bottleneck of the KRMN algorithm is its network size grows with the number of processed data. To overcome this severe drawback, a quantization scheme should be used to curb the growth of network. 3. Proposed algorithms 3.1. VPKRMN algorithm An unsuitable mixing parameter selection will lead to a performance degradation of the KRMN algorithm. To circumvent this problem, the mixing parameter λ should be automatically adjusted. Here, we use λ(n) instead of λ to derive the variable mixing parameter algorithm. Considering λ(n) to minimize the mixed-norm error of the KRMN algorithm at each iteration cycle, we obtain ∂J(n) ∂λ(n)   ∂ λ(n)E e2 (n) + [1 − λ(n)] E {|e(n)|} = ∂λ(n)  2 ∂ λ(n)e (n) + [1 − λ(n)] |e(n)| ≈ ∂λ(n) λ(n + 1) = (14) where λ(n) is restricted in [0, 1]. Then, we add a scaling factor γ to (14) to control the steepness of J(n). As a result, an adaptive update rules for KRMN algorithm is obtained, and we name the new algorithm the VPKRMN-Algorithm 1:   VPKRMN − Algorithm 1 : λ(n + 1) = λ(n) + γ E {|e(n)|} − E e2 (n)  ≈ λ(n) + γ |e(n)| − e2 (n) . (15) From (15), the mixing parameter is adjusted by switching the two types of error norm. When |e(n)| > e2 (n), the mixing parameter tends to one, the KLMS algorithm plays a dominate role of the filter. When |e(n)| < e2 (n), the mixing parameter tends to zero, the KLAD algorithm plays a dominate role of the filter. 4 1 0.9 λ=0.2 0.8 λ=0.4 0.7 J(n) 0.6 0.5 λ=0.6 λ=0 λ=0.8 0.4 0.3 0.2 λ=1 0.1 0 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 e(n) Figure 1: The cost functions with different mixing parameter settings. The cost function of the VPKRMN algorithm is a unimodal function (See Fig. 1). The unimodal character is preserved for λ(n) chosen in (0, 1), that is, the second term of (15) keeps a very small value. Hence, the adaptation of VPKRMN Algorithm 1 is very sensitive to the choice of γ. To avoid this limitation, a new adaptive update approach, called VPKRMN-Algorithm 2, is proposed for adapt the mixing parameter of the KRMN algorithm.   λ(n + 1) = δλ(n) + θp2 (n) VPKRMN − Algorithm 2 :  p(n) = βp(n − 1) + (1 − β)e(n)e(n − 1) (16) where δ and β are the exponential weighting parameters in the range of [0, 1], which control the quality of estimation of the algorithm, θ > 0 is a positive constant, and p(n) is a low-pass filtered estimation of e(n)e(n − 1). Note that the mixing parameter has a fixed value when δ = 1 and θ = 0. There are two reasons that account for the use of p(n) in the update of λ(n): (1) the error autocorrelation e(n)e(n − 1) is generally a good measure of the proximity to the optimum [34]; (2) the environment is divided into two cases by error autocorrelation e(n)e(n − 1): no impulsive environment and impulsive environment. The objective of VPKRMN-Algorithm 2 is to ensure the large λ(n) when the VPKRMN Algorithm 2 is far from the optimum with decreasing λ(n). The large value of λ(n) leads to the l2 -norm error plays critical role, which provides a more accurate final solution and less misadjustment under no impulsive noise environment. Conversely, when algorithm suffers severely from the outlier problems, λ(n) is small and the l1 -norm error offers a stable convergence characteristic to KRMN algorithm. 3.2. QVPKRMN algorithm The QVPKRMN algorithm incorporates the idea of quantization into the VPKRMN algorithm to provide an efficient learning performance under impulse interference. In general, the quantization scheme is similar to the sparsification with NC method [6]. In fact, they almost have the same computational complexity. The main difference between the quantization scheme and NC method is the quantization scheme utilizes the redundant data to locally update the coefficient of the closest center. The quantization method can be summarized as a learning strategy: the input space is quantized, if the current quantized input has already been assigned a center, no new center will be added, but the coefficient of that center will be updated through merging a new coefficient [17]. 5 The feature vector ϕ(n) in quantization scheme can be expressed as:    Ω(0) = 0   e(n) = d(n) − ΩT (n − 1)ϕ(n)     Ω(n) = Ω(n − 1) + µ [2λ(n)e(n) + (1 − λ(n)) sign {e(n)}] Q [ϕ(n)] (17) where Q [·] is a quantization operator in feature space F . Owing to the high dimensionality of feature space, the quantization scheme is usually used in input space U. Therefore, the learning rule of QVPKRMN algorithm in U can be given as:    f =0   0 e(n) = d(n) − fn−1 (u(n))     f =f n n−1 + µ [2λ(n)e(n) + (1 − λ(n)) sign {e(n)}] κ (Q[ϕ(n)], ·) (18) where Q [·] is a quantization operation in input space U. Throughout this paper, the notation ϕq (n) is replaced by the notation Q [ϕ(n)], uq (n) = Q [u(n)], Cj (n − 1) is the jth element of C(n − 1), k·k is the Euclidean norm in feature space F , and εU is the threshold of the distance. For εU = 0, the QVPKRMN algorithm reduces to the VPKRMN algorithm. The proposed QVPKRMN algorithm is summarized in Table 1. Table 2 summarizes the computational complexity of the algorithms, where N is the training times, M is the length of the filter, L(L < N ) is elements of index set. With an affordable computation complexity, the VPKRMN algorithm behaves much better than the KLMS and KLAD algorithms under the impulse noise environment. Since the QKLMS and QVPKRMN algorithms are developed by using quantization scheme, these algorithms have lower computation complexity than those of the KLMS, KLAD, KRMN and VPKRMN algorithms. 4. Convergence analysis In this section, we establish the energy conservation relation (ECR) [17, 35] for the QVPKRMN algorithm and analyze its mean convergence behavior. The convergence property of QVPKRMN algorithm is difficult to analyze exactly, so the theorem in [36, 37] and the independence assumption [38] are introduced throughout the analyses. 4.1. Energy conservation relation Consider the adaptation of QVPKRMN algorithm in RKHS Ω(n) = Ω(n − 1) + µ [2λ(n)e(n) + (1 − λ(n)) sign {e(n)}] ϕq (n). (19) We define the weight deviation vector V (n) and the second moment of the misalignment vector η(n) of the QVPKRMN algorithm as: V (n) , Ω(n) − Ωopt  η(n) , E V (n)V T (n) . 6 (20) Table 1: Proposed QVPKRMN algorithms. Proposed QVPKRMN algorithms Initialization: choose step size µ, bandwidth parameters of kernel h a1 (1) = µd2 (1)sign {d(1)}, C(1) = [u(1)] f1 = a1 (1)κ(u(1), ·), λ(1) = 0.5 Computation: While u(n), d(n) n > 1 is available do (1) Compute the output of the adaptive filter: y(n) = size(C(n−1)) P aj (n)κ(u(n + 1), u(j)) j=1 (2) Compute the error: e(n) = y(n)d(n) (3) Compute the distance between u(n) and C(n − 1) dis(u(n), C(n − 1)) = min 1≤j≤size(C(n−1)) ku(n) − Cj (n − 1)k (4) If dis (u(n), C(n − 1)) ≤ εU , keep the codebook unchanged: C(n + 1) = C(n) and quantize u(n) to the closest center through updating the coefficient of that center: aj ∗ (n) = aj ∗ (n + 1) + µ [2λ(n)e(n) + [1 − λ(n)] sign {e(n)}] where j ∗ = min 1≤j≤size(C(n−1)) ku(n) − Cj (n − 1)k otherwise, assign a new center and corresponding new coefficient: C(n + 1) = [C(n), u(n + 1)] a(n) = [a(n − 1), µ [2λ(n)e(n) + (1 − λ(n))sign{e(n)}]] Then, using two new update rule of mixing parameter     Algorithm 1 : λ(n + 1) = λ(n) + γ |e(n)| − e2 (n)     λ(n + 1) = δλ(n) + θp2 (n)  Algorithm 2 :     p(n) = βp(n − 1) + (1 − β)e(n)e(n − 1) end while Table 2: Summary of the computational complexity. Algorithms KLMS Computation(training) Memory(training) Computation(test) Memory(test) 2 O(N ) O(N ) O(N ) 2 O(N ) KLAD O(N ) O(N ) O(N ) O(N ) KRMN O(N 2 ) QKLMS VPKRMN QVPKRMN O(N ) O(N ) O(N ) 2 O(L) O(L) O(L) 2 O(N ) O(N ) O(N ) 2 O(L) O(L) O(L) O(L ) O(N ) O(L ) 7 where Ωopt is the optimal weight vector. From (19) and (20), the update formulation of the weight deviation vector of QVPKRMN algorithm can be expressed as: V (n + 1) = V (n) − µ [2λ(n)e(n) + [1 − λ(n)] sign {e(n)}] ϕq (n). ∆ (21) ∆ Then, we define the a posterior error ep (n) = V T (n)ϕ (u(n)) and a priori error ea (n) = V T (n−1)ϕ (u(n)). It can be shown that their a priori and a posteriori errors are related via ep (n) = ea (n) − µ [2λ(n)e(n) + [1 − λ(n)] sign {e(n)}] ϕq (n) (22) = ea (n) − µ [2λ(n)e(n) + [1 − λ(n)] sign {e(n)}] κ (uq (n), u(n)) . Combining (21) and (22) yields V (n) = V (n − 1) + [ep (n) − ea (n)] ϕq (n) . κ (uq (n), u(n)) (23) Squaring both sides of (23), we get T  ϕq (n) V T (n)V (n) = V (n − 1) + [ep (n) − ea (n)] κ (uq (n), u(n))   ϕq (n) . · V (n − 1) + [ep (n) − ea (n)] κ (uq (n), u(n)) (24) Rearranging (24), we have 2 kV (n)kF + e2a (n) 2 2 = kV (n − 1)kF [κ (uq (n), u(n))] e2p (n) + 2 + βq [κ (uq (n), u(n))] where k·kF is the norm in feature space F , and βq = (25) 2[ep (n)−ea (n)]{V (n−1)ϕq (n)κ(uq (n),u(n))−ea (n)} [κ(uq (n),u(n))]2 As can be seen, (25) of QVPKRMN algorithm is the same form as the QKLMS algorithm. When the quantization size goes to zero, βq → 0, the ECR expression for QKLMS algorithm is obtained 2 2 kV (n)kF + e2a (n) = kV (n − 1)kF + e2p (n). (26) 4.2. Mean convergence In this subsection, the mean convergence analysis of weight vector is performed. Taking the mathematical expectation of (21) and using independence assumption [38], we obtain E {V (n + 1)} = E {V (n)} − µE {[2λ(n)e(n) + [1 − λ(n)]sign {e(n)}] ϕq (n)} (27) = E {V (n)} − {2µλ(n)E [e(n)ϕq (n)] + µ [1 − λ(n)] E [sign {e(n)} ϕq (n)]} . According to [36, 37], the second term of the right hand side in (27) can be expressed as: r 2 1 E [e(n)ϕq (n)] . E [sign {e(n)} ϕq (n)] ≈ π σe Substituting (28) into (27), we arrive ( E [V (n + 1)] ≈ E [V (n)] − 2µλ(n)E [e(n)ϕq (n)] + µ [1 − λ(n)] " ≈ E [V (n)] 1 − 2µλ(n) + µ [1 − λ(n)] 8 r r 2 1 E [e(n)ϕq (n)] π σe #   2 1 E ϕTq (n)ϕq (n) . π σe (28) ) (29) where e(n) ≈ ϕTq (n)V (n). It is easily observed that V (n) will converge to zero vector as n → ∞ if and only if the step size satisfies the following inequality ( 0< 2µλ(n) + µ [1 − λ(n)] r 2 1 π σe ) Hence, we obtain   E ϕTq (n)ϕq (n) < 2. 2 0<µ< 2λ(n) + [1 − λ(n)] . q 2 1 π σe Rϕϕ q 2 1 π σe λmax (30) (31)   where Rϕϕ = E ϕTq (n)ϕq (n) . It is easy to see that the mean convergence condition of the QVPKRMN algorithm is 2 0<µ< 2λ(n) + [1 − λ(n)] (32) where λmax is the maximum eigenvalues of Rϕϕ . Since λmax < tr (Rϕϕ ), where tr (Rϕϕ ) denotes the trace of the autocorrelation matrix Rϕϕ , a more rigorous condition can be gained 0<µ< 2 q 2λ(n) + [1 − λ(n)] π2 √ζ1 min where ζmin (33) tr (Rϕϕ )  = E d2 (n) − RTϕd Ωopt , and Rϕd is the cross-correlation vector of ϕq (n) and d(n). The optimal weight vector can be expressed as: Ωopt = R−1 ϕϕ Rϕd . (34) From formula (19), we get o n T E (Ωopt + V (n + 1)) (Ωopt + V (n + 1)) o n T + µ2 Rϕϕ = E (Ωopt + V (n)) (Ωopt + V (n))  + µE (Ωopt + V (n)) ϕTq (n)K(n) n o T + µE ϕq (n)(Ωopt + V (n)) K(n) (35) where K(n) = 2λ(n)e(n) + [1 − λ(n)] sign {e(n)}. Thus, (35) can be expressed with the form of the second moment of the misalignment vector  η(n + 1) = η(n) + µ2 Rϕϕ + µE V (n)ϕTq (n)K(n)  + µE ϕq (n)V T (n)K(n) . (36) Introducing (35) to (36) and using the independence assumption [38], (36) can be given as: η(n + 1) = η(n) + µ2 Rϕϕ  + µE V (n)ϕTq (n) [2λ(n)e(n) + [1 − λ(n)] sign {e(n)}]  + µE ϕq (n)V T (n) [2λ(n)e(n) + [1 − λ(n)] sign {e(n)}]  = η(n) + µ2 Rϕϕ + 2λ(n)µE V (n)ϕTq (n)e(n)  + [1 − λ(n)] µE V (n)ϕTq (n)sign {e(n)}  + 2λ(n)µE ϕq (n)V T (n)e(n)  + [1 − λ(n)] µE ϕq (n)V T (n)sign {e(n)} . 9 (37) Using the theorem in [36, 37], the fourth term of equation (37) can be respectively simplified as follows:  E V (n)ϕTq (n)sign {e(n)}   = E E V (n)ϕTq (n)sign {e(n)} |V (n) ( ) r  T 2 1 = E V (n) E ϕq (n)e(n)|V (n) π σe|Ω(n) ) ( r i h 1 2 T T Rϕd − [Ωopt + V (n)] Rϕϕ = E V (n) π σe|Ω(n) ( ) r r 2 2 1 1 T = − E V (n)V (n)Rϕϕ =− η(n)Rϕϕ . π σe|Ω(n) π σe Similarity, the simplified form of sixth term of (37) can be obtained r  2 1 T E ϕq (n)V (n)sign {e(n)} = − Rϕϕ η(n). π σe (38) (39) To calculate the third term and the fifth term of (37), we have   E V T (n)ϕq (n)e(n) = E ϕq (n)V T (n)e(n)  ∆ ≈ E e2 (n) = σe2 . (40) Substituting (38), (39) and (40) in (37) will yield 2 η(n + 1) = η(n) + µ Rϕϕ + 4λ(n)µσe2 " r + µ [1 − λ(n)] − 2 1 η(n)Rϕϕ π σe " r # # 2 1 + µ [1 − λ(n)] − Rϕϕ η(n) π σe ( ( ) ) r r 2 1 2 1 2 = η(n) I − µ [1 − λ(n)] Rϕϕ + Rϕϕ µ I − µ [1 − λ(n)] η(n) π σe π σe (41) + 4µλ(n)σe2 I. Furthermore, (41) can be decomposed into a scalar form. The matrix M is defined as an orthonormal matrix of the autocorrelation matrix Rϕϕ . Pre- and Post-multiplying both side of (41) by M and MT , given ( ) r 2 1 ξ(n + 1) = ξ(n) I − µ [1 − λ(n)] Λ π σe ) ( (42) r 2 1 2 2 ξ(n) + 4µλ(n)σe I. + Λ µ I − µ [1 − λ(n)] π σe where ξ(n) is a symmetric matrix, ξ(n) = MT (n)η(n)M(n), Λ = MT (n)Rϕϕ M(n), and Λ is a diagonal matrix and its elements λi (i − 1, 2, ..., M ) are eigenvalues of matrix Rϕϕ . A scalar form of (41) can be obtained as: ξij (n + 1) = ( 1 − µ [1 − λ(n)] 2 + µ λi τ (i − j) + r ) 2 1 [λi + λj ] ξij (n) π σe 4µλ(n)σe2   1, where ξij (n) is the (i, j)th element of ξ(n), and τ (i − j) =  0, 10 if i = j otherwise . (43) 5. Simulation results                                            Figure 2: Block diagram of the kernel adaptive identification. To demonstrate the effectiveness of the proposed algorithms in impulsive noise environments, simulation studies are carried out for nonlinear system identification problem. In the following simulations, the software of Matlab R2013a is used to program the experiments under the computer environment of AMD (R) A10 CPU 2.1 GHz. The block diagram of the kernel adaptive system identification is plotted in Fig. 2. The goal of nonlinear system identification is to employ pairs of {u(n), d(n)} inputs and addictive noise v(n) to fit a function that maps an arbitrary system input into an appropriate output. The model coefficients at n moment an (n) are adjusted by the error signal e(n). The nonlinear system contains a linear filter and a memoryless nonlinearity. The linear system impulse response is generated by [28] H(z) = 0.1 + 0.2z −1 + 0.3z −2 + 0.4z −3 + 0.5z −4 + 0.4z −5 + 0.3z −6 + 0.2z −7 + 0.1z −8 (44) and the nonlinearity is given as d(n) = r(n) − 0.9r2 (n) + v(n). 0 10 Testing MSE VPKRMN−Algorithm 1 VPKRMN−Algorithm 1 VPKRMN−Algorithm 2 VPKRMN−Algorithm 2 −1 10 −2 10 0 5000 10000 15000 n Figure 3: The effect of the parameters on VPKRMN algorithms with c = 0.2, σI = σG = 0.02. (Keys: blue line, µ = 0.1, γ = 0.00001; red line, µ = 0.1, γ = 0.00005; black line, µ = 0.1, θ = 0.01, δ = 0.97, β = 0.98; azury line, µ = 0.1, θ = 0.05, δ = 0.97, β = 0.98.) 11 KLMS KLAD KRMN VPKRMN−Algorithm 1 VPKRMN−Algorithm 2 0 Testing MSE 10 −1 10 −2 10 0 5000 10000 15000 n Figure 4: Learning curves of KLMS, KLAD KRMN and VPKRMN algorithms for nonlinear system identification with c = 0.2, σI = σG = 0.02. (Keys: KLMS, µ = 0.1, h = 0.05; KLAD, µ = 0.05, h = 0.01; KRMN, µ = 0.1, h = 0.1, λ = 0.3; VPKRMN-Algorithm 1, µ = 0.1, h = 0.05; VPKRMN-Algorithm 2, µ = 0.1, h = 0.1.) 0 10 Testing MSE QKLMS VPKRMN−Algorithm 1 VPKRMN−Algorithm 2 QVPKRMN−Algorithm 1 QVPKRMN−Algorithm 2 −1 10 −2 10 0 5000 10000 15000 n Figure 5: Learning curves of QLMS, VPKRMN and QVPKRMN algorithms for nonlinear system identification with c = 0.2, σI = σG = 0.02. (Keys: QKLMS, εU = 1, h = 0.05; QVPKRMN-Algorithm 1, εU = 0.5; QVPKRMN-Algorithm 2, εU = 0.5.) 5.1. Test under impulsive noise environment with BG model In this example, the impulsive noise is modeled by Bernoulli-Gaussian (BG) distribution [36] with probability function c and the root deviation σI . The white Gaussian noise (WGN) with zero mean and variance σu2 = 1 is used as the input signal. The White Gaussian noise is a zero mean with root deviation σG . A segment of 15000 samples are used as the training data and another 1000 samples as the test data. Simulation results are obtained by 50 Monte Carlo trials. Firstly, the effect of the parameter on proposed VPKRMN algorithms are studied. Fig. 3 plots the effect of the update parameter for algorithm. It can be seen from this figure that the VPKRMN algorithm 1 achieves the fast convergence rate under γ = 0.00005 as compared to γ = 0.00001 of VPKRMN. And, the θ = 0.01 of KRMN algorithm 2 obtains the faster convergence speed than that of θ = 0.05. For this reason, the γ = 0.00005 and θ = 0.01 are selected for proposed VPKRMN algorithm 1 and 2, respectively. Figs. 4 and 5illustrate the learning curves of existing algorithms. All the bandwidth parameters of kernelbased algorithms are set to 0.1. It observed from Fig. 4 that the proposed algorithms outperform the other algorithms in terms of convergence rate and steady-state error under the impulsive noise. The KRMN with 12 2500 Network size 2000 1500 1000 500 QVPKRMN−Algorithm 1 QVPKRMN−Algorithm 2 0 0 5000 10000 15000 n Figure 6: Network size growth of the QVPKRMN algorithms. VPKRMN−Algorithm 1 VPKRMN−Algorithm 1 VPKRMN−Algorithm 2 VPKRMN−Algorithm 2 0 Testing MSE 10 −1 10 −2 10 0 5000 10000 15000 n Figure 7: The effect of the parameters for KRMN algorithms with α-stable noise (α=1.4, SNR=15dB). (Keys: blue line, µ = 0.1, γ = 0.00001; black line, µ = 0.1, γ = 0.00003; red line, µ = 0.1, θ = 0.01, δ = 0.97, β = 0.98; azury line, µ = 0.1, θ = 0.05, δ = 0.97, β = 0.98.) λ = 0.3 has fast convergence rate against impulse noise, but high MSE as compared to VPKRMN algorithms. Moreover, the VPKRMN-Algorithm2 achieves better performance than the VPKRMN-Algorithm1. Fig. 5 shows the performance of the proposed two algorithms based on quantization scheme, and the network size growth curves of QVPKRMN algorithms are plotted in Fig. 6. As can be seen, the proposed QVPKRMN algorithms achieve faster convergence rate and lower MSE as compared to QKLMS algorithm, and slightly slow down the convergence rate as compared to the VPKRMN algorithms. Owing to using the quantization scheme, the QVPKRMN algorithms produce about 2000 network size in nonlinear system identification, which reduces the computational burden. 5.2. Test under impulsive noise environment with α-stable distribution model In the second example, the WGN is employed as the input signal, and the nonlinear system model in first experiment is continued to use. An impulsive noise can be modeled as the symmetric α-stable distribution. A standard symmetric α-stable (SαS) distribution has the form [39] α ϕSαS (t) = exp {−m|t| } 13 (45) KLMS KLAD KRMN VPKRMN−Algorithm 1 VPKRMN−Algorithm 2 0 Testing MSE 10 −1 10 −2 10 0 5000 10000 15000 n Figure 8: Learning curves of KLMS, KLAD and VPKRMN algorithms for nonlinear system identification with α-stable noise (α=1.4, SNR=15dB). (Keys: KLMS, µ = 0.1, h = 0.05; KLAD, µ = 0.05, h = 0.01; KRMN, µ = 0.1, h = 0.1, λ = 0.3; VPKRMN-Algorithm 1, µ = 0.1, h = 0.1; VPKRMN-Algorithm 2, µ = 0.1, h = 0.1.) QKLMS VPKRMN−Algorithm 1 VPKRMN−Algorithm 2 QVPKRMN−Algorithm 1 QVPKRMN−Algorithm 2 0 Testing MSE 10 −1 10 −2 10 0 5000 10000 15000 n Figure 9: Learning curves of QLMS, VPRMN and QVPKRMN algorithms for nonlinear system identification with α-stable noise (α=1.4, SNR=15dB). (Keys: QKLMS, εU = 1, h = 0.1; QVPKRMN-Algorithm 1, εU = 0.01; QVPKRMN-Algorithm 2, εU = 0.1.) where 0 < α ≤ 2 is a characteristic exponent, which indicates a peaky and heavy tailed distribution and likely more impulsive noise, and m > 0 is dispersion of the noise. In following simulation studies, α = 1.4 is used, which is well model the radio frequency interference (RFI) for the embedded wireless data transceivers [40]. In addition, the signal-to-noise ratio (SNR) of the α-stable noise is defined as: [41, 42]: SNR , σu2 . m (46) To demonstrate the effect of the variable parameter on the proposed algorithms, Fig. 7 shows the VPKRMN algorithms with different parameter settings. As can be found, a tiny change of the parameters cause a large change of the performance, and the appropriate selection of the parameters are γ = 0.0003, θ = 0.01. Fig. 8 illustrates a comparison with the LMS, RMN, KLMS, KLAD, KRMN and VPKRMN algorithms for nonlinear system identification in α-stable noise. Obviously, the KRMN algorithm has worse results than two VPKRMN algorithms because it is based on fixed mixing parameter, and the proposed VPKRMN algorithms achieve improved performance. Finally, we evaluate the performance of the QKLMS, VPKRMN 14 2500 Network size 2000 1500 1000 500 0 QVPKRMN−Algorithm 1 QVPKRMN−Algorithm 2 0 5000 10000 15000 n Figure 10: Network size growth of the QVPKRMN algorithms. and QVPKRMN algorithm, as shown in Fig. 9. As can be seen, the proposed QVPKRMN algorithms have similar identification performance, and superior performance in the presence of α-stable noise as compared to QKLMS algorithm. Fig. 10 shows the network size growth of QVPKRMN algorithms. One can see that the network size of QVPKRMN algorithm decreases to about 10% by sacrificing a little performance, which reduces the computational complexity. From the experiment results of the above two examples, the proposed VPKRMN algorithms demonstrate the improved performance than the existing algorithms, and the performance of QVPKRMN algorithm is close to VPKRMN algorithm with less computational complexity. Also, the robustness of the proposed algorithms is confirmed by simulating various population sizes and different bandwidth parameters. The proposed VPKRMN algorithm 1 and VRKRMN algorithm 2 have similar misadjustment and convergence speed under the slightly impulsive process. By using the error autocorrelation e(n)e(n − 1), the VPKRMN algorithm 2 obtains a faster convergence rate than VPKRMN algorithm 1 in highly impulsive case. We conclude that all the proposed algorithms for nonlinear system identification can provide a satisfying result in impulsive interference. 6. Conclusion Two VPKRMN algorithms and their quantization form (QVPKRMN algorithms) are proposed for nonlinear system identification under impulsive noises. The VPKRMN algorithms effectively solve the problem of mixing parameter selection. Then, to address the problem of computational intensive of the VPKRMN algorithm, the quantization scheme is introduced to the VPKRMN algorithms to generate a QVPKRMN algorithm. Moreover, the convergence behavior of the QVPKRMN algorithms is analyzed. Simulations results showed that the proposed VPKRMN algorithms are superior to the KLMS, KLAD and KRMN algorithms, and the QVPKRMN algorithm preserves the robustness performance under the impulsive interference with low computational complexity. Acknowledgment The work partially supported by the National Science Foundation of P.R. China (Grant: 61571374, 61271340, 61433011). The first author would also like to acknowledge the China Scholarship Council (CSC) for providing him with financial support to study abroad (No. 201607000050). 15 References References [1] B. Schölkopf, A. J. Smola, Learning with kernels: support vector machines, regularization, optimization, and beyond, MIT press, 2002. [2] H. Ishida, Y. Oishi, K. Morita, K. Moriwaki, T. Y. Nakajima, Development of a support vector machine based cloud detection method for MODIS with the adjustability to various onditions, Remote Sensing of Environment 205 (2018) 390–407. [3] B. Schölkopf, A. Smola, K.-R. Müller, Nonlinear component analysis as a kernel eigenvalue problem, Neural computation 10 (5) (1998) 1299–1319. [4] R. Fezai, M. Mansouri, O. Taouali, M. F. Harkat, N. Bouguila, Online reduced kernel principal component analysis for process monitoring, Journal of Process Control 61 (2018) 1–11. [5] W. Liu, J. C. Principe, S. Haykin, Kernel adaptive filtering: a comprehensive introduction, Vol. 57, John Wiley & Sons, 2011. [6] Y. Engel, S. Mannor, R. Meir, The kernel recursive least-squares algorithm, IEEE Transactions on Signal Processing 52 (8) (2004) 2275–2285. [7] S. Van Vaerenbergh, J. Via, I. Santamarı́a, A sliding-window kernel RLS algorithm and its application to nonlinear channel identification, in: IEEE International Conference on Acoustics, Speech and Signal Processing, IEEE, 2006, pp. 789–792. [8] W. Liu, I. Park, Y. Wang, J. C. Prı́ncipe, Extended kernel recursive least squares algorithm, IEEE Transactions on Signal Processing 57 (10) (2009) 3801–3814. [9] J. D. A. Santos, G. A. Barreto, An outlier-robust kernel rls algorithm for nonlinear system identification, Nonlinear Dynamics 90 (3) (2017) 1707–1726. [10] W. Liu, P. P. Pokharel, J. C. Principe, The kernel least-mean-square algorithm, IEEE Transactions on Signal Processing 56 (2) (2008) 543–554. [11] A. Takeuchi, M. Yukawa, K.-R. Müller, A better metric in kernel adaptive filtering, in: 24th European Signal Processing Conference (EUSIPCO), IEEE, 2016, pp. 1578–1582. [12] W. Wang, S. Wang, G. Qian, B. Yang, Kernel least mean square with tracking, in: 36th Chinese Control Conference (CCC), IEEE, 2017, pp. 5100–5104. [13] W. Liu, J. C. Prı́ncipe, Kernel affine projection algorithms, EURASIP Journal on Advances in Signal Processing 2008 (1) (2008) 784292. [14] F. Albu, K. Nishikawa, Low complexity kernel affine projection-type algorithms with a coherence criterion, in: International Conference on Signals and Systems (ICSigSys), IEEE, 2017, pp. 87–91. 16 [15] J. M. Gil-Cacho, M. Signoretto, T. van Waterschoot, M. Moonen, S. H. Jensen, Nonlinear acoustic echo cancellation based on a sliding-window leaky kernel affine projection algorithm, IEEE Transactions on Audio, Speech, and Language Processing 21 (9) (2013) 1867–1878. [16] C. Richard, J. C. M. Bermudez, P. Honeine, Online prediction of time series data with kernels, IEEE Transactions on Signal Processing 57 (3) (2009) 1058–1067. [17] B. Chen, S. Zhao, P. Zhu, J. C. Principe, Quantized kernel least mean square algorithm, IEEE Transactions on Neural Networks and Learning Systems 23 (1) (2012) 22–32. [18] X. Xu, H. Qu, J. Zhao, X. Yang, B. Chen, Quantised kernel least mean square with desired signal smoothing, Electronics Letters 51 (18) (2015) 1457–1459. [19] S. Wang, Y. Zheng, S. Duan, L. Wang, H. Tan, Quantized kernel maximum correntropy and its mean square convergence analysis, Digital Signal Processing 63 (2017) 164–176. [20] A. Flores, R. C. de Lamare, Set-membership kernel adaptive algorithms, in: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2017, pp. 2676–2680. [21] H. Zhao, Y. Yu, S. Gao, X. Zeng, Z. He, A new normalized LMAT algorithm and its performance analysis, Signal Processing 105 (2014) 399–409. [22] S. Guan, Z. Li, Nonparametric variable step-size LMAT algorithm, Circuits, Systems, and Signal Processing 36 (3) (2017) 1322–1339. [23] E. Eweda, N. J. Bershad, Stochastic analysis of a stable normalized least mean fourth algorithm for adaptive noise canceling with a white Gaussian reference, IEEE Transactions on Signal Processing 60 (12) (2012) 6235–6244. [24] E. Eweda, A stable normalized least mean fourth algorithm with improved transient and tracking behaviors, IEEE Transactions on Signal Processing 64 (18) (2016) 4805–4816. [25] O. Tanrikulu, J. Chambers, Convergence and steady-state properties of the least-mean mixed-norm (LMMN) adaptive algorithm, IEE Proceedings-Vision, Image and Signal Processing 143 (3) (1996) 137– 142. [26] Y. Li, Y. Wang, F. Albu, Sparse channel estimation based on a reweighted least-mean mixed-norm adaptive filter algorithm, in: 24th European Signal Processing Conference (EUSIPCO), IEEE, 2016, pp. 2380–2384. [27] J. Chambers, A. Avlonitis, A robust mixed-norm adaptive filter algorithm, IEEE Signal Processing Letters 4 (2) (1997) 46–48. [28] Q. Miao, C. Li, Kernel least-mean mixed-norm algorithm, International Conference on Automatic Control and Artificial Intelligence (2012) 1285–1288. 17 [29] S. Wang, J. Feng, K. T. Chi, Kernel affine projection sign algorithms for combating impulse interference, IEEE Transactions on Circuits and Systems II: Express Briefs 60 (11) (2013) 811–815. [30] Z. Wu, J. Shi, X. Zhang, W. Ma, B. Chen, Kernel recursive maximum correntropy, Signal Processing 117 (2015) 11–16. [31] J. Zhao, H. Zhang, G. Wang, Projected kernel recursive maximum correntropy, IEEE Transactions on Circuits and Systems II: Express Briefs. [32] W. Gao, J. Chen, Kernel least mean p-power algorithm, IEEE Signal Processing Letters 24 (7) (2017) 996–1000. [33] J. Liu, H. Qu, B. Chen, W. Ma, Kernel robust mixed-norm adaptive filtering, in: International Joint Conference on Neural Networks (IJCNN), IEEE, 2014, pp. 3021–3024. [34] T. Aboulnasr, K. Mayyas, A robust variable step-size LMS-type algorithm: analysis and simulations, IEEE Transactions on Signal Processing 45 (3) (1997) 631–639. [35] A. H. Sayed, Fundamentals of adaptive filtering, John Wiley & Sons, 2003. [36] V. Mathews, S. Cho, Improved convergence analysis of stochastic gradient adaptive filters using the sign algorithm, IEEE Transactions on Acoustics, Speech, and Signal Processing 35 (4) (1987) 450–454. [37] L. R. Vega, H. Rey, J. Benesty, S. Tressens, A new robust variable step-size NLMS algorithm, IEEE Transactions on Signal Processing 56 (5) (2008) 1878–1893. [38] S. S. Haykin, Adaptive filter theory, Pearson Education India, 2008. [39] M. Shao, C. L. Nikias, Signal processing with fractional lower order moments: stable processes and their applications, Proceedings of the IEEE 81 (7) (1993) 986–1010. [40] M. Nassar, K. Gulati, M. R. DeYoung, B. L. Evans, K. R. Tinsley, Mitigating near-field interference in laptop embedded wireless transceivers, Journal of Signal Processing Systems 63 (1) (2011) 1–12. [41] H. Zhao, L. Lu, Z. He, B. Chen, Adaptive recursive algorithm with logarithmic transformation for nonlinear system identification in α-stable noise, Digital Signal Processing 46 (2015) 120–132. [42] L. Lu, H. Zhao, B. Chen, Collaborative adaptive Volterra filters for nonlinear system identification in α-stable noise environments, Journal of the Franklin Institute 353 (17) (2016) 4500–4525. Vitae Lu Lu is pursuing the Ph.D. degree in the field of signal and information processing at the School of Electrical Engineering, Southwest Jiaotong University, Chengdu, China. He is currently a Visiting Ph.D. 18 Student with the Electrical and Computer Engineering, McGill University, Montreal, QC, Canada. His research interests include adaptive signal processing, kernel methods and evolutionary computing. Haiquan Zhao was born in Henan, China, in 1974. He received the B.S. degree in applied mathematics, the M.S. degree, and the Ph.D. degree in signal and information processing from Southwest Jiaotong University, Chengdu, China, in 1998, 2005, and 2011, respectively. Since 2012, he has been a Professor with the School of Electrical Engineering, Southwest Jiaotong University. His current research interests include adaptive filtering algorithm, adaptive Volterra filter, nonlinear active noise control, nonlinear system identification, and chaotic signal processing. Badong Chen received the B.S. and M.S. degrees in control theory and engineering from Chongqing University, in 1997 and 2003, respectively, and the Ph.D. degree in computer science and technology from Tsinghua University in 2008. He was a Post-Doctoral Researcher with Tsinghua University from 2008 to 2010, and a PostDoctoral Associate at the University of Florida Computational NeuroEngineering Laboratory (CNEL) during the period October 2010 to September 2012. He is currently a professor at the Institute of Artificial Intelligence and Robotics (IAIR), Xi’an Jiaotong University. His research interests are in signal processing, information theory, machine learning, and their applications in cognitive science and engineering. He has published 2 books, 3 chapters, and over 70 papers in various journals and conference proceedings. He is an IEEE senior member and an associate editor of IEEE Transactions on Neural Networks and Learning Systems and has been on the editorial boards of Applied Mathematics and Entropy 19
3cs.SY
arXiv:1512.07450v3 [cs.NE] 4 Jan 2016 Interacting Behavior and Emerging Complexity∗ Alyssa Adams†,> , Hector Zenil‡,> , Eduardo Hermo Reyes∓ , Joost J. Joosten∓,> † Beyond Center, Arizona State University, Tempe, AZ, U.S.A. Department of Computer Science, University of Oxford, U.K. ∓ Department of Logic, History and Philosophy of Science, University of Barcelona, Spain > Algorithmic Nature Group, LABoRES, Paris, France ‡ Abstract Can we quantify the change of complexity throughout evolutionary processes? We attempt to address this question through an empirical approach. In very general terms, we simulate two simple organisms on a computer that compete over limited available resources. We implement Global Rules that determine the interaction between two Elementary Cellular Automata on the same grid. Global Rules change the complexity of the state evolution output which suggests that some complexity is intrinsic to the interaction rules themselves. The largest increases in complexity occurred when the interacting elementary rules had very little complexity, suggesting that they are able to accept complexity through interaction only. We also found that some Class 3 or 4 CA rules are more fragile than others to Global Rules, while others are more robust, hence suggesting some intrinsic properties of the rules independent of the Global Rule choice. We provide statistical mappings of Elementary Cellular Automata exposed to Global Rules and different initial conditions onto different complexity classes. Keywords: Behavioral classes; emergence of behavior; cellular automata; algorithmic complexity; information theory ∗ Corresponding authors: hector.zenil at algorithmicnaturelab.org, jjoosten at ub.edu 1 1 Introduction The race to understanding and explaining the emergence of complex structures and behavior has a wide variety of participants in contemporary science, whether it be sociology, physics, biology or any of the other major sciences. One of the more well-trodden paths is one where evolutionary processes play an important role in the emergence of complex structures. When various organisms compete over limited resources, complex behavior can be beneficial to outperform competitors. But the question remains: Can we quantify the change of complexity throughout evolutionary processes? The experiment we undertake in this paper addresses this question through an empirical approach. In very general terms, we simulate two simple organisms on a computer that compete over limited available resources. Since this experiment takes place in a rather abstract modeling setting, we will use the term processes instead of organisms from now on. Two competing processes evolve over time and we measure how the complexity of the emerging patterns evolves as the two processes interact. The complexity of the emerging structures will be compared to the complexity of the respective processes as they would have evolved without any interaction. When setting up an experiment, especially one of an abstract nature, a subtle yet essential question emerges. What exactly defines a particular process and how does the process distinguish itself from the background? For example, is the shell of a hermit crab part of the hermit crab even though the shell itself did now grow from it? Do essential bacteria that reside in larger organisms form part of that larger organism? Where is the line drawn between an organism and its environment? Throughout this paper, we assume that a process is only separate from the background in a behavioral sense. In fact, we assume processes are different from backgrounds in a physical sense only through resource management. We will demonstrate how these ontological questions are naturally prompted in our very abstract modeling environment. 2 Methods As mentioned already, we wish to investigate how complexity evolves over time when two processes compete for limited resources. We chose to represent competing processes by cellular automata as defined in the subsection below. Cellular automata (CA/CAs) represent a simple parallel computational paradigm that admit various analogies to processes in nature. In this 2 section, we shall first introduce Elementary Cellular Automata to justify how they are useful for modeling interacting and competing processes. 2.1 Elementary Cellular Automata We will consider Elementary Cellular Automata (ECA/ECAs) which are simple and well-studied CAs. They are defined to act on a one-dimensional discrete space divided into countably infinite many cells. We represent this space by a copy of the integers Z and refer to it as a tape or simply row. Each cell in the tape can have a particular color. In the case of our ECAs we shall work with just two colors, represented for example by 0 and 1 respectively. We sometimes call the distributions of 0s and 1s over the tape the output space or spatial configuration. Thus, we can represent an output space by a function r : Z → {0, 1} (also written r ∈ {0, 1}Z ) from the integers Z to the set of colors, in this case {0, 1}. Instead of writing r(i), we shall often write ri and call ri the color of the i-th cell of row r. An ECA  will act on such an output space in discrete time steps and as such can be seen as a function mapping functions to functions. In our case:  : {0, 1}Z → {0, 1}Z . If some ECA  acts on some initial row r0 we will denote the output space after t time-steps by rt so that rt+1 := (rt ). Likewise, we will denote the nth cell at time t by rnt . Our ECAs are entirely defined in a local fashion in that the color rnt+1 of cell n at time t + 1 will only depend on rnt and its t t and rn−1 (in more general terminology, we two direct neighboring cells rn−1 only consider radius-one CAs ([3])). Thus, an ECA  with just two colors in the output space is entirely determined by its behavior on three adjacent cells. Since each cell can only have two colors, there are 23 = 8 such possible triplets so that there are 28 = 256 possible different ECAs. However, we will not consider all ECAs for this experiment for computational simplicity. Instead, we only consider the 88 ECA that are non-equivalent under horizontal translation, 1 and 0 exchanges, or any combination of the two. 2.2 Interacting and competing ECAs: global rules In our experiment, the entire interaction will be modeled by a particular CA that is a combination of two ECAs and something else. So far we have 3 decided to model a process Π by an ECA  with a color 0 (white) and a non-white color 1. A process is modeled by the evolution of the white and non-white cells throughout time steps as governed by the particular ECA rule. Once we have made this choice, it is natural to consider a different process Π0 in a similar fashion. That is, we model Π0 also by an ECA 0 . To tell  and 0 apart on the same grid we choose 0 to work on the alphabet {0, 2} whereas the alphabet of  was {0, 1}. Both  and 0 will use the same symbol 0 (white). The next question is how to model the interaction between  and 0 . We will do so by embedding both CAs in a setting where we have a global CA E with three colors {0, 1, 2}. We will choose E in such a fashion that E restricted to the alphabet {0, 1} is just  while E restricted to the alphabet {0, 2} is 0 . Of course these two requirements do not determine E for given  and 0 . Thus, this leaves us with a difficult modeling choice reminiscent to the ontological question for organisms: what to do for triplets that contain all three colors or are otherwise not specified by  or 0 . Since there are 12 such triplets1 we have 312 = 531 441 different ways to define E given  and 0 . Given ECAs  and 0 as above, we call any such E that extends  and 0 a corresponding global rule. Since there are 88 unique ECAs, there are 3916 unique combinations of  and 0 , which results in 3916 × 531 441 possible globals rules. An online program illustrating interacting cellular automata via a global rule can be visited at http://demonstrations.wolfram.com/ CompetingCellularAutomata/. 2.3 Intrinsic versus evolutionary complexity Let us go back to the main purpose of the paper. We wish to study two competing processes that evolve over time and we want to measure how the complexity of the emerging patterns evolves as the two processes interact. The complexity of the emerging structures will be compared to the complexity of the respective processes as they would have evolved without interacting with each other. The structure of an experiment readily suggests itself. Pick two ECAs  and 0 defined on the alphabets {0, 1} and {0, 2} respectively. Measure the typical complexities c and c0 generated by  and 0 respectively when they 1 These 12 different triplets are {0, 1, 2}, {0, 2, 1}, {1, 0, 2}, {2, 0, 1}, {1, 2, 0}, {2, 1, 0}, {1, 1, 2}, {1, 2, 1}, {2, 1, 1}, {1, 2, 2}, {2, 1, 2}, and {2, 2, 1}. 4 are applied in an isolated setting. Next, pick some corresponding global rule E and measure the typical complexity c̃ that is generated by E. Once this ‘typical complexity’ is well-defined, the experiment can be run and c̃ can be compared to c and c0 . The question is how to interpret the results. If we see a change in the typical complexity there are three possible reasons this change can be attributed to: 1. An evolutionary process triggered by the interaction between  and 0 ; 2. An intrinsic complexity injection due the nature of how E is defined on the 12 previously non-determined tuples; 3. An intrinsic complexity injection due to scaling the alphabet from size 2 to size 3. In case of the second reason, an analogy to the cosmological constant is readily suggested. Recall that E is supposed to take care of the modeling of the background so to say, where none of  or 0 is defined but only their interaction. Thus, a possible increase of complexity/entropy is attributed in the second reason to some intrinsic entropy density of the background. We shall see how the choice of E will affect the change of complexity upon interaction. Before we can further describe the experiment, we first need to decide on how to measure complexity. 2.4 Characterizing complexity In [3] a qualitative characterization of complexity is given for dynamical processes in general and ECAs in particular. More specifically, Wolfram described four classes for complexity which can be characterized as follows: • Class 1. Symbolic systems which rapidly converge to a uniform state. Examples are ECA rules2 0, 32 and 160. • Class 2. Symbolic systems which rapidly converge to a repetitive or stable state. Examples are ECA rules 4, 108 and 218. • Class 3. Symbolic systems which appear to remain in a random state. Examples are ECA rules 22, 30, 126 and 189. • Class 4. Symbolic systems which form areas of repetitive or stable states, but which also form structures that interact with each other in complicated ways. Examples are ECA rules 54 and 110. 2 Throughout this paper we will use the Wolfram enumeration of ECA rules, see [3]. 5 It turns out that one can consider a quantitative complexity measure that largely generalizes the four qualitative complexity classes as given by Wolfram. In our definition of doing so we shall make use of De Bruijn sequences B(k, n) although there are different approaches possible. A de Bruijn sequence for n colors B(k, n) is any sequence σ of n colors so that any of the nk possible string of length k with n colors is actually a substring of σ. In this sense, de Bruijn sequences are sometimes considered to be semi-random. Using de Bruijn sequences of increasing length for a fixed number of colors we can parametrize our input and as such it makes sense to speak of asymptotic behavior. Now, one can formally characterize Wolfram’s classes in terms of Kolmogorov complexity as was done in [5]. For large enough input i and for a sufficiently long time evaluation t one can use Kolmogorov complexity to assign a complexity measure W ((i, t)) of a system  for input i and runtime t (see [5]). A typical complexity measure for  can then be defined by: W () = lim sup W ((i, t)). i,t→∞ As a first approximation, one can use cut-off values of these W () outcomes to classify  into one of the four Wolfram classes. 2.5 Experiment setup In the experiment we tested the influence of global rules (GRs) E on two interacting ECAs  and 0 . Of the 312 = 531 441 possible GRs we explored a total number of 425 600 GRs representing a share3 of 80%. For each combination of the two ECAs  and 0 with one GR, we considered 100 different initial conditions of length 26 corresponding to the first 100 (in the canonical enumeration order) de Bruijn sequences of this length over the alphabet {0, 1, 2}. Each execution was evaluated for 60 timesteps. We use the method described in Section 2.4 to determine the complexity of a CA state evolution for two ECAs with mixed neighborhoods evolving under a GR. Specifically, we used the Mathematica v. 10.0 Compress function as an approximation of the Kolmogorov complexity 4 as first suggested in [4]. 3 We aimed at simulating all GRs but due to some technical restrictions only 80% of them got executed. For as far as we know, no bias has been introduced by this. 4 An online program showing how compression characterizes cellular automata evolution can be found in http://demonstrations.wolfram.com/ CellularAutomatonCompressibility/. 6 It is important to note the number of computational hours it took to undergo this experiment. For 425 600 GRs each exploring 100 different initial conditions and 3 916 combinations of  and 0 , there were a total of 166 billion CA executions. Even with one numerical value per CA execution, the total data generated for this experiment was 1.3 terabytes on hard drive. About 100 000 hours of computational time were used total on about 200 jobs at a time, each GR on a different core, which was 200 GRs. Sometimes up to 500 jobs in parallel. Each job took about 36 hours to complete. Each complexity estimate c is normalized by subtracting the compression value of a string of 0’s of equal length. For most of the CA instances, this is a good approximation of the long-term effects of a GR. As mentioned in Section 2.4 we can use these compression values to determine/approximate Wolfram’s complexity classes according to critical thresholds values. The thresholds were trained according to the ECAs we used for the interactions. Using these compression methods for each run, we determined if the output was Class 1, 2, 3 or 4. The outcome is organized according to the complexity classes of the constituent CAs. For the best clarity, the results are represented in heat maps in the next section. 3 Results Each output class (1-4) is represented by its own heat map, so each figure will have four heat maps to account for the different classes of outputs. Each map is composed of a four-by-four grid whose axes describe the complexity classes of  and 0 , as shown. Thus, of all of the runs (ECAs  and 0 interaction, GR E, and initial condition) that yield a Class 1 output are represented under the heat map labeled Class 1. For example, in Figure 1, about 11% of all Class 1 outputs were generated by Class 1  and Class 1 0 interactions. Here the more intense color represents the more densely populated value for a particular class. 3.1 General Global Rule effects The outputs of every possible 425 600 GRs are accumulated and represented in Figure 1. In general, this figure shows the change in complexity when a GR is used to determine the interaction between two ECAs. Most of the outputs for each complexity class were generated with Class 1 ECA interactions, which was least expected. 7 Figure 1: Heat maps for the 4 classes of output types for all executions. Each heat map represents the percent of cases that resulted in the corresponding output complexity class. 8 3.2 Global Rules of interest Figure 2: Heat maps for the 4 classes of output types for GR 36983. Each heat map represents the percent of cases that resulted in the corresponding output complexity class. Figure 3: Heat maps for the 4 classes of output types for GR 72499. Each heat map represents the percent of cases that resulted in the corresponding output complexity class. Figure 4: Heat maps for the 4 classes of output types for GR 77499. Each heat map represents the percent of cases that resulted in the corresponding output complexity class. There are several interesting examples of GRs that affect the classification of behavior of the CA interaction state output. The heat maps of 9 Figure 5: Interesting outputs of various  and 0 interactions and E. Note that most of the  and 0 rules belong to complexity Class 1 or 2. three GRs in particular are shown in Figures 2-4. As shown in the figure captions, GRs are enumerated according to their position in the Tuples function. That is to say, we enumerate the GRs by generating all tuples with 3 symbols in lexicographical order. Over half of all outputs from with GR 77399 were complexity Class 4 from Class 1  and 0 interactions. The majority of complexity Class 3 outputs were generated from complexity Class 2  and 0 interactions. GR 72399 had outputs that were 65 The actual outputs for some of these GRs are shown in Figure 5. Note that most  and 0 rules are complexity Class 1 or 2. 4 Discussion It is unexpected that the biggest increases in complexity arise from  and 0 complexity Class 1 interactions. This suggests that complexity is intrinsic to the GRs rather than the ECAs themselves. We suspect that interacting Class 1 ECAs readily accept complexity through the rules of their interactions. This is not as prevalent in any of the other complexity classes of  and 10 0 interactions. Likely, if  and 0 have more complexity without interaction, then they are more robust to any complexity changes introduced by the GR. In all cases, complexity increases or remains the same by introducing an interaction rule via a GR. The most interesting cases are when Global Rules increase the complexity of the output by entire classes. There are cases where mixed neighborhoods are present and sustained throughout the output, which is a form of emergence through the interaction GR rule. Because we only used a short number of time steps per execution, it is unclear whether these mixed neighborhoods eventually die out or not, it is nonetheless a case of intermediate emergence from a Global Rule. We have found interesting cases where Global Rules seem to drastically change the complexity of an interacting CA output. Some originally Class 3 ECAs, for example, were found to be too fragile under most Global Rules, while some other are more resilient. Most importantly, the greatest increases in complexity occur when the interacting ECA are both Class 1, which is true for the majority of all possible Global Rules. Although we still have yet to understand the mechanisms behind these results, we are confident that further analysis will be important in understanding the emergence of complex structures. Acknowledgement We want to acknowledge the ASU Saguaro cluster where most of the computational work was undertaken. References [1] G.J. Chaitin, On the length of programs for computing finite binary sequences: Statistical considerations, Journal of the ACM, 16(1):145– 159, 1969. [2] S.J. Chandler, “Cellular Automata with Global Control” http://demonstrations.wolfram.com/ CellularAutomataWithGlobalControl/, Wolfram Demonstrations Project, Published: May 9, 2007. [3] S. Wolfram, A New Kind of Science, Wolfram Media, Champaign, IL. 2002. 11 [4] H. Zenil, Compression-based investigation of the behaviour of cellular automata and other systems, Complex Systems, (19)2, 2010. [5] H. Zenil and E. Villarreal-Zapata, Asymptotic Behaviour and Ratios of Complexity in Cellular Automata Rule Spaces, International Journal of Bifurcation and Chaos vol. 13, no. 9, 2013. 12
9cs.NE
Computational evolution of decision-making strategies Peter Kvam ([email protected]) Center for Adaptive Rationality, Max Planck Institute for Human Development Lentzeallee 94, 14195 Berlin, Germany Joseph Cesario ([email protected]) Department of Psychology, Michigan State University 316 Physics Rd, East Lansing, MI 48824, USA arXiv:1509.05646v1 [cs.NE] 18 Sep 2015 Jory Schossau ([email protected]) Department of Computer Science and Engineering, Michigan State University 428 South Shaw Rd, East Lansing, MI 48824, USA Heather Eisthen ([email protected]), Arend Hintze ([email protected]) Department of Integrative Biology, BEACON Center for the Study of Evolution in Action, Michigan State University 288 Farm Ln, East Lansing, MI 48824, USA Abstract Most research on adaptive decision-making takes a strategyfirst approach, proposing a method of solving a problem and then examining whether it can be implemented in the brain and in what environments it succeeds. We present a method for studying strategy development based on computational evolution that takes the opposite approach, allowing strategies to develop in response to the decision-making environment via Darwinian evolution. We apply this approach to a dynamic decision-making problem where artificial agents make decisions about the source of incoming information. In doing so, we show that the complexity of the brains and strategies of evolved agents are a function of the environment in which they develop. More difficult environments lead to larger brains and more information use, resulting in strategies resembling a sequential sampling approach. Less difficult environments drive evolution toward smaller brains and less information use, resulting in simpler heuristic-like strategies. Keywords: computational evolution, decision-making, sequential sampling, heuristics Introduction Theories of decision-making often posit that humans and other animals follow decision-making procedures that achieve maximum accuracy given a particular set of constraints. Some theories claim that decision-making is optimal relative to the information given, involving a process of maximizing expected utility or performing Bayesian inference (Bogacz, Brown, Moehlis, Holmes, & Cohen, 2006; Griffiths & Tenenbaum, 2006; Von Neumann & Morgenstern, 1944). Others assume that behavior makes trade-offs based on the environment, tailoring information processing to achieve sufficient performance by restricting priors (Briscoe & Feldman, 2011), ignoring information (Gigerenzer & Todd, 1999), or sampling just enough to satisfy a particular criterion (Link & Heath, 1975; Ratcliff, 1978). In most cases, mechanisms underlying the initial development of these strategies are assumed – either explicitly or implicitly – to be the result of natural and artificial selection pressures. In cognitive science research, however, the evolution of a strategy often takes a back seat to its performance and co- herence. The clarity and intuitiveness of a theory undoubtedly play an immense role, as does its ability to explain and predict behavior, but whether or not a strategy is a plausible result of selection pressures is rarely considered. To be fair, this is largely because the process of evolution is slow, messy, and often impossible to observe in organisms in the lab. Fortunately, recent innovations in computing have enabled us to model this process with artificial agents. In this paper, we propose a method of studying the evolution of dynamic binary decision-making using artificial Markov brains (Edlund et al., 2011; Marstaller, Hintze, & Adami, 2013; Olson, Hintze, Dyer, Knoester, & Adami, 2013) and investigate the evolutionary trajectories and ultimate behavior of these brains resulting from different environmental conditions. In order to demonstrate the method and investigate an interesting problem, we focus on the simple choice situation where a decision-maker has to choose whether the source of a stimulus is ’signal’ S or ’noise’ N (for preferential decisions, nonspecific choices A or B can be substituted). A similar decision structure underlies a vast array of choices that people and other animals make, including edible/inedible, healthy/sick, safe/dangerous, and so on. The task requires a decision-maker to take in and process information over time and make a decision about which source yielded that information. However, the decision maker is free to vary the amount of information it uses and processing it applies, and different theories make diverging predictions about how each of these should vary. On one hand, it may be more advantageous to use every piece of information received, feeding it through a complex processing system in order to obtain maximum accuracy. On the other, a simpler processing architecture that ignores information may be sufficient in terms of accuracy and more robust to random mutations, errors, or over-fitting. More complex models Many of the most prominent complex decision-making models fall under the sequential sampling framework (Bogacz et al., 2006; Link & Heath, 1975; Ratcliff, 1978). These models assume that a decision-making agent takes or receives samples one by one from a distribution of evidence, with each sample pointing toward the signal or noise distribution. They posit that agent combines samples to process information, for example by adding up the number favoring S and subtracting the number favoring N. When the magnitude of this difference exceeds a criterion value θ (e.g. larger than 4 / smaller than -4), a decision is triggered in favor of the corresponding choice option (+θ ⇒ S, −θ ⇒ N). This strategy implements a particular form of Bayesian inference, allowing a decisionmaker to achieve a desired accuracy by guaranteeing that the log odds of one hypothesis (S or N) over the other is at least equal to the criterion value. In these models, each piece of information collected is used to make a decision. Although organisms may not literally add and subtract pieces of information, we should expect to observe two characteristics in organisms that implement these or similar strategies. First, they should be relatively complex, storing the cumulative history of information to make their decisions. Second, they should give each piece of information they receive relatively equal weight, spreading out the weights assigned to information across a long series of inputs. more detail. Less complex models Decision task Toward the other end of the spectrum of model complexity are heuristics which deliberately ignore information in order to obtain better performance in particular environments (Brandstätter, Gigerenzer, & Hertwig, 2006; Gigerenzer & Brighton, 2009; Gigerenzer & Todd, 1999). Many of these strategies are non-compensatory, meaning that they terminate the use of information as soon as one piece of evidence clearly favors either S or N. Accordingly, a decision maker can have a relatively simple information processing architecture, as it can just copy incoming information to its output indicators to give an answer. Some of these require ordinal information about different sources of information and their validity, resulting in increased complexity (Dougherty, FrancoWatkins, & Thomas, 2008), but for the current problem we assume that all information comes from a single source. As a result of the relatively simple architecture and onepiece decision rules, we can expect to observe two characteristics in organisms that implement strategies similar to these heuristics. First, they should have relatively simple information processing architectures, favoring short and robust pathways that do little integration. Second, they should appear to give the most weight to the last piece(s) of information they receive before making their decision, yielding a relationship between the final decision and the sequence of inputs that is heavily skewed to the most recently received inputs. Of course, the real behavior of artificially evolved organisms will probably lie somewhere along the spectrum between these two poles. However, we can compare the relative leanings of different populations of organisms by varying the characteristics of the environments in which they evolve. We next describe the decision-making task and manipulations in The task that the agents had to solve was a binary decision problem, where they received information from one source S or another N. The information coming from either source included two binary numbers, and therefore could yield any of the inputs [00], [01], [10], or [11]. Source S would yield primarily 0s on the left and 1s on the right, and source N would yield primarily 1s on the left and 0s on the right. The exact proportion of these inputs was varied in order to alter the difficulty of the task. For example, an easy S stimulus would give 90% 0s (10% 1s) on the left, and 90% 1s (10% 0s) on the right. The two inputs were independent, so this would ultimately give 81% [01] inputs, 9% [11], 9% [00], and 1% [10]. In a more difficult environment, an S stimulus might have 60% 0s on the left and 60% 1s on the right, yielding 36% [01], 24% [11], 24% [00], and 16% [10]. For an N stimulus, the possible inputs would be the same, but the frequency of [01] and [10] inputs would be flipped (i.e. more 1s on the left and 0s on the right). These frequencies were not shown to the agents at the start of each trial. Instead, each trial started with a random frequency of 50%, increasing each consecutive step by 1% until the target frequency was reached. This was done in part to emulate how agents encounter stimuli in real situations (i.e. stimuli progressively come into sensory range, increasing in strength over time rather than simply appearing), but also to avoid ’sticking’ at a local maximum where agents simply copy their first input to outputs. The target frequency of 1s and 0s was manipulated to be 60 − 90% (in 5% increments), resulting in 7 difficulty levels for different populations of agents. For each decision, the agents received up to 100 inputs. Each new input constituted one time step during which the Methods We were interested in examining the strategies and evolutionary trajectories that digital agents took to solve a simple dynamic decision-making problem. To do so, we developed a binary decision-making task for the agents to solve. The fitness of an agent was defined as the number of correct decisions it made over 100 trials of the task, and the probability that it would reproduce was determined by this fitness value. Note that fitness was determined by the number of correct answers, reflecting agents’ propensity to respond together with their accuracy when they did respond - there was no fitness penalty or cost for agent complexity. Formally, the probability that it generated each child of the next generation was given by its fitness divided by the total fitness across the total population (roulette wheel selection). An agent reproduced to the next generation by creating a copy of itself with random mutations. Over the course of 10,000 generations, this selection and mutation process led to evolution of agents that could successfully perform the task, and enabled us to analyze the strategies that the evolved agents ultimately developed. agent could process that information. If an agent gave an answer by signaling [01] to indicate S or [10] to indicate N (see below), then the decision process would come to a halt, where no new inputs would be given and the agent would be graded on its final answer. An agent received 1 point toward its fitness if it gave the correct answer or 0 points if it was incorrect or if it failed to answer before 100 inputs were given. In addition to the difficulty manipulation, we included a “non-decision time” manipulation, where an agent was not permitted to answer until t time steps had elapsed (i.e. the agent had received t inputs). This number t was varied from 10 to 50 in 5-step increments, yielding 9 levels of nondecision time across different environments. Increasing t tended to make agents evolve faster, as longer non-decision time tended to allow agents to more easily implement strategies regardless of difficulty level. pings in the gate tables (e.g. it could change between any of the gates shown in Figure 1). This code consisted of 2000200,000 ’nucleotides’ and included mutation rates of 0.005% point mutations, 0.2% duplication mutations, and 0.1% deletion mutations, consistent with previous work (Edlund et al., 2011; Marstaller et al., 2013; Olson et al., 2013). More precisely, logic gates are specified by ’genes’ within this genetic code. Each gene consists of a sequence of nucleotides, numbered 1-4 to reflect the four base nucleotides present in DNA, and starts with the number sequence ’42’ followed by ’213’ (start codon), beginning at an arbitrary location within the genome. Genes are typically about 300 nucleotides long and can have ’junk’ sequences of non-coding nucleotides between them, resulting in the large size of the genomes. The first generation of Markov brain agents in each population was generated from a random seed genome. The first 100 agents were created as random variants of this seed brain using the mutation rates described above, resulting in approximately 20 − 30 random connections per agent. These 100 agents each made 100 decisions, and were selected to reproduce based on their accuracy. This process was repeated for each population for 10,000 generations, yielding 100 agents per population that could perform the decision task. Data Figure 1: Diagram of the structure of a sample Markov brain with input, processing, and output nodes (circles) with connecting logic gates (rectangles). Each gate contains a corresponding table mapping its input values (left) to output values (right). Note that our actual agents had twice the number of nodes shown here available to them. Markov brain agents The Markov brain agents (Edlund et al., 2011; Marstaller et al., 2013; Olson et al., 2013) consisted of 16 binary nodes and of directed logic gates that moved and/or combined information from one set of nodes to another (see Figure 1). Two of these nodes (1 and 2) were reserved for inputs from the environment, described above. Another two (15 and 16) were used as output nodes. These output nodes could show any combination of two binary values. When they did not read [01] (indicating S), or [10] (indicating N), the agents were permitted to continue updating their nodes with inputs until time step 100. To update their nodes at each time step, the agents used logic gates (represented as squares in Figure 1, which took x node values and mapped them onto y nodes using an x × y table. The input nodes, table, and output nodes for these gates were all specified by an underlying genetic code that each Markov brain possessed. Point, insertion, or deletion mutations in the genetic code would cause them add / subtract inputs to a gate, add / subtract outputs, or change the map- For each of the 63 conditions (7 difficulty levels × 9 nondecision times), we ran 10,000 generations of evolution for 100 different sub-populations of Markov brains, giving 6300 total populations. From each of these populations, a random organism was chosen and its line of ancestors was tracked back to the first generation. This set of agents from the last to the first generation is called the line of decent (LOD). For each of the 100 replicates per experimental conditions, all parameters (such as fitness) of agents on the LOD were averaged for each generation. In each of these LODs, we tracked the average number of connections between nodes (see Figure 1) that agents had in each condition and each generation. We refer to this property of the agents as “brain size” — the analogous properties in an organism are the number and connectivity of neurons — and we show its evolutionary trajectory in Figure 2. Finally, we took a close look at the behavior of generation 9970 – this is near the end to ensure that the generation we examined could solve the task, but slightly and somewhat arbitrarily removed from generation 10,000 to ensure that agents in this generation weren’t approaching one of the random dips in performance (i.e. random mutations from this generation were less likely to be deleterious than more recent ones). For these agents, we examined each trial to see what information they received at each time step, which step they made their decision, and which decision they made (coded as correct or incorrect). This allowed us to examine the relationship between the inputs they received and the final answer they gave, giving an estimate of the weight they assigned to each new piece of information. Materials The agents, tasks, and evolution were implemented in C++ using Visual Studio Desktop and Xcode, and the full evolution simulations were run at Michigan State University’s High Performance Computing Center. Results With the exception of high difficulty, low non-decision time conditions, most populations and conditions of agents were able to achieve essentially perfect accuracy on the decision task after 10,000 generations. However, the strategies implemented by each population varied heavily by condition. It is perhaps worth noting at this point the tremendous amount of data that our approach yields. Each condition consisted of 100 populations of 100 agents that made 100 decisions each generation, yielding 10,000 agents and 1 million decisions per generation per condition. This tremendous sample size renders statistical comparisons based on standard error, for example, essentially moot. For this reason, we present mostly examples that illustrate important findings rather than exhaustive statistical comparisons. Brain size Final brain size (number of connections among nodes) varied as a function of both stimulus difficulty and non-decision time. We focus primarily on high non-decision time conditions, as many of the low non-decision time populations — particularly in the difficult stimuli conditions — were unable to achieve the high performance of other groups. As Figure 2 shows, agents faced with the easiest conditions (10-15%) tended to have the smallest final brain size, with means of around 15 − 20 connections. Agents faced with medium difficulty environments evolved approximately 25 − 30 connections, and agent brain size in the most difficult conditions approached 35 connections and appeared to still be climbing with further generations. Perhaps more interesting, though, is the evolutionary trajectory that each of the populations in these conditions took. As shown, each group started with 25-30 connections in the initial generation, and in all of them the number of connections initially dropped for the first 200-400 generations. After that, however, the conditions appear to diverge, with the agents in the easy conditions losing even more connections, agents in the medium conditions staying approximately level, and agents in the difficult conditions adding more and more connections. Strategy use In order to examine the pattern of information use in the agents, we additionally examined the relationship between each piece of information received and the final answer given. We did so by taking the series of inputs (e.g. [00],[11],[01],[01],[11]) and assigning each one a value - information favoring S ([01] inputs) was assigned a value of +1, information favoring N ([10] inputs) was assigned a value of Figure 2: Mean number of connections in agent brains across generations for three levels of task difficulty. For the sake of comparison, the trajectories shown are all from populations with a non-decision time of 40 steps −1, and others ([00] and [11]) were assigned a value of 0. Answers favoring S were also given a value of +1 and answers favoring N a value of −1. Doing so allowed us to track the sequence of −1, 0, +1 — which we refer to as the trajectory — leading to the decision and to correlate this with the final +1 or −1 answer. The result of this analysis for the example conditions is shown in Figure 3. As shown, the trajectory correlations in the more difficult conditions tend to be flatter than those in the easy conditions, and final answers tend to correlate with a longer history of inputs. This indicates that these agents were assigning more similar weight to each piece of information they use, utilizing the full history of inputs they had received rather than just the final piece. Note that all agents appeared to use the most recent pieces of information more heavily. This will be the case for almost any model that generates the data, as the last pieces of information tend to be those that trigger the decision rule – for example, in sequential sampling this will be the piece of information that moves the evidence across the threshold – and as such will always be highly correlated with the final answer. 1 Information use also varied somewhat across levels of nondecision time, but its effect was not particularly pronounced except in the more difficult conditions (e.g. 60-70%). However, this effect is largely a consequence of agent populations’ failure to evolve to perform the task as well when stimulus 1 However, since it can sometimes take several updates / time steps to move a ’trigger’ input through the brain to the output nodes, the final piece of information will not always be perfectly correlated with the output. Figure 3: Example correlations between inputs and final decision for easy (blue), medium (purple) and difficult (red) conditions. The trajectories are time-locked on the final answer on the right side, so the last piece of information an agent received is the rightmost value, and to the left is moving backward through the trajectory. discriminability and non-decision time were low. For example, agents in the difficult, short non-decision time condition (red in left panel of Figure 3) attained accuracy of only 82%, compared to 95+% in other conditions. Higher difficulty still led to larger brains and a longer history of processing in these conditions, but its effect was less pronounced. Therefore, high values of non-decision time apparently made it easier to evolve complex strategies, likely because agents were exposed to more information before making their decisions. Discussion While agents’ strategies spanned a range of complexity, more difficult environments pushed them toward more complex strategies resembling sequential sampling while easier environments led to strategies more similar to non-compensatory heuristics. Therefore, both sequential sampling and heuristics seem to be strategies that could plausibly result from different environmental demands. However, our results run counter to the idea that heuristics are invoked when decisions are particularly difficult or choice alternatives are not easily distinguished (Brandstätter et al., 2006). The final strategies may not support the claim that organisms are primarily heuristic decision-makers (Gigerenzer & Brighton, 2009), but it still lends credence to the premise of ecological rationality on which many heuristics are based. This approach suggests that different environments (choice ecologies) lead to different decision-making strategies rather than a one-size-fits-all process. It is certainly plausible that agents in environments with mixed or changing difficulty levels converge on a single strategy, but for the moment it seems that multiple strategies can be implemented across multiple choice environments. While difficult conditions led to larger brains and more information processing, perhaps a more critical finding is that simpler choice environments led to simpler decision strategies and architectures. While this may initially seem like the other side of the same coin, this result is particularly interesting because we did not impose any penalties for larger brains. Although other researchers have suggested that metabolic costs limit the evolution of large brains (Isler & Van Schaik, 2006; Laughlin, van Steveninck, & Anderson, 1998) and can be substantial in real brains (Kuzawa et al., 2014), they were not necessary to drive evolution toward smaller brains. Instead, we suspect that the drop in brain size is a result of the agents’ response to mutations, or the mutation load imposed by the size of its genome. For example, a random mutation in the genome that connects, disconnects, or re-maps a gate is more likely to affect downstream choice-critical elements of a brain that uses more nodes and connections to process information (has a higher mutation load), particularly if it has a larger ratio of coding to non-coding nucleotides. In this case, a smaller brain would be a tool for avoiding deleterious mutations to the information processing stream. Alternatively, the minimum number of nodes and connections required to perform the task is likely lower in the easier conditions than in the more difficult ones, so mutations that reduce brain size and function might be able to persist in the easier but not the more difficult conditions. In either case, it is clear that a larger brain does not offer sufficient benefits in the easier conditions to overcome the mutation load it imposes. Another potential risk of having a larger brain is the chance of a random mutation preventing information from reaching the output nodes – with a longer chain of processing nodes being easier to interrupt or confuse than a shorter one. While the agents in more difficult conditions were evidently able to overcome such a possibility (usually answering within 20 steps of the end of non-decision time), it may be a barrier that required substantial fitness rewards to cross, which were not present in the easier conditions. We hesitate to make claims that are too broad given the scope of our study, but the finding that brain size can be limited by mutation load is provoking. This may explain why systems that are subject to mutations and selection pressures – including neurons and muscle cells – are reduced when they are unused, even when the energetic costs of maintaining the structure appear to be low. It seems a promising direction for future research to examine in-depth how mutation rate and robustness contribute to organisms’ fitness above and beyond the costs associated with metabolism. Approach We hope to have presented a method for examining questions regarding adaptation and evolution that often arise in cognitive science and psychology. Whereas previous studies have worked from a particular strategy and examined the choice environments in which it succeeds, we present a way of answering questions about how the environment can shape the evolution of a strategy. The strategies resulting from this computational evolution approach are adaptive, easily implemented in the brain, and the result of realistic natural selection pressures. Additionally, we have shown that this approach is capable of addressing important questions about existing models of simple dynamic decisions, though it could undoubtedly shed light on an array of related problems. Of course, there are limitations to this approach, many of which are computational. The agents we used had only 16 nodes, 4 of which were reserved for inputs and outputs, meaning that only 12 could be used for storing (memory) and processing information. Although more nodes could be added – and certainly an accurate model of even very simple nervous systems would have many times more – this would severely slow down the steps required for evolution. It might also lead to problems that are analogous to the over-fitting that occurs when more parameters are added to a model, though this is itself a question worth exploring. Conclusions In this paper, we presented a computational evolution framework that could be used to examine how environments lead to different behaviors. This framework allowed us to examine the strategies that might have arisen in organisms to address the problem of dynamic decision-making, where agents receive information over time and must somehow use this input to make decisions that affect their fitness. We found that both the evolutionary trajectory and the strategies ultimately implemented by the agents are heavily influenced by the characteristics of the choice environment, with the difficulty of the task being a particularly notable influence. More difficult environments tended to encourage the evolution of complex information integration strategies, while simple environments actually caused agents to decrease in complexity, perhaps in order to maintain simpler and more robust decision architectures. They did so despite no explicit costs for complexity, indicating that mutation load may be sufficient to limit brain size. Finally, we discussed these results in the context of existing models of human decision-making, suggesting that both non-compensatory strategies such as fast and frugal heuristics (Gigerenzer & Todd, 1999) and complex ones such as sequential sampling (Link & Heath, 1975) may provide valid descriptions – or at least serve as useful landmarks – of the strategies implemented by evolved agents. In doing so, we provided evidence that strategy use is environmentdependent, as different decision environments led to different patterns of information use. More generally, we have shown that a computational evolution approach integrating computer science, evolutionary biology, and psychology is able to provide insights into how, why, and when different decision-making strategies evolve. Acknowledgments This work was supported by Michigan State University’s High Performance Computing Facility and the National Science Foundation under Cooperative Agreement No. DBI0939454 and Grant No. DGE-1424871. References Bogacz, R., Brown, E., Moehlis, J., Holmes, P., & Cohen, J. D. (2006). The physics of optimal decision making: a formal analysis of models of performance in two-alternative forced-choice tasks. Psychological Review, 113(4), 700–765. Brandstätter, E., Gigerenzer, G., & Hertwig, R. (2006). The priority heuristic: Making choices without trade-offs. Psychological Review, 113(2), 409–432. Briscoe, E., & Feldman, J. (2011). Conceptual complexity and the bias/variance tradeoff. Cognition, 118(1), 2–16. Dougherty, M. R., Franco-Watkins, A. M., & Thomas, R. (2008). Psychological plausibility of the theory of probabilistic mental models and the fast and frugal heuristics. Psychological Review, 115(1), 199–211. Edlund, J. A., Chaumont, N., Hintze, A., Koch, C., Tononi, G., & Adami, C. (2011). Integrated information increases with fitness in the evolution of animats. PLoS Computational Biology, 7(10), e1002236. Gigerenzer, G., & Brighton, H. (2009). Homo heuristicus: Why biased minds make better inferences. Topics in Cognitive Science, 1(1), 107–143. Gigerenzer, G., & Todd, P. M. (1999). Simple heuristics that make us smart. New York, NY: Oxford University Press. Griffiths, T. L., & Tenenbaum, J. B. (2006). Optimal predictions in everyday cognition. Psychological Science, 17(9), 767–773. Isler, K., & Van Schaik, C. P. (2006). Metabolic costs of brain size evolution. Biology Letters, 2(4), 557–560. Kuzawa, C. W., Chugani, H. T., Grossman, L. I., Lipovich, L., Muzik, O., Hof, P. R., . . . Lange, N. (2014). Metabolic costs and evolutionary implications of human brain development. Proceedings of the National Academy of Sciences, 111(36), 13010–13015. Laughlin, S. B., van Steveninck, R. R. d. R., & Anderson, J. C. (1998). The metabolic cost of neural information. Nature Neuroscience, 1(1), 36–41. Link, S., & Heath, R. (1975). A sequential theory of psychological discrimination. Psychometrika, 40(1), 77–105. Marstaller, L., Hintze, A., & Adami, C. (2013). The evolution of representation in simple cognitive networks. Neural Computation, 25(8), 2079–2107. Olson, R. S., Hintze, A., Dyer, F. C., Knoester, D. B., & Adami, C. (2013). Predator confusion is sufficient to evolve swarming behaviour. Journal of The Royal Society Interface, 10(85), 20130305. Ratcliff, R. (1978). A theory of memory retrieval. Psychological Review, 85(2), 59–108. Von Neumann, J., & Morgenstern, O. (1944). Theory of games and economic behavior. Princeton, NJ: Princeton University Press.
9cs.NE
A NOTE ON CONVEX CHARACTERS, FIBONACCI NUMBERS AND EXPONENTIAL-TIME ALGORITHMS arXiv:1508.02598v3 [q-bio.PE] 27 Jul 2016 STEVEN KELK AND GEORGIOS STAMOULIS Abstract. Phylogenetic trees are used to model evolution: leaves are labelled to represent contemporary species (“taxa”) and interior vertices represent extinct ancestors. Informally, convex characters are measurements on the contemporary species in which the subset of species (both contemporary and extinct) that share a given state, form a connected subtree. Given an unrooted, binary phylogenetic tree T on a set of n ≥ 2 taxa, a closed (but fairly opaque) expression for the number of convex characters on T has been known since 1992, and this is independent of the exact topology of T . In this note we prove that this number is actually equal to the (2n−1)th Fibonacci number. Next, we define gk (T ) to be the number of convex characters on T in which each state appears on at least k taxa. We show that, somewhat curiously, g2 (T ) is also independent of the topology of T , and is equal to to the (n − 1)th Fibonacci number. As we demonstrate, this topological neutrality subsequently breaks down for k ≥ 3. However, we show that for each fixed k ≥ 1, gk (T ) can be computed in O(n) time and the set of characters thus counted can be efficiently listed and sampled. We use these insights to give a simple but effective exact algorithm for the NP-hard maximum parsimony distance problem that runs in time Θ(φn · n2 ), where φ ≈ 1.618... is the golden ratio, and an exact algorithm which computes the tree bisection and reconnection distance (equivalently, a maximum agreement forest) in time Θ(φ2n · poly(n)), where φ2 ≈ 2.619. 1. Introduction Phylogenetics is the science of accurately and efficiently inferring evolutionary trees given only information about contemporary species [12]. An important concept within phylogenetics is convexity. Essentially this captures the situation when, within a phylogenetic (i.e. evolutionary) tree, each biological state emerges exactly once: it should not emerge, die out, and then re-emerge. More concretely, given a phylogenetic tree and a set of states assigned to its leaves, can we assign states to the internal vertices of the tree such that each state forms a connected “island” within the tree? If this is possible, the assignment of states to the leaves is known as a convex character. In this article we present a number of results concerning the enumeration of convex characters. In Section 2 we give formal definitions and describe relevant earlier work. In Section 3 we start by showing that an earlier result counting convex characters can be simplified to a term of the Fibonacci sequence. We then seek to count convex characters with the added restriction that each state should occur on at least k leaves, proving the somewhat surprising result that (as for k = 1) tree topology is irrelevant for k = 2, and that a formulation in terms of Fibonacci numbers is again possible. We give an explicit example showing that for k ≥ 3 the topological neutrality breaks down. In Section 4 we show that for all k the size of the space can be counted in polynomial time and space using dynamic programming, which also permits listing and sampling uniformly at random, noting also that non-isomorphic trees can have exactly the same vector of space sizes (for k = 1, 2, ...). In Section 5 we give a number of algorithmic applications for NP-hard problems arising in phylogenetics that seek to quantify the dissimilarity of two phylogenetic trees. Finally, in Section 6 we briefy discuss a number of open problems arising from this work. The software associated with this article has been made publicly available. 2. Preliminaries For general background on mathematical phylogenetics we refer to [12, 5]. An unrooted binary phylogenetic X-tree is an undirected tree T = (V (T ), E(T )) where every internal vertex has degree 3 and whose leaves are bijectively labelled by a set X, where X is often called the set of 1 2 STEVEN KELK AND GEORGIOS STAMOULIS taxa (representing the contemporary species). We use n to denote |X| and often simply write tree when this is clear from the context. A character f on X is a surjective function f : X → C for some set C of states (where a state represents some characteristic of the species e.g. number of legs). We say that f is an r-state character if |C| = r. Each character naturally induces a partition of X and here we regard two characters as being equivalent if they both induce the same partition of X. An extension of a character f to V (T ) is a function h : V (T ) → C such that h(x) = f (x) for all x in X. For such an extension h of f , we denote by lh (T ) the number of edges e = {u, v} such that h(u) 6= h(v). The parsimony score of a character f on T , denoted by lf (T ), is obtained by minimizing lh (T ) over all possible extensions h of f . We say that a character f : X → C is convex on T if lf (T ) = |C| − 1. Equivalently: a character f : X → C is convex on T if there exists an extension h of f such that, for each state c ∈ C, the vertices of T that are allocated state c (by h) form a connected subtree of T . We call such an extension h a convex extension of f . See Figure 1 for an example. The convexity of a character can be tested in polynomial [7, 8] (in fact, linear [2]) time. a a b b a c c f e a a a f a c d d g e b c d g f e b c d g f e b c e d g f e b c d g g b c d d a b f g g e f e f Figure 1. For the given tree T (on 7 taxa) there are 233 convex characters in total, but only 8 in which each state appears on at least 2 taxa, and these are shown above. 1 character uses exactly 1 state (abcdef g), so g2 (T , 1) = 1, 4 characters use 2 states (ab|cdef g, abc|def g, abcd|ef g and abcde|f g), so g2 (T , 2) = 4 and 3 characters use 3 states (ab|cd|ef g, abc|de|f g and ab|cde|f g), so g2 (T , 3) = 3. For each character we have shown an extension verifying that the subtree induced by each state is connected i.e. that the character is convex. We write g(T , r) to denote the number of r-state convex characters on T and gi (T , r) (i ≥ 1) to denote the number of those characters that have the additional property that each state used by the character appears on at least i taxa. It follows from the definition of character that g1 (T , r) = g(T , r). We define: n X gi (T ) = gi (T , r). r=1 The value g1 (T ) is therefore equal to the total number of convex characters on T . For the tree shown in Figure 1, g(T ) = g1 (T ) = 233 and g2 (T ) = 8. We adopt the standard convention that the binomial coefficient nk evaluates to 1 if k = 0, and 0 if n < k and k > 0. In [13] it is proven that, for n, r ≥ 1,   2n − r − 1 g(T , r) = g1 (T , r) = . r−1 A NOTE ON CONVEX CHARACTERS, FIBONACCI NUMBERS AND EXPONENTIAL-TIME ALGORITHMS 3 Hence, (1)  n  X 2n − r − 1 . g1 (T ) = r−1 r=1 As observed in [13], the expression for g1 (T , r) (somewhat surprisingly) does not depend on the topology of T , only on the number of taxa n. Hence we can write g(n), g1 (n) and g1 (n, r) without ambiguity. 3. Fibonacci numbers and convex characters Theorem 1. The value g2 (T , r) only depends on n (i.e. the topology of T is not relevant) and for n ≥ 2, r ≥ 1 is given by the expression   n−r−1 g2 (T , r) = . r−1 Proof. We prove this by induction on n. For the base case note that for n ∈ {2, 3, 4} there is only one binary tree topology (up to relabelling of taxa) possible on n taxa and that the expression correctly evaluates to 1 when r = 1 and, when r ≥ 2, evaluates to 0 in all cases except n = 4, r = 2 when it correctly evaluates to 1. Consider then n ≥ 5, and let r be any value 2 ≤ r ≤ ⌊n/2⌋. (For r = 1 the expression correctly evaluates to 1, and for r > ⌊n/2⌋ the expression correctly evaluates to 0). Every tree on 4 or more taxa contains at least one cherry: two taxa x, y that have a common parent u where the third neighbour of u is an interior vertex. Fix such a cherry. (A similar technique is used in [13] and [14]). Observe that any convex character f on T with the property that each state appears at least twice, has f (x) = f (y). This follows from the connected-subtree definition of convexity. Now, let T ′ = T |(X \ {x}) and let T ′′ = T |(X \ {x, y}), where T |X ′ denotes the tree (on the set of taxa X ′ ) obtained from T by taking the minimum subtree connecting the elements of X ′ and then suppressing vertices of degree 2. There are two cases to distinguish. The first case is when the state f (x) = f (y) does not appear on any other taxa. There are g2 (T ′′ , r − 1) such characters. The second case is when f (x) = f (y) does appear on at least one other taxon. There are g2 (T ′ , r) such characters. Hence, (2) g2 (T , r) = g2 (T ′′ , r − 1) + g2 (T ′ , r). By the inductive hypothesis we have     (n − 2) − (r − 1) − 1 (n − 1) − r − 1 g2 (T , r) = + (r − 1) − 1 r−1     n−r−2 n−r−2 = + r−2 r−1   n−r−1 = . r−1 The last equality follows from the well-known identity known as Pascal’s Rule i.e.  n−1 k−1 , which holds for 1 ≤ k ≤ n. This completes the proof. n k  = n−1 k  +  Consequently, the total number of convex characters on a tree T with each state appearing at least twice is independent of the topology of T . Specifically, we have:  ⌊n/2⌋ n  X n − r − 1 X n−r−1 = (3) g2 (T ) = g2 (n) = r−1 r−1 r=1 r=1 Corollary 2. For even n, g2 (n) = g1 (n/2). Proof. This is immediate by observing that Equation (3) can be obtained by substituting n/2 for n in Equation (1).  4 STEVEN KELK AND GEORGIOS STAMOULIS Let F (n) denote the nth Fibonacci number. That is, F (0) = 0, F (1) = 1 and for n ≥ 2, F (n) = F (n − 1) + F (n − 2). For comprehensive background on Fibonacci numbers see [11]. Theorem 3. For n ≥ 2, g2 (n) = F (n − 1) and g1 (n) = F (2n − 1). Proof. The following identity is classical1 (n ≥ 0): F (n + 1) = ⌊n/2⌋  X k=0  n−k . k If we index k from 1 rather than 0 we obtain, F (n + 1) = ⌊n/2⌋+1  X k=1  n − (k − 1) . k−1 Now, if we replace n with n − 2: F (n − 1) = = ⌊(n−2)/2⌋+1  X k=1 n−k−1 k−1  ⌊n/2⌋   n−k−1 k−1 X k=1 = g2 (n). The expression for g1 (n) is then obtained by applying Corollary 2.  The question arises whether the values g3 (T , r) and/or g3 (T ) share the topological neutrality of their g1 and g2 counterparts. This is not the case: see Figure 2. Here g3 (T1 ) = 5, because g3 (T1 , 1) = 1, g3 (T1 , 2) = 3, g3 (T1 , 3) = 1 and g3 (T1 , r) = 0 (for r > 3). However, g3 (T2 ) = 6, because g3 (T2 , 1) = 1, g3 (T2 , 2) = 4, g3 (T2 , 3) = 1 and g3 (T2 , r) = 0 (for r > 3). i h g a a f b h b i c d e f g e c d T1 T2 Figure 2. The number of characters that are convex on T1 with each state appearing on at least 3 taxa, is not the same as the corresponding number for T2 i.e. g3 (T1 ) 6= g3 (T2 ). This is because g3 (T1 , 2) = 3 (the characters abc|def ghi, abcdef |ghi and abcghi|def ) while g3 (T2 , 2) = 4 (the characters abc|def ghi, abcd|ef ghi, abcde|f gh and abcdef |ghi). Hence topology does play a role here, contrasting with the situation for g1 and g2 . 1This is usually attributed to Édouard Lucas (1876). By applying Pascal’s Rule and some algebraic manipulation it can be proven by induction. A NOTE ON CONVEX CHARACTERS, FIBONACCI NUMBERS AND EXPONENTIAL-TIME ALGORITHMS 5 4. Computing gk (T ) (and listing its elements) with dynamic programming The results from the previous section give rise to a number of questions. Can we compute gk (T ) in polynomial time, for k ≥ 3? Also, if we want to explicitly list all the elements counted by gk (T ) (k ≥ 1), is it possible to achieve this in some reasonable total running time e.g. O(gk (T )·poly(n))? We show that, for all k ≥ 1, the answer to both questions is yes. Specifically, we show how to compute gk (T ) using dynamic programming, for each k ≥ 1. The combinatorial recurrence within the dynamic programming will also allow us to derive a polynomial-time computable bijection from {1, 2, . . . , gk (T )} to the characters counted by gk (T ). Using this bijection it is then straightforward to list (or sample) these elements. Note that this is also an advance for g1 (T ), since the recurrence used in [13] to derive g1 (n) is based on inclusion-exclusion: it is not obvious how to transform it into a bijection. We begin by rooting T by subdividing an arbitrary edge with a new vertex and (implicitly) directing all edges away from this new vertex. The new vertex becomes the root of the tree. Note that this rooting operation has no impact on the convexity of characters, and the location of the root is irrelevant; it is simply a convenience which ensures that the term “child” is well-defined. The dynamic programming works bottom-up, from the leaves towards the root. Here it is helpful to represent a character f on X as a set of non-overlapping, non-empty subsets {B1 , . . . , Bt } that partition X, where each Bi corresponds to a state. We also need some new definitions. A character f is valid for gk (T ) if – f is convex on T , and – for each Bi ∈ f , |Bi | ≥ k. Consider an ordered pair (f, B) where f is a character on X and B ∈ f . We call such a pair a character-root pair of T if – f is convex on T , and – there exists a convex extension of f in which the root of T is assigned state B. Equality between character-root pairs is defined strictly i.e. (f, B) = (f ′ , B ′ ) if and only if f = f ′ and B = B ′ . We say that a character-root pair (f, B) of T is semi-valid for gk (T ) if, for each Bi ∈ f , Bi 6= B ⇒ |Bi | ≥ k. Note that if a character-root pair (f, B) is semi-valid for gk (T ), then f is valid for gk (T ) if and only if |B| ≥ k. At each vertex u of the tree we will compute and store the following k + 1 values, where Tu is simply the subtree rooted at u: – gk (Tu ), – for each 1 ≤ m ≤ k − 1 the number h(Tu , m) which is defined as the number of characterroot state pairs (f, B) of Tu such that both the following conditions hold: (f, B) is semivalid for gk (Tu ) and |B| = m. We also store h(Tu , k) which is defined slightly differently: we replace the term |B| = m with |B| ≥ k. If u is a taxon, then: – gk (Tu ) is equal to 1 if k = 1, and 0 if k > 1, – h(Tu , m) is equal to 1 if m = 1 and 0 if m > 1. We show how to compute these values recursively, assuming the corresponding values have already been computed for Tl (the subtree rooted at the left child of u) and Tr (the subtree rooted at the right child of u). First, X (4) gk (Tu ) = gk (Tl )gk (Tr ) + h(Tl , i)h(Tr , j). 1≤i,j≤k i+j≥k The idea behind this recurrence is that characters counted by gk (Tu ) can be created in two ways: (i) by taking the union of a character from the left subtree with a character from the right subtree, and (ii) by taking a character-root pair (f, B) from the left subtree, a character-root pair (f ′ , B ′ ) from the right subtree, and then merging the root states to yield a character (f \B)∪(f ′ \B ′ )∪(B ∪B ′ ). Characters from the subtrees can only be used for (i) if they are already valid (with respect to gk ) in their subtrees. The characters f and f ′ that are used for (ii) might not be valid with respect 6 STEVEN KELK AND GEORGIOS STAMOULIS to their subtrees, but we do require that they can be combined to obtain a character that is valid for gk (Tu ). This is possible if and only if (f, B) and (f ′ , B ′ ) are semi-valid for their respective subtrees and the sum of the cardinalities of B and B ′ is at least k. See Figure 3 for an example. Tu Tu a a c b d g e h f Tl Tr i c d b g e Tl h i f Tr Figure 3. The character abc|d is not valid for g3 (Tl ), and ef g|hi is not valid for g3 (Tr ), but the character abc|dhi|ef g is valid for g3 (Tu ). The validity is obtained by allowing d and hi to merge, which was possible because they could both “reach” the roots of their respective subtrees. Second, for 1 ≤ m ≤ k − 1 we have, (5) h(Tu , m) = gk (Tl )h(Tr , m) + h(Tl , m)gk (Tr ) + X h(Tl , i)h(Tr , j). 1≤i,j≤k−2 i+j=m Note that here m ≤ k − 1, which means that the semi-valid character-root pairs (f, B) counted by this recurrence are such that f is not valid for gk (Tu ). The first two terms of the recurrence concern the situation analogous to (i) above. Specifically, in this case we assume that no states are merged, so a new semi-valid character-root pair can be created for Tu if and only if it is constructed from the combination of a valid character from one subtree, with a semi-valid character-root pair from the other. The summation term corresponds to (ii). That is, we only count combinations of character-root pairs from the two subtrees if the cardinality of the merged state is exactly m. Finally we have, X (6) h(Tu , k) = gk (Tl )h(Tr , k) + h(Tl , k)gk (Tr ) + h(Tl , i)h(Tr , j). 1≤i,j≤k i+j≥k This final recurrence is semantically very similar to the previous one. The main difference is that it counts all semi-valid character-root pairs (f, B) for gk (Tu ) such that f is valid for gk (Tu ). For a given vertex u, Equation (4) can be computed in O(k 2 ) time, assuming the values for Tl and Tr have already been computed earlier. The same time bound holds for Equations (5) (for a specific 1 ≤ m ≤ k − 1) and (6). Equation (5) has to be computed for each m, yielding a naı̈ve running time bound of O(k 3 ) (per vertex u), but this can easily be improved to O(k 2 ) by observing that a single 1 ≤ i, j ≤ k sweep over the h(Tl , i) and h(Tr , j) values can be recycled for computation of all the different h(Tu , m) values. There are 2(n − 1) vertices in the tree. This yields the following theorem. Theorem 4. Let T be an unrooted binary tree on n taxa. For each 1 ≤ k ≤ n, gk (T ) can be computed in O(k 2 · n) time and O(k · n) space. Corollary 5. Let T be an unrooted binary tree on n taxa. For each 1 ≤ k ≤ n, all the characters that are counted by gk (T ) can be generated in O(gk (T ) · k 2 · n) total time, and a character counted by gk (T ) can be sampled uniformly at random in O(k 2 · n) time and O(k · n) space. Proof. Critically, there is no inclusion-exclusion involved Equations (4), (5) and (6). This allows us to impose a canonical ordering on the characters (and character-root pairs) counted by these A NOTE ON CONVEX CHARACTERS, FIBONACCI NUMBERS AND EXPONENTIAL-TIME ALGORITHMS 7 equations. For example, within Equation (4) we can choose to place the type-(i) characters earlier in the ordering than the type-(ii) characters. Within the gk (Tl )gk (Tr ) type-(i) characters we can refine the order as follows: the first character from the left subtree combined with in turn each of the gk (Tr ) characters from the right subtree, then the second character from the left subtree combined in turn with each of the gk (Tr ) characters from the right subtree, and so on. Once a canonical ordering has been chosen and the dynamic programming has been completed, we can start at the root of T and (using the g and h values computed at all vertices of the tree) recursively backtrack to generate the uniqely defined ith character. Hence, we obtain a bijection from {1, 2, . . . , gk (T )} to the characters counted by gk (T ). The time and space requirements for backtracking through the tree (i.e. evaluating the bijection for a given element of {1, 2, . . . , gk (T )}) are dominated by the time and space requirements of executing the original dynamic program, which are O(k 2 · n) and O(k · n) respectively. This bijection can then be used to list all the characters counted by gk (T ) or to sample uniformly at random from this space.  We have implemented the dynamic programming (and the corresponding algorithms for listing and sampling) in Java and this can be downloaded from http://skelk.sdf-eu.org/convexcount. Finally within the section, for an unrooted binary tree T on n leaves, we define the g-spectrum as simply the vector (g1 (T ), g2 (T ), . . . , gn (T )). It is natural to ask whether two trees on n taxa have the same g-spectrum if and only if they are isomorphic (see e.g. [3] for related discussions of “identifiability”). Using the code above we have verified that, while this claim is true for n ≤ 10 leaves (see the software website for a proof), a counter-example exists for n = 11, see Figure 4. T2 T1 Figure 4. These two trees on 11 leaves are non-isomorphic but have the same g-spectrum: (10946, 55, 8, 3, 2, 1, 1, 1, 1, 1, 1). 5. Algorithmic applications One of the advantages of expressing g1 (n) and g2 (n) as Fibonacci numbers is that it allows us to give tight bounds on their rate of growth. This can be particularly useful when bounding the running time of algorithms. Consider√the following classical, closed-form expression for the Fibonacci numbers (n ≥ 0), where φ = 1+2 5 ≈ 1.618... is the golden ratio:   n φ 1 . F (n) = √ + 5 2 8 STEVEN KELK AND GEORGIOS STAMOULIS √ (1843) by observing that the term (−φ)−n (It is obtained from Binet’s Formula F (n) = φ −(−φ) 5 is vanishing.) Combining with Theorem 3 we obtain  2n−1  φ 1 √ + g1 (n) = , 2 5  n−1  φ 1 g2 (n) = √ + . 2 5 n −n Using asymptotic notation, it is clear that there are Θ(φ2n ) convex characters and Θ(φn ) convex characters in which each state occurs on at least two taxa. We give two examples of how these insights yield non-trivial exponential-time algorithms for two NP-hard problems arising in phylogenetics. 5.1. Computation of maximum parsimony distance. Let T1 , T2 be two unrooted binary trees on the same set of taxa X. The metric dMP (T1 , T2 ) (the maximum parsimony distance of T1 and T2 ) is defined as follows, where f ranges over all characters on X and lf (.) is as defined in Section 2: dMP (T1 , T2 ) = max |lf (T1 ) − lf (T2 )| f It is NP-hard to compute dMP [9, 6]. It can be used to quantify the dissimilarity of two phylogenetic trees and is a lower bound on the similarly NP-hard tree bisection and reconnection (TBR) distance, denoted dT BR [1]. Theorem 6. Given two unrooted binary trees T1 , T2 on the same set of taxa X, where |X| = n, dMP (T1 , T2 ) can be computed in time Θ(φn · n2 ), where φ ≈ 1.618... is the golden ratio. Proof. In [9, 6] it is proven that the optimum is achieved by some character f that is convex on T1 or T2 and where each state in the character occurs on at least two taxa. Hence simply looping through all the characters counted by g2 (T1 ) and, separately, all the characters counted by g2 (T2 ) is sufficient to locate an optimal character. Note that lf (.) can be computed in O(n2 ) time using Fitch’s algorithm2 [7] or dynamic programming. Hence, scoring each character f can easily be performed in quadratic time. The result then follows by leveraging Corollary 5.  We have implemented the dMP algorithm in Java and for an exponential-time algorithm the results are encouraging; the code is freely available at http://skelk.sdf-eu.org/convexmpdist. On a single 32-bit 1.66GHz Intel Atom (N450) processor the algorithm terminates for n = 20, 25, 30 in less than 1 second, 3 seconds and 51 seconds respectively. On a more powerful 64-bit 3.10GHz machine the previously fastest algorithm, the Integer Linear Programming (ILP) approach described in [9], took 70 seconds to terminate on 12 taxa, and stalled completely on trees with more than 16 taxa, even using state-of-the-art ILP software. The enhanced range of our software has been recently used in experiments to verify that dMP is often a very good lower bound on dMP [10]. 5.2. Computation of TBR distance and maximum agreement forests. Finally, we note that the results in this article also give an easy (although, in some cases, somewhat crude) upper bound on the number of agreement forests of two unrooted binary trees T1 , T2 on n taxa. Recall that, for an unrooted binary phylogenetic tree T on X and X ′ ⊆ X, T |X ′ is defined to be the unrooted binary phylogenetic tree on X ′ obtained by taking the minimal subtree of T that connects X ′ , and suppressing vertices of degree 2. An agreement forest is a partition of X into non-empty subsets X1 , . . . , Xk such that (i) within T1 (respectively, T2 ) the minimal connecting subtrees induced by the Xi are vertex-disjoint and (ii) for each Xi , T1 |Xi = T2 |Xi (where here equality explicitly takes the taxa into account). See [1] and recent articles such as [4] for further background on agreement forests. A maximum agremeent forest is an agreement forest with a 2Fitch’s algorithm has running time O(ns) where n is the number of taxa and s is the number of states in the character. In our context s can rise to O(n). A NOTE ON CONVEX CHARACTERS, FIBONACCI NUMBERS AND EXPONENTIAL-TIME ALGORITHMS 9 minimum number of components, and this minimum is denoted dMAF . Note that due to part (i) of the definition every agreement forest induces a character that is convex on both T1 and T2 (although not all characters that are convex on both T1 and T2 necessarily correspond to agreement forests3). Hence there are at most O(φ2n ) agreement forests, which is O(2.619n ). It is well-known that dT BR is equal to the number of components in a maximum agreement forest, minus 1 [1]. Hence, again by leveraging Corollary 5 we obtain: Theorem 7. Given two unrooted binary trees T1 , T2 on the same set of taxa X, where |X| = n, dT BR (T1 , T2 ) = dMAF (T1 , T2 ) − 1 can be computed in time Θ(2.619n · poly(n)). Moreover, all agreement forests of T1 and T2 can be listed in the same time bound. 6. Conclusion A number of interesting open problems remain. For example, can we characterize non-isomorphic trees that have the same g-spectrum? For a given k ≥ 3 and n, can we give analytical lower and upper bounds on gk (T ), ranging over the space of all trees T on n taxa? 7. Acknowledgements We thank Mike Steel for helpful discussions. References [1] B. Allen and M. Steel. Subtree transfer operations and their induced metrics on evolutionary trees. Annals of Combinatorics, 5:1–15, 2001. [2] E. Bachoore and H. Bodlaender. Convex recoloring of leaf-colored trees. Utrecht University technical report, 2006. [3] M. Bordewich, K. Huber, and C. Semple. Identifying phylogenetic trees. Discrete Mathematics, 300(1):30–43, 2005. [4] J. Chen, J-H. Fan, and S-H. Sze. Parameterized and approximation algorithms for maximum agreement forest in multifurcating trees. Theoretical Computer Science, 562:496–512, 2015. [5] A. Dress, K. Huber, J. Koolen, V. Moulton, and A. Spillner. Basic phylogenetic combinatorics. Cambridge University Press, 2012. [6] M. Fischer and S. Kelk. On the maximum parsimony distance between phylogenetic trees. Annals of Combinatorics, 20(1):87–113, 2016. [7] W. Fitch. Toward defining the course of evolution: minimum change for a specific tree topology. Systematic Zoology, 20(4):406–416, 1971. [8] J. Hartigan. Minimum mutation fits to a given tree. Biometrics, pages 53–65, 1973. [9] S. Kelk and M. Fischer. On the complexity of computing MP distance between binary phylogenetic trees. Annals of Combinatorics, 2016. To appear, preprint arXiv:1412.4076. [10] S. Kelk, M. Fischer, V. Moulton, and T. Wu. Reduction rules for the maximum parsimony distance on phylogenetic trees. Theoretical Computer Science, 2016. To appear, preprint arXiv:1512.07459. [11] T. Koshy. Fibonacci and Lucas numbers with applications. Pure and Applied Mathematics: A Wiley Series of Texts, Monographs and Tracts. Wiley, 2011. [12] C. Semple and M. Steel. Phylogenetics. Oxford University Press, 2003. [13] M. Steel. The complexity of reconstructing trees from qualitative characters and subtrees. Journal of Classification, 9(1):91–116, 1992. [14] M. Steel and Y. Fu. Classifying and counting linear phylogenetic invariants for the jukes–cantor model. Journal of Computational Biology, 2(1):39–47, 1995. Department of Data Science and Knowledge Engineering (DKE), Maastricht University, P.O. Box 616, 6200 MD Maastricht, The Netherlands E-mail address: [email protected] E-mail address: [email protected] 3However, if T = T then convex characters and agreement forests are related one-to-one. 1 2
8cs.DS
Venture: a higher-order probabilistic programming platform with programmable inference arXiv:1404.0099v1 [cs.AI] 1 Apr 2014 Vikash Mansinghka Daniel Selsam Yura Perov VKM @ MIT. EDU DSELSAM @ MIT. EDU PEROV @ MIT. EDU Abstract We describe Venture, an interactive virtual machine for probabilistic programming that aims to be sufficiently expressive, extensible, and efficient for general-purpose use. Like Church, probabilistic models and inference problems in Venture are specified via a Turing-complete, higher-order probabilistic language descended from Lisp. Unlike Church, Venture also provides a compositional language for custom inference strategies, assembled from scalable implementations of several exact and approximate techniques. Venture is thus applicable to problems involving widely varying model families, dataset sizes and runtime/accuracy constraints. We also describe four key aspects of Venture’s implementation that build on ideas from probabilistic graphical models. First, we describe the stochastic procedure interface (SPI) that specifies and encapsulates primitive random variables, analogously to conditional probability tables in a Bayesian network. The SPI supports custom control flow, higher-order probabilistic procedures, partially exchangeable sequences and “likelihood-free” stochastic simulators, all with custom proposals. It also supports the integration of external models that dynamically create, destroy and perform inference over latent variables hidden from Venture. Second, we describe probabilistic execution traces (PETs), which represent execution histories of Venture programs. Like Bayesian networks, PETs capture conditional dependencies, but PETs also represent existential dependencies and exchangeable coupling. Third, we describe partitions of execution histories called scaffolds that can be efficiently constructed from PETs and that factor global inference problems into coherent sub-problems. Finally, we describe a family of stochastic regeneration algorithms for efficiently modifying PET fragments contained within scaffolds without visiting conditionally independent random choices. Stochastic regeneration insulates inference algorithms from the complexities introduced by changes in execution structure, with runtime that scales linearly in cases where previous approaches often scaled quadratically and were therefore impractical. We show how to use stochastic regeneration and the SPI to implement general-purpose inference strategies such as Metropolis-Hastings, Gibbs sampling, and blocked proposals based on hybrids with both particle Markov chain Monte Carlo and mean-field variational inference techniques. Keywords: Probabilistic programming, Bayesian inference, Bayesian networks, Markov chain Monte Carlo, sequential Monte Carlo, particle Markov chain Monte Carlo, variational inference Acknowledgements: The authors thank Vlad Firoiu and Alexey Radul for contributions to multiple Venture implementations, and Daniel Roy, Cameron Freer and Alexey Radul for helpful discussions, suggestions, and comments on early drafts. This work was supported by the DARPA PPAML program, grants from the ONR and ARO, and Google’s “Rethinking AI” project. Any opinions, findings, and conclusions or recommendations expressed in this work are those of the authors and do not necessarily reflect the views of any of the above sponsors. 1 Contents 1 Introduction 1.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 The Venture Language 2.1 Modeling and Inference Instructions . . . . . . . . . . 2.2 Modeling Expressions . . . . . . . . . . . . . . . . . 2.3 Inference Scopes . . . . . . . . . . . . . . . . . . . . 2.4 Inference Expressions . . . . . . . . . . . . . . . . . . 2.5 Values . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6 Automatic inference versus inference programming . . 2.7 Procedural and Declarative Interpretations . . . . . . . 2.8 Markov chain and sequential Monte Carlo architectures 2.9 Examples . . . . . . . . . . . . . . . . . . . . . . . . 2.9.1 Hidden Markov Models . . . . . . . . . . . . 2.9.2 Hierarchical Nonparametric Bayesian Models . 2.9.3 Inverse Interpretation . . . . . . . . . . . . . . 3 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Stochastic Procedures 3.1 Expressiveness and extensibility . . . . . . . . . . . . . . . . . . . . . 3.2 Primitive stochastic procedures . . . . . . . . . . . . . . . . . . . . . . 3.3 The Stochastic Procedure Interface . . . . . . . . . . . . . . . . . . . . 3.3.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.2 Exposed Simulation Requests . . . . . . . . . . . . . . . . . . 3.3.3 Latent Simulation Requests and the Foreign Inference Interface 3.3.4 Optimizations for higher-order SPs . . . . . . . . . . . . . . . 3.3.5 Auxiliary State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 6 . . . . . . . . . . . . 7 8 9 10 11 13 14 14 15 16 16 17 19 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 21 22 23 23 24 24 26 26 Probabilistic Execution Traces 4.1 Definition of a probabilistic execution trace . . . . . . . . . . . . . . . . 4.2 Families . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3 Exchangeable coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4 Existential dependence and contingent evaluation . . . . . . . . . . . . . 4.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.1 Trick coin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.2 A simple Bayesian network . . . . . . . . . . . . . . . . . . . . 4.5.3 Stochastic memoization . . . . . . . . . . . . . . . . . . . . . . 4.6 Constructing PETs via forward simulation . . . . . . . . . . . . . . . . . 4.6.1 Pseudocode for EVAL, APPLY and EVAL-REQUESTS . . . . . 4.7 Undoing simulation of PET fragments . . . . . . . . . . . . . . . . . . . 4.7.1 Pseudocode for UNEVAL, UNAPPLY and UNEVAL-REQUESTS 4.8 Enforcing constraints via CONSTRAIN and UNCONSTRAIN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 27 28 28 28 28 28 29 29 31 31 34 35 37 2 5 Partitioning Traces into Scaffolds for Scalable, Incremental Inference 5.1 Motivation and Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Partitioning traces and defining scaffolds . . . . . . . . . . . . . . . . . . . . . . . 5.3 Constructing the scaffold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4 Pseudocode for constructing scaffolds . . . . . . . . . . . . . . . . . . . . . . . . 5.4.1 Absorbing at Applications . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5 Breaking down a global inference problem into collections of local inference problems 5.6 Local kernels and scaffolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.6.1 Local simulation kernels: resimulation & bottom-up proposals . . . . . . . 5.6.2 Local delta kernels and reuse of random choices . . . . . . . . . . . . . . 39 39 40 41 42 45 45 47 47 47 6 Stochastic Regeneration Algorithms for Scaffolds 6.1 Detaching along a scaffold with DETACH-AND-EXTRACT . . . . . . . . . . 6.2 Regenerating a new trace along a scaffold with REGENERATE-AND-ATTACH 6.3 Building invariant transition operators using stochastic regeneration . . . . . . 6.3.1 Weights assuming simulation kernels . . . . . . . . . . . . . . . . . . 6.3.2 Weights assuming delta kernels . . . . . . . . . . . . . . . . . . . . . 6.4 Context-independent versus context-specific inference schemes . . . . . . . . . . . . . . . 48 48 50 52 52 53 53 . . . . . . . . 53 53 54 55 56 56 56 57 57 7 8 General-purpose Inference Strategies via Stochastic Regeneration 7.1 Factorization of the acceptance ratio . . . . . . . . . . . . . . . . . . 7.2 Auxiliary variables for state-dependent stochastic selection of kernels 7.3 Metropolis-Hastings via Stochastic Regeneration . . . . . . . . . . . 7.4 Approximating optimal proposals via Stochastic Variational Inference 7.4.1 Posing the optimization problem . . . . . . . . . . . . . . . . 7.4.2 Stochastic gradient descent . . . . . . . . . . . . . . . . . . . 7.4.3 As a proposal for Metropolis-Hastings . . . . . . . . . . . . . 7.4.4 Using regen and detach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Particle-based Inference: Enumerative Gibbs and Particle Markov chain Monte Carlo 8.1 In-place mutation versus simultaneous particles in memory . . . . . . . . . . . . . 8.2 The Mix-MH operator for constructing particle-based transition operators . . . . . 8.2.1 Metropolis-Hastings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.2 State-dependent Mixtures of Kernels . . . . . . . . . . . . . . . . . . . . . 8.2.3 The MixMH operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3 Particle-based Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3.1 Generating particles by repeatedly applying a seed kernel . . . . . . . . . . 8.3.2 The MHn operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3.3 Metropolis-Hastings as special case of particle methods . . . . . . . . . . 8.3.4 Enumerative Gibbs as a special case: using different kernels for each particle 8.3.5 Particle Markov chain Monte Carlo: adding iteration and resampling . . . . 8.3.6 Pseudocode for PGibbs with simultaneous particles . . . . . . . . . . . . . 3 58 59 59 59 60 61 61 61 63 63 63 64 65 9 Conditional Independence and Parallelizing Transitions 9.1 Markov Blankets, Envelopes, and Conditional Independence . . . . . . . . . . . . 9.2 After-the-fact envelopes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3 The envelope of a scaffold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 67 67 68 10 Related Work 68 11 Discussion 11.1 Debugging and profiling probabilistic programs . . . . . . . . . . . . . . . . . . . 11.2 Inference programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 72 73 74 1. Introduction Probabilistic modeling and approximate Bayesian inference have proven to be powerful tools in multiple fields, from machine learning (Bishop, 2006) and statistics (Green et al., 2003; Gelman et al., 1995) to robotics (Thrun et al., 2005), artificial intelligence (Russell and Norvig, 2002), and cognitive science (Tenenbaum et al., 2011). Unfortunately, even relatively simple probabilistic models and their associated inference schemes can be difficult and time-consuming to design, specify, analyze, implement, and debug. Applications in different fields, such as robotics and statistics, involve differing modeling idioms, inference techniques and dataset sizes. Different fields also often impose varying speed and accuracy requirements that interact with modeling and algorithmic choices. Small changes in the modeling assumptions, data, performance/accuracy requirements or compute budget frequently necessitate end-to-end redesign of the probabilistic model and inference strategy, in turn necessitating reimplementation of the underlying software. These difficulties impose a high cost on practitioners, making state-of-the-art modeling and inference approaches impractical for many problems. Minor variations on standard templates can be out of reach for non-specialists. The cost is also high for experts: the development time and failure rate make it difficult to innovate on methodology except in simple settings. This limits the richness of probabilistic models of cognition and artificial intelligence systems, as these kinds of models push the boundaries of what is possible with current knowledge representation and inference techniques. Probabilistic programming languages could potentially mitigate these problems. They provide a formal representation for models — often via executable code that makes a random choice for every latent variable — and attempt to encapsulate and automate inference. Several languages and systems have been built along these lines over the last decade (Lunn et al., 2000; Stan Development Team, 2013; Milch et al., 2007; Pfeffer, 2001; McCallum et al., 2009). Each of these systems is promising in its own domain; some of the strengths of each are described below. However, none of the probabilistic programming languages and systems that have been developed thus far is suitable for general purpose use. Examples of drawbacks include inadequate and unpredictable runtime performance, limited expressiveness, batch-only operation, lack of extensibility, and overly restrictive and/or opaque inference schemes. In this paper, we describe Venture, a new probabilistic language and inference engine that attempts to address these limitations. Several probabilistic programming tools have sought efficiency by restricting expressiveness. For example, Microsoft’s Infer.NET system (Minka et al., 2010) leverages fast message passing 4 techniques originally developed for graphical models, but as a result restricts the use of stochastic choice in the language so that it cannot influence control flow. Such random choices would yield models over sets of random variables with varying or even unbounded size, and therefore preclude compilation to a graphical model. BUGS, arguably the first (and still most widely used) probabilistic programming language, has essentially the same restrictions (Lunn et al., 2000). Random compound data types, procedures and stochastic control flow constructs that could lead to a priori unbounded executions are all out of scope. STAN, a BUGS-like language being developed in the Bayesian statistics community, has limited support for discrete random variables, as these are incompatible with the hybrid (gradient-based) Monte Carlo strategy it uses to overcome convergence issues with Gibbs sampling (Stan Development Team, 2013). Other probabilistic programming tools that have seen real-world use include FACTORIE (McCallum et al., 2009) and Markov Logic (Richardson and Domingos, 2006); applications of both have emphasized problems in information extraction. The probabilistic models that can be defined using FACTORIE and Markov Logic are finite and undirected, specified imperatively (for FACTORIE) or declaratively (for Markov Logic). Both systems make use of specialized, efficient approximation algorithms for inference and parameter estimation. Infer.NET, STAN, BUGS, FACTORIE and Markov Logic each capture important modeling and approximate inference idioms, but there are also interesting models that each cannot express. Additionally, a number of probabilistic extensions of classical logic programming languages have also been developed (Poole, 1997; Sato and Kameya, 1997; De Raedt and Kersting, 2008), motivated by problems in statistical relational learning. As with FACTORIE and Markov Logic, these languages have interesting and useful properties, but have thus far not yielded compact descriptions of many useful classes of probabilistic generative models from domains such as statistics and robotics. In contrast, probabilistic programming languages such as BLOG (Milch et al., 2007), IBAL (Pfeffer, 2001), Figaro (Pfeffer, 2009) and Church (Goodman*, Mansinghka*, Roy, Bonowitz, and Tenenbaum, 2008; Mansinghka, 2009) emphasize expressiveness. Each language was designed around the needs of models whose fine-grained structure cannot be represented using directed or undirected graphical models, and where standard inference algorithms for graphical models such as belief propagation do not directly apply. Examples include probabilistic grammars (Jelinek et al., 1992), nonparametric Bayesian models (Rasmussen, 1999; Johnson et al., 2007; Rasmussen and Williams, 2006; Griffiths and Ghahramani, 2005), probabilistic models over worlds with a priori unknown numbers of objects, models for learning the structure of graphical models (Heckerman, 1998; Friedman and Koller, 2003; Mansinghka et al., 2006), models for inductive learning of symbolic expressions (Grosse et al., 2012; Duvenaud et al., 2013) and models defined in terms of complex stochastic simulation software that lack tractable likelihoods (Marjoram et al., 2003). Each of these model classes is the basis of real-world applications, where inference over richly structured models can address limitations of classic statistical modeling and pattern recognition techniques. Example domains include natural language processing (Manning and Schütze, 1999), speech recognition (Baker, 1979), information extraction (Pasula et al., 2002), multitarget tracking and sensor fusion (Oh et al., 2009; Arora et al., 2010), ecology (Csilléry et al., 2010) and computational biology (Friedman et al., 2000; Yu et al., 2004; Toni et al., 2009; Dowell and Eddy, 2004). However, the performance engineering needed to turn specialized inference algorithms for these models into viable implementations is challenging. Direct deployment of probabilistic program implementations in real-world applications is often infeasible. The elaborations on these models that expressive probabilistic languages enable can thus seem completely impractical. 5 Church makes the most extreme tradeoffs with respect to expressiveness and efficiency. It can represent models from all the classes listed above, partly through its support for higher-order probabilistic procedures, and it can also represent generative models defined in terms of algorithms for simulation and inference in arbitrary Church programs. This flexibility makes Church especially suitable for nonparametric Bayesian modeling (Roy et al., 2008), as well as artificial intelligence and cognitive science problems that involve reasoning about reasoning, such as sequential decision making, planning and theory of mind (Goodman and Tenenbaum, 2013; Mansinghka, 2009). Additionally, probabilistic formulations of learning Church programs from data — including both program structure and parameters — can be formulated in terms of inference in an ordinary Church program (Mansinghka, 2009). But although various Church implementations provide automatic Metropolis-Hastings inference mechanisms that in principle apply to all these problems, these mechanisms have exhibited limitations in practice. It has not been clear how to make general-purpose sampling-based inference in Church sufficiently scalable for typical machine learning applications, including problems for which standard techniques based on graphical models have been applied successfully. It also is not easy for Church programmers to override built in inference mechanisms or add new higher-order stochastic primitives. In this paper we describe Venture, an interactive, Turing-complete, higher order probabilistic programming platform that aims to be sufficiently expressive, extensible and efficient for generalpurpose use. Venture includes a virtual machine, a language for specifying probabilistic models, and a language for specifying inference problems along with custom inference strategies for solving those problems. Venture’s implementation of standard MCMC schemes scales linearly with dataset size on problems where many previous inference architectures scale quadratically and are therefore impractical. Venture also supports a larger class of primitives — including “likelihood-free” primitives arising from complex stochastic simulators — and enables programmers to incrementally migrate performance-critical portions of their probabilistic program to optimized external inference code. Venture thus improves over Church in terms of expressiveness, extensibility, and scalability. Although it remains to be seen if these improvements are sufficient for general-purpose use, unoptimized Venture prototypes have begun to be successfully applied in real-world system building, Bayesian data analysis, cognitive modeling and machine intelligence research. 1.1 Contributions This paper makes two main contributions. First, it describes key aspects of Venture’s design, including support for interactive modeling and programmable inference. Due to these and other innovations, Venture provides broad coverage in terms of models, approximation strategies for those models and overall applicability to inference problems with varying model/data complexities and time/accuracy requirements. Second, this paper describes key aspects of Venture’s implementation: the stochastic procedure interface (SPI) for encapsulating primitives, the probabilistic execution trace (PET) data structure for efficient representation and updating of execution histories, and a suite of stochastic regeneration algorithms for scalable inference within trace fragments called scaffolds. Other important aspects of Venture, including the VentureScript front-end syntax, formal language definitions, software architecture, standard library, and performance measurements of optimized implementations, are all beyond the scope of this paper. It is helpful to consider the relationships between PETs and the SPI. The SPI generalizes the notion of elementary random procedure associated with many previous probabilistic programming 6 languages. The SPI encapsulates Venture primitives and enables interoperation with external modeling components, analogously to a foreign function interface in a traditional programming language. External modeling components can represent sets of latent variables hidden from Venture and that use specialized inference code. The SPI also supports custom control flow, higher-order probabilistic procedures, exchangeable sequences, and the “likelihood-free” stochastic primitives that can arise from complex simulations. Probabilistic execution traces are used to represent generative models written in Venture, along with particular realizations of these models and the data values they must explain. PETs generalize Bayesian networks to handle the patterns of conditional dependence, existential dependence and exchangeable coupling amongst invocations of stochastic procedures conforming to the SPI. PETs thus must handle a priori unbounded sets of random variables that can themselves be arbitrary probabilistic generative processes written in Venture and that may lack tractable probability densities. Using these tools, we show how to define coherent local inference steps over arbitrary sets of random choices within a PET, conditioned on the surrounding trace. The core idea is the notion of a scaffold. A scaffold is a subset of a PET that contains those random variables that must exist regardless of the values chosen for a set of variables of interest, along with a set of random variables whose values will be conditioned on. We show how to construct scaffolds efficiently. Inference given a scaffold proceeds via stochastic regeneration algorithms that efficiently consume and either restore or resample PET fragments without visiting conditionally independent random choices. The proposal probabilities, local priors, local likelihoods and gradients needed for several approximate inference strategies can all be obtained via small variations on stochastic regeneration. We use stochastic regeneration to implement both single-site and composite Metropolis-Hastings, Gibbs sampling, and also blocked proposals based on hybrids with conditional sequential Monte Carlo and variational techniques. The uniform implementation of these approaches to incremental inference, along with analytical tools for converting randomly chosen local transition operators into ergodic global transition operators on PETs, constitutes another contribution. 2. The Venture Language Consider the following example Venture program for determining if a coin is fair or tricky: [ASSUME is_tricky_coin (bernoulli 0.1)] [ASSUME coin_weight (if is_tricky_coin (uniform 0.0 1.0) 0.5)] [OBSERVE (bernoulli coin_weight) True] [OBSERVE (bernoulli coin_weight) True] [INFER (mh default one 10)] [PREDICT (bernoulli coin_weight)] We will informally discuss this program before defining the Venture language more precisely. The ASSUME instructions induce the hypothesis space for the probabilistic model, including a random variable for whether or not the coin is tricky, and either a deterministic coin weight or a potential random variable corresponding to the unknown weight. The model selection problem is expressed via an if with a stochastic predicate, with the alternative models on the consequent and alternate branches. After executing the ASSUME instructions, particular values for is_tricky_coin and coin_weight will have been sampled, though the meaning of the program so far corresponds to a probability distribution over possible executions. 7 The OBSERVE instructions describe a data generator that produces two flips of a coin with the generated weight, along with data that is assumed to be generated by the given generator. In this program, the OBSERVEs encode constraints that both of these coin flips landed heads up. The INFER instruction causes Venture to find a hypothesis (execution trace) that is probable given the data using 10 iterations of its default Markov chain for inference1 . INFER evolves the probability distribution over execution traces inside Venture from whatever distribution is present before the instruction — in this case, the prior — closer to its conditional given any observations that have been added prior to that INFER, using a user-specified inference technique. In this example program, the resulting marginal distribution on whether or not the coin is tricky shifts from the prior to an approximation of the posterior given the two observed flips, increasing the probability of the coin being tricky ever so slightly. Increasing 10 to 100 shifts the distribution closer to the true posterior; other inference strategies, including exact sampling techniques, will be covered later. The execution trace inside Venture after the instruction is sampled from this new distribution. Once inference has finished, the PREDICT instruction causes Venture to report a sampled prediction for the outcome of another flip of the coin. The weight used to generate this sample comes from the current execution trace. 2.1 Modeling and Inference Instructions Venture programs consist of sequences of modeling instructions and inference instructions, each given a positive integer index by a global instruction counter2 . Interactive Venture sessions have the same structure. The modeling instructions are used to specify the probabilistic model of interest, any conditions on the model that the inference engine needs to enforce, and any requests for prediction of values against the conditioned distribution. The core modeling instructions in Venture are: 1. [ASSUME <name> <expr>]: binds the result of simulating the model expression <expr> to the symbol <name> in the global environment. This is used to build up the model that will be used to interpret data. Returns the value taken on by <name>, along with the index of the instruction. 2. [OBSERVE <expr> <literal-value>]: adds the constraint that the model expression <expr> must yield <literal-value> in every execution. Note that this constraint is not enforced until inference is performed. 3. [PREDICT <expr>]: samples a value for the model expression <expr> from the current distribution on executions in the engine and returns the value. As the amount of inference done since the last OBSERVE approaches infinity, this distribution converges to the conditioned distribution that reconciles the OBSERVEs. 1. This default Markov chain is a variant of the algorithm from Church (Goodman*, Mansinghka*, Roy, Bonowitz, and Tenenbaum, 2008; Mansinghka, 2009). This is a simple random scan single-site Metropolis-Hastings algorithm that chooses random choices uniformly at random from the current execution, resimulates them conditioned on the rest of the trace, and accepts or rejects the result. 2. Venture implementations also support labels for the instruction language. However, the recurrence of line numbers is reflective of the ways the current instruction language is primitive as compared to the modeling language. For example, it currently lacks control flow constructs, procedural abstraction, and recursion, let alone runtime generation and execution of statements in the instruction language. Current research is focused on addressing these limitations. 8 4. [FORGET <instruction-index-or-label>]: This instruction causes the engine to undo and then forget the given instruction, which must be either an OBSERVE or PREDICT. Forgetting an observation removes the constraint it represents from the inference problem. Note that the effect may not be visible until an INFER is performed. Venture supports additional instructions for inference and read-out, including: 1. [INFER <inference-expr>]: This instruction first incorporates any observations that have occurred after the last INFER, then evolves the probability distribution on executions according to the inference strategy described by <inference-expr>. Two inference expressions, corresponding to general-purpose exact and approximate sampling schemes, are useful to consider here: (a) (rejection default all) corresponds to the use of rejection sampling to generate an exact sample from the conditioned distribution on traces. The runtime requirements may be substantial, and exact sampling applies to a smaller class of programs than approximate sampling. However, rejection is crucial for understanding the meaning of a probabilistic model and for debugging models without simultaneously debugging inference strategies. (b) (mh default one 1) corresponds to one transition of the standard uniform mixture of single-site Metropolis-Hastings transition operators used as a general-purpose “automatic” inference scheme in many probabilistic programming systems. As the number of transitions is increased from 1 towards ∞, the semantics of the instruction approach an exact implementation of conditioning via rejection. 2. [SAMPLE <expr>]: This instruction simulates the model expression <expr> against the current trace, returns the value, and then forgets the trace associated with the simulation. It is equivalent to [PREDICT <expr>] followed by [FORGET <index-of-predict>], but is provided as a single instruction for convenience. 3. [FORCE <expr> <literal-value>]: Modify the current trace so that the simulation of <expr> takes on the value <literal-value>. Its implementation can be roughly thought of as an OBSERVE immediately followed by an INFER and then a FORGET. This instruction can be used for controlling initialization and for debugging. 2.2 Modeling Expressions Venture modeling expressions describe stochastic generative processes. The space of all possible executions of all the modeling expressions in a Venture program constitute the hypothesis space that the program represents. Each Venture program thus represents a probabilistic model by defining a stochastic generative process that samples from it. At the expression level, Venture is similar to Scheme and to Church, though there are several differences3 . For example, branching and procedure construction can both be desugared into applications of stochastic procedures — that is, ordinary combinations — and do not need to be treated as 3. We sometimes refer to the s-expression syntax (including syntactic sugar) as Venchurch, and the desugared language (represented as JSON objects corresponding to parse trees) as Venture. 9 special forms. Additionally, Venture supports a dynamic scoping construct called scope include for tagging portions of an execution history such that they can be referred to by inference instructions; to the best of our knowledge, analogous constructs have not yet been introduced in other probabilistic programming languages. Venture modeling expressions can be broken down into a few simple cases: 1. Self-evaluating or “literal” values: These describe constant values in the language, and are discussed below. 2. Combinations: (<operator-expr> <operand0-expr> ... <operandk-expr>) first evaluates all of its expressions in arbitrary order, then applies the value of <operator-expr> (which must be a stochastic procedure) to the values of all the <operand-expr>s. It returns the value of the application as its own result. 3. Quoted expressions: (quote <expr>) returns the expression value <expr>. As compared to combinations, quote suppresses evaluation. 4. Lambda expressions: (lambda <args> <body-expr>) returns a stochastic procedure with the given formal parameters and procedure body. <args> is a specific list of argument names (<arg0> ... <argk>). 5. Conditionals: (if <predicate-expr> <consequent-expr> <alternate-expr>) evaluates the <predicate-expr>, and then if the resulting value is true, evaluates and returns value of the <consequent-expr>, and if not, evaluates and returns the value of the <alternate-expr>. 6. Inference scope annotations: (scope include <scope-expr> <block-expr> <expr>) provides a mechanism for naming random choices in a probabilistic model, so that they can be referred to during inference programming. The scope include form simulates <scope-expr> and <block-expr> to obtain a scope value and a block value, and then simulates <expr>, tagging all the random choices in that process with the given scope and block. More details on inference scopes are given below. 2.3 Inference Scopes Venture programs may attach metadata to fragments of execution traces via a dynamic scoping construct called inference scopes. Scopes are defined in modeling expressions, via the special form (scope include <scope-expr> <block-expr> <expr>), that assigns all random choices required to simulate <expr> to a scope that is named by the value resulting from simulating <scope-expr> and a block within that scope that is named by the value that results from simulating <block-expr>. A single random choice can be in multiple inference scopes, but can only be in one block within each scope. Also, a random choice gets annotated with a scope each time the choice is simulated within the context of a scope include form, not just the first time it is simulated. Inference scopes can be referred to in inference expressions, thus providing a mechanism for associating custom inference strategies with different model fragments. For example, in a parameter estimation problem for hidden Markov models, it might be natural to have one scope for the hidden states, another for the hyperparameters, and a third for the parameters, where the blocks for the hidden state scope correspond to indexes into the hidden state sequence. We will see later how to 10 write cycle hybrid kernels that use Metropolis-Hastings to make proposals for the hyperparameters and either single-site or particle Gibbs over the hidden states. Inference scopes also provide a means of controlling the allocation of computational effort across different aspects of a hypothesis, e.g. by only performing inference over scopes whose random choices are conditionally dependent on the choices made by a given PREDICT instruction of interest. Random choices are currently tagged with (scope, block) pairs. Blocks can be thought of as subdivisions of scopes into meaningful (and potentially ordered) subsets. We will see later how inference expressions can make use of block structure to provide fine-grained control over inference and enable novel inference strategies. For example, the order in which a set of random choices is traversed by conditional sequential Monte Carlo can be controlled via blocks, regardless of the order in which they were constructed during initial simulation. Scopes and blocks can be produced by random choices; <scope-expr>s and <block-expr>s are ordinary Venture modeling expressions4 . This enables the use of random choices in one scope to control the scope or block allocation of random choices in another scope. The random choices used to construct scopes and blocks may be auxiliary variables independent of the rest of the model, or latent variables whose distributions depend on the interaction of modeling assumptions, data and inference. At present, the only restriction is that inference on the random choices in a set of blocks cannot add or remove random choices from that set of blocks, though the probability of membership can be affected. Applications of randomized, inference-driven scope and block assignments include variants of cluster sampling techniques, beyond the spin glass (Swendsen and Wang, 1986) and regression (Nott and Green, 2004) settings where they have typically been deployed. The implementation details needed to handle random scopes and blocks are beyond the scope of this paper. However, we will later see analytical machinery that is sufficient for justifying the correctness of complex transition operators involving randomly chosen scopes and blocks. Venture provides two built-in scopes5 : 1. default — This scope contains every random choice. Previously proposed inference schemes for Church as well as concurrently developed generic inference schemes for variants of Venture correspond to single-line inference instructions acting on this default, global scope. 2. latents — This scope contains all the latent random choices made by stochastic procedures but hidden from Venture. Using this scope, programmers can control the frequency with which any external inference systems are invoked, and interleave inference over external variables with inference over the latent variables managed by Venture. 2.4 Inference Expressions Inference expressions specify transition operators that evolve the probability distribution on traces inside the Venture virtual machine. This is in contrast to instructions, which extend a model, add data, or initiate inference using a valid transition operator. Venture provides several primitive forms for constructing transition operators that leave the conditioned distribution invariant, each of which is a valid inference expression. In each of these 4. Our current implementations restrict the values of scope and block names to symbols and integers for simplicity, but this restriction is not intrinsic to the Venture specification. 5. Some implementations so far have merged the default and latent scopes, or triggered inference over latents automatically after every transition over the default scope. 11 forms, scope must be a literal scope, and block must either be a literal block within that scope, or the keyword one or all. The “selected” set of random choices on which each inference expression acts is given by the specified scope and block. If the block specification is all, then the union of all blocks within the scope is taken. If the block specification is one, then one block is chosen uniformly at random from the set of all blocks within the given scope. The core set of inference expressions in Venture are as follows: 1. (mh <scope> <block> <#-transitions>) — Propose new values for the selected choices either by resimulating them or by invoking a custom local proposal kernel if one has been provided. Accept or reject the results via the Metropolis-Hastings rule, accounting for changes to the mapping between random choices and scopes/blocks using the machinery provided later in this paper. Repeat the whole process #-transitions times. 2. (rejection <scope> <block> <#-transitions>) — Use rejection sampling to generate an exact sample from the conditioned distribution on all the selected random choices. Repeat the whole process #-transitions times, potentially improving convergence if the selected set is randomly chosen, i.e. block is one. This transition operator is often computationally intractable, but is optimal, in the sense that it makes the most progress per completed transition towards the conditioned distribution on traces. All the other transition operators exposed by the Venture inference language can be viewed as asymptotically convergent approximations to it. 3. (pgibbs <scope> <block> <#-particles> <#-transitions>) — Use conditional sequential Monte Carlo to propose from an approximation to the conditioned distribution over the selected set of random choices. If block is ordered, all the blocks in the scope are sorted, and each distribution in the sequence of distributions includes all the random choices from the next block. Otherwise, each distribution in the sequence includes a single random choice drawn from the selected set, and the ordering is arbitrary. 4. (meanfield <scope> <block> <iters> <#-transitions>) — Use iters steps of stochastic gradient to optimize the parameters of a partial mean-field approximation to the conditioned distribution over the random choices in the given scope and block (with block interpreted as with mh). Make a single Metropolis-Hastings proposal using this approximation. Repeat the process #-transitions times. 5. (enumerative gibbs <scope> <block> <#-transitions>) — Use exhaustive enumeration to perform a transition over all the selected random choices from a proposal corresponding to the optimal conditional proposal (conditioned on the values of any newly created random variables). Random choices whose domains cannot be enumerated are resimulated from their prior unless they have been equipped with custom simulation kernels. If all selected random choices are discrete and no new random choices are created, this is equivalent to the rejection transition operator, and corresponds to a discrete, enumerative implementation of Gibbs sampling, hence the name. The computational cost scales exponentially with the number of random choices, as opposed to the KL divergence between the prior and the conditional (Freer et al., 2010). Both mh and pgibbs are implemented by in-place mutation. However, versions of each that use simultaneous particles to represent alternative possibilities are given by func-mh and func-pgibbs. 12 These are prepended with func to signal an aspect of their implementation: these simultaneously accessible sets of particles are implemented using persistent data structure techniques typically associated with pure functional programming. func-pgibbs can yield improvements in order of growth of runtime as compared to pgibbs, but it imposes restrictions on the selected random choices6 . There are also currently two composition rules for transition operators, enabling the creation of cycle and mixture hybrids: 1. (cycle (<inference-expr-1> <inference-expr-2>) ... <#-transitions>) — This produces a cycle hybrid of the transition operators represented by the given inference expression: each transition operator is run in sequence, and the whole sequence is repeated #-transitions times. 2. (mixture ((<w1> <inference-expr-1>) (<w2> <inference-expr-2>) ...) <#-transitions>) — This produces a mixture hybrid of the given transition operators, using the given mixing weights, that is invoked #-transitions times. This language is flexible enough to express a broad class of standard approximate inference strategies as well as novel combinations of standard inference algorithm templates such as conditional sequential Monte Carlo, Metropolis-Hastings, Gibbs sampling and mean-field variational inference. Additionally, the ability to use random variables to map random choices to inference strategies and to perform inference over these variables may enable new cluster sampling techniques. That said, from an aesthetic standpoint, the current inference language also has many limitations, some of which seem straightforward to relax. For example, it seems natural to expand inference expressions to support arbitrary modeling expressions, and thereby also support arbitrary computation to produce inference schemes. The machinery needed to support these and other natural extensions is discussed later in this paper. 2.5 Values Venture values include the usual scalar and symbolic data types from Scheme, along with extended support for collections and additional datatypes corresponding to primitive objects from probability theory and statistics. Venture also supports the stochastic procedure datatype, used for built-in and user-added primitive procedures as well as compound procedures returned by lambda. A full treatment of the value hierarchy is out of scope, but we provide a brief list of the most important values here: 1. Numbers: roughly analogous to floating point numbers, e.g. 1, 2.4, -23, and so on. 2. Atoms: discrete items with no internal structure or ordering. These are generated by categorical draws, but also Dirichlet and Pitman-Yor processes. 3. Symbols: symbol values, such as the name of a formal argument being passed to lambda, the name associated with an ASSUME instruction, or the result of evaluating a quote special form. 6. To support multiple simultaneous particles, all stochastic procedures within the given scope and block must support a clone operation for their auxiliary state storage (or have the ability to emulate it). This is feasible for standard exponential family models, but may be not be feasible for external inference systems hosted on distributed hardware. 13 4. Collections: vectors, which map numbers to values and support O(1) random access, and maps, which map values to values and support O(1) amortized random access (via a hash table that relies on the built-in hash function associated with each kind of value). 5. Stochastic procedures: these include the components of the standard library, and can also be created by lambda and other stochastic procedures. 2.6 Automatic inference versus inference programming Although Venture programs can incorporate custom inference strategies, it does not require them. Interfaces that are as automatic as existing probabilistic programming systems are straightforward to implement. Single-site Metropolis-Hastings and Gibbs sampling algorithms — the sole automatic inference option in many probabilistic programming systems — can be invoked with a single instruction. We have also seen that global sequential Monte Carlo and mean field algorithms are similarly straightforward to describe. Support for programmable inference does not necessarily increase the education burden on would-be probabilistic programmers, although it does provide a way to avoid limiting probabilistic programmers to a potentially inadequate set of inference strategies. The idea that inference strategies can be formalized as structured, compositionally specified inference programs operating on model programs is, to the best of our knowledge, new to Venture. Under this view, standard inference algorithms actually correspond to primitive inference programming operations or program templates, some of which depend on specific features of the model program being acted upon. This perspective suggests that far more complex inference strategies should be possible if the right primitives, means of combination and means of abstraction can be identified. Considerations of modularity, analyzability, soundness, completeness, and reuse will become central. and will be complicated by the interaction between inference programs and model programs. For example, inference programmers will need to be able to predict the asymptotic scaling of inference instructions, factoring out the contribution of the computational complexity of the model expressions to which a given inference instruction is being applied. Another example comes from considering abstraction and reuse. It should be possible to write compound inference procedures that can be reused across different models, and perhaps even use inference to learn these procedures via an appropriate hierarchical Bayesian formulation. Another view, arguably closer to the mainstream view in machine learning, is that inference algorithms are better thought of by analogy to mathematical programming and operations research, with each algorithm corresponding to a “solver” for a well-defined class of problems with certain structure. This perspective suggests that there is likely to be a small set of monolithic, opaque mechanisms that are sufficient for most important problems. In this setting, one might hope that inference mechanisms can be matched to models and problems via simple heuristics, and that the problem of automatically generating high-quality inference strategies will prove easier than query planning for databases, and will be vastly easier than automatic programming. It remains to be seen whether the traditional view is sufficient in practice or if it underestimates the richness of inference and its interaction with modeling and problem specification. 2.7 Procedural and Declarative Interpretations We briefly consider the relationship between procedural and declarative interpretations of Venture programs. 14 Venture code has a direct procedural reading: it defines a probabilistic generative process that samples hypotheses, checks constraints, and invokes inference instructions that trigger specific algorithms for reconciling the hypotheses with the constraints. Re-orderings of the instructions can significantly impact runtime and change the distribution on outputs. The divergence between the true conditioned distribution on execution traces and the distribution encoded by the program may depend strongly on what inference instructions are chosen and how they are interleaved with the incorporation of data. Venture code also has declarative readings that are unaffected by some of these procedural details. One way to formalize the of meaning of a Venture program is as a probability distribution over execution traces. A second approach is to ignore the details of execution and restrict attention to the joint probability distribution of the values of all PREDICTs so far. A third approach, consistent with Venture’s interactive interface, is to equate the meaning of a program with the probability distribution of the values of all the PREDICTs in all possible sequences of instructions that could be executed in the future. Under the second and third readings, many programs are equivalent, in that they induce the same distribution albeit with different scaling behavior. As the amount of inference performed at each INFER instruction increases, these interpretations coalesce, recovering a simple semantics based on sequential Bayesian reasoning. Consider replacing all inference instructions are replaced with exact sampling — [INFER (rejection default one 1)] — or a sufficiently large number of transitions of a generic inference operator, such as [INFER (mh default one 1000000)]. In this case, each INFER implements a single step of sequential Bayesian reasoning, conditioning the distribution on traces with all the OBSERVEs since the last INFER. The distribution after each INFER becomes equivalent to the distribution represented by all programs with the same ASSUMEs, all the OBSERVEs before the infer (in any order), and a single INFER. The computational complexity varies based on the ordering and interleaving of INFERs and OBSERVEs, but the declarative meaning is unchanged. Although correspondence with these declarative, fully Bayesian semantics may require an unrealistic amount of computation in real-world applications, close approximations can be useful tools for debugging, and the presence of the limit may prove useful for probabilistic program analysis and transformation. Venture programs represent distributions by combining modeling operations that sample values for expressions, constraint specification operations that build up a conditioner, and inference operations that evolve the distribution closer to the conditional distribution induced by a conditioner. Later in this paper we will see how to evaluate the partial probability densities of probabilistic execution traces under these distributions, as well as the ratios and gradients of these partial densities that are needed for a wide range of inference schemes. 2.8 Markov chain and sequential Monte Carlo architectures The current Venture implementation maintains a single probabilistic execution trace per virtual machine instance. This trace is initialized by simulating the code from the ASSUME and OBSERVE instructions, and stochastically modified during inference via transition operators that leave the current conditioned distribution on traces invariant. The prior and posterior probability distributions on traces are implicit, but can be probed by repeatedly re-executing the program and forming Monte Carlo estimates. This Markov chain architecture has been chosen for simplicity, but sequential Monte Carlo architectures based on weighted collections of traces are also possible and indeed straightforward. 15 The number of initial traces could be specified via an INFER instruction at the beginning of the program. Forward simulation would be nearly unchanged. OBSERVE instructions would attach weights to traces based on the “likelihood” probability density corresponding to the constrained random choice in each observation, and PREDICT instructions would read out their values from a single, arbitrarily chosen “active” trace. An [INFER (resample <k>)] instruction would then implement multinomial resampling, and also change the active trace to ensure that PREDICTs are always mutually consistent. Venture’s other inference programming instructions could be treated as rejuvenation kernels (Del Moral et al., 2006), and would not need to modify the weights7 . This kind of sequential Monte Carlo implementation would have the advantage that the weights could be used to estimate marginal probability densities of given OBSERVEs, and that another source of non-embarrassing parallelism would be exposed. Integrating sophisticated coupling strategies from the αSMC framework (Whiteley et al., 2013) into the inference language could also prove fruitful. Running separate Venture virtual machines is guaranteed to produce independent samples from the distribution represented by the Venture program. This distribution will typically only approximate some desired conditional. If there is no need to quantify uncertainty precisely, then Venture programmers can append repetitions of a sequence of INFER and PREDICT instructions to their program. Unless the INFER instructions use rejection sampling, this choice yields PREDICT outputs that are dependent under both Markov chain and sequential Monte Carlo architectures. It only approximates the behavior of independent runs of the program. Application constraints will determine what approximation strategies are most appropriate for each problem. 2.9 Examples Here we give simple illustrations of the Venture language, including some standard modeling idioms as well as the use of custom inference instructions. Venture has been also used to implement applications of several advanced modeling and inference techniques; examples include generative probabilistic graphics programming (Mansinghka*, Kulkarni*, Perov, and Tenenbaum, 2013) and topic modeling (Blei et al., 2003). A description of these and other applications is beyond the scope of this paper. 2.9.1 H IDDEN M ARKOV M ODELS To represent a Hidden Markov model in Venture, one can use a stochastic recursion to capture the hidden state sequence, and index into it by a stochastic observation procedure. Here we give a variant with continuous observations, a binary latent state, and an a priori unbounded number of observation sequences to model: [ ASSUME observation_noise ( scope_include ’ hypers ’ unique ( gamma 1.0 1.0) )] [ ASSUME get_state ( mem ( lambda ( seq t) ( scope_include ’ state t ( if (= t 0) ( bernoulli 0.3) ( transition_fn ( get_state seq (- t 1) ))))))] [ ASSUME get_observation 7. The only subtlety is that transition operators must not change which random choice is being constrained, as this would require changing the weight of the trace. 16 ( mem ( lambda ( seq t) ( observation_fn ( get_state seq t))))] [ ASSUME transition_fn ( lambda ( state ) ( scope_include ’ state t ( bernoulli ( if state 0.7 0.3) )))] [ ASSUME observation_fn ( lambda ( state ) ( normal ( if state 3 -3) observation_noise ))] [ OBSERVE ( get_observation 1 1) 3.6] [ INFER ( mh default one 10) ] [ OBSERVE ( get_observation 1 2) -2.8] [ INFER ( mh default one 10) ] <... > This is a sequentialized variant of the ”default” resimulation-based Metropolis-Hastings inference scheme. If all but the last INFER statement were removed, the program would yield the same stochastic transitions as several Church implementations, but with linear (rather than quadratic) scaling in the length of the sequence. Interleaving inference with the addition of observations improves over bulk incorporation of observations by mitigating some of the strong conditional dependencies in the posterior. Another inference strategy is particle Markov chain Monte Carlo. For example, one could combine Metropolis-Hastings moves on the hyperparameters, given the latent states, with a conditional sequential Monte Carlo approximation to Gibbs over the hidden states given the hyper parameters and observations. Here is one implementation of this scheme, where 10 transitions of MetropolisHastings are done on the hyper parameters for every 5 transitions of approximate Gibbs based on 30 particles, all repeated 50 times: [ INFER ( cycle (( mh hypers one 10) ( pgibbs state ordered 30 5) ) 50) ] The global particle Gibbs algorithm from (Wood et al., 2014) with 30 particles would be expressed as follows: [ INFER ( pgibbs default ordered 30 100) ] Note that Metropolis-Hastings transitions can be more effective than pure conditional sequential Monte Carlo for handling global parameters (Andrieu et al., 2010). This is because MH moves allow hyperparameter inference to be constrained by all latent states. In real-time applications, hyperparameter inference is sometimes skipped. Here is one representation of a close relative8 of a standard 30-particle particle filter that uses randomly chosen hyper parameters and yields a single latent trajectory: [ INFER ( pgibbs state ordered 30 1) ] 2.9.2 H IERARCHICAL N ONPARAMETRIC BAYESIAN M ODELS Here we show how to implement one version of a multidimensional Dirichlet process mixture of Gaussians (Rasmussen, 1999): 8. The only difference is that a particle filter exposes all its weighted particles for forming Monte Carlo expectations or for rapidly obtaining a set of approximate samples. Only straightforward modifications are needed for Venture to expose a set of weighted traces instead of a single trace and literally recover particle filtering. 17 [ ASSUME alpha ( scope_include ’ hypers 0 ( gamma 1.0 1.0) )] [ ASSUME scale ( scope_include ’ hypers 1 ( gamma 1.0 1.0) )] [ ASSUME crp ( make_crp alpha )] [ ASSUME get_cluster ( mem ( lambda ( id ) ( scope_include ’ clustering id ( crp ))))] [ ASSUME get_mean ( mem ( lambda ( cluster dim ) ( scope_include ’ parameters cluster ( normal 0 10) )))] [ ASSUME get_variance ( mem ( lambda ( cluster dim ) ( scope_include ’ parameters cluster ( gamma 1 scale ))))] [ ASSUME get_component_model ( lambda ( cluster dim ) ( lambda () ( normal ( get_mean cluster dim ) ( get_variance cluster dim ))))] [ ASSUME get_datapoint ( mem ( lambda ( id dim ) (( get_component_model ( get_cluster id dim )))))] [ OBSERVE ( get_datapoint 0 0) 0.2] <... > ; default resimulation - based Metropolis - Hastings scheme [ INFER ( mh default one 100) ] The parameters are explicitly represented, i.e. “uncollapsed”, rather than integrated out as they often are in practice. While the default resimulation-based Metropolis-Hastings scheme can be effective on this problem, it is also straightforward to balance the computational effort differently: [ INFER ( cycle (( mh hypers one 1) ( mh parameters one 5) ( mh clustering one 5) ) 1000) ] On each execution of this cycle, one hyperparameter transition, five parameter transitions (each to both parameters of a randomly chosen cluster), and five cluster reassignments are made. Which hyperparameter, parameters and cluster assignments are chosen is random. As the number of data points grows, the ratio of computational effort devoted to inference over the hyperparameters and cluster parameters to the cluster assignments is higher than it is for the default scheme. Note that the complexity of this inference instruction, as well as the computational effort ratio, depend on the number of clusters in the current trace. It is also straightforward to use a structured particle Markov chain Monte Carlo scheme over the cluster assignments: [ INFER ( mixture ((0.2 ( mh hypers one 10) ) (0.5 ( mh parameters one 5) ) (0.3 ( pgibbs clustering ordered 2 2) )) 100) ] Due to the choice of only 2 particles for the pgibbs inference strategy, this scheme closely resembles an approximation to blocked Gibbs over the indicators based on a sequential initialization of the complete model. Also note that despite the low mixing weight on the clustering scope in the mixture, this inference program allocates asymptotically greater computational effort to inference over the cluster assignments than the previous strategy. This is because the pgibbs transition operator is guaranteed to reconsider every single cluster assignment. 18 2.9.3 I NVERSE I NTERPRETATION We now describe inverse interpretation, a modeling idiom that is only possible in Turing-complete languages. Recall that Venture modeling expressions are easy to represent as Venture data objects, and Venture models can invoke the evaluation and application of arbitrary Venture stochastic procedures. These Scheme-like features make it straightforward to write an evaluator — perhaps better termed a simulator — for a Turing-complete, higher-order probabilistic programming language. This application highlights Turing-completeness and also embodies a new potentially appealing path for solving problems of probabilistic program synthesis. In less expressive languages, learning programs (or structure) requires custom machinery that goes beyond what is provided by the language itself. In Venture, the same inference machinery used for state estimation or causal inference can be brought to bear on problems of probabilistic program synthesis. The dependency tracking and inference programming machinery that is general to Venture can be brought to bear on the problem of approximately Bayesian learning of probabilistic programs in a Venture-like language9 . We first define some utility procedures for manipulating references, symbols and environments: [ ASSUME make_ref ( lambda (x) ( lambda () x))] [ ASSUME deref ( lambda (x) (x))] [ ASSUME incremental_initial_environment ( lambda () ( list ( dict ( list ( quote bernoulli ) ( quote normal ) ( quote plus ) ( quote times ) ( quote branch )) ( list ( make_ref bernoulli ) ( make_ref normal ) ( make_ref plus ) ( make_ref times ) ( make_ref branch )))))] [ ASSUME extend_env ( lambda ( outer_env syms vals ) ( pair ( dict syms vals ) outer_env ))] [ ASSUME find_symbol ( lambda ( sym env ) ( if ( contains ( first env ) sym ) ( lookup ( first env ) sym ) ( find_symbol sym ( rest env ))))] The most interesting of these are make ref and deref. These use closures to pass references around the trace, using an idiom that avoids unnecessary growth of scaffolds. Consider an execution trace in which a value that is the argument to make ref becomes the principal node of a transition. Only those uses of the reference to the value that have been passed to deref will become resampling nodes. The value of the reference is unchanged, though the value the reference refers to is not. This permits dependence tracking through the construction of complex data structures. 9. Although we are still far from a study of the expressiveness of probabilistic languages via definitional interpretation (Reynolds, 1972; Abelson and Sussman, 1983), it seems likely that probabilistic programming formulations of probabilistic program synthesis — inference over a space of probabilistic programs, possibly including inference instructions, and an interpreter for those programs — will be revealing. 19 Given this machinery, it is straightforward to write an evaluator for a simple function language that has access to arbitrary Venture primitives: [ ASSUME incremental_venture_apply ( lambda ( op args ) ( eval ( pair op ( map_list deref args )) ( get_empty_environment )))] [ ASSUME incremental_apply ( lambda ( operator operands ) ( incremental_eval ( deref ( lookup operator 2) ) ( extend_env ( deref ( lookup operator 0) ) ( deref ( lookup operator 1) ) operands )))] [ ASSUME incremental_eval ( lambda ( expr env ) ( if ( is_symbol expr ) ( deref ( find_symbol expr env )) ( if ( not ( is_pair expr )) expr ( if (= ( deref ( lookup expr 0) ) ( quote lambda )) ( pair ( make_ref env ) ( rest expr )) (( lambda ( operator operands ) ( if ( is_pair operator ) ( incremental_apply operator operands ) ( incremental_venture_apply operator operands ))) ( incremental_eval ( deref ( lookup expr 0) ) env ) ( map_list ( lambda (x) ( make_ref ( incremental_eval ( deref x) env ))) ( rest expr )))))))] It is also possible to generate the input exprs using another Venture program, and use generalpurpose, Turing-complete inference mechanisms to explore a hypothesis space of expressions given constraints on the values that result. We call this the inverse interpretation approach to probabilistic program synthesis. As in Church — and contrary to (Liang et al., 2010) — inverse interpretation algorithms are not limited to rejection sampling. But while Church was limited to a single-site Metropolis-Hastings scheme, Venture programmers have more options. In Venture it is possible to associate portions of the program source (and portions of the induced program’s executions) with custom inference strategies. Here is an example expression grammar that can be used with the incremental evaluator: [ ASSUME genBinaryOp ( lambda () ( if ( flip ) ( quote plus ) ( quote times )))] [ ASSUME genLeaf ( lambda () ( normal 4 3) )] [ ASSUME genVar ( lambda (x) x)] [ ASSUME genExpr ( lambda (x) ( if ( flip 0.4) ( genLeaf ) ( if ( flip 0.8) ( genVar x) ( list ( make_ref ( genBinaryOp )) ( make_ref ( genExpr x)) ( make_ref ( genExpr x))))))] [ ASSUME noise ( gamma 5 .2) ] [ ASSUME expr ( genExpr ( quote x))] [ ASSUME f ( mem ( lambda (y) ( incremental_eval expr ( extend_env ( incremental_initial_environment ) 20 ( list ( quote x)) ( list ( make_ref y))))))] [ ASSUME g ( lambda (z) ( normal (f z) noise ))] [ OBSERVE (g 1) 10] ;f(x) = 5x + 5 [ OBSERVE (g 3) 20] <... > Indirection via references substantially improves the asymptotic scaling of programs like these. When a given production rule in the grammar is resimulated, only those portions of the execution of the program that depend on the changed source code are resimulated. A naive implementation of an evaluator would not have this property. Scaling up this approach to larger symbolic expressions and small programs will require multiple advances. Overall system efficiency improvements will be necessary. Inverse interpretation also may benefit from additional inference operators, such as Hamiltonian Monte Carlo for the continuous parameters. Expression priors with inference-friendly structures would also help. For example, a prior where resimulation recovers some of the search moves from (Duvenaud et al., 2013) may be expressible by separately generating symbolic expressions and the contents of the environment into which they are evaluated. Longer term, it may be fruitful to explore formalizations of some of the knowledge taught to programmers using probabilistic programming. 3. Stochastic Procedures Random choices in Venture programs arise due to the invocation of stochastic procedures (SPs). These stochastic procedures accept input arguments that are values in Venture and sample output values given those inputs. Venture includes a built-in stochastic procedure library, which includes SPs that construct other SPs, such as the SP make_csp which the special form lambda gets desugared to. Stochastic procedures can also be added as extensions to Venture, and provide a mechanism for incremental optimization of Venture programs. Model fragments for which Venture delivers inadequate performance can be migrated to native inference code that interoperates with the enclosing Venture program. 3.1 Expressiveness and extensibility Many typical random variables, such as draws from a Bernoulli or Gaussian distribution, are straightforward to represent computationally. One common approach has been to use a pair of functions: a simulator that maps from an input space of values X and a stream of random bits {0, 1}∗ to an output space of values Y , and a marginal density that maps from (x, y) pairs to [0, ∞). This representation corresponds to the “elementary random procedures” supported by early Church implementations. Repeated invocations of such procedures correspond to IID sequences of random variables whose densities are known. While simple and intuitive, this simple interface does not naturally handle many useful classes of random objects. In fact, many objects that are easy to express as compound procedures in Church and in Venture cannot be made to fit in this form. Stochastic procedures in Venture support a broader class of random objects: 1. Higher-order stochastic procedures, such as mem (including stochastic memorization), apply and map. Higher-order procedures may accept procedures as arguments, apply these 21 procedures internally, and produce procedures as return values. Stochastic procedures in Venture are equipped with a simple mechanism for handling these cases. In fact, it turns out that all structural changes to execution traces — including those arising from the execution of constructs that affect control flow, such as IF — can be handled by this mechanism. This simplifies the development of inference algorithms, and permits users to extend Venture by adding new primitives that affect the flow of control. 2. Stochastic procedures whose applications are exchangeably coupled. Examples include collapsed representations of conjugate models from Bayesian statistics, combinatorial objects from Bayesian nonparametrics such as the Chinese Restaurant and Pitman-Yor processes, and probabilistic sequence models (such as HMMs) whose hidden state sequences can be efficiently marginalized out. Support for these primitives whose applications are coupled is important for recovering the efficiency of manually optimized samplers, which frequently make use of collapsed representations. Whereas exchangeable primitives in Church are thunks, which prohibits collapsing many important models such as HMMs, Venture supports primitives whose applications are row-wise partially exchangeable across different sets of arguments. The formal requirement is that the cumulative log probability density of any sequence of input-output pairs is invariant under permutation. 3. Likelihood-free stochastic procedures that lack tractable marginal densities. Complex stochastic simulations can be incorporated into Venture even if the marginal probability density of the outputs of these simulations given the inputs cannot be efficiently calculated. Models from the literature on Approximate Bayesian Computation (ABC), where priors are defined over the outcome of forward simulation code, can thus be naturally supported in Venture. Additionally, a range of “doubly intractable” inference problems, including applications of Venture to reasoning about the behavior of approximately Bayesian reasoning agents, can be included using this mechanism. 4. Stochastic procedures with external latent variables that are invisible to Venture. There will always be models that admit specialized inference strategies whose efficiency cannot be recovered by performing generic inference on execution traces. One of the principal design decisions in Venture is to allow these strategies to be exploited whenever possible by supporting a broad class of stochastic procedures with custom inference over internal latent variables, hidden from the rest of Venture. The stochastic procedure interface thus serves as a flexible bridge between Venture and foreign inference code, analogous to the role that foreign function interfaces (FFIs) play in traditional programming languages. 3.2 Primitive stochastic procedures Informally, a primitive stochastic prodecure (PSP) is an object that can simulate from a family of distributions indexed by some arguments. In addition to simulating, PSPs may be able to report the logdensity of an output given an input, and may incorporate and unincorporate information about the samples drawn from it using mutation, e.g. in the case of a conjugate prior. This mutation cannot be aribitrary: the draws from the PSP must remain row-wise partially exchangeable as discussed above. A PSP may also have custom proposal kernels, in which case it must be able to return its contribution to the Metropolis-Hastings acceptance rate. For example, the PSP that simulates 22 Gaussian random variables may provide a drift kernel that proposes small steps around its previous location, rather than resampling from the prior distribution. Primitive stochastic procedures are parameterized by the following properties and behaviors: 1. isStochastic() — does this PSP consume randomness when it is invoked? 2. canAbsorbArgumentChanges() — can this PSP absorb changes to its input arguments? If true, then this PSP must correctly implement logdensity() as described below. 3. childrenCanAbsorbAtApplications() — does this PSP return an SP that implements the short-cut “absorbing at applications” optimization, needed to integrate optimized expressions for the log marginal probability of sufficient statistics in standard conjugate models. 4. value = simulate(args) — samples a value for an application of the PSP, given the arguments args 5. logp = logdensity(value, args) — an optional procedure that evaluates the log probability density10 of an output given the input arguments p psp (simulate(args) = value|args). 6. incorporate(aux, value, args) — incorporate the value stored in the variable named value into the auxiliary storage aux associated with the SP that contains the PSP. incorporate() is used to implement SPs whose applications are exchangeably coupled. While it is always sufficient to store and update the full set of values returned for each observed args, often only the counts (or some other sufficient statistics) are necessary. 7. unincorporate(aux, value, args) — remove value from the auxiliary storage, restoring it to a state consistent with all other values that have been incorporated but not unincorporated; this is done when an application is unevaluated. 3.3 The Stochastic Procedure Interface The stochastic procedure interface specifies the contract that Venture primitives must satisfy to preserve the coherence of Venture’s inference mechanisms. It also serves as the vehicle by which external inference systems can be integrated into Venture. This interface preserves the ability of primitives to dynamically create and destroy internal latent variables hidden from Venture and to perform custom inference over these latent variables. 3.3.1 D EFINITION Definition 3.1 (Stochastic procedure). A stochastic prodecure is a pair (request, output) of PSPs, along with a latent variable simulator, where 1. request returns: 10. This density is implicitly defined with respect to a PSP-specific (but argument independent) choice of dominating measure. For PSPs that are guaranteed to produce discrete outputs, the measure is assumed to be the counting measure, so logdensity is equivalent to the log probability mass function. A careful measure-theoretic treatment of Venture is left for future work. 23 (a) A list of tuples (addr, expr, env) that represent expressions whose values must be available to output before it can start its simulation. We refer to requests of this form as exposed simulation requests (ESRs), (b) A list of opaque tokens that can be interpreted by the SP as the latent variables that output will need in order to simulate its output, along with the values of the exposed simulation requests. We refer to requests of this form as latent simulation requests (LSRs). 2. The latent variable simulator responds to LSRs by simulating any latent variables requested. 3. output returns the final output of the procedure, conditioned on the inputs, the results of any of the exposed simulation requests, and the results of any latent simulation requests. 3.3.2 E XPOSED S IMULATION R EQUESTS We want our procedures to be able to pass (expr, env) pairs to Venture for evaluation, and make use of the results in some way. A procedure may also have multiple applications all make use of a shared evaluation, e.g. mem, and in these cases the procedure must take care to request the same addr each time, and the (expr, env) will only be evaluated the first time and then reused thereafter. Specifically, an ESR request of the (addr, expr, env) is handled by Venture as follows. First Venture checks the requesting SP’s namespace to see if it already has an entry with address addr. If it does not, then Venture evaluates expr in env, and adds the mapping addr → root to the SP’s namespace, where root is the root of the evaluation tree. If the SP’s namespace does contain addr, then Venture can look up root. Either way, Venture wires in root as an extra argument to the output node. 3.3.3 L ATENT S IMULATION R EQUESTS AND THE F OREIGN I NFERENCE I NTERFACE Some procedures may want to simulate and perform inference over latent variables that are hidden from Venture. Consider an optimized implementation of a hidden Markov model integrated into the following Venture program: [ ASSUME my_hmm ( make_hmm 10 0.1 5 0.2) ] [ OBSERVE ( my_hmm 0 0) 0] [ OBSERVE ( my_hmm 0 1) 1] [ OBSERVE ( my_hmm 0 2) 0] [ INFER ( mh default one 20) ] [ PREDICT ( my_hmm 0 3) ] [ PREDICT ( my_hmm 1 0) ] Here we have a constructor SP (make hmm <num-states> <transition-hyper> <num-output-symbols> <observation-hyper) that generates an (uncollapsed) hidden Markov model by generating the rows of the transition and observation matrices at random. The assumption is that make hmm is a primitive, although it would be straightforward to implement make hmm as a compound procedure in Venture, using variations of the examples presented earlier. The constructor returns a procedure bound to the symbol my hmm that permits observations from this process to be queried via (my hmm <sequence-id> <index>). The program adds a sequence fragment of length three then requests predictions for the next observation in the sequence, as well as the initial observation from an entirely new sequence. 24 This probabilistic program captures a common pattern: integrating Venture with an foreign probabilistic model fragment that can be dynamically queried and contains latent variables hidden from Venture. It is useful to partition the random choices in this program as follows. The transition and observation matrices of the HMM could be viewed as part of the value of the my hmm SP and therefore returned by the output PSP of the make hmm SP. The observations are managed by Venture as the applications of the my hmm SP. The hidden states, however, are fully latent from the standpoint of Venture, yet need to be created, updated and destroyed as invocations of my hmm are created or destroyed and as their arguments (or the arguments to make hmm) change. Venture makes it possible for procedures to instantiate latent variables only as necessary to simulate a given program. The mechanism is similar to that for exposed simulation requests, except in this case the requests–which we call latent simulation requests (LSRs)–are opaque to Venture. Venture only calls appropriate methods on the SP at appropriate times to ensure that all the bookkeeping is handled correctly. This framework is straightforward to apply to make hmm and the stochastic procedure(s) that it returns. If my hmm is queried for an observation at time t, the requestPSP can return the time t as an LSR. When Venture tells the HMM to simulate that LSR, the HMM will either do nothing if xt already exists in its internal store of simulations, or else continue simulating from its current position up until xt . The outputPSP then samples ot conditioned on the latent xt . If the application at time t is ever unevaluated, Venture will tell the HMM to detach the LSR t, which will cause the HMM to place the latents that are no longer necessary to simulate the program into a “latentDB”, which it returns to Venture. Later on, Venture may tell the HMM to restore latents from a latentDB, for example if a proposal is rejected and the starting trace is being restored. The main reason to encapsulate latent variables in this way, as opposed to requesting them as ESRs, is so that the SP can use optimized implementations of inference over their values, potentially utilizing special-purpose inference methods. For example, the uncollapsed HMM can implement forwards-filtering backwards-sampling to efficiently sample the latent variables conditioned on all observations. Such procedures are integrated into Venture by defining an “Arbitrary Ergodic Kernel” (AEKernel) which Venture may call during inference, and which is simply a black-box to Venture. Note that this same mechanism may be used by SPs that do not make latent simulation requests at all, but which have latent variables instantiated upon creation, such as a finite-time HMM or an uncollapsed Dirichlet-Multinomial. To implement this functionality, stochastic procedures must implement three procedures in addition to the procedures needed for their ESR requestor, LSR requestor and output PSPs: 1. simulateLatents(aux, LSR, shouldRestore, latentDB) — simulate the latents corresponding to LSR, using the tokens in latentDB (indexed by LSR) to find a previous value if shouldRestore is true. 2. detachLatents(aux, LSR, latentDB) — signal that the latents corresponding to the request LSR are no longer needed, and store enough information in latentDB so that the value can be recovered later. 3. AEInfer(aux) — Trigger the external implementation to perform inference over all latent variables using the contents of aux. It is often convenient for simulateLatents to store latent variables in the aux and for incorporate to store the return values of applications in tt aux, along with the arguments that produced them. 25 Examples of this use of the stochastic procedure interface can be found in current releases of the Venture system. 3.3.4 O PTIMIZATIONS FOR HIGHER - ORDER SP S Venture provides a special mechanism that allows certain SPs to exploit the ability to quickly compute the logdensity of its applications. Consider the following program: [ASSUME alpha (gamma 1 1)] [ASSUME collapsed_coin (make_beta_bernoulli alpha alpha)] [OBSERVE (collapsed_coin) False] [OBSERVE (collapsed_coin) True] <repeat 10ˆ9 times> [OBSERVE (collapsed_coin) False] [INFER] A hand-written inference scheme would only keep track of the counts of the observations, and could perform rapid Metropolis-Hastings proposals on alpha by exploiting conjugacy. On the other hand, a naive generic inference scheme might visit all one billion observation nodes to compute the acceptance ratio for each proposal. We can achieve this efficient inference scheme by letting make beta bernoulli be responsible for tracking the sufficient statistics from the applications of collapsed coin, and for evaluating the log density of all those applications as a block. Stochastic procedures that return other stochastic procedures and implement this optimization are said to be absorbing at applications, often abbreviated AAA. We will discuss techniques for implementing this mechanism in a later section. 3.3.5 AUXILIARY S TATE An SP is itself stateless, but may have an associated auxiliary store, called SPAux, that carries any mutable information. SPAuxs have several uses: 1. If an SP makes exposed simulation requests, Venture uses the SPAux to store mappings from the addresses of the ESRs to the node that stores the result of that simulation. 2. If a PSP keeps track of its sample counts or other sufficient statistics, such as the collapsedbeta-bernoulli which stores the number of trues and falses, it will store this information in the SPAux. 3. If an SP makes latent simulation requests, then all latent variables it simulates to respond to those requests are stored in the SPAux. 4. Some SPs may optionally store part of the value of the SP directly in the SPAux. This is necessary if an SP cannot easily store its value, its latents, and its outputs separately. 4. Probabilistic Execution Traces Bayesian networks decompose probabilistic models into nodes for random variables, equipped with conditional probability tables, and edges for conditional dependencies. They can be interpreted as 26 probabilistic programs via the ancestral simulation algorithm (Frey, 1997). The network is traversed in an order consistent with the topology, and each random variable is generated given its immediate parents. Bayesian networks can also be viewed as expressing a function for evaluating the joint probability density of all the nodes in terms of a factorization given by the graph structure. Here we describe probabilistic execution traces (PETs), which serve analogous functions for Venture programs and address the additional complications that arise from Turing-completeness and the presence of higher order probabilistic procedures. We also describe recursive procedures for constructing and destroying probabilistic execution traces as Venture modeling language expressions are evaluated and unevaluated. 4.1 Definition of a probabilistic execution trace Probabilistic execution traces consist of a directed graphical model representing the dependencies in the current execution, along with the stateful auxiliary data for each stochastic procedure, the Venture program itself, and metadata for existential dependencies and exchangeable coupling. We will typically identify executions and PETs, and denote them via the symbols ρ or ξ. PETs contain the following nodes: 1. One constant node for the global environment. 2. One constant node for every value bound in the global environment, which includes all builtin SPs. 3. One constant node for every call to eval on a expression that is either self-evaluating or quoted. 4. One lookup node for every call to eval that triggers a symbol lookup. 5. One request node and one output node for every call to eval that triggers an SP application. We refer to the operator nodes and operand nodes of request nodes and output nodes, but note that these are not special node types. PETs also contain the following edges: 1. One lookup edge to each lookup node from the node it is looking up. 2. One operator edge to every request node from its operator node. 3. One operator edge to every output node from its operator node. 4. One operand edge to every request node from each of its operand nodes. 5. One operand edge to every output node from each of its operand nodes. 6. One requester edge to every output node from its corresponding request node. 7. One ESR edge from the root node of every SP family to each SP application that requests it. Every node represents a random variable and has a value that cannot change during an execution. The PET also includes the SPAux for every SP that needs one. Unlike the values in the nodes, the SPAuxs may be mutated during an execution, for example to increment the number of trues for a beta-bernoulli. 27 4.2 Families We divide our traces into families: one Venture family for every assume, predict, observe directive, and one SP family for every unique ESR requested during forward simulation. Because of our uniform treatment of conditional simulation, executions of programs satisfy the following property: the structure of every family is a function of the expression only, and does not depend on the random choices made while evaluating that expression. The only part of the topology of the graph that can change is which ESRs are requested. 4.3 Exchangeable coupling Given our exchangeability assumptions for SPs, we can cite (generalized) de Finetti (Orbanz and Roy) to conclude that there is, in addition to the observed random variables explicitly represented in the PET, one latent random variable θ f for every SP f in the PET corresponding to the unobserved de Finetti measure, and one latent variable θ f ,args for every set of arguments args that f is called on, with an edge from f ’s maker-node to θ f , edges from θ f to every θ f ,args , and an edge from θ f ,args to every node corresponding to an application of the form f (args). However, each θ f and θ f ,args is marginalized out by each f by way of mutation on its SPAux, effectively introducing a hyperedge that indirectly couples the application nodes of all applications of f . We have chosen not to represent these dependencies in the graphical structure of the PET for the following reasons. First, once we integrate out θ f and θ f ,args , we can only encode these dependencies at all with directed edges once we fix a specific ordering for the applications. Second, for the orderings we will be interested in, the graph that combines both types of directed edges would be cyclic. This would complicate future efforts to develop a causal semantics for PETs and Venture programs. 4.4 Existential dependence and contingent evaluation An SP family is existentially dependent on the nodes that request it as part of an ESR, in the sense that if at any point it is not requested by any nodes, then the family would not have been computed while simulating the PET. Existential dependence is then handled with garbage collection semantics, whereby an SP family can no longer be part of the PET if it is not selected by any W nodes, and thus should be unevaluated. 4.5 Examples Here we briefly give example PETs for simple Venture programs. 4.5.1 T RICK COIN This is a variant of our running example. It defines a model that can be used for inferences about whether or not a coin is tricky, where a trick coin is allowed to have any weight between 0 and 1. The version we use includes one observation that a single flip of the coin came up heads. To keep the PET as simple as possible, we give both the program and the PET in the form where IF has already been desugared into an SP application: (IF <predicate> <consequent> <alternate>) has been replaced with (branch <predicate> (quote <consequent>) (quote <alternate>)), where branch is an ordinary stochastic procedure. 28 exp: (bernoulli weight) value: #t exp: (bernoulli weight) value: #t request: [] exp: weight value: 0.500000 request: [] exp: bernoulli value: sp:bernoulli exp: weight value: 0.614839 exp: bernoulli value: sp:bernoulli 3 3 exp: (branch_exp coin_is_tricky (quote (beta 1.000000 1.000000)) (quote 0.500000)) value: 0.500000 exp: (quote 0.500000) value: 0.500000 exp: (quote (beta 1.000000 1.000000)) value: (beta 1.000000 1.000000) exp: (branch_exp coin_is_tricky (quote (beta 1.000000 1.000000)) (quote 0.500000)) value: 0.614839 exp: 0.500000 value: 0.500000 request: [(160256752, 0.500000, <env>)] exp: coin_is_tricky value: #f exp: (beta 1.000000 1.000000) value: 0.614839 request: [(160256752, (beta 1.000000 1.000000), <env>)] exp: branch_exp value: sp:branch exp: (quote 0.500000) value: 0.500000 2 exp: (quote (beta 1.000000 1.000000)) value: (beta 1.000000 1.000000) exp: coin_is_tricky value: #t exp: branch_exp value: sp:branch request: [] 2 exp: (bernoulli 0.100000) value: #f exp: (bernoulli 0.100000) value: #t 158928592 request: [] exp: 0.100000 value: 0.100000 158928592 request: [] exp: bernoulli value: sp:bernoulli exp: 0.100000 value: 0.100000 1 exp: 1.000000 value: 1.000000 155761760 exp: bernoulli value: sp:bernoulli 1 Venture Families (a) exp: beta value: sp:beta Venture Families 158011216 (b) 158011216 Figure 1: The two different PET structures corresponding to the trick coin program. (a) An execution trace where the coin is fair. (b) An execution trace where the coin is tricky, containing additional nodes that depend existentially on the coin flip. [ASSUME coin_is_tricky (bernoulli 0.1)] [ASSUME weight (branch coin_is_tricky (quote (beta 1.0 1.0)) (quote 0.5))] [OBSERVE (bernoulli weight) true] Figure 1 shows the two PET structures that can arise from simulating this program, along with arbitrarily chosen values. 4.5.2 A SIMPLE BAYESIAN NETWORK Figure 2 shows a PET for the following program, implementing a simple Bayesian network: [ASSUME rain (bernoulli 0.2)] [ASSUME sprinkler (bernoulli (branch rain 0.01 0.4))] [ASSUME grassWet (bernoulli (branch rain (branch sprinkler 0.99 0.8) (branch sprinkler 0.9 0.00001)))] [OBSERVE grassWet True] Note that each of the Venture families in the PET corresponds to a node in the Bayesian network. A coarsened version of the PET contains the same conditional dependence and independence information as the Bayesian network would. 4.5.3 S TOCHASTIC MEMOIZATION We now illustrate a program that exhibits stochastic memoization. This program constructs a stochastically memoized procedure and then applies it three times. Unlike deterministic memoization, a stochastically memoized procedure has a stochastic request PSP which sometimes returns a previously sampled value, and sometimes samples a fresh one. These random choices follow a Pitman-Yor process. Figure 3 shows a PET corresponding to a typical simulation; note the overlapping requests. [ASSUME f (pymem bernoulli 1.0 0.1)] 29 exp: 1.000000 value: 1.000000 exp: grassWet value: #t 4 exp: (bernoulli (branch_exp rain (branch_exp sprinkler 0.990000 0.800000) (branch_exp sprinkler 0.900000 0.000010))) value: #t request: [] exp: bernoulli value: sp:bernoulli exp: (branch_exp rain (branch_exp sprinkler 0.990000 0.800000) (branch_exp sprinkler 0.900000 0.000010)) value: 0.000010 exp: 0.000010 value: 0.000010 request: [(160484384, 0.000010, <env>)] exp: branch_exp value: sp:branch exp: 0.000010 value: 0.000010 exp: (branch_exp sprinkler 0.900000 0.000010) value: 0.000010 exp: (branch_exp sprinkler 0.990000 0.800000) value: 0.800000 request: [(159586680, 0.000010, <env>)] exp: 0.900000 value: 0.900000 exp: branch_exp value: sp:branch exp: 0.000010 value: 0.000010 exp: rain value: #f exp: 0.800000 value: 0.800000 request: [(159470160, 0.800000, <env>)] exp: sprinkler value: #f exp: branch_exp value: sp:branch exp: 0.800000 value: 0.800000 exp: 0.990000 value: 0.990000 exp: sprinkler value: #f 3 exp: (bernoulli (branch_exp rain 0.010000 0.400000)) value: #f request: [] exp: (branch_exp rain 0.010000 0.400000) value: 0.400000 exp: 0.400000 value: 0.400000 exp: bernoulli value: sp:bernoulli request: [(158076656, 0.400000, <env>)] exp: 0.400000 value: 0.400000 exp: 0.010000 value: 0.010000 exp: branch_exp value: sp:branch exp: rain value: #f 2 exp: (bernoulli 0.200000) value: #f 158947952 request: [] exp: 0.200000 value: 0.200000 exp: bernoulli value: sp:bernoulli 1 Venture Families 159421304 Figure 2: The PET corresponding to a three-node Bayesian network. Each of the three numbered families correspond to a node in the Bayesian network, capturing the execution history needed to simulate the node given its parents. 30 exp: (f) value: #t exp: (f) value: #f exp: (f) value: #t request: [(0, (pymemoizedSP), <env>)] request: [(1, (pymemoizedSP), <env>)] request: [(0, <exp>, <env>)] exp: f value: sp:pymem exp: f value: sp:pymem 2 exp: f value: sp:pymem 3 exp: (pymemoizedSP) value: #f exp: (pymemoizedSP) value: #t request: [] request: [] exp: pymemoizedSP value: sp:bernoulli exp: pymemoizedSP value: sp:bernoulli 4 exp: (pymem bernoulli 1.000000 0.100000) value: sp:pymem request: [] exp: 1.000000 value: 1.000000 exp: pymem value: sp:sp_no_name exp: 0.100000 value: 0.100000 exp: bernoulli value: sp:bernoulli 1 Venture Families 158967384 159411912 Figure 3: A PET corresponding to an execution of a program with stochastic memoization. The procedure being stochastically memoized, bernoulli, is applied twice, based on the value of the requests arising in invocations of (f). [PREDICT (f)] [PREDICT (f)] [PREDICT (f)] 4.6 Constructing PETs via forward simulation Let P be a constrained program. Venture’s primary inference strategies require an execution of P with positive probability before they can even begin. Therefore the first thing we do is simply evaluate the program by interpreting it in a fairly standard way, except with all conditional evaluation handled uniformly through the ESR machinery presented above. For simplicity, we elide details related to inference scoping. 4.6.1 P SEUDOCODE FOR EVAL, APPLY AND EVAL-REQUESTS Evaluation is generally similar to a pure Scheme, but there are a few noteworthy differences. First, evaluation creates nodes for every recursive call, and connects them together to form the directed graph structure of the probabilistic execution trace. Second, there are no distinctions between primitive procedures and compound procedures. We call the top-level evaluation procedure E VAL FAM ILY to emphasize the family block structure in PETs. Third, a scaffold and a database of random choices db are threaded through the recursions, to support their use in inference, including restoring trace fragments when a transition is rejected by reusing random choices from the db. Note that an environment model evaluator (Abelson and Sussman, 1983) is used, even though the underlying language is pure, with the PET storing both the environment structure and the recursive invocations of eval. 31 E VAL FAMILY(trace, exp, env, scaffold, db) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 if isSelfEvaluating?(exp) return (0,trace.createConstantNode(exp)) elseif isQuoted?(exp) return (0,trace.createConstantNode(TextOfQuotation(exp))) elseif isVariable?(exp) sourceNode = env.findSymbol(exp) R EGENERATE(trace, exp, scaffold, FALSE, db) return (0,trace.createLookupNode(sourceNode) else weight, operatorNode = E VAL FAMILY(trace, exp.operator, env, scaffold, db) operandNodes = [] for operand in exp.operands (w, operandNode) = E VAL FAMILY(trace, operand, env, scaffold, db) weight = weight + w operandNodes.append(operandNode) (requestNode, outputNode) = trace.createApplicationNodes(operatorNode, operandNodes, env) weight = weight + A PPLY SP(trace, requestNode, outputNode, scaffold, FALSE, db) return (weight, outputNode) The call to R EGENERATE ensures that if E VAL FAMILY is called in the context of a pre-existing PET, it is traversed in an order that is compatible with the dependence structure of the program. We will sometimes refer to such orders as “evaluation-consistent”. From the standpoint of forward simulation, however, R EGENERATE can be safely ignored. A PPLY SP(trace, requestNode, outputNode, scaffold, restore?, db) 1 2 3 4 5 weight = A PPLY PSP(trace, requestNode, scaffold, restore?, db) weight = weight + E VAL R EQUESTS(trace, requestNode, scaffold, restore?, db) weight = weight + R EGENERATE ESRPARENTS(trace, outputNode, scaffold, restore?, db) weight = weight + A PPLY PSP(trace, outputNode, scaffold, restore?, db) return weight Stochastic procedures are allowed to request evaluations of expressions11 ; this enables higherorder procedures as well as the encapsulation of custom control flow constructs. For example, compound procedures do this to evaluate their body in an environment extended with their formal parameters and argument values. 11. To prevent this flexibility from introducing arbitrary dependencies, the expressions and environments are restricted to those constructible from the arguments to the procedure or the procedure that constructed it. 32 E VAL R EQUESTS(trace, node, scaffold, restore?, db) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 sp, spaux = trace.getSP(node),trace.getSPAux(node) weight = 0 request = trace.getValue(node) for id, exp, env, block, subblock in request.esrs if not spaux.containsFamily(id) if restore? esrParent = db.getESRParent(sp, id) weight = weight + R ESTORE FAMILY(trace, esrParent, scaffold, db) else (w, esrParent) = E VAL FAMILY(trace, exp, env, scaffold, db) weight = weight + w trace.registerFamily(node, id, esrParent) else weight = weight + R EGENERATE(trace, spaux.getFamily(id), scaffold, restore?, db) esrParent = spaux.getFamily(id) if not block == N ONE trace.registerBlock(block, subblock, esrParent) trace.addESREdge(esrParent, node.outputNode) for lsr in request.lsrs if db.hasLatentDBFor(sp) latentDB = db.getLatentDB(sp) else latentDB = N ONE weight = weight + sp.simulateLatents(spaux, lsr, restore?, latentDB) return weight Stochastic procedures are also allowed to perform opaque operations to produce outputs given the value of their arguments and any requested evaluations. Note that this may result in random choices being added to the record maintained by the trace. 33 A PPLY PSP(trace, node, scaffold, restore?, db) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 psp, args = trace.getPSP(node),trace.getArgs(node) if db.hasValueFor(node) oldValue = db.getValue(node) else oldValue = N ONE // Determine new value if shouldRestore newValue = oldValue elseif scaffold.hasKernelFor(node) newValue = scaffold.getKernel(node).simulate(trace, oldValue, args) else newValue = psp.simulate(args) // Determine weight if scaffold.hasKernelFor(node) weight = scaffold.getKernel(node).weight(trace, newValue, oldValue, args) else weight = 0 trace.setValue(node, newValue) psp.incorporate(newValue, args) if isSP?(newValue) PROCESS M ADE SP(trace, node, scaffold.isAAA?(node)) if psp.isRandom?() trace.registerRandomChoice(node) return weight This functionality is supported by book-keeping to link new stochastic procedures into the trace, and register any customizations they implement with the trace so that inference transitions can make use of them: P ROCESS M ADE SP(trace, node, isAAA?) 1 sp = trace.getValue(node) 2 trace.setMadeSP(node, sp) 3 trace.setValue(node, spref(sp)) 4 if not isAAA? 5 trace.setMadeSPAux(node, sp.constructSPAux()) 6 if sp.hasAEKernel() 7 trace.registerAEKernel(node) 4.7 Undoing simulation of PET fragments Venture also includes requires unevaluation procedures that are dual to the evaluation procedures described earlier. This is because PET fragments need to be removed from the trace in two situations. First, when FORGET instructions are triggered, the expression corresponding to the directive is removed. Second, during inference, changes to the values of certain requestPSPs make cause some 34 SP families to no longer be requested. In the first case, all of the random choices are permanently removed, whereas in the second case, the random choices may need to be restored if the proposal is rejected. 4.7.1 P SEUDOCODE FOR UNEVAL, UNAPPLY AND UNEVAL-REQUESTS When a trace fragment is unevaluated, we must visit all application nodes in the trace fragment so that the PSPs have a chance to unincorporate the (input,output) pairs. The operations needed to do this are essentially inverses of the simulation procedures described above, designed to visit nodes in the reverse order that evaluation does, to ensure compatibility with exchangeable coupling. Here we give pseudocode for these operations, eliding the details of garbage collection12 : U NEVAL FAMILY(trace, node, scaffold, db) 1 2 3 4 5 6 7 8 9 10 11 if node.isConstantNode? weight = 0 elseif node.isLookupNode? trace.disconnectLookup(node) weight = E XTRACT(trace, node.sourceNode, scaffold, db) else weight = U NAPPLY SP(trace, node, scaffold, FALSE, db) for operandNode in node.operandNodes weight = weight + U NEVAL FAMILY(trace, operandNode, scaffold, db) weight = weight + U NEVAL FAMILY(trace, node.operatorNode, scaffold, db) return weight To unapply a stochastic procedure, Venture must undo its random choices and also unapply any requests it generated: U NAPPLY SP(trace, node, scaffold, db) 1 2 3 4 5 weight = U NAPPLY PSP(trace, node, scaffold, db) weight = weight + EXTRACT ESRPARENTS(trace, outputNode, scaffold, db) weight = weight + U NEVAL R EQUESTS(trace, node.requestNode, scaffold, db) weight = weight + U NAPPLY PSP(trace, node.requestNode, scaffold, db) return weight To unapply a primitive stochastic procedure, we remove its random choices from the trace, unincorporate them, and update the weight accordingly. Note that we store the value in the db so that we can restore it later if necessary. 12. Previous work has anecdotally explored the possibility of preserving unused trace fragments and treating them as auxiliary variables, following the treatment of component model parameters from Algorithm 8 in (Neal, 1998). In theory, this delay of garbage collection, where multiple copies of each trace fragment are maintained for each branch point, could support adaptation to the posterior. A detailed empirical evaluation of these strategies is pending a comprehensive benchmark suite for Venture as well as a high-performance implementation. 35 U NAPPLY PSP(trace, node, scaffold, db) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 psp, args = trace.getPSP(node),trace.getArgs(node) if psp.isRandom() trace.unregisterRandomChoice(node) if isSP(trace.getValue(node)) TEARDOWN M ADE SP(trace, node, scaffold.isAAA(node)) oldValue = trace.getValue(node) psp.unincorporate(oldValue, args) if scaffold.hasKernelFor(node) weight = scaffold.getKernel(node).reverseWeight(trace, oldValue, args) else weight = 0 db.extractValue(node, oldValue) trace.clearValue(node) return weight Handling of requests during unevaluation involves two additional subtleties. First, latent random choices must be handled appropriately. Second, requests must be unevaluated only if no other application of the SP refers to them. U NEVAL R EQUESTS(trace, node, scaffold, db) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 sp, spaux = trace.getSP(node),trace.getSPAux(node) weight = 0 request = trace.getValue(node) if request.lsrs and not db.hasLatentDB(trace.getSP(node)) db.registerLatentDB(sp, sp.constructLatentDB()) for lsr in reversed(request.lsrs) weight = weight + sp.detachLatents(spaux, lsr, db.getLatentDB(sp)) for id, exp, env in reversed(request.esrs) esrParent = trace.popLastESRParent(node.outputNode) if trace.getNumberOfRequests(esrParent) == 0 trace.unregisterFamily(node, id) db.registerFamily(sp, id, esrParent) weight = weight + U NEVAL FAMILY(trace, esrParent, scaffold, db) return weight When an SP is finally destroyed — i.e. when the output PSP of its maker SP application is unapplied — then the SP can be destroyed and its auxiliary storage can be garbage collected. Because unapplication of the maker must happen after all applications of the made SP have been unapplied — due to the constraint that unevaluation visits nodes in the reverse order of evaluation and regeneration — there is no information in the auxiliary storage that needs to be preserved in the latent DB. 36 T EARDOWN M ADE SP(trace, node, isAAA?) 1 sp = trace.getMadeSP(node) 2 trace.setValue(node, sp) 3 trace.clearMadeSP(node) 4 if not isAAA? 5 if sp.hasAEKernel() 6 trace.unregisterAEKernel(node) 7 trace.clearMadeSPAux(node) 4.8 Enforcing constraints via CONSTRAIN and UNCONSTRAIN Unfortunately, the observations may not have positive probability given the program execution. Venture may not happen to sample the right data, and but if noise is added to each data generating expression to ensure positive probability, it may take a long time for Venture to incorporate the observed data at all. On the other hand, the general problem of finding a program execution for which some output has positive probability is intractable in general – it can encode SAT without even making use of contingent evaluation or special primitives – and Venture is not designed to attempt to invert such programs. Venture implements a middle ground. It is designed with sufficiently stochastic probabilistic programs in mind, and in particular is not a SAT-solver. However, in order to support common cases from Bayesian statistics, we do support inverting a simple kind of determinism. We introduce a method constrain(<directive>,<value>) that recursively walks backwards along id-on-requests edges to find the outermost application of a different kind of PSP. If the constrained value has positive probability at this PSP application, we constrain the value and we are done. Otherwise, we unevaluate the entire program and evaluate it again, in hopes that this execution will be constrainable. Simple versions of constrain and unconstrain can be implemented as follows: C ONSTRAIN(trace, node, value) 1 if node.isLookupNode? 2 return C ONSTRAIN(trace, node.sourceNode, value) 3 elseif trace.getPSP(node).isESRReference? 4 return C ONSTRAIN(trace,trace.getESRParents(node)[0], value) 5 else 6 psp, args = trace.getPSP(node),trace.getArgs(node) 7 psp.unincorporate(trace.getValue(node), args) 8 weight = psp.logDensity(value, args) 9 trace.setValue(node, value) 10 psp.incorporate(value, args) 11 if psp.isRandom? 12 trace.registerConstrainedChoice(node) 13 return weight 37 U NCONSTRAIN(trace, node) 1 if node.isLookupNode? 2 return U NCONSTRAIN(trace, node.sourceNode) 3 elseif trace.getPSP(node).isESRReference? 4 return U NCONSTRAIN(trace,trace.getESRParents(node)[0]) 5 else 6 oldValue = trace.getValue(node) 7 psp, args = trace.getPSP(node),trace.getArgs(node) 8 if psp.isRandom? 9 trace.unregisterConstrainedChoice(node) 10 psp.unincorporate(oldValue, args) 11 weight = psp.logDensity(oldValue, args) 12 psp.incorporate(oldValue, args) 13 return weight It is instructive to consider the following program: [ASSUME x_1 (bernoulli 0.5)] [ASSUME x_2 (bernoulli 0.5)] ... [ASSUME x_N (bernoulli 0.5)] [OBSERVE (xor x_1 ... x_N) True] If xor() happens to be true by chance — i.e. constrain() did not have to adjust any random choices — then the PET that was sampled is already drawn from the true conditioned distribution. However, the probability of this decreases rapidly in N and because xor() is deterministic, repeated rejection may appear to be the only option. To avoid this problem, we frequently restrict ourselves to OBSERVEs whose expressions have the form (<some-stochastic-SP> ...), i.e. they are applications of some fixed stochastic procedure. Then we can always successfully initialize and begin inference. Venture was designed with such “sufficiently stochastic” probabilistic programs in mind; see (Freer et al., 2010) for some rationale. We also note that to ensure ergodicity, periodic wholetrace proposals that attempt to restart from the prior (via the INFER directive) are sufficient, although this kind of independence sampler is unlikely to be efficient. We note that without additional restrictions on observations, the presence of constrain() could introduce significant complexity. Constraining a node X that is used in multiple places would not necessarily give us an execution with positive probability, since the probability of all children of X will now have changed. To resolve this and to simplify our algorithms, we make the following strong assumption on Venture programs: For a Venture program to be valid, it must guarantee that if a node gets constrained during inference (as opposed to from a call to observe), every node along every outgoing directed path must be either a deterministic output PSP, a null-request PSP, or a lookup node. Moreover, any observed nodes on this path must agree on the observed value. In addition to the code given above, our implementation of constrain also propagates the constrained value along these outgoing deterministic paths, but the details of the implementation are beyond the scope of this paper. 38 5. Partitioning Traces into Scaffolds for Scalable, Incremental Inference Most scalable inference algorithms for Bayesian networks take advantage of the network’s representation of conditional dependence. We have seen that probabilistic execution traces make analogous scalability gains possible for probabilistic programs written in Turing-complete, higher-order languages, by representing not just conditional dependences but also existential dependencies and exchangeable coupling. However, PETs and the SPI also introduce complexities that can be avoided in the setting of Bayesian networks. For example, in a Bayesian network, the random variables that could possibly directly depend on a given random variable are easy to identify: they are the immediate children of the node according to the direct graph structure of the network. In a PET, the value of a random choice may change which children exist. Additionally, in Bayesian networks, all nodes are equipped with conditional probability densities, but SPs that lack conditional density functions are permitted in Venture. Joint densities for chains of random choices in Venture may not be available, let alone the conditional density of a choice given its immediate descendants. Here we describe scaffolds, the mechanism used in Venture to handle these complexities. Scaffolds carve out coherent subproblems of inference in the context of a global inference problem analogously to the inference subproblems in a Bayesian network that arise by conditioning on a subset of the nodes and querying the nodes contained within it. 5.1 Motivation and Notation Let ρ be an execution and let X = {X} denote a set of nodes that we wish to propose new values for as a block. If we only propose new values to these nodes, we may end up with a trace that has probability 0. For example, consider the following program: [ASSUME x (normal 0 1)] [PREDICT (+ x 1)] If we only propose a change to the value of x from 1 to 2, and do not propagate the change through the application of +, then + will report probability 0 of sampling its output 2 given its inputs 1 and 2. Similarly, in the program [ASSUME x (normal 0 1)] [PREDICT (if (< x 0) (normal -10 1) (normal 10 1))] if we only propose a new value for x and its sign flips, then the requester for IF will report probability 0 of its old request given its new inputs. SPs that cannot report their likelihoods introduce a similar problem. In the program [ASSUME x (normal 0 1)] [PREDICT (likelihood-free-sp x)] if we only propose a new value for x, then we cannot compute the Metropolis-Hastings acceptance ratio since we have no way to account for the probability of the likelihood-free-sp’s old value given its new arguments. 39 In all three cases, the problem can be solved by proposing new values to other nodes downstream of the nodes in X. However, we do not want to resimulate the rest of the program just to make a single proposal. We want to find a middleground between rejecting at + and running the entire program. As a default, we propose new values to downstream nodes until we reach applications of stochastic PSPs for which we can definitely compute the logdensity of its original output given any new values for its inputs. We say that these nodes “absorb the flow of change” and so we call them absorbing nodes. Note that in the case of the IF statement that may switch from one branch to another, the nodes in the old branch may no longer exist in the proposed trace, and other nodes may come into existence. To simplify notation, we will not explicitly condition on the parents of a random variable in our derivations. For example, in the Bayesian network A → B, we will write P(A) for P(A) and P(B) for P(B|A). Although this would be unacceptable in the case of Bayesian networks where we marginalize and condition arbitrarily, in our discussion of probabilistic programming we will only ever compute probabilities of nodes given their parents, so that this notation will not be problematic. 5.2 Partitioning traces and defining scaffolds As before, let ρ be an execution and let X = {X} denote a set of nodes that we wish to propose new values for as a block. We call X the principal nodes of the proposal. Let ξ denote the proposed trace. We introduce the following definitions: 1. D = D(ρ, X): the nodes that will definitely still exist in ξ, whose values might change. This includes X, as well as any nodes at which we do not want to absorb at (e.g. deterministic nodes) or are unable to absorb at (e.g. applications of likelihood-free SPs). Note that we have D(ρ) = D(ξ) = D by symmetry. 2. brush(ρ) = brush(ρ, X): the nodes that may no longer be requested if D is resampled. This includes any nodes in branches that may be abandoned, or any requests whose operator may change. 3. R(ρ) = R(ρ, X) = D(ρ, X)∪brush(ρ, X): the regenerated subset of ρ. These are all the nodes that may either change value or no longer exist in ξ. Conversely R(ξ) consists of precisely the set of new values we are proposing for ξ. 4. A = A(ρ, X): the children of D, or equivalently, the “absorbing” nodes. For each absorbing, the differences between the logdensity of the output given the new value and the logdensity of the output of the old value appears in the Metropolis-Hastings acceptance ratio. 5. torus(ρ, X) = ρ \ R(ρ): the nodes in ρ that are guaranteed to still exist in ξ, and whose values are guaranteed not to change. By construction we have torus(ρ) = torus(ξ) = torus. 6. P(ρ) = P(ρ, X): all parents of all nodes in R(ρ) ∪ A, excluding any nodes in those sets. These are the nodes whose values would have been required to simulate R(ρ) from torus and to calculate the new log densities at A, but which definitely exist and whose values cannot change. 7. I(ρ): the set of all nodes that would never be referenced at all while regenerating R(ρ). 40 Figure 4: The scaffold partitions the trace into five groups: the gold nodes that will definitely still exist in the proposal trace but whose values may change (drg), the blue nodes at which we will definitely compute likelihoods (absorbing), the green nodes that may no longer exist (the brush), the dark grey nodes that are parents of nodes in these three groups (parents), and all other light grey nodes that need never be visited at all (ignored). Note that future graphs will not distinguish between these last two groups. 8. O = P(ρ) ∪ I(ρ) = P(ξ) ∪ I(ξ). Note that we may not have P(ρ) = P(ξ), since R(ρ) and R(ξ) may lookup different symbols in different environments, and may also request different simulations. These definitions give us the following partitions: ρ = D ∪ R(ρ) ∪ A ∪ O (1) ξ = D ∪ R(ξ) ∪ A ∪ O (2) We refer to the pair (D, A) as the scaffold of the set X of principal nodes. As we will see, the scaffold is the fundamental entity in the inference methods we will consider. It induces a set Ξ = Ξ[O, scaffold] of executions that can be reached by resimulating along the scaffold, consulting parents in O as needed. Moreover, for any ξ ∈ Ξ, the scaffold scaffold will yield the same set O in the factorization ξ = scaffold ∪ brush(ξ) ∪ O. 5.3 Constructing the scaffold Given a trace ρ and a set of principal nodes X, we construct the scaffold D(ρ, X) as follows. First, we walk downstream from each of the principal nodes depth-first until every path either terminates or reaches the application of a stochastic PSP for which we are certain we can compute a logdensity. As we go along, we mark nodes as D or A as appropriate. However, some nodes that we marked during this process may actually be in the brush. Therefore our second step is to recursively “disable” all requests that may change value or that may no longer exist, and to mark as brush every node that is 41 (a) (b) Figure 5: Scaffold construction. The principal node is shown in red, resampling nodes are shown in gold, absorbing nodes are shown in blue, and the brush is shown in green. (a) To construct a scaffold, Venture first walks downstream from the (red) principal node, identifying all the nodes whose values could possibly change, stopping at nodes that are guaranteed to be able to absorb the change (albeit perhaps with a probability density of 0). (b) Venture then identifies the brush — those nodes that may no longer exist depending on the values chosen for the nodes being resampled — using a separate recursion. Once the brush has been identified, the definite regeneration graph (gold) and the border (blue) are straightforward to identify. See the main text for additional details. only requested by these disabled requests. We then remove from D and A every node in the brush. Figure 5 gives an overview of this process, while Figure 6 illustrates the need for removing the brush of the source trace ρ before making a proposal. 5.4 Pseudocode for constructing scaffolds The overall construction procedure for scaffolds involves several passes, all linear in the size of the scaffold: C ONSTRUCT S CAFFOLD(trace, principalNodes, kernelInfo) 1 2 3 4 5 6 7 cDRG, cAbsorbing, cAAA = F IND C ANDIDATE S CAFFOLD(trace, principalNodes) brush = F IND B RUSH(trace, cDRG, cAbsorbing, cAAA) drg, absorbing, aaa = RemoveBrush(cDRG, cAbsorbing, cAAA, brush) border = F IND B ORDER(drg, absorbing, aaa) regenCounts = C OMPUTE R EGEN C OUNTS(trace, drg, absorbing, aaa, border, brush) lkernels = loadKernels(trace, drg, kernelInfo) return Scaffold(drg, absorbing, aaa, border, regenCounts, lkernels) Finding a candidate scaffold given a set of principal nodes involves walking downstream from each principal node until it is possible to absorb. To find the brush, Venture must count the number of times a given requested family is reachable from the DRG. If this equals the number of times 42 exp: (branch_exp (f) (quote (f)) (quote (f))) value: 1 exp: (f) value: 1 request: [(161334320, (f), <env>)] exp: branch_exp value: sp:branch exp: (f) value: 0 exp: (quote (f)) value: (f) exp: (quote (f)) value: (f) exp: f value: sp:sym_dir_mult request: [] 157180560 request: [] exp: f value: sp:sym_dir_mult 2 exp: (make_sym_dir_mult 1.000000 2.000000) value: sp:sym_dir_mult request: [] exp: 2.000000 value: 2.000000 exp: 1.000000 value: 1.000000 exp: make_sym_dir_mult value: sp:make_sym_dir_mult 1 Venture Families 159241808 Figure 6: The need for detaching the brush before making a proposal. This trace includes the invocation of make symmetric dirichlet discrete, an SP that produces SPs whose applications are exchangeably coupled. In this trace, the result of one application determines whether or not another application is made. Thus the second application is in the brush of the scaffold generated by the first application. If we do not detach the brush first, we may propose to the first application conditioned on an application that will not exist in the new trace. the family is requested, then all references to the family originate in the DRG, and it is therefore possible for the family’s existence to depend on the values of the DRG nodes. F IND B RUSH(trace, cDRG, cAbsorbing, cAAA) 1 2 3 4 5 6 7 disableCounts = map(0) disabledRequestNodes = set() brush = set() for node in cDRG if node.isRequestNode? D ISABLE R EQUESTS(trace, node, disableCounts, disabledRequestNodes, brush) return brush 43 D ISABLE R EQUESTS(trace, node, disableCounts, disabledRequestNodes, brush) 1 2 3 4 5 6 if node in disabledRequests return for esrParent in trace.getESRParents(node.outputNode) disableCounts[esrParent] = disableCounts[esrParent] + 1 if disableCounts[esrParent] == trace.getNumberOfRequests(esrParent) D ISABLE FAMILY(trace, esrParent, disableCounts, disabledRequestNodes, brush) D ISABLE FAMILY(trace, node, disableCounts, disabledRequestNodes, brush) 1 2 3 4 5 6 7 brush.insert(node) if node.isOutputNode? brush.insert(node.requestNode) D ISABLE R EQUESTS(trace, node.requestNode, disableCounts, disabledRequestNodes, brush) D ISABLE FAMILY(trace, node.operatorNode, disableCounts, disabledRequestNodes, brush) for operatorNode in node.operatorNodes D ISABLE FAMILY(trace, operatorNode, disableCounts, disabledRequestNodes, brush) The border consists of nodes where resampling stops. This includes absorbing nodes, AAA nodes, and also resampling nodes with no children, e.g. from top-level PREDICT directives: F IND B ORDER(trace, drg, absorbing, aaa) 1 2 3 4 5 border = union(absorbing, aaa) for node in drg if not hasChildInAorD(trace, node, drg, absorbing) border.insert(node) return border For stochastic regeneration to proceed correctly, the nodes in the scaffold must be annotated with counts that allow determination of when a given node is no longer referenced and can therefore be removed. 44 C OMPUTE R EGEN C OUNTS(trace, drg, absorbing, aaa, border, brush) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 regenCounts = map(0) for node in drg if node in aaa regenCounts[node] = 1 // will be added to shortly elseif node in border regenCounts[node] = trace.getChildren(node).size + 1 else regenCounts[node] = trace.getChildren(node).size // Now determine the number of times each reference to an AAA node will be regenerated for node in union(drg, absorbing) for parent in trace.getParents(node) M AYBE I NCREMENTAAAR EGEN C OUNT(parent) for node in brush if node.isLookupNode? M AYBE I NCREMENTAAAR EGEN C OUNT(node.sourceNode) elseif node.isOutputNode? M AYBE I NCREMENTAAAR EGEN C OUNT(trace.getESRParents(node)[0]) return regenCounts M AYBE I NCREMENTAAAR EGEN C OUNT(trace, node, regenCounts) 1 2 3 value = trace.getValue(node) if value.isSPRef ? and not value.makerNode == node and aaa.contains?(value.makerNode) regenCounts[value.makerNode] = regenCounts[value.makerNode] + 1 5.4.1 A BSORBING AT A PPLICATIONS As we discussed earlier, there are cases in which we do not need to visit all the absorbing nodes explicitly to account for their logdensities because the SP can keep track of its outputs and report the joint logdensity of all of its applications. If this is the case, we say that the application of the SP is absorbing at applications (AAA), and our preliminary walk to the find the scaffold can stop upon reaching an AAA node. Figure 7 gives a graphical illustration. Note that the AAA node itself is in D–not A–since its value may change. Handling AAA introduces many subtle bookkeeping challenges, and the code for computing the regenCounts for AAA nodes will seem mysterious until inspecting the special cases for AAA nodes in regen and extract in the next section. 5.5 Breaking down a global inference problem into collections of local inference problems If the principal nodes are chosen to be all the random choices in the current trace, then the border will correspond to the random choices constrained by the OBSERVEs, and the terminal resampling nodes will be any ASSUMEs that are not subsequently referred to, as well as all the PREDICTs. The definite regeneration graph will only contain random choices that are guaranteed to exist in every execution of the probabilistic program. Conditionally simulating the entire program can be reduced to conditionally simulating a completion of a torus given the absorbing nodes A. 45 (a) (b) Figure 7: Scaffolds with and without the “absorbing-at-applications” optimization. Repeated applications of stochastic procedures are common in probabilistic programs for machine learning and statistics. (a) shows a PET fragment typical of this repetition. The principal node is in red, resampling nodes are in yellow, and the constructor SP is in blue. Note the size of the scaffold, which grows linearly with the number of applications. (b) shows the same PET fragment but where the “constructor” SP (whose resulting SP is repeatedly applied) can absorb at applications. The SP is now pink, and it because it absorbs at applications (implementing the weight calculations needed for inference via information in its auxiliary storage) the scaffold size no longer scales linearly with the number of applications. 46 This equivalence is an important feature of Venture’s design. Any inference strategy that works for entire programs can be run on a program fragment, conditioned on the remainder, and vice versa. Venture can break down the problem of sampling from the conditioned distribution over the entire PET, which might be extremely high dimensional, into collections of overlapping lowerdimensional inference subproblems. Any Venture strategy applicable to the overall problem becomes applicable to each subproblem. Current research is exploring the use of Venture programs to construct custom proposals, where each PREDICT directive is mapped to a principal node or a resampling node in the border, and each OBSERVE statement is matched to an absorbing node. 5.6 Local kernels and scaffolds Later sections describe techniques for building up transition operators that efficiently modify contents of scaffolds, sampling its definite regeneration graph and brush from a distribution that leaves its conditional invariant. These transition operators are built out of stochastic proposals for individual random choices that we call local kernels. These kernels must be able to sample new values for individual random choices as well as calculate the ratio of forward and reverse transition probabilities necessary for using the sampled value as a proposal for a Metropolis-Hastings scheme. Venture supports local kernels of two types: simulation and delta kernels. 5.6.1 L OCAL SIMULATION KERNELS : RESIMULATION & BOTTOM - UP PROPOSALS A local simulation kernel for a random choice x cannot look at previous values of x when making a proposal. The simplest simulation kernel is the one that samples x according to the simulate() provided by its associated stochastic procedure. We will sometimes refer to this as a “resimulation” kernel. 1. Let π→ denote the order in which the local kernels are called during sample. Then sample returns a new value x along with the following weight: Pπ→ (x) Kπ→ (x) 2. Let π← denote the reverse of the order in which the local kernels are called during unsample. Then unsample returns the following weight: Kπ← (x) Pπ← (x) For the resimulation kernel, the weights from this local kernel are identically 1. Simulation kernels can also be conditioned on anything in torus without breaking asymptotic convergence. This is because the values of random choices in the torus are all available at proposal time and also unaffected by the transition. Simulation kernels can thus be used to make bottom-up proposals, using other values in the source trace as input. This provides one mechanism for augmenting the topdown processing that is typical in probabilistic programming with custom bottom-up information. 5.6.2 L OCAL DELTA KERNELS AND REUSE OF RANDOM CHOICES Venture also supports delta kernels. When these kernels transition the trace from x0 → x, they have access to the previous state x0 in addition to the contents of the torus. A Gaussian drift kernel, where x ∼ N (x0 , σ2 ), is a typical example. They satisfy the following contract: 47 1. sample returns a state x0 with weight: P(x)K(x → x0 ) P(x0 )K(x0 → x) 2. unsample returns 0. These kernels cannot be applied to stochastic procedures that exhibit exchangeable coupling between applications, nor can they be used in particle methods. Since they receive the source state as an input, they also provide a mechanism for re-using the old value of a random choice if its parent choices have not changed. Delta kernels can thus reduce unnecessary or undesirable resampling. 6. Stochastic Regeneration Algorithms for Scaffolds We now show how to coherently modify a trace, given a partition and a valid scaffold, using an algorithm we introduce called stochastic regeneration. Variations on stochastic regeneration, some of which can be expressed as simple parameterizations, can be used to implement a wide range of stochastic inference strategies. The simplest use of stochastic regeneration is to implement a Metropolis-Hastings transition operator for PETs as follows. First, the border of the scaffold is “detached” from the trace in an arbitrary order, and the random choices within the trace fragment that has been detached are stored in an “extract”. We will sometimes abbreviate this process as detach. Next, the border of the scaffold is “regenerated and attached” in reverse order, yielding a new trace. We will sometimes abbreviate this process as regen. This new trace can be accepted or rejected. If it is accepted, the algorithm has finished. If it is rejected, it is detached again, and the original extract is fed to the regeneration algorithm. This restores the trace to its original state. Assuming simulation and density evaluation of the constituent PSPs is constant, the runtime of this process scales with the size of the scaffold and the brush, not the size of the entire PET. Nodes that cannot influence or be influenced by the transition are never visited. PET nodes are regenerated in the opposite order that they were visited during detach; see Figure 8 for a trace that illustrates the need for this symmetry. 6.1 Detaching along a scaffold with DETACH-AND-EXTRACT Detach-and-Extract takes a scaffold and a trace, and converts the trace into a torus and a full omegaDB. At that point, the torus is ready for a new proposal, while the omegaDB then contains all the information needed to restore the original trace. The first step is to disconnect (detach) the trace fragment from the border of the scaffold. The second step is to extract all the random choices from the trace fragment and store them in the omegaDB. 48 exp: (branch_exp (eq n 1.000000) (quote (f 0.000000)) (quote (f 0.000000))) value: 1 exp: (f 0.000000) value: 1 request: [(159619384, (f 0.000000), <env>)] exp: branch_exp value: sp:branch exp: (eq n 1.000000) value: #f exp: (quote (f 0.000000)) value: (f 0.000000) exp: (quote (f 0.000000)) value: (f 0.000000) exp: (f n) value: 0 request: [] 159295920 exp: 1.000000 value: 1.000000 exp: eq value: sp:= exp: n value: 0.000000 exp: n value: 0.000000 exp: f value: sp:sym_dir_mult exp: f value: sp:sym_dir_mult 4 exp: (uniform_discrete 0.000000 2.000000) value: 0.000000 exp: (make_sym_dir_mult 1.000000 2.000000) value: sp:sym_dir_mult request: [] exp: 2.000000 value: 2.000000 exp: 0.000000 value: 0.000000 request: [] 3 159714440 request: [] request: [] exp: 0.000000 value: 0.000000 exp: uniform_discrete value: sp:uniform_discrete exp: 2.000000 value: 2.000000 2 exp: 1.000000 value: 1.000000 exp: make_sym_dir_mult value: sp:make_sym_dir_mult 1 Venture Families 159611288 160807176 Figure 8: Due to partially exchangeable SPs, detach and regen must visit nodes in opposite orders. A procedure with partial-exchangeable coupling is applied in both the absorbing border and the brush. The probability of the absorbing nodes depends on the order in which regeneration takes place. As Venture is detaching, it needs to compute the probability of the absorbing nodes that it would have calculated if it had proposed this trace by regenerating along the scaffold. In particular, if it would have visited the absorbing application first during regen, then it must visit the absorbing application last during detach. Venture solves this problem by having detach always detach all nodes in the opposite order from how regen would have generated them. D ETACH A ND E XTRACT(trace, border, scaffold) 1 2 3 4 5 6 7 8 9 10 weight = 0 db = DB() for node in reversed(border) if scaffold.isAbsorbing(node) weight = weight + D ETACH(trace, node, scaffold, db) else if node.isObservation weight = weight + U NCONSTRAIN(trace, node) weight = weight + E XTRACT(trace, node, scaffold, db) return weight, db To detach a node from the border, Venture must unincorpoate the value from the PSP that was previously responsible for generating it, adjust the weights, and then extract its parents: D ETACH(trace, node, scaffold, db) 1 2 3 4 5 6 psp, args = trace.getPSP(node),trace.getArgs(node) groundValue = trace.getGroundValue(node) psp.unincorporate(groundValue, args) weight = psp.logDensity(groundValue, args) weight = E XTRACT PARENTS(trace, node, scaffold, db) return weight To extract the parents of a node, Venture loops over the parents in the correct order: 49 E XTRACT PARENTS(trace, node, scaffold, db) 1 2 3 4 weight = E XTRACT ESRPARENTS(trace, node, scaffold, db) for parent in reversed(trace.getDefiniteParents(node)) weight = weight + E XTRACT(trace, parent, scaffold, db) return weight E XTRACT ESRPARENTS(trace, node, scaffold, db) 1 2 3 4 weight = 0 for parent in reversed(trace.getESRParents(node)) weight = weight + E XTRACT(trace, parent, scaffold, db) return weight The key idea in this process is that when a node is no longer referenced by any others — that is, its regenCount is 0 — it can be unevaluated. Also, if a node is a request node, before unevaluating it all requests it generates should be unevaluated, which may trigger the trace fragments corresponding to the request to be unevaluated if they are no longer referenced. E XTRACT(trace, node, scaffold, db) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 weight = 0 value = trace.getValue(node) if value.isSPRef ? and not value.makerNode == node and scaffold.isAAA(value.makerNode) weight = weight + E XTRACT(trace, value.makerNode, scaffold, db) if scaffold.isResampling(node) scaffold.decrementRegenCount(node) if scaffold.getRegenCount(node) == 0 if node.isLookup? trace.clearValue(node) else if node.isRequestNode? weight = weight + U NEVAL R EQUESTS(trace, node, scaffold, db) weight = weight + U NAPPLY PSP(trace, node, scaffold, db) weight = weight + E XTRACT PARENTS(trace, node, scaffold, db) return weight 6.2 Regenerating a new trace along a scaffold with REGENERATE-AND-ATTACH Regenerate-and-Attach, often abbreviated as regen, takes a scaffold, a torus, and an omegaDB, as well as a control parameter that indicates whether or not it is restoring (i.e. replaying random choices from the omegaDB)13 . The key idea in regen is that it ensures the PET is constructed via sequence of self-supporting PETs, that is PETs where all parents of a node — including all choices on which a given random choice directly depends — have been regenerated before the node itself is. 13. It also takes a map from nodes to the gradient of their log densities, for use by variational kernels; we will discuss this later in the advanced inference section. 50 R EGENERATE A NDATTACH(trace, border, scaffold, restore?, db) 1 2 3 4 5 6 7 8 9 weight = 0 for node in border if scaffold.isAbsorbing?(node) weight = weight + ATTACH(trace, node, scaffold, restore?, db) else weight = weight + R EGENERATE(trace, node, scaffold, restore?, db) if node.isObservation? weight = weight + C ONSTRAIN(trace, node, node.observedValue) return weight To attach a node from the border, one first regenerates its parents, then “attaches” the trace fragment to the current trace, making it responsible for generating the current node in the border. This involves updating the weight and also incorporating the value into the stochastic procedure that is newly responsible for generating it. ATTACH(trace, node, scaffold, restore?, db) 1 2 3 4 5 6 psp, args = trace.getPSP(node),trace.getArgs(node) groundValue = trace.getGroundValue(node) weight = R EGENERATE PARENTS(trace, node, scaffold, restore?, db) weight = weight + psp.logDensity(groundValue, args) psp.incorporate(groundValue, args) return weight R EGENERATE PARENTS(trace, node, scaffold, restore?, db) 1 2 3 4 5 weight = 0 for parent in trace.getDefiniteParents(node) weight = weight + R EGENERATE(trace, parent, scaffold, restore?, db) weight = weight + R EGENERATE ESRPARENTS(trace, node, scaffold, restore?, db) return weight R EGENERATE ESRPARENTS(trace, node, scaffold, restore?, db) 1 2 3 4 weight = 0 for parent in trace.getESRParents(node) weight = weight + R EGENERATE(trace, parent, scaffold, restore?, db) return weight Regenerating a node involves updating its regeneration counts, to ensure that after regeneration, detach can be correctly called on the new trace: 51 R EGENERATE(trace, node, scaffold, restore?, db) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 weight = 0 if scaffold.isResampling?(node) if scaffold.getRegenCount(node) == 0 weight = weight + R EGENERATE PARENTS(trace, node, scaffold, restore?, db) if node.isLookup? trace.setValue(node,trace.getValue(sourceNode)) else weight = weight + A PPLY PSP(trace, node, scaffold, restore?, db) if node.isRequestNode? weight = weight + E VAL R EQUESTS(trace, node, scaffold, restore?, db) scaffold.incrementRegenCount(node) value = trace.getValue(node) if value.isSPRef ? and not value.makerNode == node and scaffold.isAAA?(value.makerNode) weight = weight + R EGENERATE(trace, value.makerNode, scaffold, restore?, db) return weight 6.3 Building invariant transition operators using stochastic regeneration Stochastic regeneration can be used to implement a variety of inference schemes over probabilistic execution traces. In addition to modifying the trace in an undoable fashion as described above, stochastic regeneration provides numerical weights that are needed for inference schemes such as Metropolis-Hastings. 6.3.1 W EIGHTS ASSUMING SIMULATION KERNELS Assume that the local kernels used for each stochastic procedure application is a simulation kernel — that is, the value it proposes is independent of the value of the random choice in ρ. Let π be the (after the fact) regen order for ξ, which is the order in which the local kernels are called during regen, and which is also by construction the reverse of the order in which the local kernels would be called during detach. Then 1. regen returns Pπ (R(ξ), A) Kregen (R(ξ)) 2. detach returns: Kregen (R(ξ)) Pπ (R(ξ), A) where Kregen is the kernel that proposes to R(·) by calling local kernels in the regen order, and detach computes the reverse probability correctly since each local kernel is unapplied in the reverse of the order in which it would have been applied during regen. These weights can be used to implement Metropolis-Hastings transitions between ρ and ξ. Note that if all local kernels are resimulation kernels, only the “likelihood ratio” induced by the absorbing nodes is involved in the Metropolis-Hastings acceptance ratio. 52 6.3.2 W EIGHTS ASSUMING DELTA KERNELS Now assume that at least one of the local kernels is a delta kernel — that is, the value it proposes depends on the value of the random choice being modified from ρ. In this case, regen and detach do not return useful quantities individually, but their product (detachAndRegen) is the same as in the case of simulation kernels, and is suitable for implementing Metropolis-Hastings transitions: Pπξ (R(ξ), A)Kregen (R(ξ) → R(ρ)) Pπρ (R(ρ), A)Kregen (R(ρ) → R(ξ)) (3) 6.4 Context-independent versus context-specific inference schemes The contents of the scaffold and brush do not depend on the values of the principal nodes for a given transition. For example, every random choice whose existence might be affected by a value chosen for a principal node (or a descendant of a principal node that must be resampled) must be included in the brush, and regenerated during a transition, even if the particular value(s) sampled for the principal node were consistent with the old execution trace. It could be asymptotically more efficient in some circumstances to exploit context-specific independencies, both for the values of random choices and for their existence. Unfortunately, context-specificity comes at the cost of significant additional complexity. Other Turing-complete probabilistic programming systems — for example, the original implementation in (Mansinghka, 2009), as well as MIT-Church, and an earlier version of Venture — were context sensitive. Monte was context-independent but significantly more limited than Venture. Of these, we believe only the earlier version of Venture achieved the correct order of growth for all scale parameters in Latent Dirichlet Allocation, and none supported reprogrammable inference. It seems likely that scaffold construction and stochastic regeneration can be interleaved to retain the flexibility of our current architecture while achieving the efficiency of an update scheme sensitive to context-specific independencies. We leave this for future work. 7. General-purpose Inference Strategies via Stochastic Regeneration Venture’s inference programming language provides a number of primitive inference mechanisms . The effect of each strategy is the evolution of a PET according to a transition operator that leaves the posterior distribution on PETs invariant, with modifications constrained to lie within the scaffold. Each of these strategies can be applied to any scaffold, including the global scaffold. Each is implemented using stochastic regeneration, with regen and detach performing mutation of the trace in place and returning weights that can be used for any rejection steps necessary to maintain invariance. Before describing specific inference strategies, we establish some preliminary results that simplify reasoning about transition operators on PETs. 7.1 Factorization of the acceptance ratio Let ρ be a PET representing an execution and π be an ordering of its nodes. Suppose we visit each node in this order, for each SP application (i.e. random choice) computing the probability density of its output given its input, and then passing that input-output pair to the incorporate method asso53 ciated with the SP, to handle exchangeable coupling. Define Pπ (ρ) to be the total probability density of the sequence of random choices in ρ, i.e. the product of the probability densities calculated above. Recall that every SP satisfies the property that the probability of any sequence of input-output pairs is invariant under permutation. Thus we have that Pπ (ρ) = P(ρ). For an arbitrary subset Z of the nodes in ρ, we do not necessarily have Pπ1 (Z) = Pπ2 (Z). The equivalence only holds over the probability of all of ρ, not over arbitrary fragments of it. For example, consider the following program: ( assume f ( make - ccoin 1 1) ) ( predict (f)) ( predict (f)) Now let ρ be an execution with (say) both predicts true. Then the probability of the first predict depends on whether or not it is weighted before or after the second predict, but the total probability will be the same no matter which order we proceed in. However, the probabilities will be equal if Z is a prefix in both orderings, because we can view Z as the complete trace associated with some modified probabilistic program ρ0 . Proposition 7.1. Let ρ, ξ ∈ Ξ[torus, D]. Let πρ be a permutation on ρ such that πρ (P ∪ I) < πρ (R(ρ) ∪ A), and likewise for πξ . Then P(ρ) Pπρ (R(ρ), A) = P(ξ) Pπξ (R(ξ), A) (4) Proof. We can factor ρ and ξ into disjoint subsets: ˙ ρ = (P ∪ I)∪(R(ρ) ∪ A) ˙ ξ = (P ∪ I)∪(R(ξ) ∪ A) (5) (6) We have Pπρ (P ∪ I) = Pπξ (P ∪ I) since the set is a prefix in both orderings. The result follows immediately. 7.2 Auxiliary variables for state-dependent stochastic selection of kernels We would like to be able to use Metropolis-Hastings to construct valid transition operators that leave a conditional distribution P() on PETs invariant. In some cases, we also would like these transition operators to be individually ergodic. For any random choice guaranteed to be in all executions — that is, the SP application is in the DRG of the global scaffold — this is straightforward. Standard cycle or mixture hybrid kernels (see e.g. (Bonawitz, 2008; Andrieu et al., 2003; Mansinghka, 2009)) can be used to sequence Metropolis-Hastings proposals on individual variables. Alternately, if we are only interested in inference strategies that choose random choices at uniformly at random, we can make a compound Metropolis-Hastings proposal that first selects a random variable to change (using it to seed a scaffold) and then makes a within that scaffold. Because the choice of random choices involves only a simple state dependence, and it is (nearly) trivial to integrate out the number of paths from some trace ρ to another trace ξ under this sequence of choices, it is straightforward to accommodate it in a single Metropolis-Hastings step. Venture supports a broader range of transition operators: Metropolis-Hastings proposals where the proposal kernel is randomly chosen from a state-dependent distribution. Let I be a (possibly 54 infinite) set, f a stochastic function from traces to I with induced density Pf (). For every i ∈ I, let Qi be a proposal kernel on PETs, and let ρ be the current PET. Then the following cycle kernel on the extended state space Ω × I preserves π: 1. Sample i|ρ ∼ f (ρ). 2. (a) Propose ξ ∼ Qi (ρ → ξ). (b) Accept if U ∼ Uniform(0, 1) < Pf ( f (ξ) = i) P(ξ)Qi (ξ → ρ) · Pf ( f (ρ) = i) P(ρ)Qi (ρ → ξ) This scheme treats the choice of kernel as a transient auxiliary variable I that is created solely for the duration of the ordinary Metropolis-Hastings proposal that depends on it. Invariance is ensured by accounting for the effect of Qi on Pf ( f (ξ) = i). This construction can be used to straightforwardly justify sophisticated custom proposal strategies such as Algorithm 8 from (Neal, 1998) without need for any approximation. Stochastic index selection makes it straightforward to compositionally analyze richer transition operators on PETs. For example, one can select scaffolds by choosing a single principal node uniformly at random, as in typical Church implementations. Alternately, one could estimate the entropy of the conditional distributions for each random choice given its containing scaffold, and choose scaffolds with probability proportional to this estimate. One could also select principal nodes by starting with some query variable of interest — e.g. a random choice representing an important prediction or action variable — and walk the PET, favoring random choices that are “close” to the query variable. We anticipate that the evolution of the Venture inference programming language will depend on the expansion and refinement of a library of techniques for converting proposals into invariant transition operators. 7.3 Metropolis-Hastings via Stochastic Regeneration Here we describe the basic Metropolis-Hastings algorithm for incremental inference over a PET. With all of the tools we have developed so far, a generic version of this transition operator is straightforward to specify: 1. Sample a set of principal nodes principals ∼ f (ρ), and record the probability P( f (ρ) = principals) of sampling it. 2. Construct D with principals as the set of principal nodes. 3. Set α = detachAndRegen(D) 4. Calculate P( f (ξ) = principals). 5. Accept if U< P( f (ξ) = principals) ·α P( f (ρ) = principals) 55 Correctness follows immediately from 7.1 and (3), and applies to a broad class of customized kernels. Because stochastic regeneration does not visit portions of the PET that are conditionally independent of the random choices affected by the transition, it is more scalable than approaches based on transformational compilation. In particular, for typical machine learning problems on datasets of size N, the number of random choices scales linearly with N but the connections in the PET are sparse. Thus an inference sweep of N single-site Metropolis-Hastings transitions — enough to consider each random choice roughly once — requires O(N) time, as opposed to O(N 2 ) time for transformational compilers. 7.4 Approximating optimal proposals via Stochastic Variational Inference We can use the same auxiliary variable technique to lift any local proposal into one that preserves global stationarity. Here we show how to learn a local variational approximation to the posterior distribution on R(ξ) given A using regen and detach, and then wrap it in Metropolis-Hastings as discussed above. 7.4.1 P OSING THE OPTIMIZATION PROBLEM Let ρ be a trace and D a definite regeneration graph inducing torus. Let Q be some family of distributions on Ξ[torus, D] where every Qθ ∈ Q is defined by its parameters θ ∈ Θ, which control the local kernels along D, and where each Qθ reverts to the resimulation kernel on the brush. Our goal is to I-Project P(R) onto the space Q , which involves solving the following optimization problem:   Qθ (R(ξ)) minimize f (θ) = Eξ∼Qθ log P(R(ξ), A) subject to θ ∈ Θ 7.4.2 S TOCHASTIC GRADIENT DESCENT One generic approach is stochastic gradient descent. Under assumptions of differentiability that may not hold in arbitrary programs, the gradient of the objective function has the following nice form (as originally shown in (Wingate and Weber, 2013) for entire traces):   Qθ (R(ξ)) ∇θ f (θ) = ∇θ Eξ∼Qθ log P(R(ξ), A) Z Qθ (R(ξ)) = ∇θ Qθ (R(ξ)) log dξ P(R(ξ), A) ξ∈Ξ   Z Qθ (R(ξ)) = ∇θ Qθ (R(ξ)) log dξ P(R(ξ), A) ξ∈Ξ   Z Z Qθ (R(ξ)) Qθ (R(ξ)) ∇θ Qθ (R(ξ)) log Qθ (R(ξ)) ∇θ log dξ = dξ + P(R(ξ), A) P(R(ξ), A) ξ∈Ξ ξ∈Ξ (7) (8) (9) (10) (11) 56 Z = ∇θ Qθ (R(ξ)) log ξ∈Ξ Qθ (R(ξ)) dξ P(R(ξ), A) (12) Qθ (R(ξ)) dR(ξ) Qθ (R(ξ))∇θ log Qθ (R(ξ)) log P(R(ξ), A) ξ∈Ξ   Qθ (R(ξ)) = Eξ∼Qθ ∇θ log Qθ (R(ξ)) log P(R(ξ), A) Z = (13) (14) where we used that   Z Qθ (R(ξ)) Qθ (R(ξ)) ∇θ log dξ = Qθ (R(ξ)) (∇θ log Qθ (R(ξ))) dξ P(R(ξ), A) ξ∈Ξ ξ∈Ξ Z (15) Z = ∇θ Qθ (R(ξ))dξ (16) Qθ (R(ξ))dξ (17) ξ∈Ξ Z = ∇θ ξ∈Ξ =0 (18) Thus for a given setting of θ(t) , we can approximate the gradient ∇θ f (θ)|θ(t) using Monte Carlo, by forward-sampling trajectories from Qθ(t) . This is the scheme proposed in (Wingate and Weber, 2013): 1. Sample a trajectory from Qθ(t) . 2. Calculate the noisy gradient estimate δ(t) . 3. Take a noisy gradient descent step θ(t+1) := θ(t) − α(t) δ(t) . 7.4.3 A S A PROPOSAL FOR M ETROPOLIS -H ASTINGS We can solve this optimization problem as above to construct a proposal distribution Qθ? , and then sample ξ from it and accept or reject it as a single M-H proposal. Note that while solving the optimization problem, we did not inspect any nodes that were in ρ but that might not be in ξ; therefore we can use the same Qθ? to compute the probability of the reverse transition. 7.4.4 U SING REGEN AND DETACH The stochastic regeneration recursions can be easily extended to carry additional metadata that enables them to evaluate local gradients for each random choice, and store this information as a local kernel that modifies additional state capturing variational parameters. Once this is done, it turns out that both the stochastic gradient optimization of the variational approximation and the use of it as a proposal can be expressed using stochastic regeneration: 1. Construct D. 2. Detach ρ. 3. For every application node in D whose operator cannot change and that can compute gradients, initialize variational parameters to the current arguments. 57 4. Loop until satisfied with the quality of the variational approximation: (a) Attach variational kernels to D to define Qθ(t) . (b) Call regen to sample ξ ∼ Qθ(t) , and to compute   P(ξ, A) log Qθ(t) (R(ξ)) as well as the local gradients {∇θ log Qθ(t) (R(ξ))}. (c) Take a gradient step. (d) Detach. 5. restore ρ to compute αρ := P(R(ρ),A) Qθ? (R(ρ)) . 6. detach 7. regen to sample ξ and compute αξ := P(R(ξ),A) . Qθ? (R(ξ)) 8. Accept if U ∼ Uniform(0, 1) < P( f (ξ) = i) αξ · P( f (ρ) = i) αρ 8. Particle-based Inference: Enumerative Gibbs and Particle Markov chain Monte Carlo Many inference strategies involve weighing multiple alternative states against one another as a way of approximating optimal Gibbs-type transition operators over portions of the hypothesis space. Gibbs sampling on a discrete variable in a Bayesian network can be implemented by enumerating all possible values for the variable, substituting each into the network, evaluating the joint probability of each network configuration, then renormalizing and sampling. Normalized importance sampling for approximating Gibbs involves proposing multiple values for a variable from its prior, weighting by its likelihood, and then normalizing over and sampling from the sampled set. Each of these techniques can be applied iteratively to larger subsets of variables. For example, discrete Gibbs can be iterated over a sequence of variables to generate a proposal from a single-particle particle filter, or extended via enumeration (and perhaps dynamic programming) to larger subsets of random variables. Normalized importance sampling can be also iterated to yield sequential importance sampling with resampling, and embedded within a Markov chain to yield particle Markov Chain Monte Carlo techniques. It can be useful to think of these inference strategies as instantiations of the idea of weighted speculative execution technique within inference programming. An ensemble of executions is represented, evolved and stochastically weighted and selected via the absorbing nodes, although ultimately only one of these executions will be used. We refer to each alternative state as a particle, and implement Gibbs sampling, particle Markov chain Monte Carlo, and other techniques using general machinery for particle-based inference methods. This allows us to develop common techniques for handling the dependent random choices in the brush — both in the proposal and in the analysis — that apply to all these inference strategies. 58 8.1 In-place mutation versus simultaneous particles in memory Particle methods also provide new opportunities for time-space tradeoffs in inference. Although the semantics of particle methods involve multiple alternative traces, it may not always be desirable, or even possible, to represent alternate traces in memory simultaneously. Consider scaffolds whose definite regeneration graph or brush invokes a complex external simulator. If we cannot modify the code of this simulator, or it relies on an expensive external compute resource that cannot be multiplexed, it will be necessary to implement particle methods via mutation of a single PET in place. Otherwise, it is likely to be more efficient to represent all particles in memory simultaneously, sharing a common PET and scaffold as their source, and cloning the auxiliary states of stochastic procedures that are referred to from multiple particles. This choice also enables the parallel evolution of collections of particles via multiple independent threads, synchronizing whenever resampling occurs. Venture supports both in-place mutation and simultaneous representation of particles. This is necessary to recover asymptotic scaling that is competitive with custom sequential Monte Carlo techniques. In this paper, we focus on the transition operators associated with particle methods, and give pseudocode for implementing them in terms of a generic API for simultaneous particles. The data structures needed to implement this API efficiently, and the algorithm analysis for efficient implementations, involve novel applications of techniques from functional programming and persistent data structures and are beyond the scope of this paper. 8.2 The Mix-MH operator for constructing particle-based transition operators Although the auxiliary variable technique introduced previously is sufficient for lifting a local proposal into a global Metropolis-Hastings proposal, it will be helpful to introduce a variant of the technique that is closed under application — i.e. the kernel that results from applying the technique is a valid proposal kernel that can be parameterized, randomly selected, and used as a proposal in a subsequent application. This will give us significant additional flexibility in explaining and justifying our particle-based inference schemes. 8.2.1 M ETROPOLIS -H ASTINGS Suppose we have some distribution π of interest on Ω. Let K be a proposal kernel on Ω. We can create a new kernel MH(K, π) based on it that satisfies detailed balance with respect to π by the Metropolis-Hastings rule. Instead of representing it as a black-box kernel, we will keep around its constituent elements, so we have MH(K, π) = (K, λ) where λ(ρ → ξ) = π(ξ)K(ξ → ρ) and where applying (K, λ) involves: 1. Sample ξ from K(ρ, ξ). 2. Let U ∼ [0, 1]. Accept the new trace ξ iff U< λ(ρ → ξ) λ(ξ → ρ) 59 (19) It is straightforward to show that this transition operator preserves detailed balance with respect to π, and therefore leaves π invariant. Ergodic convergence can be proved under fairly mild assumptions (Andrieu et al., 2003). 8.2.2 S TATE - DEPENDENT M IXTURES OF K ERNELS We would like to be able to construct compound kernels that are themselves stochastically chosen, and certify detailed balance of the compound kernel given easily verified conditions on the component kernel(s) and the stochastic kernel selection rule. Let P() be a density of interest, such as the conditioned density on resampling nodes given the absorbing nodes in a scaffold. Let I be a (possibly infinite) set, and f (ρ) = i be a stochastic function from traces to I, with Pf (i) the density of i induced by f . We will use f (ρ) to select a kernel Qi = (Ki , αi ) at random given the trace ρ (and also evaluate Pf ( f (ξ) = i) for reversing the transition). Let Ki (ρ → ξ) be a transition operator on traces that can be simulated from, and let αi (ρ → ξ) be a procedure that evaluates the factors out of which the Metropolis-Hastings acceptance ratio is constructed: αi (ρ → ξ) = P(ξ)Ki (ξ → ρ) Consider the following kernel, parameterized by scalars λi (ρ → ξ) to be determined: 1. Sample i ∼ f (ρ). This will be used to choose a kernel Ki . 2. Sample ξ ∼ Ki (ρ → ξ). 3. Let U ∼ [0, 1]. Accept ξ iff U< λi (ρ → ξ) λi (ξ → ρ) Our goal is to define λi (ρ → ξ) such that this kernel preserves detailed balance, analogously to the Metropolis-Hastings kernels described above. We assume that the stochastic function f satisfies the following symmetry condition: Pf ( f (ρ) = i)Ki (ρ → ξ) > 0 ⇐⇒ Pf ( f (ξ) = i)Ki (ξ → ρ) > 0 This expresses the constraint that if you can reach ξ from ρ by selecting Qi according to f (ρ) then applying it, then with nonzero probability, reaching ρ from ξ is possible by selecting the same Qi according to f (ξ) then applying it. Let I(ρ) = {i ∈ I : Pf ( f (ρ) = i) > 0} be the set of indices (kernels) reachable from ρ, and I(ρ, ξ) = I(ρ) ∩ I(ξ) be the set of indices (kernels) reachable from both ρ and ξ. To establish detailed balance for this kernel, it suffices to show that P(ρ) ∑ Pf ( f (ρ) = i)Ki (ρ → ξ)λi (ρ → ξ) = P(ξ) i∈I(ρ) Pf ( f (ξ) = i)Ki (ξ → ρ)λi (ξ → ρ) (20) ∑ i∈I(ξ) Due to our symmetry condition, i ∈ I(ρ) but i ∈ / I(ξ) implies that Ki (ρ → ξ) = 0. Therefore detailed balance is satisfied if and only if the following holds: P(ρ) ∑ Pf ( f (ρ) = i)Ki (ρ → ξ)λi (ρ → ξ) = P(ξ) i∈I(ρ,ξ) ∑ Pf ( f (ξ) = i)Ki (ξ → ρ)λi (ξ → ρ) i∈I(ρ,ξ) (21) 60 By hypothesis, we have that each Qi = (Ki , αi ) satisfies detailed balance with respect to P: P(ρ)Ki (ρ → ξ)αi (ρ → ξ) = P(ξ)Ki (ξ → ρ)αi (ξ → ρ) for all i (22) Let λi (ρ → ξ) = αi (ρ → ξ)Pf ( f (ξ) = i). If we multiply both sides of (22) by Pf ( f (ρ) = i)Pf ( f (ξ) = i) then sum both sides over i ∈ I(ρ, ξ), we establish (21). 8.2.3 T HE M IX MH OPERATOR We define the MixMH operator as follows. MixMH takes as input stochastic index sampler f and a set of base kernels {Ki , αi }i∈I parameterized by the index i sampled from the index sampler: MixMH( f , {(Ki , αi )}i∈I ) Now let λi (ρ → ξ) = αi (ρ → ξ)P( f (ξ) = i). To apply MixMH( f , {(Ki , αi )}i∈I ) we: 1. Sample i from f (ρ). 2. Apply (Ki , λi ). 8.3 Particle-based Kernels Our goal here is to define and sketch correctness arguments for particle-based transition operators. We first describe particle sets obtained from repeated applications of a single kernel. Our formulation incorporates a boosted acceptance ratio that is designed to recover Metropolis-Hastings in the special case where there is only one new particle generated.14 8.3.1 G ENERATING PARTICLES BY REPEATEDLY APPLYING A SEED KERNEL Suppose we have a (torus, D) pair, and we attach local simulation kernels to the nodes in D to define a distribution K on Ξ[torus, D]. By (1), regenAndAttach will return wξ = P(R(ξ), A) K(ξ) Let ϒ = {(ξi , wi ) : i = 1, . . . , n} be a multiset of weighted particles in Ξ. We can define the kernel Kϒ (· → ·) on ϒ as follows: nds(ξ j , i)w j Kϒ (ξi → ξ j ) = w−i where nds(ξ j , i) is the number of duplicates of ξ j in ϒ \ ξi .The α-factor for Kϒ in isolation is: α(ξi → ξ j ) P(ξ j )Kϒ (ξ j → ξi ) = α(ξ j → ξi ) P(ξi )Kϒ (ξi → ξ j ) = i , j)wi P(R(ξ j ), A) nds(ξ w− j P(R(ξi ), A) nds(ξ j ,i)w j w−i (23) (24) 14. This boosting comes at the cost of subtle ergodicity violations if this is the only transition operator used and the hypothesis space contains the right symmetries. Simpler particle-based analyses — where all particles, including the current trace, are resampled from — are possible and indeed straightforward. These yield Boltzmann-style acceptance rules that self-transition more frequently and therefore explore the space less efficiently, but avoid ergodicity issues. 61 We can sample the kernel ϒ starting from ξi by generating n − 1 additional particles from Kregen . The probability of generating ϒ is the probability of sampling the n − 1 other particles, multiplied by the number of distinct orders in which they could have been sampled: P(ϒ|ξi ) = (n − 1)! ∏ Kregen (ξk ) ∏γ∈unique(ϒ\ξi ) nds(γ, i)! k6=i (25) Claim 8.1. P(ϒ|ξ j ) nds(ξ j , i) ∏k6= j Kregen (ξk ) = P(ϒ|ξi ) nds(ξi , j) ∏k6=i Kregen (ξk ) Proof. P(ϒ|ξ j ) = P(ϒ|ξi ) (n−1)! ∏γ∈unique(ϒ\ξ j ) nds(γ, j)! ∏k6= j Kregen (ξk ) (n−1)! ∏γ∈unique(ϒ\ξi ) nds(γ,i)! ∏k6=i Kregen (ξk ) (26) ∏γ∈unique(ϒ\ξi ) nds(γ, i)! ∏k6= j Kregen (ξk ) ∏γ∈unique(ϒ\ξ j ) nds(γ, j)! ∏k6=i Kregen (ξk ) nds(ξ j , i)!nds(ξi , i)! ∏k6= j Kregen (ξk ) = nds(ξi , j)!nds(ξ j , j)! ∏k6=i Kregen (ξk ) nds(ξ j , i) ∏k6= j Kregen (ξk ) = nds(ξi , j) ∏k6=i Kregen (ξk ) = (27) (28) (29) Thus once we apply MixMH, the α-factor becomes: nds(ξi , j)wi α(ξi → ξ j ) P(R(ξ j ), A) w− j P(ϒ|ξ j ) = nds(ξ ,i)w α(ξ j → ξi ) P(R(ξi ), A) w−ij j P(ϒ|ξi ) = i , j)wi P(R(ξ j ), A) nds(ξ nds(ξ j , i) ∏k6= j K(ξk ) w− j nds(ξ j ,i)w j nds(ξi , w−i P(R(ξ j ), A) ww−i j ∏k6= j K(ξk ) w P(R(ξi ), A) w−ij ∏k6=i K(ξk )   P(R(ξi ), A) = = (30) P(R(ξi ),A) K(ξi ) P(R(ξ j ), A) P(R(ξi ), A) w−i = w− j ∏k6= j K(ξk ) w− j  P(R(ξ j ),A) K(ξ j ) w−i (31) j) ∏k6=i K(ξk )  (32) (33) ∏k6=i K(ξk ) (34) 62 8.3.2 T HE MHn OPERATOR This pattern, where a multiset of weighted particles is generated from a “seed” kernel K and then sampled from, is a sufficiently common operation that we give it a name: MHn (K) = MixMH(ϒ ∼ K, Kϒ ) We will use variations on it throughout this section. The multiset functions as the index for MixMH, and the sampling step where a particle is chosen functions as the base proposal kernel. 8.3.3 M ETROPOLIS -H ASTINGS AS SPECIAL CASE OF PARTICLE METHODS Metropolis-Hastings with a simulation kernel is actually just the special case MH2 , since in this case w−ρ = wξ . Parallelizable extensions of this kind of “locally independent” Metropolis-Hastings scheme are also natural. For example, one could make multiple proposals and weigh them against against one another. This could be viewed as an importance sampling approximation to an optimal Gibbs proposal over a scaffold, where the proposal can be any simulation kernel, including one that conditions on downstream information. That said, the restriction to simulation kernels is significant. Gaussian drift kernels, for example, are not permitted. 8.3.4 E NUMERATIVE G IBBS AS A SPECIAL CASE : USING DIFFERENT KERNELS FOR EACH PARTICLE There are many ways to generate weighted particle sets ϒ. For instance, we may want to allow a different kernel Ki for each particle ξi . We can represent enumeration over some set of variables this way: each distinct tuple in the Cartesian product of the domains of the variables will correspond to a different kernel. Suppose there are some variables in D that we want to enumerate over, yielding n total combinations. Assume that ρ has the first combination. We can sample ϒ from ρ as follows: call regen n − 1 times as before to generate n − 1 new particles, but passing Ki for ξi , where Ki is the same as K except with deterministic kernels replacing the old kernels on the nodes that we are enumerating. The analysis is even simpler in this case because the particles are distinct. Once we apply MixMH, the α-factor becomes: α(ξi → ξ j ) P(ξ j )Kϒ (ξ j → ξi ) ∏k6= j Kk (ξk ) = α(ξ j → ξi ) P(ξi )Kϒ (ξi → ξ j ) ∏k6=i Kk (ξk ) P(R(ξ j ), A) ww−i j ∏k6= j Kk (ξk ) = w P(R(ξi ), A) w−ij ∏k6=i Kk (ξk )  = P(R(ξ j ), A) as above. 63 P(R(ξ j ),A) K(ξ j ) w−i (36)  ∏k6= j Kk (ξk ) w− j  P(R(ξ j ), A) w−i = w− j P(R(ξi ),A) K(ξi ) (35)  (37) ∏k6=i Kk (ξk ) (38) 8.3.5 PARTICLE M ARKOV CHAIN M ONTE C ARLO : ADDING ITERATION AND RESAMPLING It is often useful to iterate particle-based methods, and to resample collections of particles based on their weights, stochastically allocating particles to regions of the execution space that appear locally promising. This insight is at the heart of particle filtering, i.e. sequential importance sampling with resampling (Doucet et al., 2001), as well as more sophisticated sequential Monte Carlo techniques. For example, (Andrieu et al., 2010) introduces particle Markov chain Monte Carlo methods, a family of techniques for using sequential Monte Carlo to make sensible proposals over subsets of variables as part of larger MCMC schemes. We can also use conditional SMC to generate ϒ. As above, suppose we have ρ and D, with local simulation kernels attached to the nodes in D, which determines a simulation kernel K that generates samples by propagating along D via regen. Suppose we group the sinks into T groups, which in turn partitions all of D into T groups D1 , . . . , DT according to the regen recursion. We will refer to K on Rt as Kt . To establish notation, first let us consider a simpler case: basic sequential Monte Carlo, instead of conditional SMC. We can propose along R1 from K1 , and then (inductively) propagate to time t by sampling independently from a proposal kernel, (m) (m) (m) Mt (at , Rt (ξt )) w (m) = t−1 Kt (Rt (ξt )) wt−1 where (m) (m) wt−1 = P(Rt−1 (ξt−1 ), At−1 ) m ) Kt−1 (ξt−1 (m) is the return value from regen on the t − 1st group of sinks, and where Rt (ξt ) is assumed to be (m) (at ) (m) conditioned on ξt ’s parent particle ξt−1 . It is instructive to view this sampling procedure as a way of generating a single sample of all auxiliary variables Γ from the density n ψ(Γ) = (m) m=1 (m) n T n (m) (39) (m) wt−1 (m) Kt (Rt (ξt )) w t−1 t=2 m=1 (m) T n ∏ K1 (R1 (ξ1 )) ∏ ∏ m=1 T = (m) ∏ K1 (R1 (ξ1 )) ∏ ∏ m=1 = n t=2 m=1 n = T ∏ K1 (R1 (ξ1 )) ∏ ∏ Mt (at , ξt ) t=2 m=1 n 1 ! n ∏ ∏ wt−1 ∏ t=2 m=1 m=1 (40) (m) (m) P(Rt−1 (ξt−1 ), At−1 )Kt (Rt (ξt )) (41) (m) Kt−1 (Rt−1 (ξt−1 ))wt−1 (m) K1 (R1 (ξ1 )) T n ∏∏ t=2 m=1 (m) (m) P(Rt−1 (ξt−1 ), At−1 )Kt (Rt (ξt )) (m) (42) Kt−1 (Rt−1 (ξt−1 )) Now suppose we perform the conditional SMC sweep to sample Γ. First we sample an index bt ∈ {1, . . . , n} for each time step t = 1, . . . , T −1 for the index of the forced resampling of the source particle ξi . Then we sample the other particles as in standard CSMC. The probability of Γ starting from ξi is precisely: 64 ! T P(Γ; ξi ) = ∏ wt−1 t=2 ! T = ∏ wt−1 t=2 ψ(Γ) (b ) (bt ) (b ) T nT −1 K1 (R1 (ξ1 1 )) ∏t=2 wt−1 Kt (Rt (ξt t )) (i) wT ψ(Γ) nT −1 P(R(ξ(i) ), A) (43) (44) However, we want to index our kernel with multiset semantics for the complete particles as we did above, and so we index by the equivalence class Γ where Γ ≡ Γ0 if they agree on everything up until the last time step, and then agree on the same multiset of complete particles. We have P(Γ; ξi ) = P(Γ; ξi ) (n − 1)! ∏γ∈unique(Γ(T ) \ξi ) nds(γ, i)! and thus when we apply MixMH, the α-factor becomes15 α(ξi → ξ j ) P(ξ j )KΓ (ξ j → ξi )P(Γ|ξ j ) = α(ξ j → ξi ) P(ξi )KΓ (ξi → ξ j )P(Γ|ξi ) P(ξ j )KΓ (ξ j → ξi )nds(ξ j , i)P(Γ|ξ j ) = P(ξi )KΓ (ξi → ξ j )nds(ξi , j)P(Γ|ξi ) P(ξ j )KΓ (ξ j → ξi )nds(ξ j , i)P(R(ξ(i) ), A)w j P(ξi )KΓ (ξi → ξ j )nds(ξi , j)P(R(ξ( j) ), A)wi K (ξ j → ξi )nds(ξ j , i)w j = Γ KΓ (ξi → ξ j )nds(ξi , j)wi = (45) (46) (47) (48) = nds(ξi , j)wi nds(ξ j , i)w j w− j nds(ξ j ,i)w j nds(ξi , j)wi w−i (49) = w−i w− j (50) We can think of this as a strict generalization of MHn where the kernel is split into a sequence of kernels. This scheme is called PGibbs in the Venture inference programming language, as it enables the approximation of blocked Gibbs sampling over arbitrary scaffolds. Cycles and mixtures of PGibbs with other kernels recovers a wide range of existing particle Markov chain Monte Carlo schemes as well as novel algorithms. 8.3.6 P SEUDOCODE FOR PGibbs WITH SIMULTANEOUS PARTICLES Here we give pseudocode for implementing the PGibbs transition operator using an interface for simultaneous particles. The particle interface permits particles to be constructed from source PETs or from one another, sharing the maximum amount of state possible. Stochastic regeneration is used (i) 15. When it is clear we are referring to the final weight, we will refer to wT as wi to be consistent with the previous sections. 65 to prepare the source trace ρ for conditional SMC, to populate the array of particles, and to calculate weights. PG IBBS(trace, border, scaffold, P) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 T = border.length rhoWeights = [T ] rhoDBs = [T ] for t = T to 1 rhoWeights[t], rhoDBs[t] = D ETACH A ND E XTRACT(trace, border[t], scaffold) particles = [P] particleWeights = [P] for p = 0 to P particles[p] = PARTICLE(trace) particleWeights[0] = R EGENERATE A NDATTACH(particles[0], border[1], scaffold, TRUE, rhoDBs[1]) for p = 1 to P particleWeights[p] = R EGENERATE A NDATTACH(particles[p], border[1], scaffold, FALSE, NIL) newParticles = [P] newParticleWeights = [P] for t = 2 to T newParticles[0] = PARTICLE(particles[0]) newParticleWeights[0] = R EGENERATE A NDATTACH(newParticles[0], border[t], scaffold, TRUE, rhoDBs[t]) for p = 1 to P parentIndex = S AMPLE C ATEGORICAL(M AP E XP(particleWeights)) newParticles[p] = PARTICLE(particles[parentIndex]) newParticleWeights[p] = R EGENERATE A NDATTACH(newParticles[p], border[t], scaffold, FALSE, NIL) particles = newParticles particleWeights = newParticleWeights finalIndex = S AMPLE C ATEGORICAL(M AP E XP(particleWeights[1 : P])) weightMinusXi = L OG S UM E XP(particleWeights. R EMOVE(finalIndex)) weightMinusRho = L OG S UM E XP(particleWeights. R EMOVE(0)) alpha = weightMinusRho − weightMinusXi return particles[finalIndex], alpha This implementation illustrates the use of versions of regen and detach that act on particles, and the initialization of particles using parent particles as well as source traces. All but the last four lines are simply constructing the randomly chosen kernel to be applied, corresponding to proposing to replace the current race with the contents of a particle. 9. Conditional Independence and Parallelizing Transitions Here we briefly describe the conditional independence relationships that are easy to extract from probabilistic execution traces. This analysis clarifies the potential long-range dependencies in a 66 PET introduced by the choice to make probabilistic closures into first-class objects in the language. These independence relationships could be used to support probabilistic program analyses and also to justify the correctness of parallelized kernel composition operators in the inference programming language. They also expose parallelism that is distinct from particle-level parallelism. 9.1 Markov Blankets, Envelopes, and Conditional Independence In Bayesian networks, the Markov Blanket of a set of nodes provides a useful characterization of important conditional independencies. These independencies permit parallel simulation of Markov chain transition operators. In constrast, while the scaffold provides a useful factorization of the logdensity of a trace, it does not permit parallel simulation of transition operators. In this section, we show how to formulate a notion of locality that permits parallel transitions on PETs. We first review Markov Blankets in Bayesian networks. Let D and D0 be sets of nodes in a Bayesian network, A the children of D, and P the parents of D ∪ A excluding D and A. Then if / proposals can be made on the two regeneration graphs in parallel, although D0 ∩ (D ∪ A ∪ P) = 0, both proposals may read the same values. We refer to A ∪ P as the Markov blanket of D, and it can be thought of as the set of all nodes that one must read from (but not write to) when resampling D and computing the new probabilities of its children A. The situation becomes more complicated in the case of PETs for two main reasons. First, the parents P = P(ρ) ∪ P(ξ) are not known a priori, since we do not know P(ξ) until we have simulated ξ to see what nodes it reads from. Second, the SP auxiliary states that will be mutated are not known a priori, since the PSPs being applied in R(ξ) are not known. 9.2 After-the-fact envelopes Let S(ρ) denote the SP auxiliary states for SPs with PSPs that are applied in R(ρ). Define MB(ρ) = P(ρ) ∪ A (51) to be the set of all nodes that are read from but not written to, as before for Bayesian networks. Now define the scope of a proposal Scope(ρ) = R(ρ) ∪ S(ρ) (52) to be the set of all parts of the PET that are written to or created while generating ρ from torus, and define the envelope Envelope(ρ) = MB(ρ) ∪ Scope(ρ) (53) to be the set of all nodes that must be read from or written to during the proposal. Note that in general there may be some nodes in the scope that are written to but never read from, e.g. terminal resampling nodes and the sufficient statistics for uncollapsed SPs, but we will ignore this distinction for now. We can make the following claim: suppose the proposal ρ → ξ temporally overlaps with the proposal ρ0 → ξ0 , and that by some luck,   Envelope(ρ, ξ) ∩ Scope(ρ0 , ξ0 ) ∪ Envelope(ρ0 , ξ0 ) ∩ Scope(ρ, ξ) = 0/ (54) 67 then the transitions will not have clashed. In words, this says that the two proposals will not clash as long as no node that one proposal reads differs in the two traces of the other proposal. If the transitions have not clashed, then they could have been simulated simultaneously. 9.3 The envelope of a scaffold Define: Envelope[D] = [ Envelope(ξ) (55) Scope(ξ) (56) ξ∈Ξ[D] Scope[D] = [ ξ∈Ξ[D] That is, Envelope[D] and Scope[D] correspond to the union of the trace-specific envelope and scope, taken over all possible completions of torus. Then two proposals on D and D0 are guaranteed not to clash if (Envelope[D] ∩ Scope[D0 ]) ∪ (Envelope[D0 ] ∩ Scope[D]) = 0/ (57) Now let R and R0 be the random variables corresponding to the results of invocations of regen (including the weights) on scaffolds with definite regeneration graphs D and D0 . One approach to formalizing the relationship between conditional independence and parallel simulation in Venture would be to first try to show that (57) holds if and only if R ⊥ R0 , and second to try to show that random variables can safely be simulated in parallel if and only if they are conditionally independent. When probabilistic programmers can identify two scaffolds for which (57) holds, they can schedule transitions on them simultaneously. Also, in some circumstances speculative simultaneous simulation may be beneficial, for example if the clash is on small, cheap procedures or an approximate transition that ignores some dependency can be used as a proposal for a serial transition. It may also be possible to use this notion of envelope to predict beforehand which transitions we can run simultaneously. To do this, we will need to bound the Markov blanket and the scope. One option would be to statically analyze Venture program fragments, implementing a kind of “escape analysis” for random choices. Another would be to introduce language constructs and/or stochastic procedures that came with hints about Markov blanket composition. Future work will explore the efficacy of these and other techniques in practice. 10. Related Work Venture builds on and contains complementary insights to a number of other probabilistic programming languages, implementations, and inference techniques. For example, Monte (Bonawitz and Mansinghka, 2009) was the first commercially developed prototype interpreter for a Church-like probabilistic programming language. Its architecture is the most direct antecedent of Venture. For example, it included a scaffold-like decomposition of local transitions in terms of “proposal blankets” for single-variable changes. Preliminary prototypes of Venture (Wu, 2013) and of an approximate multi-core scheme for Church-like languages (Perov and Mansinghka, 2012) also introduced early, ad-hoc versions of some of the ideas in this paper, all in an attempt to control the asymptotic cost of each inference transition. 68 The most salient inefficiency in many Church systems has been that for typical machine learning problems with N datapoints, a single sweep of inference over all random variables requires runtime that scales as O(N 2 ). This quadratic scaling, combined with the absolute constant factors, has made it impossible to apply these systems beyond hundreds of datapoints. For “lightweight” implementations based on direct transformational compilation or augmented interpretation, such as Bher (Wingate et al., 2011b), this is a basic consequence of the approach: the entire program must be re-simulated to make a change to a single latent variable. A variety of approaches have been tried to mitigate this problem. For earlier implementations such as the first prototype implementation of Church (Mansinghka, 2009) atop the Blaise system (Bonawitz, 2008), the MIT-Church implementation, and also Monte, ad-hoc attempts were made to control the scope of re-simulation, approximating the Venture notions of PETs and their partition into scaffolds. Similarly, the Shred implementation of Church (Yang et al., 2012) uses techniques from program analysis coupled with just-in-time compilation to try to avoid the re-simulation involved in lightweight implementations of Metropolis-Hastings and also reduce constant-factor overhead. These approaches exhibit different limitations with respect to generality, scalability, absolute achievable efficiency and runtime predictability. For example, the Blaise implementation of Church and MIT-Church both exhibited quadratic scaling in some cases, and involved substantial runtime overheads. The program analysis and compilation techniques used in Shred incur runtime costs that can be significant in absolute terms and difficult to predict a priori, and additionally impose constraints on the integration of custom primitives into the language. Finally, all of these approaches have only applied directly to single-site Metropolis-Hastings approaches to inference. Outside of Church-like languages, BLOG (Milch et al., 2007) — the first open-universe probabilistic programming language — is of particular relevance. For example, the original BLOG inference engine was based on infinite contingent Bayesian networks (ICBN) (Milch et al., 2005). ICBNs characterize the valid factorizations of their probability distribution and represent dependencies whose existence is contingent on the values of other variables. They thus provide a data structure for BLOG that serves an analogous function to PETs in Venture. Also, the Gibbs sampling algorithm introduced for BLOG (Arora et al., 2012) exploits a decomposition of a possible world that is related to the scaffolds used to define the scope of inference in Venture. However, BLOG is not higher-order — it does not support random variables that are themselves probabilistic procedures — nor does it provide a programmable inference mechanism. BLOG also does not support collapsed primitives that exhibit exchangeable coupling between applications, primitives that lack calculable probability densities, or primitives that can create and destroy latent variables while providing their own external inference mechanism. The Figaro system for probabilistic programming (Pfeffer, 2009) shares several goals with Venture, although the differences in language semantics and inference architecture are substantial. Venture is a stand-alone virtual machine, while Figaro is based on an embedded design, implemented as a Scala library. Figaro models are represented as data objects in Scala, and model construction and inference both proceed via calling into a Scala API. This enables Figaro to leverage the mature toolchain for Scala and the JVM, and avoids the need to reimplement difficult but standard programming language features from scratch. It remains to be seen how to apply this machinery to models that depend on higher-order probabilistic procedures, primitives exhibiting exchangeable coupling, and likelihood-free primitives, or how to enable users to extend Figaro with custom model fragments equipped with arbitrary inference schemes. IBAL (Pfeffer, 2001), an ancestor of Figaro 69 based on an embedding of stochastic choice into the functional language ML, is arguably more similar to Venture in terms of its interface. However, the inference strategies around which IBAL was designed imply restrictions on stochastic choices that are analogous to Infer.NET. Our hybrid inference primitives embody generalizations of other proposals for global inference in probabilistic programs. The idea of global mean-field inference in probabilistic programs via stochastic optimization, implemented using repeated re-simulation of the program, was first proposed in (Wingate and Weber, 2013). Similarly, a global implementation of particle Gibbs via conditional sequential Monte Carlo, implemented via repeated forward re-simulation, has been independently and concurrently developed in (Wood et al., 2014). Both of these approaches were implemented using the “lightweight” scheme proposed in (Wingate et al., 2011b). They thus lack dependency tracking and will exhibit unfavorable asymptotic scaling. The mean field and conditional SMC schemes developed in this paper are implemented via stochastic regeneration, building on PETs for efficiency and supporting the full SPI. The techniques from this paper can also be used on subsets of the random choices in a probabilistic program and composed with other inference techniques, yielding hybrid inference strategies that may be asymptotically more efficient than homogeneous ones. To the best of our knowledge, Venture is the first probabilistic programming platform to support a compositional inference programming language with multiple computationally universal primitives. It is also the only probabilistic programming system to support higher-order probabilistic procedures as first class objects, and in particular the implementation of arbitrary higher-order probabilistic procedures with external, per-application latent variables as ordinary primitives. Venture is also the only probabilistic programming system that integrates exact and approximate sampling techniques based on standard Markov chain, sequential Monte Carlo and variational inference. 11. Discussion We have described Venture, along with the key concepts, data structures and algorithms needed to implement a scalable interpreter with programmable inference. Venture includes a common stochastic procedure interface that supports higher-order probabilistic procedures, “likelihood-free” procedures, procedures with exchangeable coupling between their invocations, and procedures that maintain external latent variables and supply external inference code. We have seen how the probabilistic execution traces on which Venture is based generalize key ideas from Bayesian networks, and support partitioning traces into scaffolds corresponding to well-defined inference subproblems. We have defined stochastic regeneration algorithms over scaffolds, and shown how to use them to implement multiple general-purpose inference strategies. We have also given example Venture programs that illustrate aspects of the modeling and inference languages that it provides. The coverage of Venture in terms of models, inference strategies and end-to-end problems needs to be carefully assessed. It remains to be seen whether the current set of inference strategies are truly sufficient for the full range of problems arising across the spectrum from Bayesian data analysis to large-scale machine learning to real-time robotics. However, expanding the set of inference primitives may be straightforward. One strategy rests on the analogies between PETs and Bayesian networks. For example, it may be possible to define analogues of junction trees for fragments of PETs by conditioning on random choices with existential dependencies. Along similar lines, extensions to the enumerative gibbs inference instruction could leverage the insights from (Koller et al., 1997). Variants based on message passing techniques such as expectation propagation might 70 also be fruitful. Other inference techniques that we could build using PETs and the SPI include slice sampling and Hamiltonian Monte Carlo. Of these, Hamiltonian Monte Carlo has already seen some use in probabilistic programming (Stan Development Team, 2013; Wingate et al., 2011a). However, these implementations do not apply to arbitrary Venture scopes, where proposals can trigger the resampling of other random choices due to the presence of “likelihood-free” primitives or the brush. Generalizations of Hamiltonian and slice methods that use the auxiliary variable machinery presented in this paper are straightforward and are currently included in development branches of the Venture system. Although some real-world applications have been successfully implemented using unoptimized prototypes, the performance surface of Venture is largely uncharacterized. This is intentional. Our understanding of probabilistic programming is more limited than our understanding of functional programming when Miranda was introduced (Turner, 1985). Standard graphical models from machine learning correspond to short nested loops that usually can be fully unrolled before execution, and therefore only exercise small subsets of the capabilities of typical expressive languages. It thus seems premature to focus on optimizing runtime performance until better foundations have been established. The theoretical principles needed to assess the tradeoffs between inference accuracy, asymptotic scaling, memory consumption, and absolute runtime are currently unclear. Point-wise comparisons of runtime are easy to make, but just as easily mislead. Even small changes in problem formulation, dataset size or accuracy can lead to large changes in runtime. The asymptotic scaling of runtime for many probabilistic programming systems is unknown, and for some problems the relevant scale parameters are hard to identify. Systematic comparisons of inference strategies that control for implementation constants and scale thresholds have yet to be performed. Although Venture’s asymptotic scaling competitive with hand-optimized samplers in theory, a rigorous empirical assessment is needed. Mathematically rigorous cost models are in our view a precondition for comparative benchmarking and thus could be an important research direction. Additionally, once the relationships between the asymptotic scaling of forward simulation and the asymptotic scaling of typical inference strategies has been characterized, it will be possible to search for effective equivalences between custom inference strategies and model transformations. Performance engineering research for Venture can build on standard techniques for higher-order languages as well as exploitation of the structure of probabilistic models and inference strategies. Immediate opportunities include runtime specialization via type hints and/or scaffold contents, compilation of the transition operators arising from specific inference instructions, and optimizations to the SPI to minimize copying of data. PETs could be used to support a memory manager that exploits locality of reference along PETs — and conditional independence, more generally — to improve cache efficiency. For example, one could pre-calculate the envelope of a specific inference instruction, and pre-fetch the entire PET fragment. Similar optimizations will be beneficial for high-performance multicore implementations. There are also opportunities for asymptotic savings if deterministic sub-segments of PETs can be compressed; in principle, the memory requirements of inference should scale with the amount of randomness consumed by the model, not by its execution time. Preliminary explorations of some of these ideas have yielded promising results (Perov and Mansinghka, 2012; Wu, 2013) but much work remains to be done. In practice, we believe the judicious migration of performance sensitive regions of a probabilistic program into hand-optimized inference code will be as important as runtime and compiler optimizations. 71 11.1 Debugging and profiling probabilistic programs Venture makes interactive modeling and inference possible by providing an expressive modeling language with scalable programmable inference. By removing the need to do integrated co-design of probabilistic models, inference algorithms and their implementations, Venture removes one of the main bottlenecks in building probabilistic computing systems. This is far from the only bottleneck, however. Design, validation, debugging and optimization of probabilistic programs — both model programs and inference programs — still requires expertise. The promise of probabilistic programming is that instead of having to learn multiple nonoverlapping fields, future modelers will be able to learn a coherent body of probabilistic programming principles. They should be able to use standard probabilistic programming tools and workflows to navigate the design, validation, debugging and optimization process. This could evolve analogously to how programmers learn to design, test, debug and optimize traditional programs using a mix of intuitive modeling, mathematical analysis and experimentation supported by sophisticated tools. Venture implements a stochastic semantics for inference that is designed to cohere with Bayesian reasoning via exact sampling in the limit of infinite computation. This facilitates the development of mathematically rigorous debugging strategies. First, the precise gap between exact and approximate sampling can sometimes be measured on small-scale examples; this provides a crucial control for larger-scale debugging. Second, as the amount of computation devoted to inference increases, the quality of the approximation can only increase. As a result, at potentially substantial computational cost, it is always possible to reduce the impact of approximate inference (as distinct from modeling issues or data issues). There is even a path forward at scales where the true distribution is effectively impossible to sample from or even roughly characterize. For example, the Bernstein-von Mises family of theorems gives general conditions under which posteriors concentrate onto hypotheses that satisfy various invariants. These can be exploited by debugging tools: a tool could increase the amount of synthetic data and inference used to test a probabilistic program until the posterior concentrates. At this point the probabilistic program begins to resemble a problem with a “single right answer”, significantly simplifying debugging and testing as compared to regimes where the true distribution can be arbitrarily spread out or multi-modal. Many more tools and techniques will no doubt be needed to tease apart the impact of model mismatch, inadequate or inappropriate data and insufficient inference. Traditional profilers may also have natural analogs in probabilistic programming. A profiler for a Venture program might track “hot spots” — scaffolds (and their source code locations) where a disproportionate fraction of runtime is spent — as well as “cold spots”: scaffolds where transitions are rejected with unusual frequency. Such cold spots might serve as useful warning signs of limited, purely local convergence of inference. As Venture maintains an execution trace, it should be possible to link an execution trace location with the source code location responsible for it, and potentially help probabilistic programmers identify and modify unnecessarily deterministic or constrained model fragments. Both together could help Venture programmers identify the portions of their program that ought to be moved to optimized foreign inference code. 72 11.2 Inference programming The Venture inference programming language has many known limitations that would be straightforward to address. For example, it currently lacks an iteration construct, a notion of procedural abstraction (so inference strategies can be parameterized and reused), the ability to dynamically evaluate modeling expressions to stochastically generate scope and block names, or an analogue of eval that gives inference programs the ability to programmatically execute new directives in the hosting Venture virtual machine. Additionally, its current set of inference primitives cannot themselves be implemented via compound statements in the language; to support this, an analogue of stochastic regeneration could potentially be exposed. Future work will address these issues by reintroducing the full expressiveness of Lisp into the inference expression language and determining what extensions are needed to capture the interactions between modeling and approximate inference. More substantial inference programming extensions may also be fruitful. For example, it may be fruitful to incorporate support for inference procedures: reusable proposal schemes for scaffolds that match pre-specified patterns, where the proposal mechanism is implemented by another Venture program. The scaffold’s contents constitute the formal arguments. Each random choice in the border could be mapped to an OBSERVE, and the new values for resampling nodes could each come from a PREDICT. If inference in the proposal program is assumed to have converged, it should be possible to construct a valid transition operator. This mechanism would enable Venture programmers to use probabilistic modeling and approximate inference to design inference strategies, turning every modeling idiom in Venture into a potential inference tool. A similar mechanism could facilitate the use of custom Venture programs as the skeleton for variational approximations. Finally, the inference programming language needs to be extended to relax the constraints of soundness. Instead of restricting programmers to transition operators that are guaranteed by construction to leave the conditioned distribution on traces invariant, the language should permit experts to introduce arbitrary transition operators on traces. It will be interesting to develop probabilistic programs that automate aspects of inference while going beyond traditional formulations of interpretation and compilation. For example, it is theoretically possible to develop probabilistic programs that work as inference optimizers. Such programs would take a Venture program as input, including only placeholder INFER instructions, and produce as output a new Venture program with INFER instructions that are likely to perform better. These programs could also transform the modeling instructions and the instructions introducing the data to improve computational or inferential performance. Depending on architecture, these programs could be viewed as inference planners, compilers, or an integrated combination of the two. Probabilistic models, approximate inference and Bayesian learning could all be deployed to augment engineering of the planning algorithm, compiler transformations or even the objective function used to summarize inference performance. Machine learning schemes for algorithm selection (Xu et al., 2008) and reinforcement-learning-based meta-computation (Lagoudakis and Littman, 2001) can be viewed as natural special cases. Venture also makes it possible to integrate a “control” probabilistic program into an interpreter running another probabilistic program. The control program could intercept and modify the inference instructions in the running Venture program based on approximate, model-based inferences from live performance data, and perhaps influence the machine code generation process invoked by a just-in-time compiler. 73 Inference programming also supports the integration of “approximate” compilers based on probabilistic modeling and approximately Bayesian learning. Consider a probabilistic program with a specific set of OBSERVEs and PREDICTs. The expressions in the OBSERVEs define a space of possible inputs, each corresponding to a set of literal values, one per observation. The expressions in the PREDICTs define a space of possible outputs, each corresponding to an assignment of a sampled value to each PREDICT. A compiler for Venture might generate an equivalent program that is restricted to precisely the given pattern of inputs and outputs, optimizing the implementation of any INFER instructions in the program given this restriction. It is also possible to use modeling and inference to approximately emulate the program, by writing a probabilistic program that models p({PREDICTs}|{OBSERVEs}), with parameters and/or structure estimated from data that is generated from the original probabilistic program. This can be viewed as approximate compilation implemented via inference, where the target language is given by the hypothesis space of the model in the emulator program. Both the creation of this kind of emulator and its use as a proposal would be natural to integrate as additional inference instructions. Probabilistic programming systems should ultimately support the complete spectrum from blackbox, truly automatic inference to highly customized inference strategies. So far Venture has focused on the extremes. Other points on the spectrum will require extensions to Venture’s instruction language to encode specifications for exact and approximate inference within particular runtime and/or accuracy parameters. Developing a suitable specification language and cost model to enable precise control over the scope of automatic inference is an important challenge for future research. 11.3 Conclusion The similarities between non-probabilistic, Turing-complete, higher-order programming languages are striking given the enormous design space of such languages. Lisp, Java and Python support many of the same programming idioms and can be used to simulate one another without changing the asymptotic order of growth of program runtime. Each of these languages also presents similar foreign interfaces for interoperation with external software. They have all been used for data analysis and for system building, deployed in fields ranging from robotics to statistics. They have also all been used for machine intelligence research grounded in probabilistic modeling and inference. One measure of their flexibility and interchangeability is provided by their use in probabilistic programming research. Each of these languages has been used to implement expressive probabilistic programming languages. For example, multiple versions of Church and Venture have been written in Python, Lisp and Java, and BLOG has been implemented in both Java and Python. We do not know if it will be possible to attain the flexibility, extensibility and efficiency of Lisp, Java or Python in any single probabilistic programming language, especially due to the complexity of inference. Thus far, most probabilistic programming languages have opted for a narrower scope. Most are not Turing-complete or higher-order, only support a subset of standard approximate inference strategies, and have no notion of inference programming. Translating between sufficiently expressive probabilistic programming languages without distorting the asymptotic scaling of forward simulation will be difficult, especially given the lack of a standard cost model. Providing faithful translations that do not distort the asymptotic scaling of inference — in runtime, in accuracy, or both — will be harder still. Despite these challenges, it may be possible to develop probabilistic languages that can be used to specify and solve probabilistic modeling and approximate inference problems from many fields. 74 Ideally such languages would be able to cover the modeling idioms and inference strategies from fields such as robotics, statistics, and machine learning, while also meeting key representational needs in cognitive science and artificial intelligence. We hope Venture, and the principles behind its design and implementation, represent a significant step towards the development of a probabilistic programming platform that is both computationally universal and suitable in practice for generalpurpose use. References Harold Abelson and Gerald Jay Sussman. Structure and Interpretation of Computer Programs. 1983. Christophe Andrieu, Nando De Freitas, Arnaud Doucet, and Michael I Jordan. An introduction to MCMC for machine learning. Machine learning, 50(1-2):5–43, 2003. Christophe Andrieu, Arnaud Doucet, and Roman Holenstein. Particle Markov chain Monte carlo methods. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 72(3):269–342, 2010. Nimar S Arora, Stuart J Russell, Paul Kidwell, and Erik B Sudderth. Global seismic monitoring as probabilistic inference. In NIPS, pages 73–81, 2010. Nimar S Arora, Rodrigo de Salvo Braz, Erik B Sudderth, and Stuart Russell. Gibbs Sampling in Open-Universe Stochastic Languages. arXiv preprint arXiv:1203.3464, 2012. James K Baker. Trainable grammars for speech recognition. The Journal of the Acoustical Society of America, 65(S1): S132–S132, 1979. C M Bishop. Pattern Recognition and Machine Learning. Springer, 2006. D. M. Blei, A. Y. Ng, and M. I. Jordan. Latent Dirichlet Allocation. Journal of Machine Learning Research, 3:993–1022, 2003. Keith Bonawitz and Vikash Mansinghka. Monte: An interactive system for massively parallel probabilistic programming, 2009. Keith A Bonawitz. Composable Probabilistic Inference with Blaise. 2008. Katalin Csilléry, Michael GB Blum, Oscar E Gaggiotti, and Olivier François. Approximate Bayesian computation (ABC) in practice. Trends in ecology & evolution, 25(7):410–418, 2010. Luc De Raedt and Kristian Kersting. Probabilistic inductive logic programming. Springer, 2008. Pierre Del Moral, Arnaud Doucet, and Ajay Jasra. Sequential Monte Carlo Samplers. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 68(3):411–436, 2006. A Doucet, N de Freitas, and N Gordon, editors. Sequential Monte Carlo Methods in Practice. Springer-Verlag, 2001. Robin D Dowell and Sean R Eddy. Evaluation of several lightweight stochastic context-free grammars for RNA secondary structure prediction. BMC bioinformatics, 5(1):71, 2004. David Duvenaud, James Robert Lloyd, Roger Grosse, Joshua B Tenenbaum, and Zoubin Ghahramani. Structure discovery in nonparametric regression through compositional kernel search. arXiv preprint arXiv:1302.4922, 2013. Cameron Freer, Vikash Mansinghka, and Daniel Roy. When are probabilistic programs probably computationally tractable? NIPS Workshop on Advanced Monte Carlo Methods with Applications, 2010. Brendan J Frey. Bayesian networks for pattern classification, data compression, and channel coding. PhD thesis, Citeseer, 1997. 75 N. Friedman and D. Koller. Being Bayesian about network structure: A Bayesian approach to structure discovery. Machine Learning, 50:95–126, 2003. Nir Friedman, Michal Linial, Iftach Nachman, and Dana Pe’er. Using bayesian networks to analyze expression data. Journal of computational biology, 7(3-4):601–620, 2000. A. Gelman, J. B. Carlin, H. S. Stern, and D. B. Rubin. Bayesian data analysis. Chapman & Hall, New York, 1995. Noah Goodman and Joshua Tenenbaum. Probabilistic Models of Cognition, 2013. Noah D Goodman*, Vikash K Mansinghka*, Daniel M Roy, Keith Bonowitz, and Joshua B Tenenbaum. Church: a language for generative models. In Uncertainty in Artificial Intelligence, May 2008. Peter J Green, Nils Lid Hjort, and Sylvia Richardson. Highly Structured Stochastic Systems. Oxford University Press, 2003. Thomas L Griffiths and Zoubin Ghahramani. Infinite latent feature models and the Indian buffet process. In NIPS, volume 18, pages 475–482, 2005. Roger Grosse, Ruslan R Salakhutdinov, William T Freeman, and Joshua B Tenenbaum. Exploiting compositionality to explore a large space of model structures. arXiv preprint arXiv:1210.4856, 2012. David Heckerman. A tutorial on learning with Bayesian networks. Springer, 1998. Frederick Jelinek, John D Lafferty, and Robert L Mercer. Basic methods of probabilistic context free grammars. Springer, 1992. Mark Johnson, Thomas L Griffiths, and Sharon Goldwater. Adaptor grammars: A framework for specifying compositional nonparametric Bayesian models. Advances in Neural Information Processing Systems, 19:641, 2007. Daphne Koller, David McAllester, and Avi Pfeffer. Effective Bayesian inference for stochastic programs. In AAAI/IAAI, pages 740–747, 1997. Michail G Lagoudakis and Michael L Littman. Learning to select branching rules in the DPLL procedure for satisfiability. Electronic Notes in Discrete Mathematics, 9:344–359, 2001. P. Liang, M. I. Jordan, and D. Klein. Learning programs: A hierarchical Bayesian approach. In International Conference on Machine Learning (ICML), pages 639–646, 2010. D. J. Lunn, A. Thomas, N. Best, and D. Spiegelhalter. WinBUGS - A Bayesian modelling framework: Concepts, structure, and extensibility. Statistics and Computing, 10(4):325–337, October 2000. Christopher D Manning and Hinrich Schütze. Foundations of statistical natural language processing. MIT press, 1999. Vikash Mansinghka. Natively Probabilistic Computation. PhD thesis, Massachusetts Institute of Technology, Cambridge, MA, June 2009. Vikash Mansinghka*, Tejas D Kulkarni*, Yura N Perov, and Josh Tenenbaum. Approximate bayesian image interpretation using generative probabilistic graphics programs. In Advances in Neural Information Processing Systems, pages 1520– 1528, 2013. VK Mansinghka, C Kemp, TL Griffiths, and JB Tenenbaum. Structured priors for structure learning. In In Proceedings of the 22nd Conference on Uncertainty in Artificial Intelligence (UAI, 2006. Paul Marjoram, John Molitor, Vincent Plagnol, and Simon Tavaré. Markov chain Monte Carlo without likelihoods. Proceedings of the National Academy of Sciences, 100(26):15324–15328, 2003. Andrew McCallum, Karl Schultz, and Sameer Singh. FACTORIE: Probabilistic programming via imperatively defined factor graphs. In Advances in Neural Information Processing Systems (NIPS), 2009. 76 Brian Milch, Bhaskara Marthi, David Sontag, Stuart Russell, Daniel L Ong, and Andrey Kolobov. Approximate inference for infinite contingent Bayesian networks. In Proc. 10th AISTATS, pages 238–245, 2005. Brian Milch, Bhaskara Marthi, Stuart Russell, David Sontag, Daniel L Ong, and Andrey Kolobov. BLOG: Probabilistic models with unknown objects. Introduction to statistical relational learning, page 373, 2007. TP Minka, JM Winn, JP Guiver, and DA Knowles. http://research.microsoft.com/infernet, 2010. Infer.NET 2.4. Microsoft Research Cambridge. See Radford M. Neal. Markov Chain Sampling Methods for Dirichlet Process Mixture Models. Technical Report, September 1998. David J Nott and Peter J Green. Bayesian variable selection and the Swendsen-Wang algorithm. Journal of computational and Graphical Statistics, 13(1), 2004. Songhwai Oh, Stuart Russell, and Shankar Sastry. Markov chain Monte Carlo data association for multi-target tracking. Automatic Control, IEEE Transactions on, 54(3):481–497, 2009. Peter Orbanz and Daniel M Roy. Bayesian Models of Graphs, Arrays and Other Exchangeable Random Structures. Hanna Pasula, Bhaskara Marthi, Brian Milch, Stuart Russell, and Ilya Shpitser. Identity uncertainty and citation matching. In Advances in Neural Information Processing Systems, pages 1401–1408, 2002. Yura Perov and Vikash Mansinghka. Exploiting conditional independence for efficient, automatic multicore inference for Church, 2012. Avi Pfeffer. IBAL: A Probabilistic Rational Programming Language. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 733–740, 2001. Avi Pfeffer. Figaro: An object-oriented probabilistic programming language. Charles River Analytics Technical Report, page 137, 2009. David Poole. The independent choice logic for modelling multiple agents under uncertainty. Artificial intelligence, 94 (1):7–56, 1997. Carl Edward Rasmussen. The infinite gaussian mixture model. In NIPS, volume 12, pages 554–560, 1999. CE Rasmussen and CKI Williams. Gaussian processes for machine learning. Adaptive computation and machine learning, 2006. John C Reynolds. Definitional interpreters for higher-order programming languages. In Proceedings of the ACM annual conference, pages 717–740. ACM, 1972. Matthew Richardson and Pedro Domingos. Markov logic networks. Machine Learning, 62(1-2):107–136, 2006. D Roy, V Mansinghka, N Goodman, and J Tenenbaum. A stochastic programming perspective on nonparametric Bayes. In ICML Nonparametric Bayes Workshop, volume 22, page 26, 2008. Stuart J. Russell and Peter Norvig. Artificial Intelligence: A Modern Approach. Prentice Hall, Englewood Cliffs, NJ, 2nd edition, 2002. Taisuke Sato and Yoshitaka Kameya. Prism: a language for symbolic-statistical modeling. In IJCAI, volume 97, pages 1330–1339. Citeseer, 1997. Stan Development Team. Stan Modeling Language Users Guide and Reference Manual, Version 2.0, 2013. URL http: //mc-stan.org/. Robert H Swendsen and Jian-Sheng Wang. Replica monte carlo simulation of spin glasses. Physical Review Letters, 57 (21):2607–2609, 1986. 77 Joshua B Tenenbaum, Charles Kemp, Thomas L Griffiths, and Noah D Goodman. How to grow a mind: structure, statistics and abstraction. science, 331(6022):1279–1285, 2011. Sebastian Thrun, Wolfram Burgard, Dieter Fox, et al. Probabilistic Robotics. MIT Press, 2005. Tina Toni, David Welch, Natalja Strelkowa, Andreas Ipsen, and Michael PH Stumpf. Approximate Bayesian computation scheme for parameter inference and model selection in dynamical systems. Journal of the Royal Society Interface, 6 (31):187–202, 2009. David A Turner. Miranda: A non-strict functional language with polymorphic types. In Functional Programming Languages and Computer Architecture, pages 1–16. Springer, 1985. N. Whiteley, A. Lee, and K. Heine. On the role of interaction in sequential Monte Carlo algorithms. ArXiv e-prints, September 2013. David Wingate and Theo Weber. Automated variational inference in probabilistic programming, 2013. David Wingate, Noah Goodman, Andreas Stuhlmueller, and Jeffrey M Siskind. Nonstandard interpretations of probabilistic programs for efficient inference. In Advances in Neural Information Processing Systems, pages 1152–1160, 2011a. David Wingate, Andreas Stuhlmueller, and Noah D Goodman. Lightweight implementations of probabilistic programming languages via transformational compilation. In Proceedings of the International Conference on Artificial Intelligence and Statistics, 2011b. Frank Wood, Jan-Willem van de Meent, and Vikash Mansinghka. A new approach to probabilistic programming inference. In Artificial Intelligence and Statistics, 2014. Jeff Wu. Reduced traces and JITing in Church, 2013. Lin Xu, Frank Hutter, Holger H Hoos, and Kevin Leyton-Brown. SATzilla: Portfolio-based Algorithm Selection for SAT. J. Artif. Intell. Res.(JAIR), 32:565–606, 2008. Lingfeng Yang, Yi-Ying Yeh, Noah Goodman, and Pat Hanrahan. Just-in-time Compilation of MCMC for Probabilistic Programs, 2012. Jing Yu, V Anne Smith, Paul P Wang, Alexander J Hartemink, and Erich D Jarvis. Advances to Bayesian network inference for generating causal networks from observational biological data. Bioinformatics, 20(18):3594–3603, 2004. 78
6cs.PL
Toward Smart Power Grids: Communication Network Design for Power Grids Synchronization 12-E-SMG-1366 Hojjat Salehinejad1, Farhad Pouladi2, and Siamak Talebi1,3 1 Electrical Engineering Department, Shahid Bahonar University of Kerman, Kerman, Iran 2 Scientific-Applied Faculty of Post and Telecommunications, Tehran, Iran 3 Advanced Communications Research Institute, Sharif University of Technology, Tehran,Iran [email protected]; [email protected]; [email protected] Key Words: Ant Colony System (ACS), communication network, Laplacian matrix, smart grid, swing equation. I. INTRODUCTION Synchronization in power grids is one of the critical issues in system stability. A network in synchrony will be able to deliver stable electric power after a disturbance and avoid large scale black out. The synchronization for the power networks can be provided with or without assist of communication infrastructures. In case of none-communication network, physical links like transmission lines are used to couple generators. In this case, the synchronicity is determined using the ‘swing equation’ and the largest eigenvalue of the Laplacian matrix of the power network [1]. On the other side, in case of having an available communication infrastructure, the measurements of generators are sent to other generators for feedback control. In this case, the synchronicity is determined by the largest eigenvalue of the weighted sum of the Laplacian matrices of the power network and the communication network [1]. In [1], the communication network is modeled using ‘swing equation’ and Laplacian matrices of the power and communication networks. In [2], the reduce admittance ABSTRACT In smart power grids, keeping the synchronicity of generators and the corresponding controls is of great importance. To do so, a simple model is employed in terms of swing equation to represent the interactions among dynamics of generators and feedback control. In case of having a communication network available, the control can be done based on the transmitted measurements by the communication network. The stability of system is denoted by the largest eigenvalue of the weighted sum of the Laplacian matrices of the communication infrastructure and power network. In this work, we use graph theory to model the communication network as a graph problem. Then, Ant Colony System (ACS) is employed for optimum design of above graph for synchronization of power grids. Performance evaluation of the proposed method for the 39-bus New England power system versus methods such as exhaustive search and Rayleigh quotient approximation indicates feasibility and effectiveness of our method for even large scale smart power grids. 1 Toward Smart Power Grids: Communication Network Design for Power Grids Synchronization 27th International Power System Conference matrix impact, on the synchronization of power networks is studied. In [3] and [4], a continuum model is proposed for the power network and partial differential equations are employed for describing the dynamics. In [1], the derivation of the generator couple is similar to those of [3] and [4], while impact of network topology is also considered. The main difference between [1] and existing literature is that it has considered communication network in its model. In graph models, the problem of finding a route with specific characteristics is considered as a NP-hard problem, which needs considering all the possible routes [5]. Metaheuristic algorithms are of great interest for solving large scale problems and particularly the NP-hard ones in academia and industry [5, 6]. The crucial benefits of mentioned nondeterministic structures are their less complexity, low computational load, and few time usage for solving complex problems, which may have no derivatives, in a short period [5,6,8,9]. This paper investigates a new method initiated on Ant Colony System (ACS) for optimum design and planning of communication infrastructure for power grids synchronization using the communication network graph in [1]. In this method, it is attempted to study how to design the topology of the communication network such that the power network can be efficiently synchronized. The rest of the paper is organized as follows. In Section II a short survey on ACS is introduced. Next, a model for power grid is presented in section III. Then, the proposed method is discussed in section IV. Later, performance of proposed method is evaluated in section V. At the end, the paper is concluded by mentioning some research challenges in Section VI. discharge on the journey path. The ACS endeavors to work base on some artificial characteristics, which are adapted from real ants with the addition of visibility, discrete time, and memory. This computational approach to problem solving has many applications in complex problems such as in wireless communications [8] and intelligent systems [9]. According to [5] and [9], a typical ACS algorithm is consisted of the following steps, interested readers are referred to [9] for more details: Problem Graph Depiction: The artificial ants are mostly employed for solving discrete problems in discrete environments. This can result to considering the discrete problems as graphs with a set of N nodes and R routes. Ants Distribution Initializing: In order to move ants through the graph, some ants must first be placed on a set of random origin nodes. In most applications, trial and errors as well as nodes density in the region define the number of ants. Ants Probability Distribution Rule: When an ant wants to move from a node to another, a node transition rule is necessary. The transition probability of ant k from node i to node j is given by α β ⎧ τ ij ) (η ij ) ( ⎪⎪ α β p ijk = ⎨ ∑ (τ ih ) (η ih ) h ∉tabu ⎪ 0 ⎪⎩ j ∉ tabu k (1) k otherwise where τ ij and η ij are the pheromone intensity and the cost of direct route between nodes i and j, respectively; relative importance of τ ij and η ij are controlled by parameters α and β respectively; tabuk is set of unavailable edges for ant k. Update Global Trail: When every ant has assembled a solution at the end of each cycle, the intensity of pheromone is updated by a pheromone trail updating rule, which is given by II. ANT COLONY SYSTEM In natural world, ants find shortest path from a food source to their nest by using information of pheromone liquid that they τ ijnew = (1 − ρ )τ ijold + m ∑ Δτ k =1 2 k ij (2) Toward Smart Power Grids: Communication Network Design for Power Grids Synchronization 27th International Power System Conference angle δ i and neighboring nodes [1]. For the transmission line between two adjacent nodes i ~ k in the network, the flowing current is given where 0 < ρ < 1 is a constant parameter named pheromone evaporation. The amount of pheromone laid on the route between nodes i and j by ant k is ⎧Q ⎪ Δτ ijk = ⎨ f k ⎪0 ⎩ if route (i, j ) is traversed by th the k ant ( at the current cycle) Ii ,k = (3) otherwise (5) where Z = R + jX is the impedance of the transmission line, Ei and Ek are the voltages of the generators i ~ k and their magnitudes are considered constant as Ei = Ve jδ . By considering the Kirchoff’s current law and defining Y as the shunt admittance, the current conveying from the voltage source to the node is defined as where Q is a constant parameter and f k is the cost value of the found solution by ant k. Stopping Procedure: In typical algorithms, ACS procedure is completed by arriving to a predefined number of runs. i III. SYSTEM MODEL To plan an optimum communication network, first we need to model the communication synchronized power grid. To do so, first a power grid including the dynamics of generators is presented as a graph. Then, the synchronization with communication infrastructure is studied. I i = ∑ I i , k + YEi = i~k 1 ∑ ( Ei − Ek ) + YEi Z i~k Therefore, the electric power Pei determined as Pei ({δ i }i ~ j ,i = j ) = Re[ Ei I i* ] = (6) is V2 Z 2 ⎡ ⎤ (7) × ⎢ R ( N i − ∑ cos(δ i − δ k )) − X ∑ sin(δ i − δ k ) ⎥ i~k i~k ⎣ ⎦ 2 V X (δ i − δ k ) + V 2 Re [Y ] + V 2 Re [Y ] ≈ − 2 ∑ Z i~k A. Graph Modeled Power Network A typical power grid is consisted of some geographical sites as well as transmission lines. To model such network as a graph G , we can consider each site as a node A and each transmission line as an edge B of graph G < A, B > . In this model, with respect to [3], each generator supplies a time-varying current and time-varying power, using a constant voltage. Each site is equipped with one synchronous generator and two adjacent generators i and j is denoted by i ~ j [1]. Using the ‘swing equation’ denoted in [7], status of each generator can be described by the rotor rotation angle δ as M δ&& + Dδ& = Pm − Pe Ei − Ek Z where Ni is the number of neighbors of node i and δ i − δ k is assumed too small to keep just the first order term in the Taylor’s expansion. The dynamics of the power grid with respect to (4) and (7) is defined as 2 V X M δ&&i + Dδ&i = Pmi + 2 Z ∑ (δ i − δ k ) − V 2 Re [Y ] (8) i~k This equation clearly illustrates that the dynamics of different nodes i and k are coupled by the difference of rotation angles δ i and δ k [1]. By letting ξi = δ i − δ 0 for the mechanical power Pm we have (4) where M is the rotor inertia constant, D is the mechanical damping constant, Pm is the mechanical power , and Pe is the electrical power defined as a function of the Pm (δ i ) = Pm (ξ i ) = V 2 Re[Y ] + hξ i + o(ξ i ) 3 (9) Toward Smart Power Grids: Communication Network Design for Power Grids Synchronization 27th International Power System Conference i≤N ⎧ξ xi = ⎨ i ⎩θ i otherwise (11) To summarize the dynamics in (10) according to [1] we have x&i (t ) = ci xi + σ ij ∑ aij x j (12) i≠ j Fig. 1. An illustration of the virtual network with first order dynamics [1]. where for all nodes in the virtual network G% we have where h is a negative constant. This means that the generator will linearly increases its mechanical power with rate h , proportionally to the frequency drop, with higher order error, when the generator operates close to the stable point δ 0 [1]. i~ j ⎧1 aij = ⎨ ⎩0 otherwise (13) and B. Communication based Synchronized Model By considering a communication network available for the power grid synchronization, except for two physically adjacent generators 1 [1], the second order equation in (8) is converted to a first order one by setting θ = δ& as ⎧ xi ∈ set of ξ → ⎪ if ⎨ ⎪⎩ xi ∈ set of θ → σ ij = 0; ci = 1 σ ij = ψ ; ci = − D M (14) where ⎧ V2X ⎪ 2 Z ⎪ ψ =⎨ 2 ⎪ h − V XN 2 ⎪M Z ⎩ ⎧ ξ&i = θ i ⎪ (10) 1 D V2X ⎨& = − − + h ( x ) (ξ i − ξ k ) θ ξ θ ∑ ∑ i i j 2 ⎪ M M j >i i~k Z ⎩ j in G% corresponds to another generator (15) j in G% is the angle node of the same generator The system dynamics with communication infrastructure is rewritten in matrix form as where j > i in the communication network means the generators i and j are adjacent; h ∑ (ξ i − x j ) denotes the control at & ˆ x=Fx (16) where j >i generator i , where the information form adjacent generators in the communication network is employed. In [1], the variables θi and δ i are considered as being in different virtual nodes although they actually belong to the same real node. Therefore, the nodes corresponding to θi and δ i are called the ‘acceleration node’ and the ‘angle node’, respectively. Such a virtual network is denoted by G% as in Fig. 1 [1]. The variable xi is defined as ⎛ 0 ⎜ F̂= ⎜ h ˆ V 2 X ˆ Lp 2 ⎜ M Lc + Z ⎝ ⎞ ⎟ D ⎟ − ˆI ⎟ M ⎠ ˆI (17) where L̂ p is the Laplacian matrix of the power network and L̂c is that of the communication network. After applying unitary transformation, decoupled system state x% , can be achieved by solving the characteristic equation V 2 XM λi M && x%i + D x%&i − (h + ) x%i = 0 ∀ i = 1,..., N (18) 2 Z 1. Since, in practice, the information transmission may not catch up with the physical disturbance propagation [1]. 4 Toward Smart Power Grids: Communication Network Design for Power Grids Synchronization 27th International Power System Conference communication link. The generators can obtain the measurement information of all other generators connected to the center, which means if j > i , if and only if generators j and i are both connected to the control center. In the proposed method, all possible combinations of generators are considered connected to the center. Then, as the ACS moves forward, useless links are omitted and at the final stage the optimum links remain. By having in mind the brief ACS descriptions in section II, different steps of the proposed algorithm in pseudocode is illustrated in Fig. 3 and described as follow. Initialize: In this step, initial values of the algorithm parameters such as the number of ants, evaporation coefficients, and the network topology are loaded. Locate Ants: Ants are located on the control center node in this stage. We call the ant which has not been blocked in a junction an alive ant. Since each ant can traverse each node once, an ant is blocked in a node when it has no chance of continuing its transition toward another node and has no possible route to move backward. Construct Probability: When an active ant arrives to a node, the probability of moving from node i to node j for ant k is determined based on its cost function as which is φ1,2 = − D ± D 2 + 4M (h + V 2 XM Z 2M −2 λi ) (19) The condition for synchronization of the power network is that φ1,2 < 0 . In this case, we can have the following proposition using [1]. Proposition 1: The necessary and sufficient condition for the synchronization of the power network model with communication infrastructure in (10) is λmax < −h Z 2 (20) MV 2 X where λmax is the maximum eigenvalue of the matrix h ˆ V2X ˆ Lc + Lp [1]. 2 M Z This matrix can be considered as a Laplacian matrix L̂pc of the following weighted graph ⎧⎪ −ωij (Lˆ pc )ij = ⎨ ω ik ⎪⎩∑ k i≠ j i= j (21) where ⎧ h ⎪− M ⎪ ωij = ⎨ 2 V X ⎪ 2 ⎪⎩ Z i> j (22) ⎧ ⎪⎪ pijk = ⎨ ⎪ ⎪⎩ i~ j IV. OPTIMUM COMMUNICATION NETWORK DESIGN In this section, we introduce the proposed method for optimum design of communication infrastructure. In this method, we have employed ACS for selecting the communication lines in the network using the corresponding graph model of network in Fig. 2. In the understudy topology, there is a control center and a number of generators each connected to the center with a ∑ τ ijα ϑij − β τ ϑ h∉tabuk ih ij j ∉ tabuk −β (23) otherwise 0 where ϑij = λmax { and τ ij is h ˆ V2X ˆ Lcij + Lp ij } 2 M Z the direct route (24) pheromone intensity on the edge between nodes i and j ; α and β controls importance of τ ij and ϑij respectively and are set to 2 [5,6]; tabuk is set of blocked edges. 5 Toward Smart Power Grids: Communication Network Design for Power Grids Synchronization 27th International Power System Conference Update Tabu List: In this step, the edge that ant k has chosen is added to the tabu list to not be selected again and its probability not be calculated anymore. If ant k is blocked in a node, it is omitted from the active ant list. In other words, this step kills the blocked or arrived ant in the current iteration. Update Local Pheromone: The traditional ACS pheromone system is consisted of two main rules [5] where one is applied whilst constructing solutions (local pheromone update rule) and the other is applied after all ants have finished constructing a solution (global pheromone update rule) [5, 9]. After node selection, the pheromone amount of the edge between nodes i and j is updated for ant k as Fig. 2. An illustration of the power network and communication network, which can be modeled to a graph problem. Procedure ACS approach for communication network design Initialize For each iteration Locate Ants For each ant While ant is active Construct Probability Select Edge Update Tabu list End While Update Local Pheromone Next ant Update Global Pheromone Next iteration Select best direction End ACS approach for communication network design τijnew =τijold + γ ϑij (26) where γ > 1 is the amount of award. The edge with less cost ϑ obtains more pheromone award. Global Pheromone Update: The traditional step of each completed loop is global pheromone updating defined as Fig.3. Proposed ACS approach for communication network design in pseudocode. τ ijn ew = ρ τ ijo ld (27) Select Edge: A random parameter 0 ≤ q ≤ 1 with uniform probability is compared with the parameter 0 ≤ Q ≤ 1 which is usually fixed to 0.9 [5, 6]. The comparison result between Q and q picks up one of the two selection methods for the active ant to continue its route to the next node as where 0 < ρ < 1 is evaporation. Select Best Direction: After m iterations, edges with the maximum amount of pheromone are selected as the communication links as ( ) ( ) where ni * is list of connected nodes to the ⎧ arg M ax p ihk ⎪ j=⎨ k ⎪⎩ R oulette W heel p ih q>Q ni * = arg Max (τ ij ) ∀ i = 1,..., N . (25) (28) node i . otherw ise V. NUMERICAL RESULTS In this section, performance of the proposed ACS approach for communication infrastructure planning in power grids is evaluated. To do so, the 39- If q > Q , active ant selects edge with the highest probability, otherwise, Roulette Wheel rule is selected to choose the next node through probabilities. 6 Toward Smart Power Grids: Communication Network Design for Power Grids Synchronization 27th International Power System Conference bus New England power system is used with the assumption that each node has an identical generator [1]. The system is run in 30 iterations with 15 initial ants, γ = 10 , ρ = 0.9 , Q = 0.9 , α = 2 , and β = 2 , where the parameters are set based on trial and error as well as system performance [5,6]. A desktop computer with Intel Core2Quad Q8300 2.5GHz CPU and 3GB of RAM is employed for simulations in MATLAB 2010b environment. In Fig. 4, performance of the proposed ACS approach is compared with the greedy method in [1] as well as random selection of communication links. Clearly, small largest eigenvalue results in easier synchronization of the system [1]. As this figure illustrates, the random method has the least performance, i.e. has the major largest eigenvalue, among other two methods. This is while the ACS has significant difference from the other two methods and has the best performance. It shows that as the number of links increases and the topology gets more complex, the ACS is still the pioneer among all approaches for different number of links. In Fig. 5, performance of the greedy algorithms presented in [1], which are the exhaustive search in each step and the Rayleigh quotient based selection, and the proposed ACS approach in this paper are compared. As it is stated in [1], the greedy algorithms have almost very similar behavior in front of different number of links. However, the proposed ACS approach due to its heuristic characteristic as well as search behavior of artificial ants [5, 6], has the smallest eigenvalues for different number of links. This is while the exhaustive search can only guarantee the optimality in each step but cannot guarantee the global optimality [1]. As it is illustrated in Fig. 6, the proposed ACS method has the most normalized cost average in the first cycles of algorithm. As the number of iterations increases, the system arrives to a stable value and by the Fig. 4. Comparison of the ACS, greedy exhaustive search, and random communication link selection for different number of links. Fig. 5. Comparison of the ACS, greedy exhaustive search in each step, and the greedy Rayleigh quotient approach for different number of links. Fig. 6. Normalized cost of the ACS approach in different iterations. Fig. 7. Normalized cost of the ACS approach for different number of ants. 7 Toward Smart Power Grids: Communication Network Design for Power Grids Synchronization 27th International Power System Conference infrastructure planning using different techniques such as fuzzy controllers. 14th cycle it is converged. This convergence, which happened not so earlynot so late, demonstrates that the parameters of the ACS algorithm are set reasonably [5]. In another evaluation, performance of the ACS approach is evaluated for different number of ants as in Fig. 7. The result demonstrates that by considering one ant in the systems as the link finder, average of normalized costs is so high. However, by activating more ants, average of costs decreases while converging to a specific value of cost average. REFERENCES [1] H. Li and Z. Han, “Synchronization of Power Networks Without and With Communication Infrastructures,” Smart Grid Communications, IEEE Conference on, 2011. [2] F. D¨orfler and F. Bullo, “Spectral analysis of synchronization in a lossless structure-preserving power network model,” Smart Grid Communications, IEEE Conference on, 2010. [3] J. Thorp, C. Seyler, and A. Phadke, “Electromechanical wave propagation in large electric power systems,” IEEE Trans. Circuits & Systems: Fundamental, Theory and Applications, vol.45, 1998. VI. CONCLUSION AND FURTHER CHALLENGES This paper investigates a new algorithm based on Ant Colony System (ACS) for optimum communication network planning in power grids and develops them toward the smart grids. To do so, the dynamics of power network is first described by ‘swing equation’. Since we are using communication infrastructures for the synchronization purpose, the synchronization is determined by weighted sum of the Laplacian matrices of the power network and the communication network together. In order to have an optimum design of the communication network, the above topology is mapped to a graph problem and by employing the ACS algorithm, optimum communication links are selected. The simulation results show performance of the proposed method versus typical approaches in the literature. In real scenarios, as the scale of network grows, typical algorithms are unable for optimum network planning due to the NPhard characteristics of problem. However, it has been demonstrated in the literature that the heuristic methods are wellperformed tools for such problems. Though, this paper introduces the ACS approach for such problem, several challenges are remained such as incorporating difference of generators into the model and management of [4] M. Parashar, J. S. Thorp, and C. E. Seyler, “Continuum modeling of electromechanical dynamics in large-scale power systems,” IEEE Trans. Circuits and Systems–I: Regular Papers, vol.51, pp.1848-1858, 2004. [5] H. Salehinejad, S. Talebi, “Dynamic Fuzzy Logic-Ant Colony System Based Route Selection System,” Applied Computational Intelligence & Soft Computing, Vol. 2010. [6] M. Dorigo, LM. Gambardella, “Ant Colony System: A Cooperating Learning Approach to the Traveling Salesman Problem,” IEEE Trans. Evol. Computing, 1997. [7] J. Machowski, J. W. Bialek, and J. R. Bumby, Power System Dynamics: Stability and Control, 2nd edition, Wiley, 2008. [8] H. Salehinejad, F. Pouladi, and S. Talebi, “A Metaheuristic Approach to Spectrum Assignment for Opportunistic Spectrum Access,” 17th IEEE International Conference on Telecommunications, pp. 234-238, 2010. [9] H. Salehinejad, F. Pouladi, and S. Talebi, “Intelligent Navigation of EmergencyVehicles,” Proc. 4th Int. Conf. on Developments on e-Systems Engineering, Dubai, 2011. 8
9cs.NE
History-sensitive versus future-sensitive approaches to security in distributed systems Alejandro Hernandez and Flemming Nielson Department of Informatics, Technical University of Denmark {aher,nielson}@imm.dtu.dk We consider the use of aspect-oriented techniques as a flexible way to deal with security policies in distributed systems. Recent work suggests to use aspects for analysing the future behaviour of programs and to make access control decisions based on this; this gives the flavour of dealing with information flow rather than mere access control. We show in this paper that it is beneficial to augment this approach with history-based components as is the traditional approach in reference monitor-based approaches to mandatory access control. Our developments are performed in an aspect-oriented coordination language aiming to describe the Bell-LaPadula policy as elegantly as possible. Furthermore, the resulting language has the capability of combining both history- and future-sensitive policies, providing even more flexibility and power. 1 Introduction Distributed Systems are designed to manage large amounts of information, so they must be secured [9] to provide confidentiality for the information managed by them. The emerging Aspect-Orientation [1] field has been targeted to some security approaches [8]. Recently, a framework named AspectKB [7] has been proposed, with which is possible to model process calculi-like distributed systems and to capture security properties in a realistic way, attaching security policies to each location and then combining the relevant security policies when an interaction between locations takes place. The way of expressing security policies in the AspectKB framework refers to the traditional nondistributed information-flow [11] style of assuring security, which statically analyses the possible behaviours of the system in order to avoid any potential misuse in the future. In AspectKB, this is exploited by making access control decisions dynamically, yet not considering any state of the locations but possibly some potential future behaviour. In this paper, we shall consider a multilevel access control policy [6], the Bell-LaPadula model [2], and show some complications when trying to capture such a policy in a distributed framework in general, and in particular in a framework whose security policies focus on looking to the future, since such multilevel policies are better suited for past analysis of how the system reached its current state. We then propose an extension to the AspectKB framework, allowing to express also policies that look to the past. We do this by adding the notion of a localised state to the locations modelled in the extended framework and allowing the security policies to access those states to make their decisions. With this, multilevel policies as the Bell-LaPadula policy can be easily captured, and we show how. Since the original AspectKB framework was already intended to combine different security policies, with the extension done in this paper both policies that look to the past and policies that look to the future can be expressed and even combined. This not only benefits when trying to capture specific policies (such as the Bell-LaPadula one), but it also allows us to model every policy in its original way, providing more flexibility to the resulting extended framework. Bliudze, Bruni, Grohmann, Silva (Eds.): Third Interaction and Concurrency Experience (ICE 2010) EPTCS 38, 2010, pp. 29–43, doi:10.4204/EPTCS.38.5 c A. Hernandez and F. Nielson This work is licensed under the Creative Commons Attribution License. 30 History- vs. future-sensitive approaches to security in distributed systems Moreover, we shall argue that for some situations, expressing a policy in its original way could more precisely capture what is intended, and this insight would mean that our extended framework is more powerful as well. We shall start discussing this latter issue in the remainder of this Section (Subsection 1.1). In Section 2 we present a review of the Bell-LaPadula policy in its original formulation, and then we assess the challenges of adapting it to a distributed setting. Section 3 gives a brief review of the Logic used for dealing with the combination of policies. In Section 4 we present our extended framework, and show how to precisely capture the Bell-LaPadula policy. We also discuss why the resulting framework is more flexible than the existing one. In Section 5 we conclude. 1.1 Limitations of looking to the future The framework we shall be dealing with throughout this paper is the formal language AspectKB. In that framework, which follows a process algebraic approach, the processes are modelled as actions taking place in specific locations, and interacting with other locations modelled as well. Furthermore, security policies can also be modelled, following an aspect-oriented manner. The policies can express their intentions by analysing the continuation (namely the process after the current action) of the involved processes, so that it is possible to know in advance what a process might do in the future. This reflects an information-flow style of providing security. However, this information-flow style is not as adequate as it was for sequential programs. Indeed, since the only process that can be statically analysed is the one that continues after the current action, all the possible outcomes that may occur due to other processes could not be predicted. This means that, when deciding whether to allow the interaction to happen or not, it is necessary to look to the future of just one process, and this can lead to two possible ways of obtaining imprecise decisions, either over-approximation or under-approximation. For understanding what over-approximation is, let us assume we pessimistically expect that a particular action done by a process could, because of other processes we do not know about, lead to an insecure state. In this situation we may disallow the process to execute that action, but in some cases there might be no other process performing anything that would lead to an insecure state. For understanding what under-approximation is, let us assume we optimistically expect that a particular action done by a process will not lead to an insecure state because the very same process will not perform another related action that leads to such a state. In this situation we may allow the process to execute that action, but in some cases there might be some other process that makes the system reach some insecure state, due to some interactions that could have been avoided if the action was disallowed. Let us discuss a simple example, without going into syntactic and semantic details, but still thinking about distributed processes and policies. Let us think about a security policy where we have different security levels, and every location is assigned to some level. We do not want any information to be leaked from any security level to lower ones. Then, we should allow a process, running in a given location, to read data from another location, as long as the following two conditions are met: first, the other location, where the data is right now, is in a security level not higher than the one where the process is running; second, the process will not try, in the future, to write information to locations with security levels lower than the level of the location where the data is right now, since this writing may be influenced by the reading previously done. Let us assume now a particular situation where we have 4 locations (say A, B,C and D), and 3 security levels (say 1, 2 and 3). Let us assume the security levels are ordered as their values in natural numbers (3 > 2 > 1). Let us assume that location A is in security level 1, locations B and C are both in security level 2, and location D is in security level 3. Figure 1 contains three cases of such a situation, showing A. Hernandez and F. Nielson D t1 read t2 write B 31 Sec. level 3 D Sec. level 2 C E t2 write t2 Sec. level 2 write t1 read B C Sec. level 1 A Sec. level 3 B C Sec. level 1 (b) Secure case, but overapproximation would incorrectly disallow it. Sec. level 3 t3 Sec. level 2 write t1 read Sec. level 1 A (a) Insecure case, detectable by information-flow approach. D A (c) Insecure case, but underapproximation would incorrectly allow it. Figure 1: Examples of situations that might happen. the locations and their security levels in different layers. Illustrated in Figure 1a, there is a process in location D, that tries to read information from location B at t1, and then tries to write some information to location A at t2. This process should clearly be forbidden, because it does not meet the second condition of the policy we are trying to capture (although it meets the first one). Of course this can be done following the information-flow approach, looking to the future at t1, since we know that the process trying to read from B will try to write to A, and this should not be allowed. However, let us think about another case, illustrated in Figure 1b. Let us say that the process running in location D whose first action is to read information from location B at t1 then tries to write some information to location C (in the same level as B) at t2. This does meet the second condition of the policy, since the only information the process could write to C is what it has read from B. Therefore, this should be allowed. But let us consider the next extension to the example, illustrated in Figure 1c. Assume there is a fifth location E that is in security level 3. Assume there is a process running in E that writes some information to D at t2, after the process running in D has read from B at t1. In this case, the future writing to C by the process running in D (which in this case will be done at t3) should be forbidden, because it might be influenced by the new information learned by location D at t2. Anyway, since the process that writes to C is not the same as the one running in D, the process algebraic way of modelling does not permit us to know in advance (at t1) that this will happen. If we had taken an approach of looking to the past, then we would have checked the insecure operation of writing to C right in the moment of the writing (at t3), and we would have known that some information from E would be leaked, and therefore we would avoid the write operation. We could take the information-flow approach using over-approximation, and always avoid this type of write operation (e.g. from D to C, since the former is in level 3 while the latter in level 2), but that would be very imprecise (and restrictive), since sometimes there is nothing insecure in doing this write operation, as shown in the case of Figure 1b. Taking the information-flow approach using underapproximation would mean allowing the process in D to perform the read and the subsequent write, since that write operation is not insecure. This will be secure enough in the case of Figure 1b, but not in the case of Figure 1c. So, we have found some possible situations where using an information-flow approach in a distributed setting is not completely precise, and therefore another approach might be taken, for instance 32 History- vs. future-sensitive approaches to security in distributed systems looking to the past. In the rest of this work, we shall be studying how to deal with looking to the past, and how to extend our distributed-systems framework to achieve this. We shall see that the resulting framework allows us to combine both approaches, therefore obtaining the advantages of both of them. In particular, we shall see that the simple example we have seen is just one possible instance of something that can easily (and more precisely) be captured by the Bell-LaPadula policy. 2 Assessment of the Bell-LaPadula model In Subsection 1.1 we saw that for distributed systems the information-flow approach is not as adequate as it was for sequential programs. In this Section, we review another approach, the Bell-LaPadula (BLP for short) policy, and discuss the challenges of using it in a distributed setting, but aiming to show that this can be as adequate as in its original formulation. 2.1 The Operating System view of BLP The BLP model is the most traditional Mandatory Access Control model. Here we briefly introduce it, inspired by [6], abstracting some unnecessary details that do not contribute to our study. State. The computer system will be checked for security by looking into its state. For representing it, some sets must be introduced: • S is the set of subjects that may use the information stored in the system, • O is the set of objects (pieces of information) stored in the system, • A = {read, write} is the set of operations a subject may do over an object, • L is a lattice of security levels. Every state of the system is composed of a set of tuples of the form (s, o, a) (each tuple would mean that subject s is doing an a operation over object o), and of a tuple of functions ( fS , fC , fO ) with types S → L, S → L and O → L. The functions are supposed to be total functions, and they will give, respectively, the maximum security level a subject can have (its clearance), the current security level a subject has1 , and the security level an object has (its classification). Formally, a state (B, F) ∈ B × F , where F = ( fS , fC , fO ), and where: • B = P(S × O × A) • F = (S → L) × (S → L) × (O → L) Policies. The BLP model specifies two properties that every state should meet in order to be considered secure. • ss-property2 . A state (B, F) satisfies this property iff ∀(s, o, a) ∈ B : a = read =⇒ fS (s) ≥ fO (o). This means that each object being read by a subject should be in a level not higher than the level the subject is able to reach, which is usually called no read-up. 1A subject can log into the system with a lower security level than its corresponding clearance. Once it did so, that security level cannot be changed until it logs in again. 2 For “simple security”. A. Hernandez and F. Nielson 33 • ?-property3 . This property consists of two parts. A state (B, F) satisfies the first part (let us name it ?-property.1) of this property iff ∀(s, o, a) ∈ B : a = write =⇒ fO (o) ≥ fC (s). This means that each object being written by a subject should be in a level not lower than the level the subject is currently in, which is usually called no write-down. On the other side, a state (B, F) satisfies the second part (let us name it ?-property.2) of this property iff ∀(s, o, a) ∈ B : a = write =⇒ [∀(s, o0 , a0 ) ∈ B : a0 = read =⇒ fO (o) ≥ fO (o0 )]. This means that if a specific subject (note the use of the same s in both quantifications) is operating with many objects, some being read and some being written, then no object being read could be in a higher level than any object being written. This prevents the subject to read some high-level object and then write a low-level one. A state is said to be secure if it satisfies both properties. 2.2 The challenges of distribution The BLP model was originally meant for Operating Systems. These have a particular feature: they are centralised, meaning that a central controller (i.e. the Operating System) takes care of everything that happens in the system. In particular it can control (and in some cases restrict) the processes that try to access resources. Moreover, one key concept needed for checking BLP policy compliance is the state, and since Operating Systems have a centralised state, they can do the calculations for knowing whether the BLP policy is met or not. Lack of central controller. In a distributed setting we do not have any central controller; many locations run in parallel and share information, but no location can know what other locations are doing. Therefore, once a location is allowed access to some resource, there is no way other locations can forbid it from doing whatever it wants with the resource. In particular, there is no notion of state, processes interact and synchronise, but no central entity knows what has happened in the whole system so far. It should be clear that a distributed framework is not trivially able to meet security properties that were originally developed for simpler systems, as for example centralised ones or sequential programs. In the case of Information-Flow approach, we have seen some simple examples where we can lose precision. In the case of BLP, in the next Subsection we propose an extension that will help us to adapt the policy to a distributed setting. 2.3 Extending BLP The original formulation of the BLP policy relies on three functions, two of which can be applied to every subject and one to every object. They can be computed by the Operating System every time an action is to be executed, to check whether the resulting state will still be secure, and then decide whether to allow the action or not. Here we propose an extension to their domains to have common signatures, since in a setting without a central controller we might want to call any of them with any possible entity of the system without distinguishing between objects and subjects. We also propose a fourth function which captures information about the past interactions for each entity. Later in the paper we will see that this latter function can be used to have a form of localised state. 3 Read “star property”. In some formulations of the BLP model, this property only consists of the first part because the ss-property uses fC instead of fS , and then the second part is just a consequence. However, that kind of formulation is again too restrictive, since a subject cannot perform read operations in levels up to its clearance, but just up to the level it has logged in. 34 History- vs. future-sensitive approaches to security in distributed systems The types of the three existing functions are then changed to S ∪ O → L for all of them, and their definitions are extended in a straighforward way as follows: ∀o ∈ O, s ∈ S : fS (o) = fO (o) ∧ fC (o) = fO (o) ∧ fO (s) = fS (s) We call the new function fH since it keeps track of (a part of) the history of the system. When we apply this function to a particular input subject (resp. object) we should learn what kinds of interactions the subject (resp. object) has been involved in during the past, therefore the output of the function would be a kind of current state of the argument subject (resp. object). To capture this notion of state, the function will not be fixed once and for all, as the original three functions were. Indeed, the output of this function will be: • For a particular subject: the least upper bound of the security levels of all the objects read by the subject so far. • For a particular object: the least upper bound of the security levels of all the subjects that have written to the object so far. Formally, this can be expressed as follows (assuming (B, ( fS , fC , fO , fH )) to be some “virtual” global state that depends on the interactions that have happened and (B0 , ( fS , fC , fO , fH0 )) the next one): ∀(s, o, a) ∈ B : ( ( (a = read =⇒ fH (s) ≥ fO (o)) ∧ (a = write =⇒ fH (o) ≥ fC (s)) ) ∧ 0 0 0 0 0 0 0 0 ∀(s , o , a ) ∈ B : ( fH (o ) ≥ fH (o) ∧ fH (s ) ≥ fH (s)) ) This means that every time an interaction takes place, changing the state from (B, ( fS , fC , fO , fH )) to some (B0 , ( fS , fC , fO , fH0 )), the output of fH0 for some input may be higher than or equal to that of fH . It can actually be higher depending on the values of the entities read/written, for keeping the resulting least upper bound we expect to have. Indeed, a very simple result tells us that for every set L : t(L ) = t{t(L \ {a}), a} (∀a ∈ L ) (1) And we should also observe that t(0) / = ⊥. We shall use these four functions to capture this extended version of BLP in a distributed setting. 3 A brief review of Belnap Logic For granting access according to some security policy, the traditional boolean values (tt and ff) are enough: tt grants while ff denies access. However, for a distributed setting, where policies might be contradictory (or not sufficiently informative), those two values might not be enough. We shall consider an extension to the Boolean Logic proposed by Belnap [3], which has been used for combining security policies [4]. In this extension to the boolean logic, two more values are considered: ⊥ and > (read “bottom” and “top”). The traditional tt would mean “the policy accepts the interaction” whereas the traditional ff would mean “the policy does not accept the interaction”. Since different locations might aim at different security properties, their policies could be contradictory or they may lack information about some particular interaction. These situations can be represented by the two extra values that we have: ⊥ meaning “no decision” and > meaning “contradiction” or “conflict”. With this set of values, which we will call here Four (i.e. Four = {⊥, tt, ff, >}), it is possible to extend the usual boolean operations (∧ and ∨) and to define new ones (⊗ and ⊕). For obtaining that, the set Four is equipped with two partial orderings, say ≤k and ≤t , as shown in Figure 2. A. Hernandez and F. Nielson 35 > tt @ ≤k tt @ @@ @ @ @ @ ff ≤t ⊥ @ @> @@ @ @ @ @ ⊥ ff Figure 2: The Belnap bilattice Four: ≤k and ≤t . The usual boolean ∧ is extended as computing the greatest lower bound in the ≤t lattice, and the usual ∨ as computing the least upper bound (thereby obtaining the same results as in boolean logic if the operands belong to {tt, ff}). Analogously, the new operators over Four can be defined as computing the greatest lower bound (the ⊗ operator) and the least upper bound (the ⊕ operator), both in the ≤k lattice.4 The negation operator ¬ is extended by leaving the two new values unchanged (i.e. ¬⊥ = ⊥ and ¬> = >), and the implication ⇒ is extended as follows:  p1 ⇒ p2 = p2 if p1 ≤k tt tt otherwise ∀p1 , p2 ∈ Four Another useful operator is the priority >, which returns the first operand unless it is ⊥, in which case it returns the second operand. This would always consider what the first operand suggests unless it has no decision, in which case the second operand is considered. 4 Aspect-oriented framework for security As mentioned, the AspectKB [7] framework allows us to express location-based systems in a processcalculus-oriented manner. This is achieved by extending the KLAIM [10] coordination [5] language. These located processes interact with other locations when they try to gather (or put) information from (or into) them (maybe themselves), which are usually named tuple spaces. The possibility of attaching to each location (regardless of whether it is a process or a tuple location) some security policy, which will govern the interactions the location may be involved in, turns the AspectKB language into an aspectoriented language. Then, whenever an interaction takes place, the relevant policies are considered by the semantics to either grant or deny the interaction, using the four-valued Belnap Logic for deciding in a consistent way. In this Section, an extension to that framework is made, mixing all process locations and tuple locations into just entity locations, and attaching to them more aspects than just the security policies. The extra information attached to each location refers to security levels in the sense of a multilevel security policy. Moreover, the mechanisms of the language explicitly keep track of some information (at a certain level of abstraction) regarding the interactions that have taken place, giving the flavour of a localised 4 Notice that this could also be done by just extending the “truth tables” of the usual boolean operators and defining new ones for the new operators. That would mean, however, having not just 2 truth tables with 4 cells each (as in Boolean Logic) but 4 truth tables with 16 cells each, making it difficult to remember what each operator produces. 36 History- vs. future-sensitive approaches to security in distributed systems N ∈ Net P ∈ Proc a ∈ Act `, `λ ∈ Loc w ∈ Annot lst ∈ LocSt γ ∈L → − N ::= N1 || N2 | l ::w P | l ::w h l i P ::= P1 | P2 | ∑i ai .Pi | ∗P → − → − → − a ::= out( ` ) @ ` | in(`λ ) @ ` | read(`λ ) @ ` ` ::= u | l `λ ::= ` | !u w ::= < lst, pol > lst ::= < γ S , γ C , γ H , γ O > le f t implicit Table 1: AspectKB+ Syntax – Nets, Processes, Actions and States. state, which the semantics of the language keep updated5 . Besides this, one can write Aspects using the extra information, which is basically the output of the functions mentioned in Subsection 2.3 (considering that every entity location can be either a subject and/or an object in the whole system, so every location can be a potential input to all those functions). Then, this will allow us to capture, among others, the BLP policies without losing precision. Following this informal introduction to our extension, which we shall call AspectKB+ due to its enhanced features, we shall present its formalities. 4.1 Syntax and Semantics Syntax. The syntax of the AspectKB+ language is given in Tables 1 and 2. Table 1 gives the syntax for nets, the basic modules that can be described in the language. A net is a parallel composition of located processes and/or located tuples (data), together with an annotation explained below. Each process can be a parallel composition of processes, a non-deterministic choice between processes following an action, or a replicated process. A process not performing any action shall be written 0. The allowed actions are reading from a location (in and read, resp. with or without deleting the data read) or writing to it (out). Every location has an annotation w, whose first part (lst) is intended to keep track of the interactions that the location has been involved in. This localised state consists of 4 pieces of information (namely γ S , γ C , γ H , and γ O ) that are elements of the lattice L of security levels (introduced in Subsection 2.1). Since every location can be input to the four functions fS , fC , fH and fO , and since the result of evaluating them is in L, we can keep attached to each location the result of evaluating each of those four functions. The second part (pol) of the annotation in every location is the actual security policy governing the location, which has to be expressed using the syntax of Table 2. The policy can be a Belnap combination of policies, a boolean value, or a single aspect. This latter consists of a cut (the action, together with its continuation, to be trapped by the aspect), a condition cond (a boolean applicability condition) and a recommendation rec (a four-valued Belnap Logic advice for the aspect). To define an aspect, one may refer to the security levels stored in the trapped interaction or to a single value from the lattice L. To do the former, one can write an aspect naming some of the five syntactic names (Ss ,Cs , Hs , Ot , Ht ) specified in the category v ∈ Lev, which will later be matched by the semantics to the specific values kept in the trapped interaction. To do the latter, one can write an aspect providing a specific value from L, as the category v ∈ Lev permits (by having γ among its choices). Finally, the occurs-in operator, which can be easily defined in a compositional way, checks whether the action occurs in the continuation process. 5 As one can argue, having information inside the locations, namely the tuples, also gives us the flavour of state, yet that is information that changes according to what processes do, so we cannot rely on that information for guaranteeing any property. A. Hernandez and F. Nielson 37 pol ∈ Pol pol ::= asp | ¬pol | pol ⊕ pol | pol ⊗ pol | pol ⇒ pol | pol > pol | pol ∧ pol | pol ∨ pol | true | false asp ∈ Asp asp ::= [rec if cut : cond] cut ∈ Cut cut ::= ` :: at . X − → − → → − at ∈ Actt at ::= out( `t ) @ ` | in(`tλ ) @ ` | read(`tλ ) @ ` rec ∈ Rec rec ::= `1 = `2 | ¬rec | rec ⊕ rec | rec ⊗ rec | rec ∧ rec | rec ∨ rec | rec ⇒ rec | true | false | a occurs-in X | v1 ≥ v2 cond ∈ Cond cond ::= `1 = `2 | ¬cond | cond1 ∧ cond2 | cond1 ∨ cond2 | true | false | a occurs-in X v ∈ Lev v ::= Ss | Cs | Hs | Ot | Ht | γ `t ::= ` | `tλ ::= ` | Table 2: AspectKB+ Syntax - Aspects for Security Policies. Semantics. The semantics is given by a one-step reduction relation. It makes use of a structural congruence on nets (defined in Table 4), and also of an operation match, for matching input patterns to actual data, which could easily be defined in an inductive way by the structure of its arguments. The reaction rules (defined in Table 3) prescribe how the system may evolve in the presence of some process location and some target location. In the “where” lines of each rule, the boolean condition b is obtained by evaluating the security policies of the locations involved in the computation using the evaluation function [[.]] (formally defined in Subsection 4.2). This is done to either allow or disallow the process to compute, and for this it also makes use of the function grant (also formally defined in Subsection 4.2) for turning four-valued Belnap truth values into boolean truth values. If the action was disallowed the involved process simply terminates, thereby becoming just a 0; otherwise the process evolves as the next paragraphs explain. In the case of a read or in action, the process location ls is subject to a substitution, using the result of the matching done with the match operation. Moreover, the localised state of that location might be modified, changing the historic component of its annotation by the least upper bound of the previous value and the security level of the target location lt . This follows the suggestion of Equation (1). In the case of an out action, the data is stored in the target location. However, this is not done directly, but actually another “virtual” location is created, with a special localised state. This is intended to permit the virtual location holding the pre-existing process Q to keep running as it was, without being interfered with. The virtual location now holding the data has an historic component on the annotation that is the least upper bound of the previous value in the location lt and the security level of the process location ls that has written the data. Of course it is possible that the value is the same as in the original lt . To simulate the log-in of a subject in a lower level than its clearance, a process can be annotated with a value for γ C lower that the γ S . This value will then never change, just as the γ S and γ O components of the localised state. Note also that the security policy annotating each location never changes either. 4.2 Meaning of policies and granting access In the “where” lines of each semantic rule there is a check that tells whether the interaction should be allowed. For this purpose, the policies of both locations taking part in the interaction are combined using the Belnap operator ⊕, and the result of the evaluation by the operator [[.]] is passed to the function grant. The function grant is defined by grant(p) = p ≤k tt, for all p in Four. The aim of granting access 38 History- vs. future-sensitive approaches to security in distributed systems N1 → N10 N1 || N2 → N10 || N2 N≡M M → M0 M0 ≡ N 0 N → N0 → − − λ wt → (ls ::ws read( (` )@lt .P + · · · ) || (lt :: h l i) → − → → − − 0 ls ::ws Pθ || lt ::wt h l i if b ∧ match(`λ ; l ) = θ → → − ls ::ws 0 || lt ::wt h l i if ¬b O H C S (δ ∈ {s,t}); where wδ =<< γδ , γδ , γδ , γδ >, polδ >, → − and where b = grant([[pols ⊕ polt ]](ls :: read(`λ )@lt .P, < γsS , γsC , γtO , γsH , γtH >)); and where w0s =<< γsS , γsC , (γsH t (γtO t γtH )), γsO >, pols > . → − − λ )@l .P + · · · ) || (l ::wt h→ (ls ::ws in(`( l i) t t → − → − 0 if b ∧ match(`λ ; l ) = θ ls ::ws Pθ → → − ls ::ws 0 || lt ::wt h l i if ¬b (δ ∈ {s,t}); where wδ =<< γδS , γδC , γδH , γδO >, polδ >, → − λ and where b = grant([[pols ⊕ polt ]](ls :: in(` )@lt .P, < γsS , γsC , γtO , γsH , γtH >)); and where w0s =<< γsS , γsC , (γsH t (γtO t γtH )), γsO >, pols > . → − (ls ::ws out((l )@lt .P + · · · ) || (lt ::wt Q) − 0 → ls ::ws P || lt ::wt h l i || lt ::wt Q if b → ls ::ws 0 || lt ::wt Q if ¬b S C H O where wδ =<< γδ , γδ , γδ , γδ >, polδ >, (δ ∈ {s,t}); → − and where b = grant([[pols ⊕ polt ]](ls :: out( l )@lt .P, < γsS , γsC , γtO , γsH , γtH >)); and where wt0 =<< γtS , γtC , (γtH t (γsC t γsH )), γtO >, polt > . Table 3: Reaction Semantics of AspectKB+ . whenever the result is less than or equal to tt is for doing so not only if both policies agree with this, but also if some of the policies lack some decision, because this would mean that it does not actually forbid the interaction. This is related to the use of ⊕ for combining the policies, and the aim is that whenever the policies are contradictory, the result of the evaluation by [[.]] gives > ∈ Four, thereby denying access as long as at least one policy has evidence that the interaction should be disallowed6 . The evaluation function [[.]] (Table 5) is defined inductively on the structure of the (infix) policy. The base cases are when the policy is just a constant (true or false) and when it is just an aspect (i.e. it belongs to Asp). In this latter case, the first (postfix) parameter, a specific action with continuation, is checked against the cut of the aspect, a generic action with continuation, using the function check, which could easily be defined in an inductive way by the structure of its arguments. This is achieved using a function extract, which produces the list of literals that occur in an action with continuation in a way that, for instance, extract(` :: out(`t1 , · · · , `tn )@`0 .X) = [`, out, `t1 , · · · , `tn , `0 , X], which is done by just pattern matching the components of the given parameter and then pushing them into a list. The function check determines whether there is a substitution θ that can be performed in the cut that matches the parameter given to the [[.]]. This is needed because the cut can possibly consist of variables for representing the locations and even the arguments of the action in the cut may not be specified. If there is such θ , the rec 6 This follows a conservative principle, as to actually grant access there should not be any policy at all denying the interaction. A. Hernandez and F. Nielson 39 l ::w P1 | P2 ≡ l ::w P1 || l ::w P2 l ::w ∗P ≡ l ::w P | ∗P l ::w P ≡ l ::w P || l ::w 0 N1 ≡ N2 N || N1 ≡ N || N2 Table 4: Structural Congruence. [[[rec if cut : cond]]](l :: a . P, < γS , γC , γO , γHs , γHt >) =   case check( extract(cut) ; extract(l :: a . P)) of   fail : ⊥    0   [((rec θ )θ )] if [(cond θ )]   θ:   ⊥ if ¬[(cond θ )] 0 where θ = [γS /Ss , γC /Cs , γO /Ot , γHs /Hs , γHt /Ht ] [[¬pol]](N, Γ) = ¬([[pol]](N, Γ)) [[pol1 φ pol2 ]](N, Γ) = ([[pol1 ]](N, Γ)) φ ([[pol2 ]](N, Γ)), (φ ∈ {⊕, ⊗, ⇒, >, ∧, ∨}) [[true]](N, Γ) = tt [[false]](N, Γ) = ff Table 5: Meaning of Policies in Pol for AspectKB+. and the cond are substituted using it to determine the result. This is achieved using the usual two-valued meaning [(cond)], which could be straightforwardly adapted to a four-valued meaning [(rec)]. Due to the semantics of Table 3, the first parameter will always be the actual action taking place. The second (postfix) parameter (consisting of five values in the lattice L) is used to produce another special substitution (θ 0 ) that is also used (together with θ ) to determine the result of the recommendation rec. Due to the semantics of Table 3, the security levels annotated in the actual interacting locations are given here. Indeed, those taken from the target location are the ones identifying the classification (γtO ) of the location and the historic annotation (γtH ). Those taken from the process location are the ones identifying the clearance (γsS ) and the current level (γsC ) of the location and the historic annotation (γsH ). It should be noticed that, while the θ substitutes according to some checking performed between the cut and the first parameter (the actual action), the θ 0 substitutes according to the five syntactic names prescribed by the syntax of Table 2, in the v ∈ Lev meta-variable. Therefore, when describing a system in AspectKB+, these syntactic names could be used to describe recommendations (rec) that will later be used to check actual security levels of the interacting locations, as already pointed in Subsection 4.1. 4.3 Capturing BLP in AspectKB+ Having developed our formal framework, we shall show how the extended BLP policy of Subsection 2.3 can be elegantly captured. We shall also show that we can easily decide which cases of the example in Subsection 1.1 are secure and which are not, without losing any precision, unlike the information-flow approach. Remember that AspectKB+ is a process calculus, and even though in the original formulation of BLP the compliance of a state with the policy is checked in every state, we can just check if a transition might take us to an “insecure state”. Also remember that AspectKB+ provides us with the possibility, when describing aspects, of writing in the recommendation rec the five syntactic names we have mentioned, 40 History- vs. future-sensitive approaches to security in distributed systems which later will be substituted by the evaluation function [[.]]. So basically using those distinctive names we aim to capture the BLP policy. The first aspects. Let us focus first on the ss-property, which prescribes that a subject cannot read an object that has higher security level than itself. The operations that can read information from other locations are the read and the in actions. So the aspects that capture the ss-property are the following:   Ss ≥ Ot if ls :: read(−)@lt .P : true (2)  Ss ≥ Ot if ls :: in(−)@lt .P : true  (3) Note that each aspect is trapping a particular operation, without caring about the parameters and with a trivial applicability condition. Whenever some of these aspects trap an action, the recommendation will be considered, granting access only if the security level of the subject is not lower than that of the object, since the two names Ss and Ot will then be replaced by the corresponding security levels of the actual interacting locations, thanks to Tables 3 and 5. For the ?-property.1, which prescribes that a subject cannot write any object that has lower security level than the level the subject is currently in, we have to follow a similar approach. Considering that the write operations are the out and the in (since deleting data is a form of write, because some implicit information could be communicated), the aspects are as follows:   Ot ≥ Cs if ls :: out(−)@lt .P : true (4)  Ot ≥ Cs if ls :: in(−)@lt .P : true  (5) Whenever some of these aspects trap an action, the recommendation will only grant access if the security level of the object is not lower than the one the subject is currently in (note the use of Cs instead of Ss ). The ?-property.2. Now let us consider the ?-property.2, which was basically the one that initiated the proposal made in this paper, due to the difficulty of capturing it precisely in a distributed setting. Note that the semantics of AspectKB+ will keep track of the least upper bound of the security levels of the objects read by a particular subject location, because it updates it whenever the subject reads something that is not lower than the current value. A similar observation can be done for the object locations. Let us consider a subject location, which might have read some high information as long as its security level allows it (otherwise either aspect (2) or (3) would have denied it). Any subsequent write to a low location must be denied, and in principle either aspect (4) or (5) might decide this, unless the subject is logged into the system with a low security level. In any case, using the localised state that we have in the subject location, and making use of the Hs syntactic name provided by the syntax for expressing aspects, we define the following aspects:   Ot ≥ Hs if ls :: out(−)@lt .P : true (6)  Ot ≥ Hs if ls :: in(−)@lt .P : true  (7) They can be understood in a very similar way as aspects (4) and (5), with the difference being that they are considering the localised state of the subject location, instead of the level where the subject has logged into the system. A. Hernandez and F. Nielson 41 Analogous considerations can be done for an object location, and we can define the following aspects for finishing to capture the whole BLP policy:  Ss ≥ Ht if ls :: read(−)@lt .P : true  Ss ≥ Ht if ls :: in(−)@lt .P : true   (8) (9) Combining the aspects. After defining these eight aspects, the idea is to combine and attach them to every location, so every time an interaction is to take place, the semantics will consider all the aspects before allowing the interaction to happen. Since the BLP model says that a state is secure if both properties are satisfied, then we need to make sure that none of the aspects representing the properties detects a possible insecure interaction, as that would mean that at least one of the properties is not satisfied. For capturing this situation, again the Belnap operator that must be used to combine the aspects for attaching them to the locations is ⊕. Now we are ready to state our first Lemma: Lemma 4.1 If a distributed system is insecure in the sense of Section 2.3, then some of the aspects from (2) to (9) will deny the insecure interaction. For the converse we need to make an extra observation, discussed in the following Sub-subsection. 4.3.1 Initialising the historic value The aspects just defined will check, among other values, the historic component γ H attached to each location, and that value will be kept updated by the semantics. But, initially, one must give a particular value for the component. The chosen value will not affect the correctness of the aspects detecting insecure interactions, but to fulfil our requirement that we should not lose any precision while doing so (unlike the information-flow approach) the value should be ⊥ ∈ L. This follows the suggestion of Equation (1) and the observation just after it. Now we are ready to state our the converse of the previous Lemma: Lemma 4.2 If some of the aspects from (2) to (9) deny an interaction, then the hypothetical resulting global state, if the interaction was actually allowed, is insecure in the sense of Section 2.3. Furthermore, we can now easily verify that the three examples of Figure 1 are precisely captured. In particular, it should be taken into account what could happen after the process in location E writes to location D (Figure 1c). For the process in D to be actually influenced by this, it must explicitly read the data, since the semantics of AspectKB+ will put it in another “virtual” location, with a higher historic component. So if the process is influenced, then at t3 the aspects (actually aspect (6)) will prevent the write to C, otherwise the write will be allowed. 4.4 A very simple example Let us now consider a very simple example to show how to combine looking to the future and to the past. Assume an airline has a database where information about the passengers is kept. The historic component of the database location is initialised to ⊥ ∈ L so any process could read from it, but after 42 History- vs. future-sensitive approaches to security in distributed systems some data is written, only some processes could do so, according to the security level of the data written. The aspect that prescribes this is:   clearanceu ≥ historyAirlineDB (10) if u :: read(pass, −)@AirlineDB.P : true As one can notice, this is a special case of aspect (8), but it is written like this here to emphasise the example. One of the process locations that will not be allowed to read data from the database due to the previous aspect is the Government, whose clearance should not be enough to satisfy the rec of the aspect. Indeed, the historic component of the database should be high enough since the data written in there might be sensitive for the passengers. However, in times of heightened security due to probable threats, the Government should be able to audit the passengers, therefore allowing it to read the database is necessary. Anyway, this should be allowed as long as the Government will not, later, give the passengers’ data to the Press, to keep satisfying the right to privacy of the passengers. The following aspect prescribes this:   ¬(out(data)@PressRelease occurs-in P) (11) if Government :: read(pass, data)@AirlineDB.P : test(threatlevel, high)@AirlineDB This is just a little aspect that looks to the future7 , where we see how this is achieved. In the presence of this aspect, the Government will be allowed to perform the read action, as long as there is a tuple < threatlevel, high > in the Airline database (i.e. the Airline was already notified of the heightened security situation), and also as long as the Government process trying to read the data will not leak the data to the Press in the future. But one of the conditions is set in the cond of the aspect whereas the other in the rec. The reason is related to the fact that this aspect is a temporary one, and the aim is to combine it with the previous one. Moreover, the combination should be done in a way that the Government should actually be allowed to read the database, although the pre-existing aspect (aspect (10)) might deny this. Therefore, the operator needed for combining the two aspects is the priority >, and then the whole security policy for the Airline database would be (11) > (10)8 . With that, if the process location is the Government and the heightened security situation is declared, then aspect (11) will be considered. Otherwise, either the action will not be trapped by the aspect (if the process location is not the Government) or the condition cond will be ff (if the threat level is not high), resulting in both cases in a ⊥ ∈ Four for aspect (11), considering then the aspect (10). This example, even though it is very simple, clearly shows three features of our framework: • The use of Aspects for security allows us to temporarily modify a distributed system without having to dig into the bussiness logic of the processes. • The use of the four-valued Belnap Logic allows us to easily combine policies, providing flexibility for the aspect-oriented framework. • The combination of looking to the past and to the future provides even more flexibility, giving the power to express exactly what is intended, for precisely satisfying some properties. 7 In [12] there are many more realistic examples that look to the future, in the Electronic Health Records domain. that using this policy with the priority, the aspect (11) could even remain there, instead of just being a temporary one, since it will be ignored in most of the cases, as long as the tuple < threatlevel, high > is removed after the situation is normalised. 8 Note A. Hernandez and F. Nielson 43 While the first two features where already present in the AspectKB framework (and in particular the first one is widely used in the aspect-orientation community), the third one is a very powerful add-on provided by the new AspectKB+ framework. 5 Conclusion We have studied the problem of enforcing multilevel security in a distributed system as precisely as possible. An information-flow approach poses the problem of having to “guess” what processes in other locations may do, thereby losing some precision. Therefore, we have extended an existing framework to deal with a notion of localised state, which has given us the power to access information about the past performance of the system, thereby allowing us to capture the Bell-LaPadula policy with precision. The resulting framework provides a way to combine policies that look to both the future and the past due to the four-valued Belnap Logic. This gives flexibility to the framework, by capturing precisely what is intended by the security policies. This also gives more power than the previous framework of [7]. Acknowledgements. This work was partially funded by the Danish Strategic Research Council (project 2106-06-0028) “Aspects of Security for Citizens” and partially by the EU Integrated Project SENSORIA (contract 016004). We would like to thank Alan Mycroft for his comments on an early version of this paper. Finally, we really appreciated the comments from all the reviewers, they were very helpful. References [1] G. Kiczales et al. (1997): Aspect-oriented programming. In: ECOOP97, LNCS, 1241, Springer, pp. 220–242. [2] D. E. Bell & L. J. LaPadula (1973): Secure computer systems: mathematical foundations. Technical Report, MITRE Corp. [3] N. D. Belnap (1977): How a computer should think. In: Contemporary Aspects of Philosophy, Oriel Press, pp. 30–56. [4] G. Bruns & M. Huth (2008): Access-Control Policies via Belnap Logic: Effective and Efficient Composition and Analysis. In: CSF08, IEEE Computer Society, pp. 163–176. [5] D. Gelernter & N. Carriero (1992): Coordination Languages and their Significance. Communications of the ACM 35(2), pp. 96–107. [6] D. Gollmann (1999): Computer security. Wiley. [7] C. Hankin, F. Nielson & H. Riis Nielson (2009): Advice from Belnap Policies. In: CSF09, IEEE Computer Society, pp. 234–247. Available at http://doi.ieeecomputersociety.org/10.1109/CSF.2009.8. [8] C. Hankin, F. Nielson, H. Riis Nielson & F. Yang (2008): Advice for Coordination. In: COORDINATION08, LNCS, 5052, Springer, pp. 153–168. Available at http://dx.doi.org/10.1007/978-3-540-68265-3_ 10. [9] J. McCune, S. Berger, R. Caceres, T. Jaeger & R. Sailer (2006): Shamon: A System for Distributed Mandatory Access Control. ACSAC . [10] R. De Nicola, G. Ferrari & R. Pugliese (1998): Klaim: A Kernel Language for Agents Interaction and Mobility. IEEE Trans. on Soft. Engineering 24(5), pp. 315–330. [11] A. Sabelfeld & A. C. Myers (2003): Language-Based Information-Flow Security. IEEE J. Selected Areas in Communications 21(1), pp. 5–19. [12] F. Yang, C. Hankin, F. Nielson & H. Riis Nielson (2010): Aspects-oriented Access Control of Tuple Spaces. Manuscript submitted to a journal .
6cs.PL
A New Numerical Abstract Domain Based on Difference-Bound Matrices arXiv:cs/0703073v2 [cs.PL] 16 Mar 2007 Antoine Miné École Normale Supérieure de Paris, France, [email protected], http://www.di.ens.fr/~ mine Abstract. This paper presents a new numerical abstract domain for static analysis by abstract interpretation. This domain allows us to represent invariants of the form (x − y ≤ c) and (±x ≤ c), where x and y are variables values and c is an integer or real constant. Abstract elements are represented by Difference-Bound Matrices, widely used by model-checkers, but we had to design new operators to meet the needs of abstract interpretation. The result is a complete lattice of infinite height featuring widening, narrowing and common transfer functions. We focus on giving an efficient O(n2 ) representation and graph-based O(n3 ) algorithms—where n is the number of variables—and claim that this domain always performs more precisely than the well-known interval domain. To illustrate the precision/cost tradeoff of this domain, we have implemented simple abstract interpreters for toy imperative and parallel languages which allowed us to prove some non-trivial algorithms correct. 1 Introduction Abstract interpretation has proved to be a useful tool for eliminating bugs in software because it allows the design of automatic and sound analyzers for real-life programming languages. While abstract interpretation is a very general framework, we will be interested here only in discovering numerical invariants, that is to say, arithmetic relations that hold between numerical variables in a program. Such invariants are useful for tracking common errors such as division by zero and out-of-bound array access. In this paper we propose practical algorithms to discover invariants of the form (x − y ≤ c) and (±x ≤ c)—where x and y are numerical program variables and c is a numeric constant. Our method works for integers, reals and even rationals. For the sake of brevity, we will omit proofs of theorems in this paper. The complete proof for all theorems can be found in the author’s MS thesis [12]. Previous and Related Work. Static analysis has developed approaches to automatically find numerical invariants based on numerical abstract domains representing the form of the invariants we want to find. Famous examples are the lattice of intervals (described in, for instance, Cousot and Cousot’s ISOP’76 paper [4]) and the lattice of polyhedra (described in Cousot and Halbwachs’s POPL’78 paper [8]) which represent respectively invariants of the form (v ∈ [c1 , c2 ]) and (α1 v1 + · · · + αn vn ≤ c). Whereas the interval analysis is very efficient—linear memory and time cost—but not very precise, the polyhedron analysis is much more precise but has a huge memory cost—exponential in the number of variables. Invariants of the form (x−y ≤ c) and (±x ≤ c) are widely used by the modelchecking community. A special representation, called Difference-Bound Matrices (DBMs), was introduced, as well as many operators in order to model-check timed automata (see Yovine’s ES’98 paper [14] and Larsen, Larsson, Pettersson and Yi’s RTSS’97 paper [10]). Unfortunately, most operators are tied to modelchecking and are of little interest for static analysis. Our Contribution. This paper presents a new abstract numerical domain based on the DBM representation, together with a full set of new operators and transfer functions adapted to static analysis. Sections 2 and 3 present a few well-known results about potential constraint sets and introduce briefly the Difference-Bound Matrices. Section 4 presents operators and transfer functions that are new—except for the intersection operator— and adapted to abstract interpretation. In Section 5, we use these operators to build lattices, which can be complete under certain conditions. Section 6 shows some practical results we obtained with an example implementation and Section 7 gives some ideas for improvement. 2 Difference-Bound Matrices Let V = {v1 , . . . , vn } be a finite set a variables with value in a numerical set I (which can be the set Z of integers, the set Q of rationals or the set R of reals). We focus, in this paper, on the representation of constraints of the form (vj − vi ≤ c), (vi ≤ c) and (vi ≥ c), where vi , vj ∈ V and c ∈ I. By choosing one variable to be always equal to 0, we can represent the above constraints using only potential constraints, that is to say, constraints of the form (vj − vi ≤ c). From now, we will choose v2 , . . . , vn to be program variables, and v1 to be the constant 0 so that (vi ≤ c) and (vi ≥ c) are rewritten (vi − v1 ≤ c) and (v1 − vi ≤ −c). We assume we now work only with potential constraints over the set {v1 , . . . , vn }. Difference-Bound Matrices. We extend I to I = I∪{+∞} by adding the +∞ element. The standard operations ≤, =, +, min and max are extended to I as usual (we will not use operations, such as − or ∗, that may lead to indeterminate forms). Any set C of potential constraints over V can be represented uniquely by a n× n matrix in I—provided we assume, without loss of generality, that there does not exist two potential constraints (vj − vi ≤ c) in C with the same left member and different right members. The matrix m associated with the potential constraint set C is called a Difference-Bound Matrix (DBM) and is defined as follows:  △ c if (vj − vi ≤ c) ∈ C, mij = +∞ elsewhere . Potential Graphs. A DBM m can be seen as the adjacency matrix of a directed graph G = (V, A, w) with edges weighted in I. V is the set of nodes, A ⊆ V 2 is the set of edges and w ∈ A 7→ I is the weight function. G is defined by:  (vi , vj ) ∈ /A if mij = +∞, (vi , vj ) ∈ A and w(vi , vj ) = mij if mij 6= +∞ . We will denote by hi1 , . . . , ik i a finite set of nodes representing a path from node vi1 to node vik in G. A cycle is a path such that i1 = ik . V-Domain and V 0 -Domain. We call the V-domain of a DBM m and we denote by D(m) the set of points in In that satisfy all potential constraints: △ D(m) = {(x1 , . . . , xn ) ∈ In | ∀i, j, xj − xi ≤ mij } . Now, remember that the variable v1 has a special semantics: it is always equal to 0. Thus, it is not the V-domain which is of interest, but the V 0 -domain (which is a sort of intersection-projection of the V-domain) denoted by D0 (m) and defined by: △ D0 (m) = {(x2 , . . . , xn ) ∈ In−1 | (0, x2 , . . . , xn ) ∈ D(m)} . We will call V-domain and V 0 -domain any subset of In or In−1 which is respectively the V-domain or the V 0 -domain of some DBM. Figure 1 shows an example DBM together with its corresponding potential graph, constraint set, V-domain and V 0 -domain. P Order. The ≤ order on I induces a point-wise order P on the set of DBMs: △ m P n ⇐⇒ ∀i, j, mij ≤ nij . This order is partial. It is also complete if I has least-upper bounds, i.e, if I is R or Z, but not Q. We will denote by = the associated equality relation which is simply the matrix equality. We have m P n =⇒ D0 (m) ⊆ D0 (n) but the converse is not true. In particular, we do not have D0 (m) = D0 (n) =⇒ m = n (see Figure 2 for a counter-example).      v2 −v2 v3 (a)   −v3   v2 − v3 ≤ 4 ≤ −1 ≤ 3 ≤ −1 ≤ 1 v1 v2 v3 v1 +∞ 4 3 (b) v2 −1 +∞ +∞ v3 −1 1 +∞ v1 −1 4 (c) −1 v2 v3 3 1 000000000000 111111111111 1010 000000000000 111111111111 000000000000000 111111111111111 000000000 111111111 000000000000 111111111111 10111111111 000000000 000000000000 111111111111 10111111111 000000000 000000000000 111111111111 10111111111 000000000 (e) 000000000000 111111111111 10111111111 000000000 000000000000 111111111111 10111111111 111111111111111 000000000000000 000000000 000000000000 111111111111 1010 000000000000 111111111111 000000000000 111111111111 1010 000000000000 111111111111 000000000000 111111111111 10 v3 v3 (d) v1 v2 v2 Fig. 1. A constraint set (a), its corresponding DBM (b) and potential graph (c), its V-domain (d) and V 0 -domain (e). (a) v1 v2 v3 v1 +∞ 4 3 v2 −1 +∞ +∞ v3 −1 1 +∞ (b) v1 v2 v3 v1 0 5 3 v2 −1 +∞ +∞ v3 −1 1 +∞ (c) v1 v2 v3 v1 0 4 3 v2 −1 0 +∞ v3 −1 1 0 Fig. 2. Three different DBMs with the same V 0 -domain as in Figure 1. Remark that (a) and (b) are not even comparable with respect to P. 3 Closure, Emptiness, Inclusion and Equality Tests We saw in Figure 2 that two different DBMs can represent the same V 0 -domain. In this section, we show that there exists a normal form for any DBM with a non-empty V 0 -domain and present an algorithm to find it. The existence and computability of a normal form is very important since it is, as often in abstract representations, the key to equality testing used in fixpoint computation. In the case of DBMs, it will also allows us to carry an analysis of the precision of the operators defined in the next section. Emptiness Testing. We have the following graph-oriented theorem: Theorem 1. A DBM has an empty V 0 -domain if and only if there exists, in its associated potential graph, a cycle with a strictly negative total weight.  Checking for cycles with a strictly negative weight is done using the well-known Bellman-Ford algorithm which runs in O(n3 ). This algorithm can be found in Cormen, Leiserson and Rivest’s classical algorithmics textbook [2, §25.3]. Closure and Normal Form. Let m be a DBM with a non-empty V 0 -domain and G its associated potential graph. Since G has no cycle with a strictly negative weight, we can compute its shortest path closure G ∗ , the adjacency matrix of which will be denoted by m∗ and defined by:  ∗ △    mii = 0, N −1 X ∗ △ m = min mik ik+1 if i 6= j .   1≤N  ij hi=i1 ,i2 ,...,iN =ji k=1 The idea of closure relies on the fact that, if hi = i1 , i2 , . . . , iN = ji is a path PN −1 from vi to vj , then the constraint vj − vi ≤ k=1 mik ik+1 can be derived from m by adding the potential constraints vik+1 − vik ≤ mik ik+1 , 1 ≤ k ≤ N − 1. This is an implicit potential constraint which does not appear directly in the DBM m. When computing the closure, we replace each potential constraint vj − vi ≤ mij , i 6= j in m by the tightest implicit constraint we can find, and each diagonal element by 0 (which is indeed the smallest value vi − vi can reach). In Figure 2 for instance, (c) is the closure of both the (a) and (b) DBMs. Theorem 2. 1. m∗ = inf P {n | D0 (n) = D0 (m)}. 2. D0 (m) saturates m∗ , that is to say: ∀i, j, such that m∗ij < +∞, ∃(x1 = 0, x2 , . . . , xn ) ∈ D(m), xj − xi = m∗ij .  Theorem 2.1 states that m∗ is the smallest DBM—with respect to P—that represents a given V 0 -domain, and thus the closed form is a normal form. Theorem 2.2 is a crucial property to prove accuracy of some operators defined in the next section. Any shortest-path graph algorithm can be used to compute the closure of a DBM. We suggest the straightforward Floyd-Warshall, which is described in Cormen, Leiserson and Rivest’s textbook [2, §26.2], and has a O(n3 ) time cost. Equality and Inclusion Testing. The case where m or n or both have an empty V 0 -domain is easy; in all other cases we use the following theorem—which is a consequence of Theorem 2.1: Theorem 3. 1. If m and n have non-empty V 0 -domain, D0 (m) = D0 (n) ⇐⇒ m∗ = n∗ . 2. If m and n have non-empty V 0 -domain, D0 (m) ⊆ D0 (n) ⇐⇒ m∗ P n.  Besides emptiness test and closure, we may need, in order to test equality or inclusion, to compare matrices with respect to the point-wise ordering P. This can be done with a O(n2 ) time cost. Projection. We define the projection π|vk (m) of a DBM m with respect to a variable vk to be the interval containing all possible values of v ∈ I such that there exists a point (x2 , . . . , xn ) in the V 0 -domain of m with xk = v: △ π|vk (m) = {x ∈ I | ∃(x2 , . . . , xn ) ∈ D0 (m) such that x = xk } . The following theorem, which is a consequence of the saturation property of the closure, gives an algorithmic way to compute the projection: Theorem 4. If m has a non-empty V 0 -domain, then π|vk (m) = [−m∗k1 , m∗1k ] (interval bounds are included only if finite). 4  Operators and Transfer Functions In this section, we define some operators and transfer functions to be used in abstract semantics. Except for the intersection operator, they are new. The operators are basically point-wise extensions of the standard operators defined over the domain of intervals [4]. Most algorithms presented here are either constant time, or point-wise, i.e., quadratic time. Intersection. Let us define the point-wise intersection DBM m ∧ n by: △ (m ∧ n)ij = min(mij , nij ) . We have the following theorem: Theorem 5. D0 (m ∧ n) = D0 (m) ∩ D0 (n).  stating that the intersection is always exact. However, the resulting DBM is seldom closed, even if the arguments are closed. Least Upper Bound. The set of V 0 -domains is not stable by union1 so we introduce here a union operator which over-approximate its result. We define the point-wise least upper bound DBM m ∨ n by: △ (m ∨ n)ij = max(mij , nij ) . m ∨ n is indeed the least upper bound with respect to the P order. The following theorem tells us about the effect of this operator on V 0 -domains: Theorem 6. 1. D0 (m ∨ n) ⊇ D0 (m) ∪ D0 (n). 2. If m and n have non-empty V 0 -domains, then (m∗ ) ∨ (n∗ ) = inf {o | D0 (o) ⊇ D0 (m) ∪ D0 (n)} P and, as a consequence, D0 ((m∗ ) ∨ (n∗ )) is the smallest V 0 -domain (with respect to the ⊆ ordering) which contains D0 (m) ∪ D0 (n). 3. If m and n are closed, then so is m ∨ n.  Theorem 6.1 states that D0 (m ∨ n) is an upper bound in the set of V 0 -domains with respect to the ⊆ order. If precision is a concern, we need to find the least upper bound in this set. Theorem 6.2—which is a consequence of the saturation property of the closure—states that we have to close both arguments before applying the ∨ operator to get this most precise union over-approximation. If one argument has an empty V 0 -domain, the least upper bound we want is simply the other argument. Emptiness tests and closure add a O(n3 ) time cost. 1 V 0 -domains are always convex, but the union of two V 0 -domains may not be convex. Widening. When computing the semantics of a program, one often encounters loops leading to fixpoint computation involving infinite iteration sequences. In order to compute in finite time an upper approximation of a fixpoint, widening operators were introduced in P. Cousot’s thesis [3, §4.1.2.0.4]. Widening is a sort of union for which every increasing chain is stationary after a finite number of iterations. We define the point-wise widening operator ▽ by:  △ mij if nij ≤ mij , (m▽n)ij = +∞ elsewhere . The following properties prove that ▽ is indeed a widening: Theorem 7. 1. D0 (m▽n) ⊇ D0 (m) ∪ D0 (n). 2. Finite chain property: ∀m and ∀(ni )i∈N , the chain defined by: ( △ = m, x0 △ xi+1 = xi ▽ni , is increasing for P and ultimately stationary. The limit l is such that l Q m and ∀i, l Q ni .  The widening operator has some intriguing interactions with closure. Like the least upper bound, the widening operator gives more precise results if its right argument is closed, so it is rewarding to change xi+1 = xi ▽ni into xi+1 = xi ▽(ni ∗ ). This is not the case for the first argument: we can have sometimes D0 (m▽n) D0 ((m∗ )▽n). Worse, if we try to force the closure of the first argument by changing xi+1 = xi ▽ni into xi+1 = (xi ▽ni )∗ , the finite chain property (Theorem 7.2) is no longer satisfied, as illustrated in Figure 3. Originally [4], Cousot and Cousot defined widening over intervals ▽ by: △ [a, b] ▽ [c, d] = [e, f ], where: △ e =  a −∞ if a ≤ c, elsewhere, △ f =  b +∞ if b ≥ d, elsewhere . The following theorem proves that the sequence computed by our widening is always more precise than with the standard widening over intervals: Theorem 8. If we have the following iterating sequence: ( ( △ [y0 , z0 ] = m∗ , x0 △ xk+1 = xk ▽(nk ∗ ), △ = π|vi (m), △ [yk+1 , zk+1 ] = [yk , zk ] ▽ π|vi (nk ), v1 v1 i+1 i+1 1 △ △ m = ni = 1 1 v2 v3 i+1 v2 1 x2i = 2i+1 v2 x2i+1 = v3 v1 2i+1 2i 2i 1 v3 1 v1 2i+1 i+1 1 2i+1 v2 1 2i+2 2i+2 1 v3 1 Fig. 3. Example of an infinite strictly increasing chain defined by x0 = m∗ , xi+1 = (xi ▽ni )∗ . then the sequence (xk )k∈N is more precise than the sequence ([yk , zk ])k∈N in the following sense: ∀k, π|vi (xk ) ⊆ [yk , zk ] .  Remark that the technique, described in Cousot and Cousot’s PLILP’92 paper [7], for improving the precision of the standard widening over intervals ▽ can also be applied to our widening ▽. It allows, for instance, deriving a widening that always gives better results than a simple sign analysis (which is not the case of ▽ nor ▽). The resulting widening over DBMs will remain more precise than the resulting widening over intervals. Narrowing. Narrowing operators were introduced in P. Cousot’s thesis [3, §4.1.2.0.11] in order to restore, in a finite time, some information that may have been lost by widening applications. We define here a point-wise narrowing operator △ by:  nij if mij = +∞, △ (m△n)ij = mij elsewhere . The following properties prove that △ is indeed a narrowing: Theorem 9. 1. If D0 (n) ⊆ D0 (m), then D0 (n) ⊆ D0 (m△n) ⊆ D0 (m). 2. Finite decreasing chain property: ∀m and for any chain (ni )i∈N decreasing for P, the chain defined by: ( △ = m, x0 △ xi+1 = xi △ni , is decreasing and ultimately stationary.  0 Given a sequence (nk )k∈N such that the chain (D (nk ))k∈N is decreasing for the ⊆ partial order (but not (nk )k∈N for the P partial order), one way to ensure the best accuracy as well as the finiteness of the chain (xk )k∈N is to force the closure of the right argument by changing xi+1 = xi △ni into xi+1 = xi △(ni ∗ ). Unlike widening, forcing all elements in the chain to be closed with xi+1 = (xi △ni )∗ poses no problem. Forget. Given a DBM m and a variable vk , the forget operator m\vk computes a DBM where all informations about vk are lost. It is the opposite of the projection operator π|vk . We define this operator by:   min(mij , mik + mkj ) if i 6= k and j 6= k, △ 0 if i = j = k, (m\vk )ij =  +∞ elsewhere . The V 0 -domain of m\vk is obtained by projecting D0 (m) on the subspace − → orthogonal to I− v→ k , and then extruding the result in the direction of vk : Theorem 10. D0 (m\vk ) = {(x2 , . . . , xn ) ∈ In−1 | ∃x ∈ I, (x2 , . . . , xk−1 , x, xk+1 , . . . , xn ) ∈ D0 (m)}.  Guard. Given an arithmetic equality or inequality g over {v2 , . . . , vn }—which we call a guard—and a DBM m, the guard transfer function tries to find a new DBM m(g) the V 0 -domain of which is {s ∈ D0 (m) | s satisfies g}. Since this is, in general, impossible, we will only try to have: Theorem 11. D0 (m(g) ) ⊇ {s ∈ D0 (m) | s satisfies g}.  Here is an example definition: Definition 12. 1. If g = (vj0 − vi0 ≤ c) with i0 6= j0 , then:  min(mij , c) △ (m(vj0 −vi0 ≤c) )ij = mij if i = i0 and j = j0 , elsewhere . The cases g = (vj0 ≤ c) and g = (−vi0 ≤ c) are settled by choosing respectively i0 = 1 and j0 = 1. 2. If g = (vj0 − vi0 = c) with i0 6= j0 , then: △ m(vj0 −vi0 =c) = (m(vj0 −vi0 ≤c) )(vi0 −vj0 ≤−c) . The case g = (vj0 = c) is a special case where i0 = 1. 3. In all other cases, we simply choose: △ m(g) = m .  In all but the last—general—cases, the guard transfer function is exact. Assignment. An assignment vk ← e(v2 , . . . , vn ) is defined by a variable vk and an arithmetic expression e over {v2 , . . . , vn }. Given a DBM m representing all possible values that can take the variables set {v2 , . . . , vn } at a program point, we look for a DBM, denoted by m(vk ←e) , representing the possibles values of the same variables set after the assignment vk ← e. This is not possible in the general case, so the assignment transfer function will only try to find an upper approximation of this set: Theorem 13. D0 (m(vk ←e) ) ⊇ {(x2 , . . . , xk−1 , e(x2 , . . . , xn ), xk+1 , . . . , xn ) | (x2 , . . . , xn ) ∈ D0 (m)} .  For instance, we can use the following definition for m(vi0 ←e) : Definition 14. 1. If e = vi0 + c, then: (m(vi0 ←vi0 +c) )ij   mij − c △ = mij + c  mij if i = i0 , j 6= j0 , if i 6= i0 , j = j0 , elsewhere . 2. If e = vj0 + c with i0 6= j0 , then we use the forget operator and the guard transfer function: △ m(vi0 ←vj0 +c) = ((m\vi0 )(vi0 −vj0 ≤c) )(vj0 −vi0 ≤−c) . The case e = c is a special case where we choose j0 = 1. 3. In all other cases, we use a standard interval arithmetic to find an interval [−e− , e+ ], e+ , e− ∈ I such that [−e− , e+ ] ⊇ e(πv2 (m), . . . , πvn (m)) and then we define: (m(vi0 ←e) )ij  + e △ e− =  (m\vi0 )ij if i = 1 and j = i0 , if j = 1 and i = i0 , elsewhere .  In all but the last—general—cases, the assignment transfer function is exact. Comparison with the Abstract Domain of Intervals. Most of the time, the precision of numerical abstract domains can only be compared experimentally on example programs (see Section 6 for such an example). However, we claim that the DBM domain always performs better than the domain of intervals. To legitimate this assertion, we compare informally the effect of all abstract operations in the DBM and in the interval domains. Thanks to Theorems 5 and 6.2, and Definitions 12 and 14, the intersection and union abstract operators and the guard and assignment transfer functions are more precise than their interval counterpart. Thanks to Theorem 8, approximate fixpoint computation with our widening ▽ is always more accurate than with the standard widening over intervals ▽ and one could prove easily that each iteration with our narrowing is more precise than with the standard narrowing over intervals. This means that any abstract semantics based on the operators and transfer functions we defined is always more precise than the corresponding interval-based abstract semantics. 5 Lattice Structures In this section, we design two lattice structures: one on the set of DBMs and one on the set of closed DBMs. The first one is useful to analyze fixpoint transfer between abstract and concrete semantics and the second one allows us to design a meaning function—or even a Galois Connection—linking the set of abstract V 0 -domains to the concrete lattice P({v2 , . . . , vn } 7→ I), following the abstract interpretation framework described in Cousot and Cousot’s POPL’79 paper [5]. DBM Lattice. The set M of DBMs, together with the order relation P and the point-wise least upper bound ∨ and greatest lower bound ∧, is almost a lattice. It only needs a least element ⊥, so we extend P, ∨ and ∧ to M⊥ = M ∪ {⊥} in an obvious way to get ⊑, ⊔ and ⊓. The greatest element ⊤ is the DBM with all its coefficients equal to +∞. Theorem 15. 1. (M⊥ , ⊑, ⊓, ⊔, ⊥, ⊤) is a lattice. 2. This lattice is complete if (I, ≤) is complete (Z or R, but not Q).  There are, however, two problems with this lattice. First, we cannot easily assimilate this lattice to a sub-lattice of P({v2 , . . . , vn } 7→ I) as two different DBMs can have the same V 0 -domain. Then, the least upper bound operator ⊔ is not the most precise upper approximation of the union of two V 0 -domains because we do not force the arguments to be closed. Closed DBM Lattice. To overcome these difficulties, we build another lattice based on closed DBMs. First, consider the set M∗⊥ of closed DBMs M∗ with a least element ⊥∗ added. Now, we define a greatest element ⊤∗ , a partial order relation ⊑∗ , a least upper bound ⊔∗ and a greatest lower bound ⊓∗ in M∗⊥ by:  △ 0 if i = j, ∗ ⊤ ij = +∞ elsewhere .  △ either m = ⊥∗ , ∗ m ⊑ n ⇐⇒ or m 6= ⊥∗ , n 6= ⊥∗ and m P n .  if n = ⊥∗ , m △ ∗ m⊔ n = n if m = ⊥∗ ,  m ∨ n elsewhere .  ∗ △ ⊥ if m = ⊥∗ or n = ⊥∗ or D0 (m ∧ n) = ∅, m ⊓∗ n = ∗ (m ∧ n) elsewhere . Thanks to Theorem 2.1, every non-empty V 0 -domain has a unique representation in M∗ ; ⊥∗ is the representation for the empty set. We build a meaning function γ which is an extension of D0 (·) to M∗⊥ :  △ ∅ if m = ⊥∗ , γ(m) = 0 D (m) elsewhere . Theorem 16. 1. (M∗⊥ , ⊑∗ , ⊓∗ , ⊔∗ , ⊥∗ , ⊤∗ ) is a lattice and γ is one-to-one. 2. If d (I, ≤) is complete, this lattice is complete and γ is meet-preserving: T ∗ γ( X) = {γ(x) | x ∈ X}. We can—according to Cousot and Cousot [6, Prop. 7]—build a canonical Galois Insertion: γ −−→ − −− M∗⊥ P({v2 , . . . , vn } 7→ I) ← −− −→ α where the d abstraction function α is defined by: α(D) = ∗ { m ∈ M∗⊥ | D ⊆ γ(m) }.  The M∗⊥ lattice features a nice meaning function and a precise union approximation; thus, it is tempting to force all our operators and transfer functions to live in M∗⊥ by forcing closure on their result. However, we saw this does not work for widening, so fixpoint computation must be performed in the M⊥ lattice. 6 Results The algorithms on DBMs presented here have been implemented in OCaml and used to perform forward analysis on toy—yet Turing-equivalent—imperative and parallel languages with only numerical variables and no procedure. We present here neither the concrete and abstract semantics, nor the actual forward analysis algorithm used for our analyzers. They follow exactly the abstract interpretation scheme described in Cousot and Cousot’s POPL’79 paper [5] and Bourdoncle’s FMPA’93 paper [1] and are detailed in the author’s MS thesis [12]. Theorems 1, 3, 5, 6, 11 and 13 prove that all the operators and transfer functions we defined are indeed abstractions on the domain of DBMs of the usual operators and transfer functions on the concrete domain P({v2 , . . . , vn } 7→ I), which, as shown by Cousot and Cousot [5], is sufficient to prove soundness for analyses. Imperative Programs. Our toy forward analyzer for imperative language follows almost exactly the analyzer described in Cousot and Halbwachs’s POPL’78 paper [8], except that the abstract domain of polyhedra has been replaced by our DBM-based domain. We tested our analyzer on the well-known Bubble Sort and Heap Sort algorithms and managed to prove automatically that they do not produce out-of-bound error while accessing array elements. Although we did not find as many invariants as Cousot and Halbwachs for these two examples, it was sufficient to prove the correctness. We do not detail these common examples here for the sake of brevity. Parallel Programs. Our toy analyzer for parallel language allows analyzing a fixed set of processes running concurrently and communicating through global variables. We use the well-known nondeterministic interleaving method in order to analyze all possible control flows. In this context, we managed to prove automatically that the Bakery algorithm, introduced in 1974 by Lamport [9], for synchronizing two parallel processes never lets the two processes be at the same time in their critical sections. We now detail this example. The Bakery Algorithm. After the initialization of two global shared variables y1 and y2, two processes p1 and p2 are spawned. They synchronize through the variables y1 and y2, representing the priority of p1 and p2, so that only one process at a time can enter its critical section (Figure 4). Our analyzer for parallel processes is fed with the initialization code (y1 = 0; y2 = 0) and the control flow graphs for p1 and p2 (Figure 5). Each control graph is a set of control point nodes and some edges labeled with either an action performed when the edge is taken (the assignment y1 ← y2 + 1, for example) or a guard imposing a condition for taking the edge (the test y1 6= 0, for example). The analyzer then computes the nondeterministic interleaving of p1 and p2 which is the product control flow graph. Then, it computes iteratively the abstract invariants holding at each product control point. It outputs the invariants shown in Figure 6. The state (2, c) is never reached, which means that p1 and p2 cannot be at the same time in their critical section. This proves the correctness of the Bakery algorithm. Remark that our analyzer also discovered some non-obvious invariants, such as y1 = y2 + 1 holding in the (1, c) state. y1 = 0; y2 = 0; (p1) while true do y1 = y2 + 1; while y2 6= 0 and y1 > y2 do done; - - - critical section - - y1 = 0; done (p2) while true do y2 = y1 + 1; while y1 6= 0 and y2 ≥ y1 do done; - - - critical section - - y2 = 0; done Fig. 4. Pseudo-code for the Bakery algorithm. a 0 y1 ← y2 + 1 1 y1 ← 0 y2 6= 0 and y1 > y2 y2 = 0 or y1 ≤ y2 2 y2 ← y1 + 1 critical section (p1) b y2 ← 0 y1 6= 0 and y2 ≥ y1 y1 = 0 or y2 < y1 c critical section (p2) Fig. 5. Control flow graphs of processes p1 and p2 in the Bakery algorithm. (0, a) y1 = 0 y2 = 0 (0, b) y1 = 0 y2 ≥ 1 (0, c) y1 = 0 y2 ≥ 1 (1, a) y1 ≥ 1 y2 = 0 (1, b) y1 ≥ 1 y2 ≥ 1 (1, c) y1 ≥ 2 y2 ≥ 1 y1 − y2 = 1 (2, a) y1 ≥ 1 y2 = 0 (2, b) y1 ≥ 1 y2 ≥ 1 y1 − y2 ∈ [−1, 0] (2, c) ⊥ Fig. 6. Result of our analyzer on the nondeterministic interleaving product graph of p1 and p2 in the Bakery algorithm. 7 Extensions and Future Work Precision improvement. In our analysis, we only find a coarse set of the invariants held in a program since finding all invariants of the form (x − y ≤ c) and (±x ≤ c) for all programs is non-computable. Possible losses of precision have three causes: non-exact union, widening in loops and non-exact assignment and guard transfer functions. We made crude approximations in the last—general—case of Definitions 12 and 14 and there is room for improving assignment and guard transfer functions, even though exactness is impossible. When the DBM lattices are complete, there exists most precise transfer functions such that Theorems 11 and 13 hold, however these functions may be difficult to compute. Finite Union of V 0 -domains. One can imagine to represent finite unions of V 0 -domains, using a finite set of DBMs instead of a single one as abstract state. This allows an exact union operator but it may lead to memory and time cost explosion as abstract states contain more and more DBMs, so one may need from time to time to replace a set of DBMs by their union approximation. The model-checker community has also developed specific structures to represent finite unions of V-domains, that are less costly than sets. Clock-Difference Diagrams (introduced in 1999 by Larsen, Weise, Yi and Pearson [11]) and Difference Decision Diagrams (introduced in Møller, Lichtenberg, Andersen and Hulgaard’s CSL’99 paper [13]) are tree-based structures made compact thanks to the sharing of isomorphic sub-trees; however existence of normal forms for such structures is only a conjecture at the time of writing and only local or path reduction algorithms exist. One can imagine adapting such structures to abstract interpretation the way we adapted DBM in this paper. Space and Time Cost Improvement. Space is often a big concern in abstract interpretation. The DBM representation we proposed in this paper has a fixed O(n2 ) memory cost—where n is the number of variables in the program. In the actual implementation, we decided to use the graph representation—or hollow matrix—which stores only edges with a finite weight and observed a great space gain as most DBMs we use have many +∞. Most algorithms are also faster on hollow matrices and we chose to use the more complex, but more efficient, Johnson shortest-path closure algorithm—described in Cormen, Leiserson and Rivest’s textbook [2, §26.3]—instead of the Floyd-Warshall algorithm. Larsen, Larsson, Pettersson and Yi’s RTSS’97 paper [10] presents a minimal form algorithm which finds a DBM with the fewest finite edges representing a given V 0 -domain. This minimal form could be useful for memory-efficient storing, but cannot be used for direct computation with algorithms requiring closed DBMs. Representation Improvement. The invariants we manipulate are, in term of precision and complexity, between interval and polyhedron analysis. It is interesting to look for domains allowing the representation of more forms of invariants than DBMs in order to increase the granularity of numerical domains. We are currently working on an improvement of DBMs that allows us to represent, with a small time and space complexity overhead, invariants of the form (±x± y ≤ c). 8 Conclusion We presented in this paper a new numerical abstract domain inspired from the well-known domain of intervals and the Difference-Bound Matrices. This domain allows us to manipulate invariants of the form (x − y ≤ c), (x ≤ c) and (x ≥ c) with a O(n2 ) worst case memory cost per abstract state and O(n3 ) worst case time cost per abstract operation (where n is the number of variables in the program). Our approach made it possible for us to prove the correctness of some nontrivial algorithms beyond the scope of interval analysis, for a much smaller cost than polyhedron analysis. We also proved that this analysis always gives better results than interval analysis, for a slightly greater cost. Acknowledgments. I am grateful to J. Feret, C. Hymans, D. Monniaux, P. Cousot, O. Danvy and the anonymous referees for their helpful comments and suggestions. References [1] F. Bourdoncle. Efficient chaotic iteration strategies with widenings. In FMPA’93, number 735 in LNCS, pages 128–141. Springer-Verlag, 1993. [2] T. Cormen, C. Leiserson, and R. Rivest. Introduction to Algorithms. The MIT Press, 1990. [3] P. Cousot. Méthodes itératives de construction et d’approximation de points fixes d’opérateurs monotones sur un treillis, analyse sémantique de programmes. Thèse d’état ès sciences mathématiques, Université scientifique et médicale de Grenoble, France, 1978. [4] P. Cousot and R. Cousot. Static determination of dynamic properties of programs. In Proc. of the 2d Int. Symposium on Programming, pages 106–130. Dunod, Paris, France, 1976. [5] P. Cousot and R. Cousot. Systematic design of program analysis frameworks. In ACM POPL’79, pages 269–282. ACM Press, 1979. [6] P. Cousot and R. Cousot. Abstract interpretation and application to logic programs. Journal of Logic Programming, 13(2–3):103–179, 1992. [7] P. Cousot and R. Cousot. Comparing the Galois connection and widening/narrowing approaches to abstract interpretation, invited paper. In PLILP’92, number 631 in LNCS, pages 269–295. Springer-Verlag, August 1992. [8] P. Cousot and N. Halbwachs. Automatic discovery of linear restraints among variables of a program. In ACM POPL’78, pages 84–97. ACM Press, 1978. [9] L. Lamport. A new solution of dijkstra’s concurrent programming problem. Communications of the ACM, 8(17):453–455, August 1974. [10] K. Larsen, F. Larsson, P. Pettersson, and W. Yi. Efficient verification of real-time systems: Compact data structure and state-space reduction. In IEEE RTSS’97, pages 14–24. IEEE CS Press, December 1997. [11] K. Larsen, C. Weise, W. Yi, and J. Pearson. Clock difference diagrams. Nordic Journal of Computing, 6(3):271–298, October 1999. [12] A. Miné. Representation of two-variable difference or sum constraint set and application to automatic program analysis. Master’s thesis, ENS-DI, Paris, France, 2000. http://www.eleves.ens.fr:8080/home/mine/stage_dea/. [13] J. Møller, J. Lichtenberg, R. Andersen, H., and H. Hulgaard. Difference decision diagrams. In CSL’99, volume 1683 of LNCS, pages 111–125. Springer-Verlag, September 1999. [14] S. Yovine. Model-checking timed automata. In Embedded Systems, number 1494 in LNCS, pages 114–152. Springer-Verlag, October 1998.
6cs.PL
arXiv:1712.04946v1 [hep-th] 13 Dec 2017 The Hodge Numbers of Divisors of Calabi-Yau Threefold Hypersurfaces Andreas P. Braun,a Cody Long,b Liam McAllister,c Michael Stillman,d and Benjamin Sungb a Mathematical Institute, University of Oxford, Oxford OX2 6GG, UK Department of Physics, Northeastern University, Boston, MA 02115, USA c Department of Physics, Cornell University, Ithaca, NY 14853, USA d Department of Mathematics, Cornell University, Ithaca, NY 14853, USA b [email protected], [email protected], [email protected], [email protected], [email protected] We prove a formula for the Hodge numbers of square-free divisors of Calabi-Yau threefold hypersurfaces in toric varieties. Euclidean branes wrapping divisors affect the vacuum structure of Calabi-Yau compactifications of type IIB string theory, M-theory, and F-theory. Determining the nonperturbative couplings due to Euclidean branes on a divisor D requires counting fermion zero modes, which depend on the Hodge numbers hi (OD ). Suppose that X is a smooth Calabi-Yau threefold hypersurface in a toric variety V , and let D be the restriction to X of a square-free divisor of V . We give a formula for hi (OD ) in terms of combinatorial data. Moreover, we construct a CW complex PD such that hi (OD ) = hi (PD ). We describe an efficient algorithm that makes possible for the first time the computation of sheaf cohomology for such divisors at large h1,1 . As an illustration we compute the Hodge numbers of a class of divisors in a threefold with h1,1 = 491. Our results are a step toward a systematic computation of Euclidean brane superpotentials in Calabi-Yau hypersurfaces. December 15, 2017 Contents 1 Introduction 2 Notation and Preliminaries 2.1 Polytopes and toric varieties 2.2 The Picard group of X . . . 2.3 The ravioli complex R . . . 2.4 The puff complex P . . . . 2 . . . . 4 4 5 7 10 . . . . 12 12 14 16 18 4 Interpretation and an Example 4.1 Contractible graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Example with h1,1 = 491 . . . . . . . . . . . . . . . . . . . . . . . . . . 18 18 19 5 Conclusions 23 A Mayer-Vietoris Complexes A.1 Distributive lattices of ideals and Mayer-Vietoris complexes of ideals A.2 Mayer-Vietoris sequences corresponding to divisors and curves . . . A.3 Background on spectral sequences . . . . . . . . . . . . . . . . . . . A.4 The hypercohomology spectral sequence . . . . . . . . . . . . . . . . . . . 25 26 29 31 32 of Strata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 33 36 38 39 42 45 56 C Computation of h2 for a Divisor on a 2-Face C.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C.2 Relating h2 to toric data . . . . . . . . . . . . . . . . . . . . . . . . . . C.3 Computation of h2 for a face-limited divisor . . . . . . . . . . . . . . . 61 61 63 64 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Spectral Sequence Computation of Hodge 3.1 Hypercohomology spectral sequence . . . . 3.2 Cohomology from stratification . . . . . . 3.3 Proof that ζ = 0 . . . . . . . . . . . . . . 3.4 Generalization to Calabi-Yau n-folds . . . . . . . . . . . . . . . . . . . . . . . . . . . Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B Stratification of Toric Hypersurfaces and Hodge Numbers B.1 Toric varieties and stratification . . . . . . . . . . . . . . . . B.2 Toric varieties and divisors from polytopes . . . . . . . . . . B.3 Resolution of singularities . . . . . . . . . . . . . . . . . . . B.4 Computing Hodge-Deligne numbers of strata . . . . . . . . . B.5 Calabi-Yau hypersurfaces and reflexive polytopes . . . . . . B.6 Topology of subvarieties of Calabi-Yau threefolds . . . . . . B.7 Hodge numbers hi,0 of toric divisors of Calabi-Yau n-folds . . –1– . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 Introduction Compactifications of type IIB string theory on orientifolds of Calabi-Yau threefolds, and of F-theory on Calabi-Yau fourfolds, provide important classes of four-dimensional effective theories with N = 1 supersymmetry. The vacuum structure of these theories depends on the potential for the Kähler moduli, which parameterize the sizes of holomorphic submanifolds in the Calabi-Yau manifold. In particular, in lieu of a potential for the Kähler moduli, positive vacuum energy in four dimensions induces an instability of the overall volume of the compactification, and therefore realistic particle physics and cosmology require a computation of the Kähler moduli potential. In principle a minimum of the Kähler moduli potential could be created by competition among purely perturbative corrections to the leading-order Kähler potential. However, at present the best-understood constructions of metastable vacua require contributions to the superpotential for the Kähler moduli [1–3]. Because of the nonrenormalization theorem, such terms are necessarily nonperturbative, resulting from Euclidean branes wrapping cycles — in fact, divisors — in the compact space.1 An important goal is to determine which divisors in a Calabi-Yau threefold or fourfold support nonperturbative superpotential terms from Euclidean branes. Witten has shown [4] that Euclidean M5-branes on a smooth effective vertical divisor D of a smooth Calabi-Yau fourfold Y give a nonvanishing contribution to the superpotential whenever D is rigid, meaning that the Hodge numbers h0,i (D) = hi (D, OD ) = hi (OD ) obey h0,0 (D) = 1 , h0,1 (D) = 0 , h0,2 (D) = 0 , h0,3 (D) = 0 , (1.1) which we abbreviate as h• (OD ) = (1, 0, 0, 0). Rigidity corresponds to the absence of massless bosonic deformations, and implies that the only zero modes of the Dirac operator on the M5-brane are the two universal Goldstino modes that result from the supersymmetries broken by the M5-brane. In more general circumstances — when either D or Y is singular, when fluxes are included on D or in Y , or when D is a divisor of a Calabi-Yau threefold — the conditions for a nonperturbative superpotential are more subtle, but the Hodge numbers are still essential data. For this reason, a long-term aim is to compute the Hodge numbers of effective divisors D of Calabi-Yau manifolds X. This is most manageable in the case of smooth Calabi-Yau hypersurfaces in toric varieties: for many years there have been computational algorithms and implementations for computing sheaf cohomology of coherent sheaves on toric varieties [5–7], and there are faster implementations for computing sheaf cohomology of line bundles on toric varieties [8, 9]. Unfortunately, all of these implementations fail to finish for even very modest sizes of h1,1 (X), say for 1 Strong gauge dynamics in four dimensions, arising on seven-branes wrapping four-cycles in the compact space, is an alternative. The geometric requirements on such cycles closely parallel those in the Euclidean brane case, and we will refer only to the latter in this work. 2 h1,1 (X) & 10 − 20. A key open problem in computational algebraic geometry is to find algorithms that work for effective divisors or coherent sheaves on all Calabi-Yau threefold hypersurfaces arising from the Kreuzer-Skarke database [10] of four-dimensional reflexive polytopes, as well as for Calabi-Yau fourfold hypersurfaces in toric varieties. We will not arrive at a fully general answer. However, in the important special case in which X is a smooth Calabi-Yau threefold hypersurface in a toric variety V , b on V , we will establish and D is the restriction to X of a square-free effective divisor D i a formula for h (OD ) in terms of the combinatorial data of V . This formula, given in Theorem 3, is our main result. As we will see, Theorem 3 allows one to read off the Hodge numbers of many divisors by inspection, and moreover it is straightforward to turn this formula into an algorithm that computes the Hodge numbers of any squarefree divisor of any X arising from the Kreuzer-Skarke database [11]. The principal tools in our proof are stratification; the hypercohomology spectral sequence of the Mayor-Vietoris complex (3.2); and a correspondence that we establish between square-free divisors of a Calabi-Yau hypersurface and CW complexes constructed on a triangulation of the associated reflexive polytope ∆◦ . Let us briefly summarize these results. Stratification is the decomposition of an n-dimensional toric variety V into tori, and leads to the extremely simple expressions (2.9) for the Hodge numbers h0,i of particular subvarieties of a Calabi-Yau hypersurface X ⊂ V . Among these subvarieties are prime toric divisors Di , i = 1, . . . h1,1 (X) + 4, each of which corresponds to a lattice point of a reflexive polytope ∆◦ , as we review in §2. The Hodge numbers h• (ODi ) and the intersections of the prime toric divisors Di are fully specified by elementary combinatorial data: namely, by the simplicial complex T corresponding to a triangulation of ∆◦ , together with the number of lattice points interior to each face of ∆. A square-free divisor D is the union of a collection of distinct prime toric diviP sors, D = Di . In order to compute hi (OD ), one can appropriately combine the data characterizing the constituent prime toric divisors Di ⊂ D. To achieve this, we establish (in Appendix A) that the Mayer-Vietoris sheaf sequence associated to D is exact, and we then examine the corresponding hypercohomology spectral sequence. Formally, these methods are already sufficient to derive an expression for hi (OD ), but we find it valuable to carry out the computation, and to express the result, in terms of a particular CW complex P that encodes the data of T and ∆. The construction of P amounts to attaching cells to each j-face Θ◦ of ∆◦ , in a manner determined by the number of lattice points in the relative interior of the dual face Θ ⊂ ∆. A square-free divisor D naturally determines a subcomplex PD ⊂ P, and via the hypercohomology spectral sequence we are able to relate the sheaf cohomology of OD to the cellular homology of the CW complex PD . In particular, we show that hi (OD ) = hi (PD ), proving the theorem. A subtlety in applying our results to compute Euclidean brane superpotentials is 3 P that any nontrivial sum D = Di of prime toric divisors Di that is rigid is necessarily reducible, and involves normal crossing singularities where the Di intersect. Normal crossing singularities present no obstacle to defining and computing hi (OD ), but they do complicate the connection between hi (OD ) and the number of fermion zero modes: new zero modes can appear at the intersection loci. Thus, the Hodge numbers hi (OD ) that we compute here mark a significant step toward computing the superpotential, but do not provide a final answer. Systematically counting the fermion zero modes associated to normal crossing singularities, along the lines of [12–14], is an important problem for the future [15]. The organization of this paper is as follows. In §2 we set notation and recall elementary properties of Calabi-Yau hypersurfaces in toric varieties. Then, given any square-free divisor D in a Calabi-Yau threefold hypersurface, we define a corresponding ∆-complex RD and a CW complex PD , which are constructed so that their homology encodes the sheaf cohomology of OD . In §3 we prove our main result, Theorem 3, which asserts that hi (OD ) = hi (PD ). In §4 we illustrate our findings in the example of a Calabi-Yau threefold with h1,1 = 491. We conclude in §5. Appendix A defines Mayer-Vietoris complexes and proves some relevant properties. In Appendix B we review key results from stratification. In Appendix C we directly compute h2 (OD ) by counting lattice points in dual faces, for the special case where TD is restricted to a single 2-face of ∆◦ . 2 Notation and Preliminaries 2.1 Polytopes and toric varieties In this paper we consider Calabi-Yau threefold hypersurfaces X in simplicial toric varieties V , as studied by Batyrev in [16] (see also Appendix B for a review). Such Calabi-Yau threefolds are constructed from pairs (∆, ∆◦ ) of four-dimensional polytopes with vertices in Z4 , obeying h∆, ∆◦ i ≥ −1 . (2.1) A pair of d-dimensional lattice polytopes obeying (2.1) is called a reflexive pair. In each dimension, there are only a finite number of reflexive pairs, up to equivalence, and those in dimension d ≤ 4 have all been enumerated [10, 17–19]. Given a four-dimensional reflexive polytope ∆◦ , we choose a fine star regular2 triangulation Tb of ∆◦ . Since each simplex in Tb contains the origin, this triangulation determines a fan Σ, and V := PΣ is the corresponding simplicial toric variety. If F is a generic linear combination of the monomials in the Cox ring of V that correspond to Fine means that the triangulation uses all the lattice points of ∆◦ . A triangulation is star if the origin, which is the unique interior point of a reflexive polytope ∆◦ , is contained in every fourdimensional simplex of Tb . Regularity is a condition that ensures that V , and hence X, is projective. 2 4 the lattice points of the polytope ∆, then X := {F = 0} ⊂ V is a smooth Calabi-Yau threefold hypersurface in V [16]. The toric variety V in general has pointlike orbifold singularities, but for generic F , X does not intersect these points, and is smooth. We denote the set of faces of ∆◦ by F, the set of faces of dimension at most j by F(≤ j), and the sets of vertices, edges, and 2-faces by F(0), F(1), and F(2), respectively. For each face Θ◦ in F, there is a unique face Θ of ∆ defined by hΘ◦ , Θi = −1 . (2.2) Given any face Θ ⊂ ∆, we denote by `∗ (Θ) the number of lattice points in the relative interior of Θ. Given a face Θ◦ of ∆◦ , we define its genus g(Θ◦ ) by g(Θ◦ ) := `∗ (Θ) , (2.3) i.e. we define the genus of Θ◦ to be the number of interior lattice points of its dual face. If σ ∈ Tb is a simplex, we define the corresponding minimal face, minface(σ), to be the lowest-dimensional face of ∆◦ containing all of the lattice points {pI | I ∈ σ}. We define µ(σ) := dim minface(σ) (2.4) and g(σ) := g(minface(σ)) . 2.2 (2.5) The Picard group of X For each nonzero lattice point pI on ∆◦ there is an associated ray of the fan Σ and a corresponding homogeneous coordinate zI of the toric variety V . We may hence b I given by {zI = 0} with the point pI . This notion can associate the toric divisor D be extended to each d-simplex σ of the triangulation Tb by associating σ with the subvariety Vσ := {zI = 0 ∀ pI ∈ σ}. Let Xσ := Vσ ∩ X be the intersection of Vσ with the Calabi-Yau hypersurface X. This intersection is nonzero if and only if the simplex σ is contained in a 2-face of ∆◦ . It is therefore useful to define T := {σ ∈ Tb | µ(σ) ≤ 2} , (2.6) which omits simplices of Tb that pass through facets (3-faces) of ∆◦ , and so correspond to varieties that do not intersect a generic Calabi-Yau hypersurface X. In a slight abuse of language, we refer to T as a triangulation. The Hodge numbers of subvarieties Xσ ⊂ X are given by rather simple formulae [20–22]. For the divisor DI of X associated with a lattice point σ = pI ∈ T we find 5 (see Appendix B)     1, 0, g(σ)  0 • h (ODI ) = for µ(σ) = 1 . 1, g(σ), 0    1 + g(σ), 0, 0 2 (2.7) In particular, divisors associated with points interior to 2-faces Θ◦ with g(Θ◦ ) > 0 are reducible. A 1-simplex σ of T connecting a pair of lattice points pI and pJ in T corresponds to the intersection CIJ := DI ∩ DJ , with Hodge numbers •  h (OC ) = 1, g(σ) 1 + g(σ), 0   for µ(σ) = 1 . 2 (2.8) That is, a curve C associated with a 1-simplex σ interior to a 1-face Θ◦ is irreducible, of genus g(Θ◦ ). A curve C associated with a 1-simplex σ interior to a 2-face Θ◦ is a union of g(Θ◦ ) + 1 disjoint P1 s. A 2-simplex σ of T containing three lattice points pI , pJ , pK corresponds to the intersection of three divisors, and the corresponding Xσ consists of 1 + g(σ) points. The above results can be summarized as h0,i (Xσ ) = δi,0 + δi,2−µ(σ) g(σ) . (2.9) This result easily generalizes to arbitrary dimension: see Appendix B.7. Moreover, in Appendix B.6 we give similar, albeit slightly more complicated and triangulationdependent, formulae for the Hodge numbers h1,1 (Xσ ). As shown in [16], the Hodge numbers of a Calabi-Yau hypersurface X obey simple combinatorial relations as well. In particular, the rank of the Picard group of X satisfies X X h1,1 (X) = `∗ (Θ◦ ) − 4 + `∗ (Θ◦ )`∗ (Θ) (2.10) Θ◦ ∈F (≤2) Θ◦ ∈F (2) We can identify N := h1,1 (X) + 4 divisors that obey four linear relations, and that generate the Picard group of X. First of all, the K ≤ N divisors of X associated with lattice points in F(≤ 1) are irreducible. However, the divisor associated with a lattice point pr interior to a 2-face Θ◦ has 1 + g(Θ◦ ) irreducible, connected components, which we denote Drα , α ∈ {0, . . . , g(Θ◦ )}. The N irreducible divisors g(p ) K+1 0 0 {D1 , . . . DK , DK+1 , . . . , DK+1 , DK+2 , . . .} (2.11) then generate the Picard group of X. These can be written collectively as {DIα } with 6 α ∈ {0, . . . , δµ(pI ),2 g(pI )}, but we will reindex them as as Di ∈ {D1 , . . . , DN }. We refer to the Di as prime toric divisors, even though the Drα do not all descend from prime toric divisors on V unless X is favorable, i.e. unless g(Θ)g(Θ◦ ) = 0 ∀ Θ◦ ∈ F(2). By (2.7), a prime toric divisor Di associated with a lattice point σ has     1, 0, g(σ)  0 • h (ODi ) = 1, g(σ), 0 for µ(σ) = 1 .    1, 0, 0 2 (2.12) For any subset G ⊆ {1, . . . , N }, there is an associated divisor D (possibly reducible) defined by X D= Di . (2.13) i∈G We call such a D a square-free divisor. The main purpose of this work is to analyze square-free divisors. 2.3 The ravioli complex R Every d-simplex of T contained in the relative interior of a face Θ◦ ∈ F(j) for j ≤ 2 gives rise to a closed subvariety Xσ ⊂ X, of complex dimension 2 − d. When j ≤ 1, the simplices in T correspond to irreducible subvarieties of X. However, for j = 2, i.e. when σ ∈ T is in the interior of a 2-face Θ◦ of ∆◦ , the simplex σ corresponds to a subvariety of X that has 1 + g(Θ◦ ) connected (and irreducible) components, which we denote by Xσα , for α = 0, . . . , g(σ). The intersection structure of the Xσ is determined by ∆ and by the triangulation T of ∆◦ . For each 2-face Θ◦ of ∆◦ , we can choose an ordering of α = 0, . . . , g(σ) such that for any σ, λ ∈ T with minface(σ) = minface(λ) = Θ◦ , we have Xσα ∩ Xλβ = ( α δ αβ Xσ∪λ ∅ if σ ∪ λ ∈ T , if σ ∪ λ ∈ /T . (2.14) Next, for each τ ∈ T with µ(τ ) ≤ 1, the intersection structure with the Xσα can be written as ( Xτα∪σ if τ ∪ σ ∈ T , Xτ ∩ Xσα = (2.15) ∅ if τ ∪ σ ∈ /T . 7 Finally, for τ, ω ∈ T , with µ(τ ) ≤ 1 and µ(ω) ≤ 1, we have: Xτ ∩ X ω = ( Xτ ∪ω if τ ∪ ω ∈ T , ∅ if τ ∪ ω ∈ /T . (2.16) In the special case of (2.16) in which τ ∪ ω ∈ T with minface(τ ∪ ω) = Θ◦ ∈ F(2), we can write (2.16) as g(Θ◦ ) X τ ∩ Xω = [ Xτα∪ω . (2.17) α=0 We will now define a complex, called the ravioli complex R, that accounts for the intersection structure (2.14)-(2.17). Recall that the simplices in T correspond to subvarieties in X, which are possibly disconnected and reducible. The cells in R will correspond to the connected, irreducible components of the subvarieties in X encoded by T . The ravioli complex is defined by n o n o R = Xσα σ ∈ T , µ(σ) = 2, 0 ≤ α ≤ g(σ) ∪ Xτ τ ∈ T , µ(τ ) ≤ 1 , (2.18) which as a set consists of the irreducible connected components of intersections of the Di . The elements of R that have dimension 2 − i in X are called i-cells. For those X such that3 g(Θ◦ ) = 0 ∀ Θ◦ ∈ F(2), R is the simplicial complex T . In the general case, the i-cells of R are the same as the i-simplices of T , except that each simplex σ in the interior of any two-dimensional face Θ◦ ∈ F(2) of ∆◦ is replaced by 1 + g(Θ◦ ) disjoint copies of itself, after which the (1-cell) boundaries of the 1 + g(Θ◦ ) disjoint copies T |Θ◦ of are identified with each other. In general R is not a simplicial complex, but instead a ∆-complex:4 the 1-cells and 2-cells are not necessarily uniquely specified by the 0-cells that contain them (in X). The homology and cohomology complexes of R are readily obtained, and the fact that R is not always a simplicial complex does not present difficulties for computation or visualization. The origin of the name should be clear from Figures 1 and 2. We can now associate a ∆-complex RD ⊆ R to any square-free divisor D as follows. The points of RD are the points (0-cells) pi in R corresponding to the divisors appearing in (2.13). Some pairs (triples) of the pi may in general be connected by 1cells (2-cells) of the complex R. The pi , together with the set of all 1-cells and 2-cells connecting them, therefore define a unique subcomplex RD ⊆ R. An example is shown in the bottom image in Figure 1. The points in RD correspond to codimension-one Notice that even in the case where X is favorable, i.e. obeying g(Θ)g(Θ◦ ) = 0 ∀ Θ◦ ∈ F(2), R is not necessarily equal to T . 4 The term ∆-complex is standard in topology, and the symbol ∆ appearing in the name should not be confused with the polytope ∆. See [23] for background on ∆-complexes and CW complexes. 3 8 11 00 00 11 1 0 0 1 11 00 00 11 1 0 0 1 0 1 1 0 0 1 0 1 11 00 00 11 00 11 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 00 11 1 0 0 1 11 00 00 11 1 0 0 1 0 1 1 0 0 1 0 1 11 00 00 11 00 11 11 00 00 11 00 11 11 00 11 00 11 00 11 00 11 00 00 11 00 11 11 00 11 00 11 00 11 00 11 00 00 11 1 0 0 1 11 00 00 11 1 0 0 1 0 1 11 00 00 11 00 11 11 00 00 11 00 11 11 00 11 00 11 00 11 00 1 0 0 1 0 1 11 00 00 11 00 11 11 00 11 00 11 00 11 00 Figure 1: The simplicial complex T defined by a triangulation, the corresponding ravioli complex R, and the subcomplex RD ⊂ R associated to a divisor D. The upper figure shows two adjacent two-dimensional faces, separated by a thick line, and a triangulation. The middle figure shows the associated ravioli complex in case the genus of the face on the left is zero and the genus of the face on the right is one, so that R has two sheets over the right face. The lower figure shows RD for D the union of the four irreducible divisors Di associated with the points pi colored in red. Figure 2: The ravioli complex over a 2-face. subvarieties in X, while the higher-dimensional cells in RD encode intersections among the prime toric divisors. One-cells in RD correspond to the intersections of pairs of divisors in X, which are irreducible curves in X, and 2-cells in RD correspond to triple intersections of divisors in X. Note that a 2-cell in RD always corresponds to a single point on X. 9 2.4 The puff complex P Let us briefly recapitulate. By (2.9), the sheaf cohomology of prime toric divisors and their intersections is fully specified once one knows the simplicial complex T determined by the triangulation T , together with the genera g(Θ◦ ) of the faces Θ◦ ∈ F(≤ 2). For Θ◦ a 2-face, the number g(Θ◦ ) records the extent to which subvarieties corresponding to simplices contained in Θ◦ are reducible. By promoting the simplicial complex T to the ∆-complex R, we have encoded the information about reducibility directly in the complex R. Heuristically, viewed as sets of data about subvarieties, n  o ◦ ◦ R ↔ T , g(Θ )|Θ ∈ F(2) . (2.19) The next step is to account for the data of {g(Θ◦ )|Θ◦ ∈ F(≤ 1)}. In close analogy to the promotion T → R, we now define a CW complex5 P that encodes this data. Heuristically, n  o ◦ ◦ (2.20) P ↔ T , g(Θ )|Θ ∈ F(≤ 2) , which is made precise by the following: Definition 1 The puff complex P is a CW complex constructed from the ravioli complex R as follows. For each vertex v and edge e in R, we have natural inclusions v ,− → S 2 and e ,− → S 2 , where the latter inclusion induces a cellular structure on S 2 with g(v) W 2 an S 1 attached to each interior 0-cell. These induce the inclusions iv : v ,− → S and ie : e ,− → X where X is defined by the following pushout diagram: g(e) qi=1 e i=1 e ϕ X qi S 2 Then P is defined by the following pushout diagram: qi vi qj ej qiv qie qi W S 2 qj X 5 R ϕ P The property of CW complexes that is relevant here is that a 2-cell or 1-cell σ can be attached to the complex by a map that identifies the boundary ∂σ with a 0-cell in the complex. 10 ◦[0] Θ1 ◦[1] ◦[0] Θ1 Θ3 00 11 00 11 00 11 00 11 000000000 111111111 00 ◦[2] 11 00 11 000000000 111111111 000000000 111111111 Θ 000000000 111111111 000000000 111111111 000000000 ◦[1] 111111111 000000000 Θ2 111111111 000000000 111111111 000000000 111111111 ◦[1] 000000000 111111111 000000000 111111111 Θ3 000000000 111111111 000000000 111111111 000000000 111111111 00 11 000000000 111111111 00 11 00 11 11 00 00 11 00 11 ◦[0] Θ2 00 11 11 00 00 11 11 00 00 11 00 11 00 11 11 00 00 11 00 11 11 00 00 11 00 11 00 11 111111111 000000000 00 11 00 11 000000000 111111111 00 11 00 11 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 00 11 11 00 00 11 00 11 00 11 111111111 000000000 00 11 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 00 11 11 00 00 11 Figure 3: The layers of the puff complex for a single face. The lower row shows the three layers of P for the triangulated face at the top of the figure. In this example, ◦[1] the genera of the various faces are such that g(Θ◦[2] ) = 1, g(Θi ) = (2, 1, 0) and ◦[0] g(Θi ) = (1, 3, 0). The CW complex PD associated to a divisor D is the sub-complex of P corresponding to the image of RD under the morphism ϕ. In other words, to construct PD , we attach a bouquet of g(v) two-spheres to each vertex v ∈ G, a bouquet of g(p) circles to each point p ∈ G interior to an edge, and a bouquet of cylinders S 1 × I to each connected component of RD restricted to e. These spaces are glued together along their common points: in particular, the g(e) cylinders over a complete edge e ⊂ RD are pinched down into the two vertices bounding e, forming a collection of g(e) two-dimensional voids. We will find it useful to divide the puff complex into layers: Definition 2 For 0 ≤ j ≤ 2, the jth layer P (j) of the puff complex P is the subset of P resulting from replacing points interior to each (2 − j)-face Θ◦ of ∆◦ with g(Θ◦ ) j-spheres, replacing 1-simplices interior to Θ◦ with g(Θ◦ ) cylinders S j × I, etc., as in Definition 1. In particular, P (0) = R. Figure 3 contains a sketch of the different layers of the puff complex for a simple example. The homology of PD is readily obtained. Distinct connected components of PD can be examined separately, so we may take h0 (PD ) = 1 without loss of generality. Contributions to h1 (PD ) come from 11 (a) One-cycles in RD . (b) For each edge e ⊂ ∆◦ , a bouquet of g(p) cylinders over each connected component of RD e that is strictly interior to e. Contributions to h2 (PD ) come from (c) Two-cycles in RD . (d) The bouquet of g(e) pinched cylinders over each edge e such that RD e = e. (e) The bouquet of g(v) two-spheres over each vertex v, i.e. the layer P (2) . We will prove in §3.1 that these classes of contributions are in one-to-one correspondence with classes of contributions in the hypercohomology spectral sequence that computes the cohomology of OD . In other words, the homology of the CW complex PD encodes the cohomology of OD . As we will see, this correspondence has both computational and heuristic utility. 3 Spectral Sequence Computation of Hodge Numbers In this section we will prove our principal result: P Theorem 3 Let D = i∈G Di , with G ⊆ {1, . . . , N }, be a square-free divisor, as defined above. Denote by P and PD the puff complex and the associated subcomplex determined by D, respectively. Then the Hodge numbers hi (OD ) are given by hi (OD ) = hi (PD ) . (3.1) In the rest of this section, we prove Theorem 3. Our method is based on examining a hypercohomology spectral sequence. We use stratification to identify the E 2 page of that sequence, and then we use a somewhat indirect argument to show that the differentials on the E 2 page are all zero. 3.1 Hypercohomology spectral sequence To start, let {D1 , . . . , DN } be the prime toric divisors, as defined in §2. Given a square-free toric divisor in X, X D := Di , i∈G we will compute the cohomology h• (OD ). Suppose that we have indexed the Di so P that D = ri=1 Di , for some r ≤ N . 12 The set of divisors {D1 , . . . , Dr } is dimensionally transverse on a smooth variety, and so by Proposition 20 of Appendix A, the generalized Mayer-Vietoris sequence 0 −→ OD −→ r M ODi −→ i=1 M ODij −→ i<j M ODijk −→ 0 (3.2) i<j<k is an exact sequence of sheaves. The hypercohomology spectral sequence of this complex will allow us to compute the cohomology of OD . Define F0 := r M ODi , i=1 F1 := M ODij , i<j F2 := M ODijk . (3.3) i<j<k r (D) The complex of sheaves (3.2) gives rise to a hypercohomology spectral sequence Ep,q with differentials denoted by r r r δp,q→p 0 ,q 0 : Ep,q (D) −→ Ep0 ,q 0 (D) where p0 = p + r and q 0 = q − r + 1. The first page of this spectral sequence is H 2 (F0 ) H 1 (F0 ) H 0 (F0 ) α β 0 0 H 1 (F1 ) 0 H 0 (F1 ) γ H 0 (F2 ) 1 1 These are indexed such that Ep,q (D) = H q (Fp ), and e.g. δ0,1→1,1 = α. 2 The second page, E (D) reads 0 H 2 (F0 ) 0 0 ker α coker α 0 0 0 0 coker γ 0 ζ 0 ker β ker γ/ Im β 2 where ζ := δ0,1→2,0 is the only nonzero differential on this page. The third and final page reads 13 0 H 2 (F0 ) 0 0 0 0 ker ζ coker α 0 0 0 ker β ker γ/ Im β coker ζ 0 On this page, the differential δ 3 is zero. The hypercohomology spectral sequence converges to the cohomology of OD , meaning that dim H 0 (OD ) = dim ker β (3.4) dim H 1 (OD ) = dim ker ζ + dim(ker γ/ Im β) (3.5) dim H 2 (OD ) = dim H 2 (F0 ) + dim coker α + dim coker ζ (3.6) In the special case that ζ is the zero map, we have the simplifications dim H 1 (OD ) = dim ker α + dim(ker γ/ Im β) (3.7) dim H 2 (OD ) = dim H 2 (F0 ) + dim coker α + dim coker γ (3.8) If the dependence on a divisor is understood, we denote the maps α, β, γ, ζ, while if a divisor D is specified, we write α(D), etc. 3.2 Cohomology from stratification We now use stratification to compute the cohomology organized by the spectral sequence given above. Proposition 4 Row 0 of page 1 of the hypercohomology spectral sequence has coho(0) mology equal to that of the ravioli complex RD = PD , that is: dim ker β = dim H0 (RD ) (3.9) dim(ker γ/ Im β) = dim H1 (RD ) (3.10) dim coker γ = dim H2 (RD ) (3.11) Note: this can be written more succinctly in the form (0) 2 dim Ep,0 (D) = dim Hp (PD ) , (3.12) 2 for all p, even though dim Ep,0 (D) can only be nonzero for p = 0, 1, or 2. Proof. We will identify row 0 of page 1 of the spectral sequence with the cohomology complex of the complex RD . In the following, we will take RD (k) to be the set of kfaces of RD . In particular, RD (0) is the set of 0-dimensional faces (i.e. the Di ), RD (1) 14 is the set of 1-dimensional faces, i.e. one for each connected component of a Di ∩ Dj . The same connected component C cannot occur for two different such intersections, as otherwise the component C would be contained in the intersection of three or four of the Di , which cannot happen since the Di are dimensionally transverse. Finally, RD (2) corresponds to the points in each of the triple intersections of three divisors, Di ∩ Dj ∩ Dk . Again, each point can only appear in one such intersection. Notice that RD is a Delta complex whose cohomology sequence is 0 −→ r M δ i=1 δ M 0 CeDi −→ 1 CeC −→ M CeP −→ 0 (3.13) P ∈RD (2) C∈RD (1) where the maps δa are as follows: if C is a connected component of Di ∩ Dj , for i < j, (δ0 (eDi ))C = −1 and (δ0 (eDj ))C = 1, and all other components are zero. If C is an irreducible component of Di ∩ Dj , and if P is a point of the intersection Di ∩ Dj ∩ Dk , (for i < j), then (δ1 (eC ))P = sign(i, j, k). Now consider the complex of row 0 in page 1 of the spectral sequence. We see L L Lr 0 0 that H 0 (F0 ) = P ∈RD (2) CeP , C∈RD (1) CeC , and H (F2 ) = i=1 CeDi , H (F1 ) = i.e. the complex 0 H 0 (F0 ) β H 0 (F1 ) γ H 0 (F2 ) 0 and the complex (3.13) have the same terms. Under this identification, it is easy to see that the maps are identical too, as the morphisms β and γ are canonically induced by the generalized Mayer-Vietoris complex.  Proposition 4 generalizes to all rows of the second page: Proposition 5 For all p and q, (q) 2 dim Ep,q (D) = dim Hp+q (PD ). Proof. We have already established the case q = 0. For q = 2, we note that stratification gives h2 (ODi ) = 0, unless Di corresponds to a vertex of ∆◦ having positive genus g(v). Summing these gives X dim H 2 (F0 ) = g(v) , (3.14) v∈F (0) v∈D proving Proposition 5 for q = 2. The case q = 1 can be established by noting that row 1 of page 1 of the spectral sequence, because of stratification, breaks up into a direct sum of complexes, summed 15 over all edges e in ∆◦ intersecting D:  M M Cg(e) ⊗ 0 −→  C −→ τ ∈TD∩e (0),τ ∈∂e / e∈F (1) M C −→ 0 (3.15) τ ∈TD∩e (1) The complex in parentheses is simply the relative cochain complex for the pair (TD∩e , TD∩∂e ). Note that because e is an edge, (TD∩e , TD∩∂e ) = (RD∩e , RD∩∂e ). Therefore the p-th cohomology (for p = 0, 1) of the direct sum of complexes L (3.15) is the relative cohomology e Cg(e) ⊗ H p (RD∩e , RD∩∂e ), which is the same as L g(e) (1) ⊗ Hp (RD∩e , RD∩∂e ), which by the definition of P equals Hp+1 (PD ).  eC 2 (D) then Corollary 6 Given a square-free divisor D, if ζ(D) = δ0,1→2,0 ζ(D) = 0 ⇐⇒ dim H i (OD ) = dim Hi (PD ) for all i 3.3 (3.16) ⇐⇒ dim H 1 (OD ) = dim H1 (PD ) (3.17) ⇐⇒ dim H 2 (OD ) = dim H2 (PD ) . (3.18) Proof that ζ = 0 In this section we will prove the following, which also, by Corollary 6, proves Theorem 3. Lemma 7 If D corresponds to a square-free divisor on X, then the map 2 ζ(D) = δ0,1→2,0 (D) is the zero map. We prove this result by first splitting D = A + B1 + . . . + Bm , where A and the Bi are all square-free divisors having disjoint support. Consider the subgraph TD (≤ 1). Suppose that this graph has m components each of which is contained in the strict interior of an edge of ∆◦ . Let Bi be the sum of the divisors corresponding to the lattice points in the ith component. Let A be the sum of the rest of the divisors. P Define Aj := A + ji=1 Bi , with A0 = A and Am = D. Lemma 8 The map ζ(A) is zero, and so dim H i (OA ) = dim Hi (PA ). (1) 2 Proof. The map is zero, because its domain, E0,1 (A), which has dimension dim H1 (PA ), is zero by construction.  Lemma 9 dim H 2 (OBi ) = 0. Proof. If the given component Bi consists of divisors associated to lattice points interior to an edge of genus g, then h• (OBi ) = (1, g, 0).  16 Lemma 10 Fix an integer 1 ≤ i ≤ m, and let C := Ai−1 ∩ Bi . Then C is a curve whose irreducible components are all P1 ’s, two intersect in at most one point, and H 1 (OC ) = 0. Proof. C := Ai−1 ∩ Bi is a collection of rational curves C = ∪M CM . As a complex, C corresponds to a collection of 1-simplices σM , each with one end in RBi , and the other end in an edge not containing RBi , in a 2-face, or in a vertex. Two curves CM and CN intersect if and only if σM and σN share a 2-simplex σM,N . Now notice that h1 (OC ) 6= 0 is possible only if the CM intersect in such a way as to form a nontrivial loop. But this would require a closed loop in which three or more 1-simplices σ1 , . . . , σK are connected by 2-simplices σ1,2 , . . . , σK−1,K , σK,1 . However, because RBi is in the strict interior of an edge, no such loop can be formed. Thus, h1 (OC ) = 0.  Lemma 11 dim H 2 (OA ) = dim H 2 (OD ). Proof. We prove this by induction. Consider Ai = Ai−1 + Bi , where A0 = A. We use induction to show that dim H 2 (OA ) = dim H 2 (OAi ), for i = 0, . . . , m. The statement is trivial for i = 0, and for i > 0, we use the long exact sequence associated to the Mayer-Vietoris sequence 0 −→ OAi −→ OAi−1 ⊕ OBi −→ OC −→ 0. combined with Lemmas 9 and 10. One obtains that H 2 (OAi ) = H 2 (OAi−1 ).  Lemma 12 dim H 2 (PA ) = dim H 2 (PD ). Proof. There are four cases where an element of H 2 can appear: D contains a vertex of genus > 0, D contains an entire edge of genus > 0, D contains at least two full sheets over a 2-face, and finally, D contains a collection of full 2-faces that contain a void. Since by definition, not all of the lattice points of the edge containing Bi are in D, and Bi contains no vertices, PA and PD must have the same H 2 .  Proof of Lemma 7, and therefore also the proof of Theorem 3. Lemmas 8, 11, and 12 imply that dim H 2 (PD ) = dim H 2 (PA ) = dim H 2 (OA ) = dim H 2 (OD ), and so by Corollary 6, we have ζ(D) = 0, proving the theorem.  17 3.4 Generalization to Calabi-Yau n-folds Most aspects of the computation above generalize immediately to Calabi-Yau n-folds. Definition 13 Let V be a simplicial toric variety of dimension n + 1, and let X be a smooth Calabi-Yau n-fold hypersurface in V , with Nn := h1,1 (X) + n + 1. Let P D = i∈G Di , with G ⊆ {1, . . . , Nn }, be a square-free divisor. The construction of Definition 1 generalizes to any n ≥ 3, and we denote by P and PD the puff complex and the associated subcomplex determined by D, respectively. We note that the constructions of P and PD are immediate because our results on the stratification of subvarieties Xσ ⊂ X apply for any n. Next, the Mayer-Vietoris sequence generalizing (3.2) contains n+1 nonzero terms. The resulting hypercohomology spectral sequence again converges to the cohomology of OD , and it is easy to see that row zero of page one of the hypercohomology spectral sequence again has cohomology equal to that of RD . Moreover, Proposition 5 holds for arbitrary n [15]. However, for n > 3 there are more diagonal maps, generalizing ζ. If for every n the diagonal maps were shown to be identically zero, as we have proved in the case n = 3, then we would have hi (OD ) = hi (PD ) for all n. However, examining the diagonal maps directly in the same manner as done for ζ in §3.3 would be somewhat involved. We defer a proper analysis of these maps to [15], and state here only the following: Conjecture 14 Let V , X, D, and PD be as in Definition 13, and set n = 4. For each such D there exists k(D) ∈ Z≥0 such that h• (OD ) = h• (PD ) − (0, k(D), k(D), 0) . (3.19) We show in [15] that k(D) = 0 for all D except those obeying a rather restrictive condition. 4 Interpretation and an Example We will now briefly discuss the interpretation of our result, and then illustrate the utility of our formula in an example. 4.1 Contractible graphs Equation 3.1 depends on only certain topological properties of the complex PD , and so two complexes that correspond to divisors that are related to one another by deformations that preserve these topological properties will have identical Hodge numbers. This leads to a useful tool in enumerating certain divisors, as we now explain. 18 Given a triangulation T and the associated simplicial complex T , a square-free divisor D determines a unique simplicial complex TD ⊂ T , as well as a corresponding CW complex PD ⊂ P. The number N of distinct square-free divisors of a given 1,1 puff complex P that corresponds to a Calabi-Yau hypersurface X is 2h (X)+4 . Thus, the task of working out the Hodge numbers of all possible square-free divisors appears formidable for large h1,1 (X). Fortunately, as we will now explain, Theorem 3 implies that square-free divisors fall into equivalence classes. Suppose that we begin with a set of lattice points G1 ⊂ {1, . . . , N }, which define a square-free divisor D1 and a CW complex PG1 . Now let us add or remove one lattice point from G1 , so that G1 is changed to some G2 ⊂ {1, . . . , N }. This operation uniquely defines a new square-free divisor D2 and a new CW complex PG2 . By Equation (3.1), if hi (PG1 ) = hi (PG2 ) then hi (OD1 ) = hi (OD2 ). We define a single contraction to be the operation of changing D1 → D2 by adding or removing a lattice point, as specified above, in such a way that hi (PD1 ) = hi (PD2 ). We define a contraction to be an arbitrary composition of single contractions. Contraction is an equivalence relation on square-free divisors, and all members of a contraction equivalence class have the same Hodge numbers. However, two divisors with the same Hodge numbers do not necessarily belong to the same contraction equivalence class. As an example consider the complex in Figure 4, with associated divisor D = D1 + D2 + D3 , where the Di correspond to the points pi . Equation 3.1 gives h• (OD ) = (1 + g(f ), 0, 0), where f is the 2-face containing G. We can perform a contraction by, for example, first deleting p2 , and then deleting p3 , as in Figure 4. We will present a more involved example in §4.2. D2 → D1 D2 D1 D3 D3 → D1 D1 D3 D1 Figure 4: Example of contraction for a graph D = D1 + D2 + D3 , where the Di correspond to points interior to a 2-face. First D2 is removed, and then D3 is removed. 4.2 Example with h1,1 = 491 As a demonstration of the utility of Theorem 3 we will now calculate the Hodge numbers of some nontrivial divisors in a Calabi-Yau hypersurface X in a toric variety corresponding to a fine, star, regular triangulation of the largest polytope ∆◦L in the Kreuzer-Skarke database [10]. In this example, ∆◦L is the convex hull of the columns 19 dim 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 face # int pts genus 0 1 0 1 1 0 2 1 1 3 1 3 4 1 6 0,1 83 0 0,2 6 0 0,3 2 0 0,4 1 0 1,2 6 0 1,3 2 0 1,4 1 0 2,3 0 1 2,4 0 2 3,4 0 6 0,1,2 246 0 0,1,3 82 0 0,1,4 41 0 0,2,3 6 0 0,2,4 3 0 0,3,4 1 0 1,2,3 6 0 1,2,4 3 0 1,3,4 1 0 2,3,4 0 1 Table 1: The genera of all vertices, edges, and faces in ∆◦L . of  −1  −1 B= −1 −1 −1 −1 −1 83 −1 −1 6 −1 −1 2 −1 −1  1  −1  −1 −1 (4.1) There are 679 nonzero lattice points in ∆◦ , but 184 of these are interior to 3-faces, and therefore do not intersect X. X has h1,1 (X) = 491, with 495 toric divisors. The face structure is very simple: ∆◦L is a 4-simplex, with vertices indexed by {0,1,2,3,4}, corresponding to the columns of B. There are 5 vertices, 10 edges, and 10 triangles in ∆◦L , before triangulation. The genera of all the faces are given in Table 1. Consider the largest 2-face f` in ∆◦L , which is a triangle with vertices labeled by {0, 1, 2}. This face has 344 lattice points, as shown in Figure 5. All (sub)faces 20 1 0 2 Figure 5: The largest 2-face f` in the largest four-dimensional reflexive polytope ∆◦L , with a ‘banded’ complex show in red. The Hodge numbers of the corresponding divisor D are easily read off as h• (OD ) = (1, 9, 1). except vertex 2 have genus 0, and so the puff complex P restricted to f` is simply T restricted to f` , with a single S 2 cell attached to vertex 2. If we choose, excluding point 2, a set of points on f` whose corresponding complex is connected and has no cycles, then the corresponding divisor will be rigid. The divisor corresponding to the entire face has Hodge numbers (1, 0, 1). The divisor corresponding to the line connecting 2 to 4 has Hodge numbers (1, 0, 9). As a more nontrivial example we can form simplicial complexes with cycles on f` and ∆◦L . The complex defined by taking all the boundary points of f` , and none of the interior points, defines a divisor with Hodge numbers (1, 1, 1). In Figure 5 we show a triangulation6 of f` , and a choice of a more complicated subcomplex containing cycles. Theorem 3 allows us to easily read off the Hodge numbers of the corresponding divisor D as h• (OD ) = (1, 9, 1). 6 This triangulation of f` is for illustrative purposes; the triangulation we have obtained from a regular triangulation of ∆◦L is difficult to render. 21 1 1 1 2 0 3 (a) 1 2 0 3 2 0 (b) 3 (c) 2 0 3 (d) Figure 6: A facet of ∆◦L defined by the points {0, 1, 2, 3}, with several complexes drawn on it. The red points indicate lattice points included in the complex, and the red lines cycle-less paths that may include additional lattice points. The red triangles indicate entire faces that are included in the complex. Note that the point 1, corresponding to (−1, −1, −1, 83), has been scaled in, for visualization purposes. In complex (d) we have included the entire boundary of the facet. From Theorem 3 the Hodge numbers of the corresponding divisors are h• (ODa ) = (1, 0, 0), h• (ODb ) = (1, 1, 3), h• (ODc ) = (1, 0, 5), and h• (ODd ) = (1, 0, 6). As a further example of contraction, one can consider complexes that are not just restricted to a single 2-face, as depicted in Figure 6. Here we show a facet of ∆◦L defined by the points {0, 1, 2, 3}, with several complexes drawn on it. The red points indicate lattice points included in the complex, and the red lines cycle-free paths that may include additional lattice points. The red triangles indicate entire faces that are included in the complex. Note that the point 1, corresponding to (−1, −1, −1, 83), has been scaled in, for visualization purposes. In complex (d) we have included the entire boundary of the facet. From Theorem 3 we can immediately read off the Hodge numbers of the corresponding divisors as h• (ODa ) = (1, 0, 0), h• (ODb ) = (1, 1, 3), h• (ODc ) = (1, 0, 5), and h• (ODd ) = (1, 0, 6). 22 5 Conclusions In this work we have computed the Hodge numbers hi (OD ) of square-free divisors D of Calabi-Yau threefold hypersurfaces in toric varieties. Given the data of a simplicial complex T corresponding to a triangulation of a four-dimensional reflexive polytope ∆◦ , we constructed a CW complex P that simultaneously encodes data about ∆◦ and its dual ∆. The construction of P involves attaching certain cells to T in a manner determined by ∆. The specification of a square-free divisor D determines a subcomplex PD , as well as an exact Mayer-Vietoris sheaf sequence. By examining the corresponding hypercohomology spectral sequence, we proved that hi (OD ) = hi (PD ). Here the left-hand side is manifestly the dimension of a sheaf cohomology group, but the right-hand side is the dimension of a simplicial (or cellular) homology group. Our results are a step forward in the study of divisors in Calabi-Yau hypersurfaces. Theorem 3 permits extremely efficient computation of the Hodge numbers of squarefree divisors in threefolds, even for h1,1  1. Conjecture 14 extends these methods to fourfolds, enlarging the range of divisors with computable Hodge numbers [15]. The ultimate goal of this work was to determine which effective divisors D of a Calabi-Yau hypersurface support Euclidean brane superpotential terms. Theorem 3 represents significant progress toward this goal, but further advances will be necessary to give a completely general answer. First of all, although the Hodge numbers hi (OD ) provide essential information about the number of fermion zero modes of a Euclidean brane on D, when D is not smooth there can be additional zero modes associated to singular loci. The only smooth, rigid, square-free divisors are the prime toric divisors Di themselves; a nontrivial square-free divisor that is rigid necessarily involves normal crossing singularities where the components Di intersect. One should therefore ask how to count fermion zero modes on a divisor with normal crossings. In some cases, normal crossings can be shown to yield no new fermion zero modes, so that rigidity is a sufficient condition for a superpotential [15]. Moreover, there exist nontrivial smooth square-free divisors D with Hodge numbers h• (OD ) = (1, 0, n), with n > 0. In a suitably magnetized Euclidean D3-brane wrapping such a divisor, worldvolume flux can lift the zero modes counted by h2 (OD ), and so lead to a superpotential term, even though D is not rigid. Finally, Donagi and Wijnholt have proposed that in general the fermion zero modes can be counted by the logarithmic cohomology of D, h•log (OD ) [12]. Verifying and applying this idea is a natural task for the future. There are many possible extensions of this work. The effects of worldvolume flux could plausibly be incorporated along the lines of [24], but including the effects of bulk fluxes appears more challenging. A further step would be to compute the Hodge numbers of effective divisors that are not square-free. Advances in these directions would allow for a truly systematic computation of the nonperturbative superpotential for the Kähler moduli in compactifications on Calabi-Yau hypersurfaces. 23 Acknowledgments We are grateful to Yuri Berest, Ralph Blumenhagen, Thomas Grimm, Jim Halverson, Tristan Hübsch, John Stout, Wati Taylor, and Timo Weigand for discussions, and we thank John Stout for creating Figure 2. We thank the organizers of String Phenomenology 2017 at Virginia Tech for providing a stimulating environment for discussing this work. L.M. thanks the theory groups at MIT, Northeastern, and Stanford for their hospitality while portions of this work were carried out. C.L. and B.S. thank the theory group at MIT for their hospitality while portions of this work were carried out. The work of A.P.B. was supported by the SCFT grant ST/L000474/1 and the ERC grant 682608 (HIGGSBNDL). The work of C.L. was supported by NSF grant PHY-1620526. The work of L.M. was supported in part by NSF grant PHY-1719877. The work of M.S. was supported in part by NSF grant DMS-1502294. The work of B.S. was supported by NSF RTG grant DMS-1645877. 24 A Mayer-Vietoris Complexes In this section we establish the exactness of the generalized Mayer-Vietoris sequence, and we give related background on spectral sequences. Given closed subvarieties (or subschemes) D and E of a variety (or scheme) X, the Mayer-Vietoris complex is 0 −→ OD∪E −→ OD ⊕ OE −→ OD∩E −→ 0 (A.1) This is always an exact sequence, and one can view it as an inclusion-exclusion sequence. If I and J are ideals in a ring R, then we have the related exact sequences: 0 −→ R/I ∩ J −→ R/I ⊕ R/J −→ R/(I + J) −→ 0 (A.2) 0 −→ I ∩ J −→ I ⊕ J −→ I + J −→ 0. (A.3) and There are well-known generalizations of these sequences to the cases where there are more than two subvarieties, ideals, or quotients, but at this level of generality the resulting Mayer-Vietoris sequence is often not exact. We will now prove exactness for several cases of interest in this paper. We start with the case of ideals. Suppose that I1 , . . . , In ⊂ R are ideals in a ring R. Define the Mayer-Vietoris ideal sequence, M = M(I1 , . . . , In ), to be the cochain complex 0 −→ (I1 ∩ . . . ∩ In ) −→ M 0 −→ M 1 −→ . . . −→ M n−1 −→ 0, (A.4) where M p := M (Ii0 + . . . + Iip ) 1≤i0 <...<ip ≤n and the differential dp : M p −→ M p+1 is defined by p (d (φ))i0 ...ip+1 = p+1 X (−1)j φi0 ...îj ...ip+1 j=0 One checks immediately that ker(d0 ) = I1 ∩ . . . ∩ In , and that the sequence (A.4) is in fact a complex. Notice that if Ii = R for all i, then M(R, . . . , R) is the (reduced) cochain complex of an (n − 1)-simplex, and so is exact. The Mayer-Vietoris quotient complex, M(R/I1 , . . . , R/In ) is the cochain complex defined in a completely analogous manner: 0 −→ R/(I1 ∩ . . . ∩ In ) −→ N 0 −→ N 1 −→ . . . −→ N n−1 −→ 0, 25 (A.5) where M N p := R/(Ii0 + . . . + Iip ) 1≤i0 <...<ip ≤n and the differential (dN )p : N p −→ N p+1 is defined by the natural quotient maps N p ((d ) (φ))i0 ...ip+1 p+1 X (−1)j φi0 ...îj ...ip+1 = j=0 Again, one checks that (A.5) is a complex, and that ker(dN )0 = R/(I1 ∩. . .∩In ). There is an exact sequence of complexes 0 −→ M(I1 , . . . , In ) −→ M(R, . . . , R) −→ M(R/I1 , . . . , R/In ) −→ 0, and coupled with the exactness of the middle complex, the complex M(I1 , . . . , In ) is exact precisely when M(R/I1 , . . . , R/In ) is exact. Finally, given closed subvarieties (or subschemes) D := D1 ∪ . . . ∪ Dn of a variety (or scheme) X, we define the Mayer-Vietoris sheaf sequence M(D1 , . . . , Dn ) in parallel to the above, taking M M(D1 , . . . , Dn )p := ODi0 ∩Di1 ∩...∩Dip . 1≤i0 <i1 <...<ip ≤n Define maps dp : M(D1 , . . . Dn )p −→ M(D1 , . . . , Dn )p+1 in the same manner as for quotients of ideals above. The resulting complex M(D1 , . . . , Dn ) has the form 0 −→ OD −→ n M i=1 A.1 ODi −→ M ODi ∩Dj −→ . . . −→ OD1 ∩D2 ∩...∩Dn −→ 0 . (A.6) i<j Distributive lattices of ideals and Mayer-Vietoris complexes of ideals The relationship between distributive lattices of ideals and the exactness of MayerVietoris ideal sequences was explained in [25]. In this section, we summarize their results, and then show that several sets of ideals of interest form distributive lattices, so that their Mayer-Vietoris sequences are exact. This is the key technical fact that will allow us to show in §A.2 that certain Mayer-Vietoris complexes of sheaves are exact. Fix a (commutative) ring R and ideals J1 , . . . , Jr of R. This set of ideals generates a lattice of ideals: the join of two ideals is their sum, and the meet of two ideals is their intersection. The smallest set of ideals that contains J1 , . . . , Jr and is closed under these two operations is the lattice of ideals generated by J1 , . . . , Jr . This lattice of ideals is called distributive if for every three ideals L1 , L2 , L3 in the 26 lattice, one has L1 ∩ (L2 + L3 ) = (L1 ∩ L2 ) + (L1 ∩ L3 ). Equivalently, the lattice is distributive if and only if for every three ideals L1 , L2 , L3 in the lattice, one has L1 + (L2 ∩ L3 ) = (L1 + L2 ) ∩ (L1 + L3 ). The importance of this notion is the following useful characterization due to Maszczyk. Theorem 15 (Maszczyk [25]) For a set of ideals J1 , . . . , Jn of a ring R, the following are equivalent: (a) The lattice of ideals generated by J1 , . . . , Jn is distributive. (b) M(J1 , . . . , Jn ) is exact, and hence so is M(R/J1 , . . . , R/Jn ). The following is a typical distributive lattice of ideals: Example 16 Let S = C[x1 , . . . , xn ] be a polynomial ring over a field F. Then the lattice of ideals generated by {(x1 ), . . . , (xn )} is the set of square-free monomial ideals of S. This is a distributive lattice of ideals. A square-free monomial ideal M ⊂ F[x1 , . . . , xn ] has a unique description as an irredundant intersection of monomial prime ideals \ hxi | i ∈ αi, M= α∈Λ where the intersection is over a uniquely defined set of subsets of [1..n]. The ideals in the intersection are the minimal primes of M . A key example is a generalization of this to the case where F = {f1 , . . . , fn } ⊂ R is a regular sequence. We will only consider the case where R is a local ring, or a graded ring, and in these cases any permutation of the fi remains a regular sequence. Given F = {f1 , . . . , fn } ⊂ R, define a map φ : Z[x1 , . . . , xn ] −→ R, where φ(xi ) = fi . In this section, we call an ideal I ⊂ R an F -square-free ideal if there is a square-free monomial ideal M such that I is generated by φ(M ), i.e., I is generated by square-free products of the fi . We will say that I is an F -square-free monomial ideal if it is F -square-free, and I can be written as the intersection \ I= hfi | i ∈ αi, α∈Λ where M= \ hxi | i ∈ αi. α∈Λ 27 Theorem 17 Let F = {f1 , . . . , fn } be a regular sequence in the maximal ideal of a local ring R. (a) The lattice of ideals generated by (f1 ), . . . , (fn ) is exactly the set of F -squarefree ideals. (b) This lattice of ideals is distributive. This follows from the following more precise lemma, which we prove by induction on n. Lemma 18 Let F = {f1 , . . . , fn } be a regular sequence. Let F 0 := {f1 , . . . , fn−1 }. The following statements hold. (a) If J is an F 0 -square-free ideal, then (J : fn ) = J. (b) If Ii , Ji are F 0 -square-free ideals, then L1 = I1 fn + I2 and L2 = J1 fn + J2 obey L1 ∩ L2 = (I1 ∩ J1 + I1 ∩ J2 + I2 ∩ J1 )fn + I2 ∩ J2 . (c) If I and J are F 0 -square-free ideals, then ((fn ) + I) ∩ ((fn ) + J) = ((fn ) + (I ∩ J)). (d) If I and J are F 0 -square-free ideals, then (Ifn + J) : fn = I + J. (e) Every F -square-free ideal is an F -square-free monomial ideal. (f ) If L1 and L2 are F -square-free ideals, then L1 ∩ L2 is also F -square-free. (g) If L1 , L2 , L3 are F -square-free ideals, then L1 ∩ (L2 + L3 ) = L1 ∩ L2 + L1 ∩ L3 . Proof. We prove this by induction. The case n = 2 is immediate. Suppose the statements are true for a number of elements of F less than n. (a) By part (e) in the induction hypothesis, \ J= hfi | i ∈ αi, α∈Λ where all the fi that appear have i < n. Then J : fn is the intersection of hfi | i ∈ αi : fn = hfi | i ∈ αi, whose intersection is again J. 28 (b) The right hand side is clearly contained in the left hand side, so let γ = α1 fn + α2 = β1 fn + β2 ∈ L1 ∩ L2 . One shows that γ is contained in the right hand side (where α1 ∈ I1 , α2 ∈ I2 , β1 ∈ J1 , β2 ∈ J2 ). Thus, (α1 − β1 )fn = β2 − α2 ∈ I2 + J2 . Therefore, by part (a), since I2 +J2 is F 0 -square-free, α1 −β1 ∈ I2 +J2 . Therefore, there exists δ1 ∈ I2 , δ2 ∈ J2 such that α1 − β1 = δ1 − δ2 . Now plug this back into the formula for γ to get δ1 fn + α2 = δ2 fn + β2 ∈ I2 ∩ J2 . Since this is already in the right hand side, we can subtract it from γ, obtaining an element γ 0 = (α1 −δ1 )fn = (β1 −δ2 )fn that is in the right hand side precisely when γ is. Since fn is a nonzero divisor, α1 − δ1 = β1 − δ2 , so γ 0 ∈ ((I1 + I2 ) ∩ (J1 + J2 ))fn . Combining this with the induction hypothesis of part(g), we have that (I1 + I2 ) ∩ (J1 + J2 ) = I1 ∩ J1 + I1 ∩ J2 + I2 ∩ J1 + I2 ∩ J2 , and therefore γ 0 , and hence γ, is contained in the right hand side. (c) This follows immediately from (b), by taking I1 = J1 = R. (d) This follows immediately from (b): (Ifn + J) ∩ (fn ) = (I + J)fn , and so (Ifn + J) : fn = I + J. (e) For any ideal I, and f ∈ R, such that I : f 2 = I : f , we always have I = (I : f ) ∩ ((f ) + I). Let L = Ifn + J be an F -square-free ideal. If I = 0, then the inductive hypothesis gives that L = J is F 0 -square-free monomial, and is therefore F -square-free. If instead, fn ∈ L, then the induction hypothesis gives a decomposition for J, and part (c) gives the required decomposition for L. Otherwise, by part (d) we have L = (I + J) ∩ ((fn ) + J). Both of these have decompositions, by (e) (for n − 1) and (c), and putting them together gives a decomposition for L. (f) This follows immediately from (b) and the inductive hypothesis for (f). (g) This follows immediately from part (b).  Proof of Theorem 17. The set of F -square-free ideals is plainly closed under addition of ideals. By Lemma 18(b), the lattice of ideals generated by (f1 ), . . . , (fn ) contains the set of F -square-free ideals. Also by Lemma 18(f), the set of F -square-free ideals is closed under intersection, and so the lattice of ideals generated by (f1 ), . . . , (fn ) is the set of F -square-free ideals. Finally, by Lemma 18(g), the lattice is distributive.  A.2 Mayer-Vietoris sequences corresponding to divisors and curves Recall that D ⊂ X is called a prime divisor on X if it is irreducible and codimension one in X. Proposition 19 If X is a simplicial toric variety, and D1 , . . . , Dr are prime torusinvariant Weil divisors on X, then the Mayer-Vietoris sheaf sequence M(D1 , . . . , Dr ) is an exact sequence of OX -modules. Proof. Let S = C[x1 , . . . , xN ] be the Cox ring of X, where Di corresponds to the ring variable xi , for 1 ≤ i ≤ r. Consider the Mayer-Vietoris ideal sequence M(x1 , . . . , xr ). This is exact, as the xi generate a distributive lattice, and so M(S/x1 , . . . , S/xr ) is also 29 exact. Sheafification of this exact sequence of graded S-modules remains exact, but this sheafification is precisely M(D1 , . . . , Dr ).  Proposition 20 If X is a smooth variety, and D1 , . . . , Dr are effective divisors, such that the intersection of each set of n of these is either empty, or has codimension n in X, then the Mayer-Vietoris sheaf sequence M(D1 , . . . , Dr ) is an exact sequence of OX -modules. Proof. We prove exactness locally. Let p ∈ X be a point. Suppose that the set of Di that pass through p is {D1 , . . . , Dn }. Localizing the Mayer-Vietoris sheaf complex at p results in the complex M(R/f1 , . . . , R/fn ), where R = OX,p , and the Cartier divisor Di is defined locally in this ring by fi . Note that, by hypothesis, the elements f1 , . . . , fn form a regular sequence in the maximal ideal of R. Therefore, Theorem 17 shows that M(f1 , . . . , fn ) is exact. This implies that M(R/f1 , . . . , R/fn ) is also exact, which implies that the original complex is exact at each p ∈ X, proving the proposition.  Notice that this proof can be generalized to the case when X is equidimensional, not necessarily smooth, and the Di are Cartier divisors with the given intersection properties. There also exists a corresponding exact sequence in the case of curves simply using the corresponding normalization without necessitating the above technology. Lemma 21 Suppose X is a smooth, projective variety and assume that C1 , . . . , Cr are smooth, irreducible curves on X, and that the corresponding closed subscheme C := C1 ∪ . . . ∪ Cr is a nodal, reducible curve with components Ci . Then the MayerVietoris sequence M(C1 , . . . , Cr ) is an exact sequence of OX -modules. Proof. By hypothesis, the normalization f : C̃ → C is the scheme corresponding to the disjoint union of the smooth, irreducible components Ci . Thus, the normalization induces the following short exact sequence M M O Ci  f ∗ OCi /OC , (A.7) OC ,− → f∗ i i where the quotient sheaf is a torsion sheaf with support precisely on the nodes of C. By considering the sequence locally at the nodes, we thus obtain the following exact sequence of OC -modules M M OC ,− → f∗ OCi  g∗ Opj , (A.8) i j where g is the natural closed immersion and pj correspond to the nodes of C. Higher direct images of closed immersions vanish and hence we obtain an exact sequence of OX -modules yielding exactness of M(C1 , ..., Cr ).  30 A.3 Background on spectral sequences We will now highlight the essential aspects of hypercohomology spectral sequences needed to follow the computation of h• (OD ). We begin by describing the basic usage of the two spectral sequences corresponding to a double complex. Instead of going into detail about filtrations, we assume that the entries of the second and subsequent pages of the spectral sequence are finite-dimensional vector spaces. We are describing “first quadrant” spectral sequences, as that is what we need for hypercohomology. We start with a bounded double complex of vector spaces, C = {C p,q , dh , dv }, i.e. a collection of vector spaces C p,q , where C p,q = 0 outside of the box 0 ≤ p ≤ M , p,q 0 ≤ q ≤ M , for some M , horizontal differentials dp,q −→ C p+1,q , and vertical h : C differentials dp,q : C p,q −→ C p,q+1 satisfying d2h = d2v = 0, and the compatibility v condition that they anticommute: dh dv = −dv dh . More generally, we could allow C p,q to be modules, or objects in some Abelian category, and the maps would then be morphisms in this category. There are two spectral sequences corresponding to C, 0 Erp,q and 00 Erp,q . We will describe the first, from which the second is easily obtained. A spectral sequence is a set of pages Erp,q , for r = 0, 1, 2, . . .. Each page is a two-dimensional array of vector spaces Erp,q , for 0 ≤ p, q ≤ M , together with a map dr . The zeroth page, E0 , starts with C p,q p,q p,q −→ C p,q+1 . in the (p, q) spot, and the map is the vertical differential dp,q 0 := dv : C The differential dr at each step satisfies the equation d2r = 0, and maps p,q p+r,q−r+1 dp,q . r : Er −→ Er The first page is obtained from the zeroth page by setting p,q−1 E1p,q := ker dp,q ). 0 / Im(d0 The differential d1 is simply the horizontal map p,q p+1,q dp,q 1 : E1 −→ E1 induced by the horizontal differential on C. In general, given the r-th page of the spectral sequence, the (r + 1)-st page is the array p,q p−r,q+r−1 Er+1 := ker dp,q ). r / Im(dr By iterating this procedure, the spectral sequence eventually converges when all terms on the page stabilize. For a first quadrant spectral sequence, the spectral sequence converges to the cohomology of the total complex E2p,q =⇒ H p+q (Tot(C)) 31 where the total complex is a complex with the n-th term defined by M Tot(C)n = C p,q p+q=n with the natural differential defined by d = dv + dh . A.4 The hypercohomology spectral sequence We now explain the utility of the spectral sequence discussed above in computing sheaf cohomology. Let us fix a smooth complex projective variety X and closed subvarieties D1 , . . . , Dn . We assume that the intersection of each set of k of the Di is either empty, or has codimension k in X, so that by Proposition 20 the corresponding Mayer-Vietoris sheaf sequence M(D1 , . . . , Dn ) is an exact sequence. Let {OD } denote a complex localized in degree 0 with the term OD and let Ln { i=1 ODi −→ . . . −→ OD1 ∩...∩Dn } denote the corresponding complex beginning in degree 0. These complexes are quasi-isomorphic and hence, as objects in the bounded derived category Db (X), they are isomorphic. We have the natural global sections functor Γ : Qcoh(X) → Vec(C), and by deriving on the right and restricting to the full subcategory of bounded complexes of quasicoherent sheaves with coherent cohomology, we obtain the induced derived functor RΓ : Db (X) → Db (Vec(C)). For a given complex of coherent sheaves F • , we then compute the hypercohomology groups, or higher derived functors, by taking cohomology, namely H i (X, F • ) = H i RΓ(F • ). In order to compute the hypercohomology groups of a complex F • , we take a quasi-isomorphic complex of injective objects given by f : F • → I • and compute the cohomology H i (Γ(I • )). Such a complex is often constructed as the total complex of the Cartan-Eilenberg resolution J •,• of F • where existence is proved by applying the horseshoe lemma. In particular, the hypercohomology H(F • ) is independent of the resolution. However, in practice, injective resolutions are usually hard to find explicitly, and so one often resorts to taking acyclic resolutions. We wish to compute i i  H (D, OD ) = H X, n M ODi −→ . . . −→ OD1 ∩...∩Dn  (A.9) i=1 In the general situation, given a scheme (X, OX ), and given a bounded-below complex F • of OX -modules, we have the following computational tool: Lemma 22 There exists a spectral sequence with   E2p,q = H p H q (X, F • ) converging to Hp+q (X, F • ). 32 B Stratification of Toric Hypersurfaces and Hodge Numbers of Strata In this appendix we will review a number of results about the stratification of toric varieties and the associated stratification of hypersurfaces. In particular, we will review how the algorithm of [26] can be used to conveniently read off the Hodge numbers of toric strata. Although all of these techniques are in principle well-known, a succinct exposition, especially in the physics literature, is lacking. We refer the reader to [27– 29] for an introduction to toric geometry. The original reference for how to compute Hodge numbers of toric strata for hypersurfaces is [26]; see [16, 20, 21] for a number of applications close to physics. B.1 Toric varieties and stratification By definition, a toric variety V is characterized as containing an open dense algebraic torus, (C∗ )n , the action of which (on itself) extends to the whole variety. One can think of an n-dimensional toric variety as a (partial) compactification of (C∗ )n in which various lower-dimensional algebraic tori are added. These can be thought of as ‘limits’ of the original (C∗ )n , so that the action of (C∗ )n on itself extends naturally to the whole variety. This structure can be summarized by a combinatorial object called a fan. A fan Σ is a collection of strongly convex rational polyhedral cones such that the face of every cone counts as a cone in the fan and two cones only intersect along a face of each. Here, strong convexity demands that no subspace (except the origin) of the ambient vector space is contained in any cone. One can think of each cone as being spanned by a finite number of rays with primitive generators vi that are elements of a lattice N . Finally, we (mostly) make the further assumption that every cone is simplicial, i.e. any d-dimensional cone is spanned by d rays. In this case the associated toric variety has at most orbifold singularities and is factorial. Let us label the (generators of the) rays of Σ by vi , with the index i running from 1 to r. The toric variety PΣ associated with the fan Σ can then be described in analogy to complex projective space by associating a homogeneous coordinate zi to each ray of the fan and forming the quotient Cr − SR PΣ = . G (B.1) This data is encoded in the fan as follows. First, the exceptional set or Stanley Reisner ideal, SR corresponds to all collections {vi |i ∈ I} for which the cones vi do not share a common higher-dimensional cone. Wheneve this is the case, the set SR contains the subspace of Cr in which the associated collection of coordinates {zi |i ∈ I} vanishes 33 simultaneously. The Abelian group G is given as the kernel of the homomorphism Y he ,v i zi → tk = zj k j (B.2) j where the ek form a basis of the lattice M . Note that this implies that whenever there is a linear relation X ai v i = 0 , (B.3) i there is an associated one-parameter subgroup C∗ ∈ G which acts as (z1 , · · · , zr ) ∼ (λa1 z1 , · · · , λar zr ) (B.4) for λ ∈ C∗ . In most cases of interest, G can be completely described through relations like this. In general, however, G may contain discrete factors. Note that the above discussion implies that the (complex) dimension of V equals the (real) dimension of N . In the following, we also use the notation PΣ to denote the toric variety determined by the fan Σ. For all zi 6= 0, the coordinates tk defined in (B.2) parameterize the open dense ∗ n (C ) giving the toric variety its name. The strata that are added to turn (C∗ )n into a nontrivial toric variety are encoded in the fan as follows. We first associate the unique zero-dimensional cone σ [0] , i.e. the origin, of Σ with (C∗ )n . Each d-dimensional cone [d] σi is naturally associated with the homogeneous coordinates zi corresponding to its generating rays. Choosing the basis ek in (B.2) from lattice vectors contained in the dual cone hσ̌, σi ≥ 0 , (B.5) which sits in σ̌ ∈ M ⊗ R, where M is the dual lattice to N , we may take a limit of (B.2) where we set the zi = 0 for all vi generating σ. As a d-dimensional cone σ [d] has d generators, this lands us on an n − d-dimensional algebraic torus Tσ[d] ' (C∗ )n−d and the definition of a fan ensures that all of these strata are consistently sewn together. Denoting the set of d-dimensional cones by [d] Σ(d) = {σk } , (B.6) we can hence stratify any toric variety according to the data of our fan as a a a a PΣ = Tσ[d] = ( ∗ )n−d . d C k d [d] σk ∈Σ(d) (B.7) [d] σk ∈Σ(d) In terms of the homogenous coordinates zi , the stratum of each cone σ is described by 34 Figure 7: The fan of P2 contains a zero-dimensional cone, three rays (one-dimensional cones) and three two-dimensional cones. setting {zi = 0 ∀vi ∈ σ} and {zi 6= 0 ∀vi ∈ / σ} . (B.8) Let us discuss some examples. The fan corresponding to P1 lives in R and is composed of three cones: the origin, a ray generated by 1 and a ray generated by −1. We recover the standard presentation from (B.1) as P1 = (z1 , z2 ) − {(0, 0)} . (z1 , z2 ) ∼ (λz1 , λz2 ) (B.9) The open dense C∗ can be described by the coordinate z1 /z2 , or equivalently by z2 /z1 , and the strata corresponding to the two one-dimensional cones are simply given by z1 = 0 and z2 = 0, respectively. We hence recover the description of P as the Riemann sphere, i.e. adding the point at infinity to C. The fan corresponding to P2 is shown in figure 7. Again, (B.1) gives the standard presentation as (z1 , z2 , z3 ) − {(0, 0, 0)} P2 = . (B.10) (z1 , z2 , z3 ) ∼ (λz1 , λz2 , λz3 ) From the fan, we can directly read off the stratification data (B.7), which becomes a a P2 = (C∗ )2 C∗ pt. (B.11) i=1..3 35 k=1..3 In the bulk of this paper we are not interested in the geometry of toric varieties per se, but rather in the geometry of algebraic subvarieties. In the simplest setting, such an algebraic subvariety Z is given as the vanishing locus of a section of a line bundle L on V . For such an algebraic subvariety Z, one easily obtains a stratification of Z from that of V if all toric strata meet Z transversely. In this case we can define an n − d − 1-dimensional stratum Zσ[d] = Z ∩ Tσ[d] (B.12) for every d-dimensional cone σ [d] ∈ Σ. The stratification of the hypersurface Z is then simply a a Z= Zσ[d] . (B.13) d σ [d] ∈Σ(d) k k Let us illustrate this in the simple example of hypersurface of degree k in P2 , i.e. we consider the vanishing locus of a homogeneous polynomial Pk (z1 , z2 , z3 ) of degree k in the homogeneous coordinates of P2 . Call the resulting curve Ck . By adjunction one finds that such a hypersurface is a Riemann surface of genus g(k) = (k − 1)(k − 2) . 2 (B.14) Let us now see how Ck is stratified using (B.13). First consider the lowest dimensional strata of P2 , which correspond to the highest-dimensional cones. Each of the three two-dimensional cones of the fan of P2 corresponds to a point on P2 that is defined by setting two of the three homogeneous coordinates to zero. For a sufficiently generic polynomial Pk , such points will never lie on Z, so that these strata do not contribute to Z (indeed they are supposed to be 2 − 2 − 1 = −1-dimensional). For each of the three one-dimensional cones in the fan of P2 , there is a stratum C∗ obtained by setting one of the three homogeneous coordinates to zero while forbidding the remaining two from vanishing. Intersecting this with Pk = 0 we find that Zσ[1] consists of k points for every one-dimensional cone σ [1] . Finally, the complex one-dimensional stratum Zσ[0] is a Riemann surface of genus g(k) with 3 · k points excised. Below, we will show how to reproduce these numbers combinatorially by using the Newton polytope of Pk . B.2 Toric varieties and divisors from polytopes In order to describe how to determine the geometry of strata Zσ[d] combinatorially, we need to present the situation of interest from a slightly different perspective. Note first that the data defining the topology of Z consists of the toric variety V (or, equivalently, the fan Σ) and a line bundle L on V . We can write the divisor class D of L in terms 36 of the toric divisors Di corresponding to rays of Σ as X D = c1 (L) = ai Di . (B.15) i The group of holomorphic sections of L is then given by a polytope ∆, known as the Newton polytope, defined by PD = {m ∈ M | hm, vi i ≥ −ai ∀ vi } . More explicitly, we may use the monomials Y hm,v i+a p(m) = zi i i (B.16) (B.17) i as a basis for the group of sections. This provides a convenient way to find the zeroth cohomology group (which is in fact the only non-vanishing one) of a divisor D (line bundle L), as this counts global holomorphic sections of L: h0 (PΣ , OPΣ (D)) = |PD ∩ M | . (B.18) Interestingly, ∆ determines both the line bundle L and (a blowdown of) the toric variety V . To any polytope ∆ in the M -lattice, we can associate its normal fan Σn (∆), which then gives rise to a toric variety PΣn (∆) along with a divisor D. This works as follows. To every face Θ[k] of the polytope ∆, we associate a cone [ σ̌n (Θ[k] ) = r · (p∆ − pΘ[k] ) (B.19) r≥0 [k] where p∆ is any point lying inside ∆ and pΘ is any point lying inside Θ[k] . The dual cones σn (Θ[k] ) (defined as in (B.5)) form a complete fan that is called the normal fan Σn (∆) of ∆. Here, k-dimensional faces Θ[k] of ∆ are associated with d = n − kdimensional cones σn (Θ[k] ). In particular, the cones of Σ(∆) of highest dimension are associated with the vertices of ∆. On the toric variety PΣn (∆) , the polytope ∆ then determines a line bundle (Cartier divisor) via a strongly convex support function ψ∆ . For each cone of maximal dimension of Σn (∆), Ψ∆ can be described by using its dual vertex mi and setting Ψ∆ |σn (mi ) (p) = hmi , pi (B.20) for each point p in σn (mi ). This also determines Ψ∆ for all cones of lower dimension. 37 The divisor D∆ = X ai Di (B.21) νi ∈Σn (1) can then be determined from Ψ∆ |σn (mi ) (νi ) = −ai ∀νi ∈ σn (mi ) . (B.22) With the numbers ai we can recover a basis for the sections of the line bundle L via (B.17). The above relations can be used to associate a cone-wise linear support function ΨD to any divisor D. If the line bundle associated to D is ample, ΨD is strongly convex (i.e. it is convex and different for each cone of maximal dimension). A toric variety is projective iff its fan is the normal fan of a lattice polytope [30]. Let us come back to the example of P2 and a hypersurface Ck determined by Pk = 0. We may write [Pk ] = kD3 . (B.23) in terms of toric divisors, so that a1 = a2 = 0 and a3 = k. As shown in figure 7, the one-dimensional cones of Σ are generated by the vectors v1 = (1, 0), v2 = (0, 1), v3 = (−1, −1). The Newton polytope ∆k corresponding to Pk = 0 has vertices (k, 0), (0, k), (0, 0) (B.24) and its integral points correspond to all monomials of homogeneity degree k by using (B.17). In particular h(k,0),(1,0)i h(k,0),(0,1)i h(k,0),(−1,−1)i+k z2 z3 p((k, 0)) = z1 B.3 = z1k , p((0, k)) = z2k , p((0, 0)) = z3k (B.25) Resolution of singularities As there is a one-to-one correspondence between faces of ∆ and cones of Σn (∆), we may write the stratification (B.13) in terms of faces of ∆ instead of cones of Σn (∆): aa Z= ZΘ[k] . (B.26) i k Θ[k] i As faces of dimension k correspond to cones of dimension d = n − k, the dimension of these strata is n − d − 1 = k − 1. Note that ∆ counts as a face of itself, which corresponds to the zero-dimensional cone of Σn (∆). We will discuss in Appendix B.4 how to determine the topology of the strata ZΘ[k] appearing above from the Newton i polytope. 38 What we have ignored so far, however, is that the normal fan Σn (∆) in most cases (contrary to our simple example) does not give rise to a smooth toric variety. In fact, the fan Σn (∆) does not even need to be simplicial.7 We are hence interested in a refinement Σ of the fan Σn (∆) in order to resolve the hypersurface Z. Fortunately, this process results in only minor modifications of the stratification (B.26) above. In a slight abuse of notation, we will use the same letter Z also for the resolved hypersurface. Under a refinement π : Σ → Σ(∆), the stratification associated with faces of ∆ becomes: a aa Z = Z∆ ZΘ[n−1] EΘ[n−k] × ZΘ[n−k] . (B.27) i i k≥2 l l l Here EΘ[n−k] is the exceptional set of the refinement of the cone in Σ(∆) associated with the face Θ[n−k] , n−k−1 a (C∗ )i . (B.28) EΘ[n−k] = i=0 For every l-dimensional cone in π −1 (σ), where σ ∈ Σ(∆) is the cone in the normal fan associated with the face Θ[n−k] , there is a corresponding stratum (C∗ )k−l in EΘ[n−k] . B.4 Computing Hodge-Deligne numbers of strata In order to compute the Hodge numbers of toric hypersurfaces and, more generally, toric strata in such hypersurfaces, we need to introduce a further piece of machinery. The strata appearing in the stratification (B.27) naturally carry a mixed Hodge structure. In very simple terms, this means that the Hodge-Deligne numbers hp,q (H k (X, C)) (B.29) can be nonzero even when p + q 6= k, see e.g. [31, 32] for a proper introduction. These data can be packed into the numbers (which we will also call Hodge-Deligne numbers in the following) X ep,q (X) = (−1)k hp,q (H k (X, C)) , (B.30) k which are convenient for a number of reasons. First of all, in case these is a pure Hodge structure, they agree (up to a sign) with the usual Hodge numbers. Secondly, they behave in the same way as the topological Euler characteristic under unions and 7 A fan is simplicial if all of its d-dimensional cones are generated by d rays. Fans that are not simplicial give rise to toric varieties with singularities that are more general than orbifold singularities. In particular, not every Weil divisor is Q-Cartier for such varieties. 39 products of spaces ep,q (X1 q X2 ) = ep,q (X1 ) + ep,q (X2 ) X ep,q (X1 × X2 ) = ep1 ,q1 (X1 ) · ep2 ,q2 (X2 ) . (B.31) (B.32) p1 +p2 =p q1 +q2 =q Hence knowing the numbers ep,q (ZΘ[k] ) (and of (C∗ )n ) is sufficient to compute the Hodge numbers of a toric hypersurface. This information has been supplied (in the form of an algorithm) in the work of [26]. Before reviewing their algorithm, let us first discuss toric varieties themselves to illustrate the method. We have that   p,q ∗ n n+p n e ((C ) ) = δp,q (−1) . (B.33) p For a smooth toric variety V , we hence see that hp,0 = 0 for p > 0. A direct consequence of this formula combined with the stratification of a toric variety read off from its fan is the standard formula [30] for the nontrivial Hodge numbers of a smooth n-dimensional toric variety. Letting |Σ(l)| denote the number of l-dimensional cones in Σ we can write   n X p+k k p,p (−1) h (PΣ ) = |Σ(n − k)| . (B.34) p k=0 For the smooth hypersurface strata ZΘ[k] , the following relations, shown in [26], allow to compute the Hodge-Deligne numbers. First we have that for p > 0 X ep,0 (ZΘ[k] ) = (−1)k−1 `∗ (Θ[p+1] ) (B.35) Θ[p+1] ≤Θ[k] where `∗ (Θ[p+1] ) counts the number of lattice points in the relative interior of Θ[p+1] and the sum runs over all faces Θ[p+1] of dimension p + 1 contained in the face Θ[k] . Note that this sum only has one term for p + 1 = k which corresponds to the face Θ[k] itself. The remaining Hodge-Deligne numbers satisfy the ‘sum rule’   X k k−1 p,q p (−1) e (ZΘ[k] ) = (−1) + ϕk−p (Θ[k] ) , (B.36) p + 1 q where the function ϕn (Θ[k] ) is defined by [k] ϕn (Θ ) := n X n+j (−1) j=1 40   k+1 ∗ ` (jΘ) . n−j (B.37) Here jΘ denotes the polytope that is found by scaling the face Θ by j. For a face of dimension k ≥ 4 there is the simple formula   X [k] ϕ1 (Θ[k−1] ) . ek−2,1 (ZΘ ) = (−1)k−1 ϕ2 (Θ[k] ) − (B.38) Θ[k−1] ≤Θ[k] Finally, for higher Hodge numbers, p + q ≥ n, one has   n p,q n+p+1 . e (ZΘ[n] ) = δp,q (−1) p+1 (B.39) By subsequent application of these formulae, one can derive combinatorial formulae for the Hodge-Deligne numbers of strata ZΘ[k] for arbitrarily high k. It is convenient to use `n (Θ) to denote the number of points on the n-skeleton of Θ. Let us derive the Hodge-Deligne numbers ep,q of strata ZΘ[k] for k ≤ 3. First,  e0,0 (ZΘ[k] ) = (−1)k−1 `1 (Θ[k] ) − 1 , (B.40) so that e0,0 (ZΘ[1] ) = `∗ (Θ[1] ) + 1, i.e. the stratum ZΘ[1] consists of `∗ (Θ[1] ) + 1 points. Hence 0 0 ep,q (ZΘ[1] ) = ∗ [1] (B.41) ` (Θ ) + 1 0 For ZΘ[2] , we immediately find e1,1 (ZΘ[2] ) = 1. Furthermore, e1,0 (ZΘ[2] ) = −`∗ (Θ[2] ), and we can write −`∗ (Θ[2] ) 1 (B.42) ep,q (ZΘ[2] ) = 1 [2] ∗ 1 − ` (Θ ) −` (Θ[2] ) Similarly, we find for ZΘ[3] that `∗ (Θ[3] ) `2 (Θ[3] ) − `1 (Θ[3] ) `1 (Θ[3] ) − 1 0 1 [3] 2 [3] 1 [3] −3 + ` (2Θ ) − 4` (Θ ) − ` (Θ ) + ` (Θ ) 0 2 [3] 1 [3] ∗ ` (Θ ) − ` (Θ ) ` (Θ[3] ) (B.43) ∗ ∗ [3] and for ZΘ[4] e0,0 (ZΘ[4] ) = 1 − `1 (Θ[4] ) e1,0 (ZΘ[4] ) = −`2 (Θ[4] ) + `1 (Θ[4] ) e2,0 (ZΘ[4] ) = −`3 (Θ[4] ) + `2 (Θ[4] ) 2,1 3 [4] 2 [4] (B.44) [4] e (ZΘ[4] ) = ` (Θ ) − ` (Θ ) − ϕ2 (Θ ) e2,2 (ZΘ[4] ) = −4 . Let us return to the example of a degree k hypersurface in P2 . The stratification of P2 has already been discussed in §B.1, see (B.11), and its Newton polytope ∆k is 41 given in (B.24). The open dense torus (C∗ )2 in P2 gives rise to an open dense subset of Ck , i.e. a Riemann surface with a number of points excised. Its Hodge-Deligne numbers are given by −`∗ (∆k ) 1 e (Z∆k ) = 1 ∗ 1 − ` (∆k ) −` (∆k ) p,q −(k − 1)(k − 2)/2 1 = 1 − 3k −(k − 1)(k − 2)/2 (B.45) The stratum associated with each of the three 1-faces Θ[1] of ∆k consists of 1 + `∗ (Θ[1] ) = k (B.46) points. We hence recover the results derived in §B.1 for this example. B.5 Calabi-Yau hypersurfaces and reflexive polytopes Let us now come back to our prime interest, which is in Calabi-Yau manifolds. We assume that the dimension of the vector space N ⊗ R containing the fan Σ is n, so that we get an n-dimensional toric variety PΣ in which we want to embedd a Calabi-Yau hypersurface of dimension n − 1. In order for a hypersurface to be Calabi-Yau, its defining polynomial must be a section of the anticanonical bundle −KPΣ . The corresponding divisor is hence given by X D−K = c1 (PΣ ) = Di . (B.47) νi ∈Σ(1) In the notation of (B.2), we hence have ai = 1 for all i and we can identify sections of −KPΣ with the set of lattice points on the polytope ∆ ≡ {m ∈ M |hm, νi ≥ −1 ∀ν ∈ Σ(1)} ⊂ M ⊗ R. (B.48) A general section of −KPΣ , the zero locus of which defines a Calabi-Yau hypersurface, is then given by X X Y hm,ν i+1 αm zi i αm p(m) = (B.49) m m∈∆ νi ∈Σ(1) for complex constants αm . In general, the vertices of the polytope ∆ are not lattice points of M and we are not guaranteed that a generic section of −KPΣ defines a smooth (or even irreducible) Calabi-Yau hypersurface. If all of the vertices of ∆ are contained in M (in which case we will call ∆ a lattice polytope), it follows that the vertices νi are all sitting on a 42 lattice polytope ∆◦ , defined by h∆, ∆◦ i ≥ −1 . (B.50) ∆◦ is called the polar dual of ∆, and ∆ and ∆◦ are called a reflexive pair [16] if they are both lattice polytopes. Any lattice polytope whose polar dual is also a lattice polytope is called reflexive. A necessary condition for reflexivity is that the origin is the unique interior point of the polytope in question. Repeating the construction of §B.2, Calabi-Yau hypersurfaces in toric varieties are naturally constructed from reflexive pairs of lattice polytopes ∆, ∆◦ . Starting from a lattice polytope ∆, we may construct its normal fan Σn (∆). In the case of a reflexive pair, this is equal to the fan over the faces of ∆◦ . Of course, such a fan does not in general define a smooth toric variety. In fact, the cones need not even be simplicial. However, there is a natural maximal projective crepant partial (MPCP) desingularization that can be found as follows. Using (B.49), it follows that any refinement of the fan Σn (∆) that only introduces rays generated by lattice points ν on ∆◦ is crepant, i.e. preserves the Calabi-Yau property. We may hence find a MPCP desingularization by a fan refinement Σ → Σn (∆) for which all lattice points νi on ∆◦ are employed and for which PΣ is a projective toric variety. This is equivalent to finding a fine regular star triangulation T of ∆. Here, fine means that all lattice points of ∆◦ are used, and star means that every simplex contains the origin as a vertex.8 Projectivity of a toric variety is equivalent to its fan being the normal fan of a lattice polytope. While the toric variety PΣn (∆) is projective by construction, this is not necessarily true for a refinement Σ → Σn (∆). Triangulations T for which the associated fan Σ(T ) is the normal fan of a polytope are called regular (or projective, coherent) in the literature, see [33] for more details. Finally, a fine triangulation is in general not sufficient to completely resolve9 all singularities of PΣn (∆) meeting a generic Calabi-Yau hypersurface of dimension ≥ 4 [16]. The reason for this is that having a fine triangulation of a 2-face of a reflexive polytope implies that the corresponding cones do not lead to any singularities. The first dimension in which singularities can persist even for fine triangulations of a reflexive polytope is n = 5, i.e. Calabi-Yau fourfolds. For Calabi-Yau threefolds we are considering four-dimensional polytopes. Here, simplices in 3-faces can lead to pointlike singularities of the ambient toric variety even for fine triangulations, but these do not meet a generic Calabi-Yau hypersurface. In contrast, the four-dimensional cones associated with three-simplices for five-dimensional polytopes lead to singularities along curves which may meet a generic Calabi-Yau fourfold hypersurface. The simplices of the triangulations are hence the cones of the fan Σ, cut off at the surface of ∆◦ . Note that one may start from any triangulation T of ∆◦ , restrict it to a triangulation T∂∆◦ of the faces of ∆◦ and then simply construct the cones over T∂∆◦ . 9 Of course, we can always find non-crepant resolutions by introducing rays generated by lattice points outside of ∆◦ . 8 43 For a pair of n-dimensional reflexive polytopes, there is a one-to-one correspondence between the faces Θ[k] of ∆ and the faces Θ◦[n−k−1] of ∆◦ defined by hΘ[k] , Θ◦[n−k−1] i = −1 . (B.51) [k] Under the resolution induced by the refinement Σ → Σn (∆), the stratum ZΘ of a Calabi-Yau hypersurface, which corresponds to a (k − 1)-dimensional subvariety of Z before resolution, is changed according to the new simplices introduced in the dual face Θ◦[n−k−1] . A simplex of dimension l (a cone of dimension l + 1) corresponds to a subvariety of Z of dimension n−l −2. Hence a simplex of dimension l that is contained in the interior of a face Θ◦[n−k−1] corresponds to a subvariety of the form ZΘ[k] × (C∗ )n−l−k−1 . (B.52) Note that vertices Θ[0] of ∆, which are dual to the faces Θ◦[n−1] of maximal dimension, correspond to −1-dimensional varieties, i.e. they do not contribute in the stratification of Z. This persists after the resolution Σ → Σn (∆). A simple intersection argument shows that none of the divisors corresponding to points ν interior to a face of maximal dimension (and hence none of the other strata corresponding to simplices interior to a face of maximal dimension) intersect a smooth Calabi-Yau hypersurface. For any face of maximal dimension (also called a facet) Θ◦[n−1] we can find a normal vector n (this is the dual vertex Θ[0] ) such that hn, Θ◦[n−1] i = −1. This means that there is a linear relation of the form X X Di + aj D j = 0 , (B.53) νi ∈Θ◦[n−1] νj ∈Θ / ◦[n−1] for some integers aj . Let us now assume we have refined Σ such that there is a point νp interior to the facet Θ◦[n−1] . The associated divisor Dp can only have a nonzero intersection with divisors Dk for which νk also lies in Θ◦[n−1] , as all others necessarily lie in different cones of the fan Σ. This means that the above relation implies X Dp · Di = 0 , (B.54) νi ∈Θ◦[n−1] where we sum over all toric divisors coming from points on Θ◦[n−1] . The Calabi-Yau P hypersurface is given as the zero locus of a section of −KPnΣ = j Dj , where we sum over all toric divisors. We now find X X Dp · Dj = Dp · Di = 0 , (B.55) j νi ∈Θ◦[n−1] by using the same argument again. Hence Dp does not meet a generic Calabi-Yau hypersurface. Correspondingly, a refinement of Σ introducing νp does not have any 44 influence on Z. As the strata corresponding to simplices of dimension ≥ 1 interior to Θ◦[n−1] can be thought of as (an open subset of) intersections of divisors, at least one of which corresponds to an interior point of Θ◦[n−1] , none of the simplices interior to a face of maximal dimension gives rise to any subvariety of PΣ meeting Z. Correspondingly, such strata do not appear in the stratification of Z. The fact that all simplices contained in faces of maximal dimension of ∆◦ do not contribute to Calabi-Yau hypersurfaces means that we can ignore such faces when constructing a triangulation of ∆◦ . Using the methods explained above, one can derive combinatorial formulas for the Hodge numbers of toric Calabi-Yau hypersurfaces [16] that do not depend on the specific triangulation chosen. For a Calabi-Yau hypersurface of dimension n − 1 which is embedded in a toric variety of dimension n we have to consider a pair of reflexive polytopes of dimension n and stratification gives X X ◦ `∗ (Θ◦[n−2] )`∗ (Θ[1] ) h1,1 (Z) = `(∆◦ ) − (n + 1) − `∗ (Θ [n−1] ) + (Θ◦[n−2] ,Θ[1] ) Θ◦[n−1] (B.56) hn−2,1 (Z) = `(∆) − (n + 1) − X `∗ (Θ[n−1] ) + hm,1 (Z) = `∗ (Θ[n−2] )`∗ (Θ◦[1] ) (B.57) (Θ[n−2] ,Θ◦[1] ) Θ[n−1] X X `∗ (Θ◦[n−m−1] )`∗ (Θ[m] ) for n − 2 > m > 1. (B.58) (Θ◦[n−m−1] ,Θ[m] ) Note that these numbers only make sense for a smooth Calabi-Yau hypersurface, which is only guaranteed without further investigation for Calabi-Yau hypersurfaces of dimension ≤ 3. Although the above formulas for h1,1 (Z) and hn−2,1 (Z) are derived using the stratification technique of [26], they have a straightforward explanation. In particular, the formula for hn−2,1 (Z) counts the number of complex structure deformations by counting the number of monomial deformations appearing in the defining equation and subtracting the dimension of the automorphism group of PΣ . Finally, the last term in (B.57) corrects for the fact that not all deformations are realized as polynomial deformations. Similarly, the formula for h1,1 (Z) counts the number of inequivalent divisors of PΣ that meet Z, with a correction term taking into account that some divisors of PΣ become reducible on Z. As is apparent from the above formulae, exchanging the roles of ∆ and ∆◦ exchanges h1,1 (Z) ↔ h2,1 (Z). This is how mirror symmetry is realized for toric CalabiYau hypersurfaces. B.6 Topology of subvarieties of Calabi-Yau threefolds In this section we describe the topology of subvarieties of Calabi-Yau hypersurfaces in toric varieties that are obtained by restricting toric subvarieties of the ambient space. 45 For ease of notation we restrict to the case n = 4, i.e. Calabi-Yau threefolds, but a similar analysis may be carried out in higher dimensions. As we have already explained, we only need to consider simplices on the 2-skeleton of ∆◦ , as strata of PΣ associated with simplices interior to 3-faces of ∆◦ do not meet a smooth Calabi-Yau hypersurface. Each l-simplex of a triangulation T of ∆◦ corresponds to a l+1-dimensional cone in the fan Σ and hence to an open stratum (C∗ )4−(l+1) in PΣ . Depending on the location of the simplex on ∆◦ , the defining equation of the Calabi-Yau hypersurface will only constrain some of the C∗ factors, while others will lie entirely in the Calabi-Yau hypersurface. The reason for this is in the resolution process Σ → Σn (∆). If we work with the singular varieties determined by Σn (∆), every k-dimensional face Θ◦[k] of ∆◦ gives rise to a stratum ZΘ[4−k−1] of dimension 2 − k. This stratum is given as a hypersurface in (C∗ )3−k . The resolution process Σ → Σn (∆), described in Appendix B.3, yields ZΘ[4−k−1] → ZΘ[4−k−1] × EΘ[4−k−1] . (B.59) The factor EΘ[4−k−1] is determined by the simplices contained in the relative interior of the face Θ◦[k] . Every l-simplex contained in the relative interior of a k-dimensional face Θ◦[k] contributes a (C∗ )k−l to EΘ[4−k−1] , and hence it contributes a stratum ZΘ[4−k−1] × (C∗ )k−l (B.60) to Z. Note that the factor ZΘ[3−k] is common to all of the strata originating from simplices contained in a chosen face. For a Calabi-Yau threefold, this correspondence is such that ZΘ is a two-dimensional variety for vertices, a curve for strata interior to 2-faces and a collection of points for strata interior to two-dimensional faces. The closed subvarieties of PΣ , and hence of Z, associated with a simplex t are found by collecting all lower-dimensional simplices u attached to t (i.e. t ⊂ ∂u) and taking the disjoint union of the associated strata. As the Hodge-Deligne numbers ep,q are additive, this provides an efficient way to find the Hodge numbers of the associated subvariety. Again, we may neglect all simplices that are contained in the relative interior of faces of maximal dimension (three in this case) as these do not contribute any strata to a Calabi-Yau hypersurface. In the following, we will explicitly write down the resulting stratifications of various subvarieties and compute their Hodge numbers. At this point, we will adopt a different notation than in the rest of this appendix. As in the main text of the paper, we focus on threefolds, and so only need to distinguish vertices, edges, 2-faces, and 3-faces (facets) of ∆◦ , which we denote by v, e, f and c, respectively. These are dual to 3-faces, 2-faces, edges and vertices on the M -lattice polytope ∆, which are consequently denoted by v ◦ , e◦ , f ◦ and c◦ . We hope this does not confuse the reader. 46 B.6.1 Vertices Let us consider a divisor Di for which the associated lattice point νi = v is a vertex. The vertex has a dual face v ◦ on ∆ that contributes an open two-dimensional stratum Zv◦ . Furthermore, there will be 1-simplices contained in edges e (dual to faces e◦ ) ending on v contributing Ze◦ as well as 1-simplices on 2-faces f (dual to 1-faces f ◦ ) contributing Zf ◦ × C∗ . Finally there are 2-simplices on faces f (dual to 1-faces f ◦ ) contributing Zf ◦ . Hence such divisors contain the irreducible hypersurface Zv◦ as an open dense set that is compactified by the other strata. Note that Zf ◦ is just a collection of `∗ (f ◦ ) + 1 points. Collecting all of these strata we find the stratification of Di to be X X  Di = Zv◦ qe≥v Ze◦ × (pt) qf ≥v Zf◦ × C∗ q pt . (B.61) With the stratification (B.61) at hand, we can start computing the Hodge numbers. For h1,0 (Di ), only the first two strata contribute and we find h1,0 (Di ) = −e1,0 (Di ) ! = − e1,0 (Zv◦ ) + X e1,0 (Ze◦ ) (B.62) e⊃v ! =− X X `∗ (e◦ ) − e⊃v `∗ (e◦ ) =0 e⊃v For h2,0 (Di ), only the first stratum in (B.61) contributes and we find h2,0 (Di ) = e2,0 (Zv◦ ) = `∗ (Zv◦ ) . (B.63) Finally, we can compute h1,1 (Di ). Here, we find a contribution from Zv◦ , computed in (B.43), as well as a contribution X e1,1 (Ze◦ ) = e⊃v X 1 (B.64) e⊃v from any edge e emanating from the vertex v. Furthermore, 1-simplices interior to any 2-face f (dual to f ◦ ) connected to the vertex v contribute X e0,0 (Zf ◦ ) × f ⊃e = X X e1,1 (C∗ ) t1 ∗ ◦ (` (f ) + 1) X t1 ⊃νi f ⊃e 47 (B.65) 1 Summing it all up, the result is h1,1 (Di ) = − 3 + `∗ (2v ◦ ) − 4`∗ (v ◦ ) − `2 (v ◦ ) + `1 (v ◦ ) X X X + 1+ (`∗ (f ◦ ) + 1) 1 B.6.2 t1 ⊃νi f ⊃v e⊃v (B.66) Simplices interior to 1-faces of ∆◦ Let us first consider divisors originating from points νi interior to edges e of ∆◦ dual to 2-faces e◦ of ∆. The 0-simplex corresponding to νi is of the form Ze◦ × C∗ , and the two 1-simplices on the edge e containing νi correspond to Ze◦ . One-simplices containing νi that are interior to 2-faces f (dual to f ◦ ) contribute `∗ (f ◦ ) + 1 points times C∗ and 2-simplices contribute `∗ (f ◦ ) + 1 points. The open and dense stratum of such divisors (which originates from the simplex νi ) is simply the product of a curve of genus `∗ (e◦ ) (with `1 (e◦ ) points excised) times a C∗ , which gets compactified by the remaining strata. We may think of these as (the open dense subsets of) the intersection of Di with ‘neighboring’ divisors. The two 1-simplices along the edge e partially compactify Ze◦ × C∗ to Ze◦ × P1 . The remaining C∗ ’s and points sit over the `1 (e◦ ) points excised in the open curve Ze◦ . We may hence think of such divisors as follows: they are flat fibrations of a P1 over a curve of genus `∗ (e◦ ). Over `1 (e◦ ) points, the fiber P1 degenerates into a chain of P1 ’s, as determined by number of 1-simplices (and 2-simplices) attached to νi lying on neighboring 2-faces f ⊃ v. To see the details of how this works first note that X `∗ (f ◦ ) + 1 . (B.67) `1 (e◦ ) = f ◦ ⊂e◦ Over each of the `∗ (f ◦ ) + 1 points that are excised due to the face f ◦ ⊂ e◦ , we find the strata corresponding to the one- (and two-) simplices interior to the dual face f ⊃ e. Hence over `∗ (f ◦ ) + 1 points, where f ◦ , f are dual faces, the generic fiber P1 of Di degenerates into a number of P’s equal to the number of 1-simplices which attached to νi and interior to f . A cartoon of this is shown in figure 8. From this analysis of the fibration structure, we expect that X X h1,1 (Di ) = 2 + (`∗ (f ◦ ) + 1) · (−1 + 1) , (B.68) t1 ⊃νi f ⊃e which will be confirmed by a direct computation using the stratification below. As explained above, the stratification of Di is X X  Di = Ze◦ × (C∗ + 2pts) qf ⊃e Zf ◦ C∗ + pt . (B.69) Here, the C∗ multiplying Ze◦ is due to νi (k = 2, l = 1), whereas the 2 points correspond 48 reducible fibre consisting of 1 t1⊃νi 1 touching P ’s P generic P1 fibre 1 + `∗(f ◦) points genus `∗(e◦) curve Figure 8: The fibration structure of a divisor Di of Z associated with a lattice point νi interior to an edge e of a four-dimensional polytope. The base is a genus g = `∗ (e◦ ) curve and the generic fiber is a P1 . For each neighboring 2-face f ⊃ e, there are 1 + `∗ (f ◦ ) points over which the fiber degenerates into as many P1 s as there are 1simplices t1 on f that contain νi . to the two 1-simplices on e containing νi (k = 2, l = 2). Each C∗ multiplying Zf ◦ corresponds to a 1-simplex containing νi that is interior to the face f and each pt corresponds to a 2-simplex containing νi that is interior to f . Again, h0,0 (Di ) = 1 as Di is irreducible. The computation for h1,0 now becomes h1,0 (Di ) = −e1,0 (Ze◦ · (C∗ q 2pts)) = `∗ (e◦ ) · (e0,0 (C∗ ) + 2e0,0 (pt)) (B.70) = `∗ (e◦ ) . We have h2,0 (Di ) = e2,0 (Di ) = 0 as no stratum contributes. Already for the highest stratum Ze◦ , we have to count interior points to 3-dimensional faces of e◦ , of which there are none. These Hodge numbers fit with the fibration structure discussed above. 49 Finally, let us compute h1,1 (Di ). Here we need  h1,1 (Di ) = e1,1 (Ze◦ ) · e0,0 (C∗ ) + 2e0,0 (pt) + e0,0 (Ze◦ ) · e1,1 (C∗ ) X X + e0,0 (Zf ◦ ) · 1 f ◦ ⊂e◦ t1 ⊃νi = 1 · (−1 + 2) + (1 − `1 (e◦ )) · 1 + X (1 + `∗ (f ◦ )) · f ◦ ⊂e◦ t1 ⊃νi ! =2− X `∗ (f ◦ ) + 1 f ◦ ⊂e◦ + X X (1 + `∗ (f ◦ )) · f ◦ ⊂e◦ X 1 (B.71) 1 t1 ⊃νi ! =2+ X (1 + `∗ (f ◦ )) · −1 + f ◦ ⊂e◦ X 1 , t1 ⊃νi as predicted from the analysis of the fibration of Di carried out above. Similarly, one may analyze curves C that correspond to 1-simplices t1 interior to a 1-face e. Here, the stratification is X Zf◦ × (pt) . (B.72) C = Ze◦ + f ◦ ⊂e◦ The stratum Ze◦ is a curve of genus `∗ (e◦ ) with a number of points excised. The second term is due to the unique 2-simplex attached to t1 on every face f ⊃ e, which consists of `∗ (f ) + 1 points for each two-dimensional face containing e. It supplies the points that compactify Ze◦ to C. It follows immediately that the genus of C is h1,0 (C) = `∗ (e◦ ) . (B.73) This fits with the fact that the union of all strata corresponding to simplices in the interior of e sits over the curve C, so that two neighboring divisors Di and Dj intersect in the common base of both their fibrations. B.6.3 Simplices interior to 2-faces of ∆◦ Again, let us first consider 0-simplices νi interior to a 2-face f of ∆◦ . The open dense subset of Di originating from νi is given by Zf ◦ × (C∗ )2 , 50 (B.74) while 1-simplices (2-simplices) containing νi compactify Di this by contributing Zf ◦ ×C∗ and Zf ◦ × (pts). All in all, the stratification of Di is ! Di = Z f ◦ × X (C∗ )2 + X C∗ + t1 ⊃νi pt . (B.75) t2 ⊃νi where t1 and t2 are simplices interior to f . As Zf ◦ is just a collection of `∗ (f ◦ )+1 points, the divisors considered here are, in general, reducible with each irreducible component being a toric variety Pstar(νi ) determined by the stratification above. Starting from the triangulation of a 2-face, we may construct the star fan w.r.t. νi to find the fan of the toric variety Pstar(νi ) , see figure 9. 11 00 00 11 1 0 0 1 11 00 00 11 1 0 0 1 11 00 00 11 00 11 1 0 0 1 0 1 νi 11 00 00 11 11 00 00 11 00 11 Figure 9: On the left hand side, the neighborhood of a lattice point νi inside a 2-face f with a triangulation. We have colored the simplices containing νi in red. These contribute to the star fan star(νi ) shown on the right hand side. From this, it immediately follows that h1,1 (Pstar(νi ) ) = −2 + X 1. (B.76) t1⊂νi The same result is easily recovered from the stratification (B.75): ! h1,1 (Di ) = e0,0 (Zf ◦ ) × X e1,1 ((C∗ )2 )) + t1 ⊃νi ∗ ◦ = (` (f ) + 1) · (−2 + X e1,1 (C∗ ) (B.77) 1) . t1⊂νi Similarly, the closed subvariety associated with each 1-simplex interior to f is ` (f ) + 1 times a P1 and the closed subvariety associated with each 2-simplex consists of `∗ (f ◦ ) + 1 points. This implies that any two (three) divisors associated with points ∗ ◦ 51 interior to f with a nonzero intersection will intersect in a collection of `∗ (f ◦ ) + 1 disjoint Ps ’s (points). B.6.4 An example Let us consider a (slightly) nontrivial example to see the above machinery at work. Consider a reflexive polytope ∆◦ with vertices v0 = [−1, −3, −9, −14] v1 = [0, −2, −6, −9] v2 = [0, 0, 0, 1] (B.78) v3 = [0, 0, 1, 0] v4 = [0, 1, 0, 0] v5 = [1, 0, 0, 0] The 3-faces of ∆◦ as well as the vertices of ∆◦ spanning them, their numbers of interior points and the dual vertices on ∆ are c0 =< v0 , v1 , v2 , v3 > `∗ (c0 ) = 0 ↔ c◦0 = [0, 8, −1, −1] c1 =< v0 , v1 , v2 , v4 > `∗ (c1 ) = 0 ↔ c◦1 = [0, −1, 2, −1] c2 =< v0 , v2 , v3 , v4 > `∗ (c2 ) = 0 ↔ c◦2 = [27, −1, −1, −1] c3 =< v1 , v2 , v4 , v5 > `∗ (c3 ) = 0 ↔ c◦3 = [−1, −1, 2, −1] c4 =< v0 , v1 , v3 , v4 , v5 > `∗ (c4 ) = 3 ↔ c◦4 = [−1, −1, −1, 1] c5 =< v2 , v3 , v4 , v5 > `∗ (c5 ) = 0 ↔ c◦5 = [−1, −1, −1, −1] c6 =< v1 , v2 , v3 , v5 > `∗ (c6 ) = 0 ↔ c◦6 = [−1, 8, −1, −1] (B.79) The 3-faces of ∆ dual to the vertices of ∆◦ are v0 ↔ v0◦ =< c◦4 , c◦1 , c◦2 , c◦0 > `∗ (v0◦ ) = 1 v1 ↔ v1◦ =< c◦4 , c◦3 , c◦0 , c◦1 , c◦6 > `∗ (v1◦ ) = 0 v2 ↔ v2◦ =< c◦5 , c◦2 , c◦0 , c◦3 , c◦1 , c◦6 > `∗ (v2◦ ) = 54 v3 ↔ v3◦ =< c◦5 , c◦2 , c◦4 , c◦0 , c◦6 > `∗ (v3◦ ) = 22 v4 ↔ v4◦ =< c◦5 , c◦2 , c◦4 , c◦1 , c◦3 > `∗ (v4◦ ) = 4 v5 ↔ v5◦ =< c◦5 , c◦6 , c◦4 , c◦3 > `∗ (v5◦ ) = 1 52 (B.80) The edges of ∆◦ and their dual 2-faces on ∆ are e0 =< v0 , v1 > `∗ (e0 ) = 0 ∗ ↔ e◦0 e◦1 e◦2 e◦3 e◦4 e◦5 e◦6 e◦7 e◦8 e◦9 e◦10 e◦11 e◦12 e◦13 e1 =< v0 , v2 > ` (e1 ) = 0 ↔ e2 =< v1 , v2 > `∗ (e2 ) = 1 ↔ e3 =< v0 , v3 > `∗ (e3 ) = 0 ↔ e4 =< v1 , v3 > ∗ ` (e4 ) = 0 ∗ ↔ e5 =< v2 , v3 > ` (e5 ) = 0 ↔ e6 =< v0 , v4 > `∗ (e6 ) = 0 ↔ e7 =< v1 , v4 > ∗ ` (e7 ) = 2 ∗ ↔ e8 =< v2 , v4 > ` (e8 ) = 0 ↔ e9 =< v3 , v4 > `∗ (e9 ) = 0 ↔ e10 =< v4 , v5 > `∗ (e10 ) = 0 ↔ e11 =< v1 , v5 > ∗ ` (e11 ) = 0 ∗ ↔ e12 =< v2 , v5 > ` (e12 ) = 0 ↔ e13 =< v3 , v5 > `∗ (e13 ) = 0 ↔ =< c◦4 , c◦1 , c◦0 > =< c◦1 , c◦2 , c◦0 =< c◦3 , c◦0 , 1, c◦6 =< c◦4 , c◦0 , c◦2 =< c◦4 , c◦6 , c◦0 =< c◦5 , c◦2 , c◦0 , c◦6 =< c◦4 , c◦1 , c◦2 =< c◦4 , c◦3 , c◦1 =< c◦5 , c◦2 , c◦1 , c◦3 =< c◦5 , c◦2 , c◦4 =< c◦5 , c◦3 , c◦4 =< c◦4 , c◦3 , c◦6 =< c◦5 , c◦6 , c◦3 =< c◦5 , c◦6 , c◦4 `∗ (e0 ) =0 ∗ > ` (e1 ) =7 > `∗ (e2 ) =0 > `∗ (e3 ) =4 > ∗ =0 ∗ ` (e4 ) > ` (e5 ) = 108 > `∗ (e6 ) =1 > ∗ =0 ∗ ` (e7 ) > ` (e8 ) = 27 > `∗ (e9 ) = 13 > `∗ (e10 ) =1 > ∗ =0 ∗ =7 ` (e11 ) > ` (e12 ) > `∗ (e13 ) =4 (B.81) ◦ and finally the 2-faces on ∆ and their dual edges on ∆ together with their numbers of interior points are f0 =< v0 , v1 , v2 > `∗ (f0 ) = 0 ∗ ↔ f0◦ f1◦ f2◦ f3◦ f4◦ f5◦ f6◦ f7◦ f8◦ f9◦ ◦ f10 ◦ f11 ◦ f12 ◦ f13 ◦ f14 f1 =< v0 , v1 , v3 > ` (f1 ) = 0 ↔ f2 =< v1 , v2 , v3 > `∗ (f2 ) = 0 ↔ f3 =< v0 , v2 , v3 > ∗ ` (f3 ) = 0 ∗ ↔ f4 =< v0 , v1 , v4 > ` (f4 ) = 0 ↔ f5 =< v0 , v2 , v4 > `∗ (f5 ) = 0 ↔ f6 =< v1 , v2 , v4 > `∗ (f6 ) = 1 ↔ f7 =< v0 , v3 , v4 > ∗ ` (f7 ) = 0 ∗ ↔ f8 =< v2 , v3 , v4 > ` (f8 ) = 0 ↔ f9 =< v1 , v4 , v5 > `∗ (f9 ) = 0 ↔ f10 =< v2 , v4 , v5 > ∗ ` (f10 ) = 0 ∗ ↔ f11 =< v1 , v2 , v5 > ` (f11 ) = 0 ↔ f12 =< v3 , v4 , v5 > `∗ (f12 ) = 0 ↔ f13 =< v1 , v3 , v5 > `∗ (f13 ) = 0 ↔ f14 =< v2 , v3 , v5 > ∗ ` (f14 ) = 0 ↔ 53 =< c◦1 , c◦0 > =< =< =< =< =< =< =< =< =< =< =< =< =< =< c◦4 , c◦0 c◦6 , c◦0 c◦0 , c◦2 c◦4 , c◦1 c◦1 , c◦2 c◦3 , c◦1 c◦4 , c◦2 c◦5 , c◦2 c◦4 , c◦3 c◦5 , c◦3 c◦3 , c◦6 c◦5 , c◦4 c◦4 , c◦6 c◦5 , c◦6 `∗ (f0 ) ∗ =2 > ` (f1 ) =0 > `∗ (f2 ) =0 > ∗ =8 ∗ ` (f3 ) > ` (f4 ) =0 > `∗ (f5 ) =2 > `∗ (f6 ) =0 > ∗ ` (f7 ) ∗ = 1 (B.82) > ` (f8 ) = 27 > `∗ (f9 ) =0 > ∗ =2 ∗ ` (f10 ) > ` (f11 ) =2 > `∗ (f12 ) =1 > `∗ (f13 ) =0 > ∗ ` (f14 ) =8 The Hodge numbers of the corresponding mirror pair of Calabi-Yau threefolds Z and Z̃ can be quickly found with these numbers by evaluating (B.56) and (B.58) h1,1 (Z) = h2,1 (Z̃) = 6 (B.83) h1,1 (Z̃) = h2,1 (Z) = 228 There is a single 2-face f6 that requires triangulation. This face, with its integral points and its bounding edges, as well as its triangulations, is shown in figure 10. Let us discuss the topology of the divisors Di that correspond to the lattice points ν7 ν7 11 00 00 11 00 11 ν5 11 00 00 11 00 11 11 00 00 11 00 11 ν1 11 00 00 11 00 11 ν2 11 00 00 11 00 11 ν6 ν5 11 00 00 11 00 11 11 00 00 11 00 11 11 00 00 11 00 11 11 00 00 11 00 11 ν3 11 00 00 11 00 11 ν4 11 00 00 11 00 11 ν1 ν2 ν6 11 00 00 11 00 11 11 00 00 11 00 11 ν7 11 00 00 11 00 11 11 00 00 11 00 11 ν1 11 00 00 11 00 11 ν2 00 11 11 00 00 11 ν6 ν5 11 00 00 11 00 11 11 00 00 11 00 11 11 00 00 11 00 11 11 00 00 11 00 11 ν3 ν4 ν7 00 11 11 00 00 11 ν5 11 00 00 11 00 11 ν3 11 00 00 11 00 11 ν4 11 00 00 11 00 11 ν1 ν2 ν6 11 00 00 11 00 11 11 00 00 11 00 11 ν3 11 00 00 11 00 11 ν4 ν7 11 00 00 11 00 11 ν5 ν6 00 11 11 00 00 11 11 00 00 11 00 11 ν1 11 00 00 11 00 11 00 11 11 00 00 11 11 00 00 11 00 11 ν2 ν3 11 00 00 11 00 11 ν4 Figure 10: The only nontrivial 2-face f6 of ∆◦ along with its five possible triangulations. 54 νi , i = 1 · · · 7 sitting on this face. They are ν1 = [0, −2, −6, −9] = v1 ν2 = [0, −2, −4, −6] ν3 = [0, 0, −2, −3] ν4 = [0, 1, 0, 0] = v4 (B.84) ν5 = [0, −1, −3, −4] ν6 = [0, 0, −1, −1] ν7 = [0, 0, 0, 1] = v2 As remarked above, some of the Hodge numbers will depend on the triangulation. Let us choose the triangulation shown in the upper left of figure 10. The divisors D1 , D4 and D7 correspond to vertices, so that we conclude h1,0 vanishes for all three and h2,0 (D1 ) = `∗ (v1◦ ) = 0 h2,0 (D4 ) = `∗ (v4◦ ) = 4 (B.85) h2,0 (D7 ) = `∗ (v2◦ ) = 54 no matter which triangulation is chosen. Let us now compute the Hodge numbers h1,1 for the triangulation in the upper left of Figure 10, for which we have to evaluate h1,1 (Di ) = − 3 + `∗ (2v ◦ ) − 4`∗ (v ◦ ) − `2 (v ◦ ) + `1 (v ◦ ) X X X 1 + 1+ (`∗ (f ◦ ) + 1) e⊃v (B.86) t1 ⊃νi f ⊃v Note that there are no other 1-faces except e2 and e7 , both of which are on f6 , with interior points and all 2-faces of ∆◦ are simplicial. Hence there can be no other 1simplices except the ones in f6 , shown in Figure 10, which contain ν1 , ν4 or ν7 and are interior to a 2-face. Furthermore, `∗ (f6◦ ) = 0. The number of edges ei containing each of the three vertices in questions is found from (B.81). #e ⊃ v1 = 5 #e ⊃ v4 = 5 (B.87) #e ⊃ v2 = 5 We finally find that h1,1 (D1 ) = 1 + 5 + 0 = 6 h1,1 (D4 ) = 77 + 5 + 1 = 83 1,1 h (D7 ) = 398 + 5 + 1 = 404 55 (B.88) Only the last number depends on the triangulation chosen. Let us now investigate the points νi interior to 1-faces ei . Here h2,0 (Di ) = 0 for all cases. We start with v5 which is contained in e2 . We hence learn from (B.81) that h1,0 (D5 ) = 0. For h1,1 , we have to evaluate ! X X 2+ (1 + `∗ (f ◦ )) · −1 + 1 (B.89) f ◦ ⊂e◦ t1 ⊃νi In the triangulation we are considering, there are three 1-simplices contained in f6 which each contribute 1 (as `∗ (f ◦ ) = 0) to h1,1 . For each 2-face apart form f6 , there can only be a single 1-simplex containing ν5 , so that we conclude h1,1 (D5 ) = 4 . (B.90) As described in general above, this means we should think of D5 as a fibration of a P1 over another P1 for which the fiber degenerates into a union of three P1 s over a single point in the base. The other two points ν2 and ν3 interior to edges are contained in the same edge, e7 , so that h1,0 (D2 ) = h1,0 (D2 ) = `∗ (e◦7 ) = 0 . (B.91) For the triangulation chosen, ν2 only connects to a single vertex inside f6 , whereas ν3 connects to two, hence h1,1 (D2 ) = 2 h1,1 (D3 ) = 3 . (B.92) Finally, there is ν6 . As it is interior to a 2-face, it is n = `∗ (f6 ) + 1 = 1 copies of a toric variety. This toric variety can be directly read off from the star fan to be the Hirzebruch surface F1 for the triangulation chosen. Hence h1,0 (D6 ) = h2,0 (D6 ) = 0 h1,1 (D6 ) = 2 . (B.93) A similar discussion can now easily be made for other triangulations. We can e.g. consider a flop taking us from the triangulation on the upper left to the one on the upper right. This will decrease the h1,1 of D3 and D6 by one, whereas the h1,1 of D5 and D4 are increased by one. B.7 Hodge numbers hi,0 of toric divisors of Calabi-Yau n-folds For Calabi-Yau manifolds of higher dimension than 3, the same technique as used above can be used to find topological data of toric divisors restricted to a Calabi-Yau hypersurface. Whereas Hodge numbers such as h1,1 (Di ) will depend on the triangulation, one can derive a remarkably simple formula for the Hodge numbers hi,0 . For a smooth 56 Calabi-Yau n−1-fold associated with a pair of n-dimensional reflexive polytopes ∆◦ , ∆ and a lattice point ν in the relative interior of a face Θ◦[n−d] of dimension n − d, the b is such that associated divisor D b = δi,d−2 `∗ (Θ[d−1] ) , hi,0 (D) (B.94) where d > 2. Here Θ[d−1] is the face of ∆ dual to the face Θ◦[n−d] containing νD . b = 1 holds for d > 2, as all such divisors are connected. Formula Furthermore, h0,0 (D) (B.94), which we will prove in the following, is the central result of this section. Note that (B.94) reduces to the corresponding relations derived above for Calabi-Yau threefolds (where n = 4). In the threefold case, divisors associated with vertices (d = 4) only have a non-vanishing h2,0 and divisors associated with points interior to edges (d = 3) only have a non-vanishing h1,0 . Let us assume that we are given a pair of reflexive polytopes ∆◦ , ∆ and a triangulation giving rise to a smooth projective toric variety10 PΣ . Let D be the toric divisor associated with a lattice point νD contained in the relative interior of a face Θ◦[n−d] of b = D ∩ Z. dimension n − d. We are interested in the Hodge numbers of a divisor D Any toric divisor D is composed of the strata associated with all cones that contain the ray over νD . As before, these descend to a subset of the strata of Z and we can b sum their Hodge-Deligne numbers ep,q to find the Hodge numbers of D. We first note that the cases d = 1 and d = 2 are trivial. In the first case, d = 1, D does not give rise to any divisor on Z, while in the second case, d = 2, the face Θ◦[n−2] is dual to a face of dimension n − (n − 2) − 1 = 1, denoted by Θ[1] , and it enjoys a stratification of the form   b = ZΘ[1] × strata of the form(C∗ )i D (B.95) so that such divisors have h0,0 = `∗ (Θ[1] ) + 1 disconnected components which are all smooth toric varieties. Hence the only nontrivial Hodge numbers of such divisors are b hi,i (D). We hence assume d > 2 in the following. Let us start by writing down the stratification of an arbitrary divisor of Z descending from a toric divisor. It is given by X  b = qk,Θ◦[n−d+k] ⊃ν ZΘ[d−1−k] × D pt. + · · · + (C∗ )n−d+k . (B.96) D As usual, (Θ◦[n−d+k] , Θ[d−1−k] ) are a pair of dual faces. For a divisor inside a face Θ◦[n−d] , only strata on neighboring faces containing Θ◦[n−d] contribute. On ∆, this can be expressed by saying that only faces Θ[d−1−k] of Θ[d−1] contribute. For each such face, the toric strata (C∗ )l in each term originate from various simplices of the 10 More generally, it is enough to for the singularities of PΣ to miss a generic Calabi-Yau hypersurface, i.e. the only cones of Σ of lattice volume ≥ 1 are sitting inside faces of maximal dimension of ∆◦ . 57 triangulation on the faces Θ◦[n−d+k] dual to Θ[d−1−k] that contain the point νD . In particular, a (C∗ )n−d+k originates from the point νD itself, a (C∗ )n−d+k−1 originates from every 1-simplex in the interior of Θ◦[n−d+k] containing νD , and so on. Finally, the highest-dimensional simplices, which are n − d + k-dimensional, give rise to points in the above expression. Our main tool in deriving (B.94) will be (B.35). Only the strata ZΘ[d−1−k] can potentially contribute to hi,0 , as ei,0 ((C∗ )k ) = 0 for i > 0. Hence we will only need to evaluate ei,0 (Θ[d−1−k] ) and e0,0 of the sum over simplices on the right-hand side of (B.96). b =0 The first conclusion that can be drawn directly from (B.35) is that hi,0 (D) whenever i > d − 2. In this case, none of the strata ZΘ[d−1−k] can contribute, as we would need to count points in faces of dimension i + 1 in the face ZΘ[d−1−k] , but even for k = 0, there are no such faces. The geometric reason for this is that any divisor associated with a point νD inside a face of dimension n − d should be thought of as an exceptional divisor originating from the resolution discussed in §B.3. Correspondingly, each such divisor is a fibration of a toric variety of dimension n − d (which degenerates over various subloci) over an irreducible manifold of dimension d−2. Hence the highest possible i for which hi,0 is nonzero is d − 2, as already established. Indeed, we do get a nonzero contribution whenever i = d − 2. In this case, only the stratum ZΘ[d−1] (i.e. k = 0) contributes and we find X  ei,0 = (−1)d−2 `∗ (Θ[d−1] ) × e0,0 (C∗ )l strata (B.97) The sum on the right hand side runs over all of the simplices on Θ◦[n−d] that contain νD , including νD itself, and the sign alternates according to the dimension of the simplex in question as e0,0 ((C∗ )l ) = (−1)l . If we neglect νD , the remaining simplices are arranged such that they form an n − d − 1-dimensional polyhedron. It can be found by intersecting the various simplices with an n − d − 1-sphere on Θ◦[n−d] centered at νD . Here, 1-simplices in the alternating sum, which contribute (−1)n−d−1 above, correspond to vertices of the polyhedon. As this polyhedron is topologically a sphere we can write X  e0,0 (C∗ )l strata = (−1)n−d−1 χ(S n−d−1 ) + (−1)n−d = 1 (B.98) b is a smooth compact manifold we can where νD contributes the second term. As D b = (−1)i ei,0 (D), b so that we have shown the case i = d − 2 of (B.94). use h0,i (D) We now proceed to show that h0,i = 0 for all 0 < i < d − 2. Depending on i, a number of strata ZΘ[k] from (B.96) contribute. Starting again from (B.96), we can 58 write b = ei,0 (D) X ei,0 (ZΘ[d−1−k] ) × e0,0 k,Θ◦[n−d+k] ⊇ν X  pt. + · · · + (C∗ )n−d+k  . (B.99) D For every term in the above sum over k, we have to find the alternating sum of all of the simplices containing νD on the face Θ◦[n−d+k] dual to Θ[d−1−k] to evaluate the various e0,0 . For k = 0, we have already found that this sum simply gives 1 by relating it to the Euler characteristic of a sphere. For higher values of k, we can essentially use a similar argument. In this case, the point νD sits on the codimension-k hyperplane of Rn−d+k defined by the face Θ◦[n−d] . Furthermore, the face Θ◦[n−d+k] will be bounded by other hyperplanes of dimension greater than or equal to n − d, so that the set of all simplices on Θ◦[n−d+k] connecting to νD will correspond to a triangulation of an open subset of a sphere of dimension n − d + k − 1. This has Euler characteristic 1 in even and −1 in odd dimensions. To fix the sign, note that points on this sphere correspond again to 1-simplices, which in turn have a factor of (−1)n−d+k−1 in the sum. As such points contribute 1 in the computation of the Euler characteristic, the contribution of the sum over simplices to e0,0 is always equal to 1. Using (B.35), we are hence led to X b = ei,0 (D) ei,0 (ZΘ[d−1−k] ) k,Θ◦[n−d+k] ⊃νD = X X (−1)d−1−k k,Θ◦[n−d+k] ⊇νD `∗ (Θ[i+1] ) . (B.100) Θ[i+1] ⊆Θ[d−1−k] Note that each face containing νD appears multiple times and with alternating signs in the above expression. In particular, a single face Θ[i+1] can appear multiple times in a single term in the sum over k. Let us consider a single such face Θ[i+1] and find how often it appears with which signs. First, note that we may equally well phrase the problem in terms of faces of ∆◦ . Given the face Θ◦[n−d] containing νD and the face Θ◦[n−i−2] dual to Θ[i+1] , the factor multiplying `∗ (Θ[i+1] ) for a fixed face of dimension [i − 1] in the above sum is then simply X (−1)d−1−k . (B.101) k,Θ◦[n−d] ⊆Θ◦[n−d+k] ⊆Θ◦[n−i−2] The contribution proportional to `∗ (Θ[i+1] ), the dual face of which is Θ◦[n−i−2] , is hence given by counting all n − d + k-dimensional faces containing Θ◦[n−d] and contained in Θ◦[n−i−2] . To compute this quantity, we again interpret this as an Euler characteristic of a topological space as follows. Consider a sphere of dimension n−i−2−(n−d)−1 = d−i−3 centered at νD and orthogonal to the face Θ◦[n−d] . All the faces contributing to the sum above, except Θ◦[n−d] itself, will give rise to a decomposition of one closed half 59 of this sphere, which has Euler characteristic +1 in any dimension. The contribution of the highest-dimensional stratum on this half-sphere has k = d − i − 2, so that it contributes (−1)i−3 to the alternating sum in (B.101). As its dimension is d − i − 3, it contributes (−1)d−i−3 to the Euler characteristic, so that the sum in (B.101), still neglecting the face Θ◦[n−d] , is (−1)d . The face Θ◦[n−d] contributes (−1)d−1 , so that these two terms always cancel and the sum (B.101) vanishes for any pair of faces Θ◦[n−d] and Θ◦[n−i−2] . Hence the sum (B.100) vanishes except when d = i − 2, when only one term in (B.101) contributes. This completes the proof of (B.94). 60 C Computation of h2 for a Divisor on a 2-Face In this appendix we give an alternative computation of h2 (OD ) in a special case (defined below). We will compute h2 (OD ) directly in terms of a counting of lattice points in ∆, arriving at a result that coincides with Theorem 3 in this subcase. This computation provides an alternative perspective to that of the spectral sequence in §3.1. C.1 Preliminaries We begin by assembling some elementary results about divisors and Calabi-Yau hypersurfaces in toric varieties. Let V be a four-dimensional simplicial toric variety, with X a b denote a divisor in V . We write D = D b ∩ X. Calabi-Yau hypersurface in V , and let D Proposition 23 Serre duality gives b = h4−i (OV (D b − X)) hi (OV (−D)) (C.1) hi (OX (−D)) = h3−i (OX (D)) (C.2) on V , and on X. b is effective. We then have h0 (OV (−D)) b = 0, and so h4 (OV (D b− Let us now assume D Pb b i are effective, we have h0 (OV (D b− X)) = 0. Using also that X = Di , where the D b = 0. In addition because D is effective we have X)) = 0, and so h4 (OV (−D)) h0 (OX (−D)) = 0 and so h3 (OX (D)) = 0. Because V is a toric fourfold, we have the relation h• (V, OV ) = (1, 0, 0, 0, 0) . (C.3) Using Serre duality (as in [27]) in the long exact sequence in cohomology induced by the Koszul sequence 0 → OV (−X) → OV → OX → 0 , (C.4) one immediately finds that the Hodge numbers h• (X, OX ) obey h• (X, OX ) = (1, 0, 0, 1) . (C.5) Similarly, we can establish: Proposition 24 For a space S, define h̃i (S) := hi (S) for i > 0, and h̃0 (S) := h0 (S) − 1. Then the following relations hold, for 0 ≤ i ≤ 3: b O b ) = hi+1 (OV (−D)) b = h3−i (OV (D b − X)) . h̃i (D, D 61 (C.6) b ⊂ V , which reads We consider the Koszul sequence for D b → OV → O b → 0 . 0 → OV (−D) D (C.7) This induces the long exact sequence in cohomology 0 b H 0 (OV (−D)) H 0 (OV ) H 0 (ODb ) b H 1 (OV (−D)) H 1 (OV ) H 1 (ODb ) b H 2 (OV (−D)) H 2 (OV ) H 2 (ODb ) b H 3 (OV (−D)) H 3 (OV ) H 3 (ODb ) b H 4 (OV (−D)) H 4 (OV ) 0 Applying (C.3) leads to the first equality in (C.6). The second equality then follows from (C.1).  b= Corollary 25 If D 6 X we have b − X)) = 0 , b O b ) = h0 (OV (D h3 (D, D (C.8) b O b ) = h0 (OV (D b − X)) = 1 . h3 (D, D (C.9) b = X we have while if D In close parallel to Proposition 24, we can show the following: Proposition 26 The following relations hold, for 0 ≤ i ≤ 2: h̃i (D, OD ) = h̃2−i (X, OX (D)) . (C.10) We consider the Koszul sequence for D ⊂ X, which reads 0 → OX (−D) → OX → OD → 0 . (C.11) Applying (C.5) and (C.2) to the long exact sequence in cohomology induced by (C.11) yields (C.10).  In particular, we have Corollary 27 h2 (D, OD ) = h0 (X, OX (D)) − 1 . 62 (C.12) C.2 Relating h2 to toric data We now state the condition that defines the special case treated in this appendix. b be a square-free divisor in V corresponding to a collection of Definition 28 Let D b face-limited if the uI are all contained in a lattice points {uI } ⊂ ∆◦ . We call D b ∩ X with D b single 2-face f of ∆◦ . We call a divisor D in X face-limited if D := D face-limited. Notice that RD contains all the layers of the ravioli complex R over f . Thus, D corresponds not just to RD ⊂ R, but also to a subcomplex TD ⊂ T . We now examine the simplicial complex TD associated to D. b be a face-limited divisor in V , let D = D b ∩ X, and let TD be the Lemma 29 Let D associated simplicial complex. Then b O b ) = hi (TD ) . hi (D, D (C.13) This follows from the spectral sequence associated to the generalized Mayer-Vietoris sequence on V given in Proposition 19 of Appendix A.  b be a face-limited divisor in V , and let D = D b ∩ X be the correCorollary 30 Let D sponding face-limited divisor in X. Then b O b) = 0 . h2 (D, D (C.14) This follows from (C.13), because h2 (TD ) = 0 for a divisor on a single 2-face.  We can now relate sections on X to sections on V : b be a face-limited divisor in V , and let D = D b ∩ X be the correLemma 31 Let D 0 0 b sponding face-limited divisor in X. Then h (X, OX (D)) = h (V, OV (D)). b which reads We tensor the Koszul sequence from V to X with O(D), b − X) → OV (D) b → OX (D) → 0 . 0 → OV (D b on V is written as D b = P ai D b i , where the D b i are A general square-free divisor D Pb b i are effective, we have the toric divisors and ai ∈ {0, 1}. Because X = Di and the D b − X)) = 0. Therefore, to show that h0 (X, OX (D)) = h0 (V, OV (D)), b that h0 (OV (D b − X)) = 0. By Serre duality, equation (C.1), we can we need to show that h1 (OV (D 3 b = 0. Consider the Koszul sequence from V to D, b equivalently show that h (OV (−D)) b → OV → O b → 0 . 0 → OV (−D) D 63 (C.15) Using (C.3) in the long exact sequence induced by (C.15), we find that b O b ) = 0 ⇒ h3 (OV (−D)) b = 0. h2 (D, D (C.16) The lemma follows upon using (C.14).  We have thus proved: b be a face-limited divisor in V , and let D = D b ∩ X be the correCorollary 32 Let D sponding face-limited divisor in X. Then b − 1. h2 (D, OD ) = h0 (V, OV (D)) (C.17) Computation of h2 for a face-limited divisor C.3 We are now equipped to calculate h2 (D, OD ) for an arbitrary face-limited divisor P D = ai Di . We first establish how to compute h0 (V, OV (D)). i b i be the Proposition 33 Let V be a toric variety corresponding to a fan Σ, let D P b b := toric divisors on V , and for ai ∈ Z, define D i ai Di . Define the polyhedra b = |P b ∩ M |. PDb = {m ∈ MR |hm, ui i ≥ −ai for all ui ∈ Σ(1)}. Then h0 (V, OV (D)) D The proof is given in Proposition 4.3.2 of [27]. b be a face-limited divisor in V , let D = D b ∩ X be the corresponding Lemma 34 Let D face-limited divisor in X, and let v be the vertices, e the complete edges, and f the complete faces included in TD .11 Then X X X b =1+ g(f ) . (C.18) g(e) + g(v) + h0 (V, OV (D)) e v f b are counted by lattice points m such that hm, ui i ≥ −ai , Proof. Sections of OV (D) 0 b can be computed by counting suitable lattice points. We consider and so h (V, OV (D)) Pb b a divisor D = Di specified by a set of points ui ∈ f , i ∈ 1, . . . , N , where f is a i 2-face. We label the points not in the set {ui } as ua . First, note that for any effective b is by assumption also divisor the origin m = ~0 corresponds to a global section. As D b are counted by lattice points m such that square-free, the additional sections of OV (D) hm, ui i ≥ −1, and hm, ua i ≥ 0. We will count these sections by including points in the set {ui } one by one, and checking how the number of sections changes. In other j+1 j b = PD b i , where j ≤ N − 1, and let D b0 = P D bi = D b +D b j+1 . words, let D i=1 i=1 11 Because TD is by assumption contained in a single 2-face f , the last sum in (C.18) will only have one term, but we find it useful to write (C.18) in a form that anticipates our result for a completely general square-free divisor. 64 b j+1 corresponds to a lattice point uj+1 . Then h0 (V, OV (D b 0 )) equals The divisor D the number of lattice points m such that hm, ui i ≥ −1, (C.19) hm, uj+1 i ≥ −1, (C.20) hm, ua i ≥ 0 . (C.21) b equals the number of lattice points m such that On the other hand, h0 (V, OV (D)) hm, ui i ≥ −1 , (C.22) hm, ua i ≥ 0 . (C.23) b 0 )) − h0 (V, OV (D)) b is the number of lattice points m such that Thus, h0 (V, OV (D hm, ui i ≥ −1, (C.24) hm, uj+1 i = −1, (C.25) hm, ua i ≥ 0 . (C.26) The points m obeying (C.25) are by definition the lattice points in the face u◦j+1 of ∆, dual to the point uj+1 . We need to count points in u◦j+1 that also satisfy (C.24) and (C.26). There are three types of ui : vertices, points interior to edges, and points interior to f . We will include them in the set {ui } in that particular order. First consider the case where uj+1 is a vertex of f . Then we need to solve hm, uj+1 i = −1 (C.27) hm, ua i ≥ 0 . (C.28) Equation (C.25) defines the facet u◦j+1 ∈ ∆(3). However, any point on the boundary of u◦j+1 has a dual in ∆◦ , defined by hm , ·i = −1, that violates (C.28). The only points that do not violate (C.28) are those in the interior of u◦j+1 , as they are dual only to b 0 )) = 1 + g(uj+1 ). uj+1 itself, and therefore h0 (V, OV (D Next let uj+1 correspond to a point internal to an edge e ∈ ∆◦ (1). The condition (C.28) is violated unless the entire edge, including the vertices bounding it, is included, since hm, uj+1 i = −1 implies that hm, uα i = −1 for any uα ⊂ e. Therefore, a divisor D corresponding to a complete edge e with vertices ua and ub has b = 1 + g(vu ) + g(vu ) + g(e) . h0 (V, OV (D)) (C.29) In a similar manner we find that including points uα internal to f in the set {ui } can only contribute to h2 if every point in the face f is included in {ui }.  65 From Corollary 32 and Lemma 34 we deduce: Corollary 35 Let D be a face-limited square-free divisor in X. Then X X X h2 (D, OD ) = g(v ◦ ) + g(e◦ ) + g(f ◦ ) . v e (C.30) f References [1] S. Kachru, R. Kallosh, A. D. Linde and S. P. Trivedi, De Sitter vacua in string theory, Phys. Rev. D68 (2003) 046005, [hep-th/0301240]. [2] V. Balasubramanian, P. Berglund, J. P. Conlon and F. Quevedo, Systematics of moduli stabilisation in Calabi-Yau flux compactifications, JHEP 03 (2005) 007, [hep-th/0502058]. [3] J. P. Conlon, F. Quevedo and K. Suruliz, Large-volume flux compactifications: Moduli spectrum and D3/D7 soft supersymmetry breaking, JHEP 08 (2005) 007, [hep-th/0505076]. [4] E. Witten, Nonperturbative superpotentials in string theory, Nucl.Phys. B474 (1996) 343–360, [hep-th/9604030]. [5] D. Eisenbud, M. Mustaţǎ and M. Stillman, Cohomology on toric varieties and local cohomology with monomial supports, J. Symbolic Comput. 29 (2000) 583–600. [6] D. R. Grayson and M. E. Stillman, “Macaulay2, a software system for research in algebraic geometry.” Available at https://faculty.math.illinois.edu/Macaulay2/. [7] G. G. Smith, NormalToricVarieties: a Macaulay2 Package for Computing with Toric Varieties, distributed with Macaulay2 . [8] R. Blumenhagen, B. Jurke, T. Rahn and H. Roschy, Cohomology of Line Bundles: A Computational Algorithm, J.Math.Phys. 51 (2010) 103525, [1003.5217]. [9] R. Blumenhagen, B. Jurke, T. Rahn and H. Roschy, Cohomology of Line Bundles: Applications, J.Math.Phys. 53 (2012) 012302, [1010.3717]. [10] M. Kreuzer and H. Skarke, Calabi-Yau Data, http: // hep. itp. tuwien. ac. at/ ~ kreuzer/ CY/ . [11] M. Stillman et al., StringTorics: a Macaulay2 Package for Computing with Toric Varieties in String Theory, to appear . [12] R. Donagi and M. Wijnholt, MSW Instantons, JHEP 06 (2013) 050, [1005.5391]. [13] R. Donagi, S. Katz and M. Wijnholt, Weak Coupling, Degeneration and Log Calabi-Yau Spaces, 1212.0553. [14] A. Clingher, R. Donagi and M. Wijnholt, The Sen Limit, Adv. Theor. Math. Phys. 18 (2014) 613–658, [1212.4505]. 66 [15] A. Braun et al., The Hodge Numbers of Divisors of Calabi-Yau Fourfold Hypersurfaces, work in progress . [16] V. V. Batyrev, Dual Polyhedra and Mirror Symmetry for Calabi-Yau Hypersurfaces in Toric Varieties, J. Alg. Geom (1996) 493–535, [alg-geom/9310003]. [17] M. Kreuzer and H. Skarke, On the classification of reflexive polyhedra, Commun. Math. Phys. 185 (1997) 495–508, [hep-th/9512204]. [18] M. Kreuzer and H. Skarke, Classification of reflexive polyhedra in three-dimensions, Adv. Theor. Math. Phys. 2 (1998) 847–864, [hep-th/9805190]. [19] M. Kreuzer and H. Skarke, Complete Classification of Reflexive Polyhedra in Four Dimensions, Adv.Theor.Math.Phys. 4 (2002) 1209–1230, [hep-th/0002240]. [20] A. Klemm, B. Lian, S. S. Roan and S.-T. Yau, Calabi-Yau fourfolds for M theory and F theory compactifications, Nucl. Phys. B518 (1998) 515–574, [hep-th/9701023]. [21] A. P. Braun, Tops as building blocks for G2 manifolds, JHEP 10 (2017) 083, [1602.03521]. [22] S. Greiner and T. W. Grimm, Three-form periods on Calabi-Yau fourfolds: Toric hypersurfaces and F-theory applications, JHEP 05 (2017) 151, [1702.03217]. [23] A. Hatcher, Algebraic topology. Cambridge University Press, Cambridge, 2002. [24] M. Bianchi, A. Collinucci and L. Martucci, Magnetized E3-brane instantons in F-theory, JHEP 12 (2011) 045, [1107.3732]. [25] T. Maszczyk, Distributive lattices and cohomology, ArXiv e-prints (Nov., 2008) , [0811.3997]. [26] V. I. Danilov and A. G. Khovanskii, Newton Polyhedra and an Algorithm for Computing Hodge–Deligne Numbers, Math. USSR. Izv. 29 (1987) 279–298. [27] D. Cox, J. Little and H. Schenck, Toric Varieties. Graduate studies in mathematics. American Mathematical Soc., 2011. [28] V. I. Danilov, The Geometry of Toric Varieties, Russ.Math.Surv. 33:2 (1978) 97. [29] M. Kreuzer, Toric geometry and Calabi-Yau compactifications, Ukr. J. Phys. 55 (2010) 613–625, [hep-th/0612307]. [30] W. Fulton, Introduction to toric varieties. Princeton University Press, Princeton, 1993. [31] C. Voisin and L. Schneps, Hodge Theory and Complex Algebraic Geometry I:. No. v. 1 in Cambridge Studies in Advanced Mathematics. Cambridge University Press, 2002. [32] C. Voisin and L. Schneps, Hodge Theory and Complex Algebraic Geometry II:. No. v. 2 in Cambridge Studies in Advanced Mathematics. Cambridge University Press, 2003. [33] J. De Loera, J. Rambau and F. Santos, Triangulations. Springer, Heidelberg Dordrecht London New York, 2010. 67
0math.AC
1 M-RWTL: Learning Signal-Matched Rational Wavelet Transform in Lifting Framework arXiv:1710.10394v1 [cs.SY] 28 Oct 2017 Naushad Ansari, Student Member, IEEE, Anubha Gupta, Senior Member, IEEE Abstract—Transform learning is being extensively applied in several applications because of its ability to adapt to a class of signals of interest. Often, a transform is learned using a large amount of training data, while only limited data may be available in many applications. Motivated with this, we propose wavelet transform learning in the lifting framework for a given signal. Significant contributions of this work are: 1) the existing theory of lifting framework of the dyadic wavelet is extended to more generic rational wavelet design, where dyadic is a special case and 2) the proposed work allows to learn rational wavelet transform from a given signal and does not require large training data. Since it is a signal-matched design, the proposed methodology is called Signal-Matched Rational Wavelet Transform Learning in the Lifting Framework (M-RWTL). The proposed M-RWTL method inherits all the advantages of lifting, i.e., the learned rational wavelet transform is always invertible, method is modular, and the corresponding M-RWTL system can also incorporate nonlinear filters, if required. This may enhance the use of RWT in applications which is so far restricted. M-RWTL is observed to perform better compared to standard wavelet transforms in the applications of compressed sensing based signal reconstruction. Index Terms—Transform Learning, Rational Wavelet, Lifting Framework, Signal-Matched Wavelet I. I NTRODUCTION T RANSFORM learning (TL) is currently an active research area and is being explored in several applications including image/video denoising, compressed sensing of magnetic resonance images (MRI), etc. [1]–[3]. Transform learning has the advantage that it adapts to a class of signals of interest and is often observed to perform better than existing sparsifying transforms such as total variation (TV), discrete cosine transform (DCT), and discrete wavelet transform (DWT) in the above said applications [1]–[3]. In general, transform learning is posed as an optimization problem satisfying some constraints that are specific to applications. Transform domain sparsity of signals is a widely used constraint along with some additional constraints on the transform to be learned, say, the minimization of Frobenius norm or the log-determinant of transform [1]. The requirement of joint learning of both the transform basis and the transform domain signal under the constraints renders the optimization problem to be non-convex with no closed form solution. Thus, in general, TL problems are solved using greedy algorithms [1]. A large number of variables (learned transform as well as transform coefficients) along with the greedy-based solution makes TL computationally expensive. Naushad Ansari and Anubha Gupta are with SBILab, Deptt. of ECE, IIITDelhi, India (emails: [email protected] and [email protected]). Naushad Ansari is financially supported by CSIR (Council of Scientific and Industrial Research), Govt. of India, for this research work. Recently, deep learning (DL) and convolutional neural network (CNN) based approaches are gaining momentum and are being used in several applications [4], [5]. In general, these methods (TL, DL, CNN) require a large amount of training data for learning. Hence, these methods may run with challenges in applications where only single snapshots of short-duration signals such as speech, music, or electrocardiogram (ECG) signals are available because a large amount of training data, required for learning transform, is absent. Hence, one uses existing transforms that are signal independent. This motivates us to look for a strategy to learn transform in such applications. Among existing transforms, although Fourier transform and DCT find use in many applications, DWT provides an efficient representation for a variety of multi-dimensional signals [6]. This efficient signal representation stems from the fact that DWT tends to capture signal information into a few significant coefficients. Owing to this advantage, wavelets have been applied successfully in many applications including compression, denoising, biomedical imaging, texture analysis, pattern recognition, etc. [7]–[11]. In addition, wavelet analysis provides an option to choose among existing basis or design new basis, thus motivates one to learn basis from a given signal of interest and/or in a particular application that may perform better than the fixed basis. Since the translates of the associated wavelet filters form the basis in l2 -space (functional space for square summable discrete-time sequences), wavelet transform learning implies learning wavelet filter coefficients. This reduces the number of parameters required to be learned with wavelet learning compared to the traditional transform learning. Also, the requirement of learning fewer coefficients allows one to learn basis from a short single snapshot of signal or from the small training data. This motivates us to explore wavelet transform learning for small data in this work. We also show that closed form solution exists for learning the wavelet transform leading to fast implementation of the proposed method without the need to look for any greedy solution. This is to note that most applications of DWT use dyadic wavelet transform, where the wavelet transform is implemented via a 2-channel filterbank with downsampling by two. In the frequency domain, this process is equivalent to decomposition of signal spectrum into two uniform frequency bands. M -band wavelet transform introduces more flexibility in analysis and decomposes a signal into M uniform frequency bands. However, some applications, such as speech or audio signal processing, may require non-uniform frequency band decomposition [12]–[14]. 2 Rational wavelet transform (RWT) can prove helpful in applications requiring non-uniform partitioning of the signal spectrum. Decimation factors of the corresponding rational filterbank (RFB) may be different in each subband and are rational numbers [15]. RWT has been used in applications. For example, RWT is applied in context-independent phonetic classification in [13]. In [16], it is used for synthesizing 10m multispectral image by merging 10m SPOT panchromatic image and a 30m Landsat Thematic Mapper multispectral image. Wavelet shrinkage based denoising is presented in [17] using signal independent rational filterbank designed by [18]. RWT is used in extracting features from images in [19] and in click frauds detection in [20]. Rational orthogonal wavelet transform is used to design optimum receiver based broadband Doppler compensation structure in [21] and for broadband passive sonar detection in [22]. A number of RWT designs have been proposed in the literature including FIR (Finite Impulse Response) orthonormal rational filterbank design [12], overcomplete FIR RWT designs [23], [24], IIR (Infinite Impulse Response) rational filterbank design [25], biorthogonal FIR rational filterbank design [26], [27], frequency response masking technique based design of rational FIR filterbank [14], and complex rational filterbank design [28]. However, so far RWT designed and used in applications are meant to meet certain fixed requirements in the frequency domain or time-domain instead of learning the transform from a given signal of interest. For example, all the above designs are signal independent and hence, the concept of transform learning from a given signal has not been used so far in learning rational wavelets. Lifting has been shown to be a simple yet powerful tool for custom design/learning of wavelet [29]. Apart from custom wavelet design/wavelet learning, lifting provides several other advantages such as a) wavelets can be designed in the spatial domain, b) designed wavelet transform is always invertible, c) the design is modular, and d) the design is DSP (Digital Signal Processing) hardware friendly from the implementation viewpoint [29]. However, the framework developed so far is used only for the custom design/learning of dyadic (or Mband) wavelets [30]–[33] and has not been used to learn the rational wavelet transform to the best of our knowledge. Moreover, the existing architecture of lifting framework cannot be extended directly to rational filterbank structure because of different sample/signal rates in subband branches. From the above discussion, we note that the lifting framework can help in learning custom-design rational wavelets from given signals in a simple modular fashion that will also be easy to implement in the hardware. Also, this may lead to the enhanced use of rational wavelet transforms in applications, similar to dyadic wavelets, which is so far restricted. Motivated by the success of transform learning in applications, the flexibility of rational wavelet transform with respect to non-uniform signal spectral splitting, and the advantages of lifting in learning custom design wavelets, we propose to learn rational wavelet transform from a given signal using the lifting framework. The proposed methodology is called, “Learning Signal-Matched Rational Wavelet Transform in the Lifting Framework (M-RWTL)”. Following are the salient contributions/significance of this work: 1) The theory of lifting is extended from dyadic wavelets to rational wavelets, where dyadic wavelet is a special case. The concept of rate converters is introduced in predict and update stages to handle variable subband sample rates. 2) Theory is proposed to learn rational wavelet transform from a given signal, where rational wavelets with any decimation ratio can be designed. 3) FIR analysis and synthesis filters are learned that can be easily implemented in hardware. 4) Closed form solution is presented for learning rational wavelet and thus, no greedy solution is required making M-RWTL computationally efficient. 5) The proposed M-RWTL can be learned from a short snapshot of a single signal and hence, extends the use of transform learning from the requirement of large training data to small data snapshots. 6) The utility of M-RWTL is demonstrated in the application of compressed sensing-based reconstruction and is observed to perform better than the existing dyadic wavelet transforms. This paper is organized as follows. In section II, we briefly describe the theory of lifting corresponding to the dyadic wavelet system and the theory of rational wavelet. In Section III, we present the proposed theory of learning M-RWTL and some learned examples. Section IV presents simulation results in the application of compressive sensing based signal reconstruction. Some conclusions are presented in section VI. Notations: Scalars are represented as lower case letters and, vectors and matrices are represented as bold lower case and bold upper case letters, respectively. II. B RIEF BACKGROUND In this section, we provide brief reviews on the theory of dyadic wavelet in lifting structure, rational wavelet system, and polyphase decomposition theory required for the explanation of the proposed work. A. Theory of Lifting in Dyadic Wavelet A general dyadic wavelet structure is shown in Fig. 1, where Gl (z) and Gh (z) are analysis lowpass and highpass filters and Fl (z) and Fh (z) are synthesis lowpass and highpass filters, respectively. Lifting is a technique for either factoring existing wavelet filters into a finite sequence of smaller filtering steps or constructing new customized wavelets from existing wavelets. In general, lifting structure has three stages: Split, Predict, and Update (Fig. 2). Split: This stage splits an input signal x[n] into two disjoint sets of even xe [n] and odd indexed xo [n] samples (Fig. 2). The original input signal x[n] is recovered fully by interlacing this even and odd indexed sample stream. Wavelet transform associated with the corresponding filterbank is also called as the Lazy Wavelet transform [29]. A Lazy Wavelet transform can be obtained from the standard dyadic structure shown in Fig. 1 by choosing analysis filters as Gl (z) = Z{gl [n]} = 1, Gh (z) = Z{gh [n]} = z and synthesis filters as Fl (z) = Z{fl [n]} = 1, Fh (z) = Z{fh [n]} = z −1 . 3 Predict: In this stage, one of the above two disjoint sets of samples is predicted from the other set of samples. For example, in Fig.2(a), odd samples are predicted from the neighboring even samples using the predictor filter P ≡ T (z). Predicted samples are subtracted from the actual samples to calculate the prediction error. This step is equivalent to applying a highpass filter on the input signal. This stage modifies the analysis highpass and synthesis lowpass filters, without altering the other two filters, according to the following relations: Gnew (z) = Gh (z) − Gl (z)T (z 2 ). h (1) Flnew (z) = Fl (z) + Fh (z)T (z 2 ). (2) input signals into non-uniform frequency bands. In general, any ith analysis branch of a rational structure is as shown in Fig.3, where Gi (z) denotes the analysis filter, qi denotes the upsampling factor, and pi denotes the downsampling factor. For example, if pi = 3 and qi = 2, the downsampling ratio in this branch is equal to 3/2. At the synthesis end, the order of downsampler and upsampler are reversed. qi Fig. 3: ith branch of rationally decimated analysis filterbank x[n] Update: In this stage, the other disjoint sample set is updated using the predicted signal obtained in the previous step via filtering through U ≡ S(z) as shown in Fig.2. In general, this step is equivalent to applying a lowpass filter to the signal and modifies/updates the analysis lowpass filter and synthesis highpass filter according to the following relations: Gnew (z) = Gl (z) + Gh (z)S(z 2 ). l (3) Fhnew (z) (4) 2 = Fh (z) − Fl (z)S(z ). One of the major advantages of lifting structure is that each stage (predict or update) is invertible. Hence, perfect reconstruction (PR) is guaranteed after every step. Also, the lifting structure is modular and non-linear filters can be incorporated with ease. x[n] x[n] g l [ n] f l [ n] f0[n] f0[n] x1[n] x[n] + g h [ n] G0 ( z ) M M F0 ( z ) G1 ( z ) M M F1 ( z ) G2 ( z ) M M F2 ( z ) GM 1 ( z ) M M FM 1 ( z ) xˆ[n] f h [ n] x2 [n] h1[n] f1[n] Fig. 1: Two Channel Biorthogonal Wavelet System Fig. 4: M-band wavelet structure a[n] q1 Gl ( z ) q2 Gh ( z ) M M x[n] x[n] + a[n] ++ Odd/ Even Split P≡T(z) xo [n] + + U≡S(z) d [ n] (a) a[n] + + U≡S(z) d [ n] P≡T(z) + + + Interleave/ combine samples from the two input streams Fl ( z ) q1 Fh ( z ) q2 𝑥[𝑛] 𝑥 [𝑛] M M d [ n] An M-channel rational filterbank is said to be critically sampled if the following relation is satisfied M −1 X i=0 xe [n] xˆ[ n ] Fig. 5: General 2-band rational wavelet structure + + pi Gi ( z ) xˆ[n] (b) Fig. 2: Stages of Lifting: Split, Predict and Update qi = 1. pi (5) This is to note that, throughout this paper, we consider all pi ’s to be equal, especially, pi = M for all i. In general, a given M -band wavelet system, as shown in Fig.4, can be converted into an equivalent 2-band rational wavelet structure of Fig.5, having downsampling ratios M q1 and M in the two branches. Note that q and q are relatively 1 2 q2 prime with each other and with M . Also, q1 + q2 = M for a critically sampled rational wavelet system. For example, the analysis filters Gj (z) and synthesis filters Fj (z) for j = 0, 1, 2, ..., q1 − 1 as shown in Fig.4 can be combined using the following equations: Gl (z) = qX 1 −1 z −iM Gi (z q1 ), (6) i=0 B. Rational Wavelet and Equivalent M-band structure M -band wavelet system has integral downsampling ratio M as shown in Fig. 4, whereas rational wavelet system has rational down-sampling ratios that allows decomposition of Fl (z) = qX 1 −1 z iM Fi (z q1 ), (7) i=0 where Gl (z) and Fl (z) are the corresponding analysis and synthesis lowpass filters of the equivalent 2-band rational 4 wavelet structure of Fig.5. Similarly, rest of the filters of both sides can be combined using the following equations: Gh (z) = qX 2 −1 z −iM Gi+q1 (z q2 ), Fh (z) = z iM Fi+q1 (z q2 ), (9) i=0 C. Polyphase Representation and Perfect Reconstruction The polyphase representation of filters is very helpful in filterbank analysis and design [34]. Consider the M-band critically sampled filterbank shown in Fig.4. Analysis filter Gi (z) can be written using type-1 polyphase representation as: M −1 X z j Ei,j (z M ), (10) j=0 where Ei,j (z) = gi (j) + gi (M + j)z M + gi (2M + j)z 2M + ... and Gi (z) = Z{gi [n]}. Synthesis filter Fi (z) can be written using type-2 polyphase representation as: Fi (z) = x̂[n] = cx[n − M − n0 ], M −1 X z −j Ri,j (z M ), (11) i=0 III. P ROPOSED M-RWTL L EARNING M ETHOD FOR S IGNAL -M ATCHED R ATIONAL WAVELET In this section, we present the proposed M-RWTL method of learning signal-matched 2-channel rational wavelet system using the lifting framework. First, we propose the extension of 2-channel dyadic Lazy wavelet transform to M -band Lazy wavelet and find its equivalent 2-channel rational Lazy filterbank. Both these structures will be used in the proposed work. A. M-band and Rational Lazy Wavelet System As explained earlier, a 2-band Lazy wavelet system divides an input signal into two disjoint signals. Similarly, on the analysis side, an M -band Lazy wavelet system divides an input signal x[n] into M disjoint sets of data samples, given by vi [n], i = 0, 1, 2, ..., M − 1, where vi [n] = x[M n + i]. At the synthesis end, these M disjoint sample sets are combined or interlaced to reconstruct the signal at the output. An Mband Lazy wavelet can be designed with the following choice of analysis and synthesis filters in Fig.4: Gi (z) =z i Fi (z) =z where Ri,j (z) = fi (j)+fi (M +j)z −M +fi (2M +j)z −2M +... and Fi (z) = Z{fi [n]}. The M-band filterbank of Fig.4 can be equivalently drawn using polyphase matrices as shown in Fig.6, where   E0,0 E0,1 ··· E0,M −1  E1,0 E1,1 ··· E1,M −1    E(z) =   , (12) .. .. .. . .   . . . . EM −1,0 EM −1,1 · · · EM −1,M −1 and R0,0 R1,0 .. .   R(z) =   RM −1,0 R0,1 R1,1 .. . ··· ··· .. . R0,M −1 R1,M −1 .. . RM −1,1 ··· RM −1,M −1    .  (13) x[n] xˆ[ n ] M M z 1 z M z M M E( z) R( z ) z 1 M z 1 z M M Fig. 6: M-band wavelet structure with PR i = 0, 1, 2, ..., M − 1. (16) (17) In order to obtain the corresponding 2-band rational Lazy wavelet system with dilation factor M q1 (Fig.5) from the M band Lazy wavelet, we use (6) and (7) to obtain lowpass analysis and synthesis filters as: Gl (z) = Fl (z) = qX 1 −1 i=0 qX 1 −1 z −iM iq1 z = z iM z −iq1 = qX 1 −1 i=0 qX 1 −1 z −iq2 , z iq2 . (18) i=0 Similarly, we use (8) and (9) to obtain the corresponding highpass analysis and synthesis filters of rational Lazy wavelet as: Gh (z) = The below relation of polyphase matrices i = 0, 1, 2, ..., M − 1, −i i=0  (15) where c is a constant and n0 is a constant delay. where Gh (z) and Fh (z) are the corresponding higpass filters of analysis and synthesis ends of Fig.5. Gi (z) = (14) yields the condition of perfect reconstruction (PR) stated as (8) i=0 qX 2 −1 R(z)E(z) = cz −n0 I, Fh (z) = qX 2 −1 i=0 qX 2 −1 i=0 z −iM z (i+q1 )q2 = z q1 q2 qX 2 −1 z −iq1 , i=0 qX 2 −1 z iM z −(i+q1 )q2 = z −q1 q2 z iq1 . (19) i=0 The above filters form the rational Lazy wavelet system equivalent of M -band Lazy wavelet transform. To learn signalmatched rational wavelet system, we start with the rational Lazy wavelet. This provides us initial filters Gl (z), Gh (z), Fl (z), and Gh (z). These filters are updated according to signal characteristics to obtain the signal-matched rational wavelet system. Analysis highpass and synthesis lowpass filters are updated in the predict stage, whereas analysis lowpass and 5 q1 1  z  q2  ...  z ( q1 1) q2 M 0th block 1st block kth block (k+1)th block a[n] ≡ x[0], ... , x [q1-1] x[M], … , x[M+q1-1] … x[(k-1)M], ... , x[(k-1)M+q1-1] x[kM], ... , x[kM+q1-1] q2 z q1q2 [1  z  q1  ...  z ( q2 1) q1 ] M d[n] ≡ x[q1], … , x[M-1] x[M+q1], … , x[2M-1] … x[(k-1)M+q1], … , x[kM-1] x[kM+q1], … , x[(k+1)M-1] … Fig. 7: Analysis side of rational Lazy wavelet; Each block consists of M samples of input signal x[n] divided into a[n] and d[n]. synthesis highpass filters are updated in the update stage. These stages are described in the following subsections. B. Predict Stage As discussed in section II-A, a 2-band Lazy wavelet system divides the input signal into two disjoint sample sets, wherein one set is required to be predicted using the other set of samples. In the conventional 2-band lifting framework with integer downsampling ratio of M=2 in both the branches (refer to Fig. 1 and 2), the output sample rate of xe [n] and xo [n] is equal. However, in a 2-band rational wavelet system, the output sample rate of two branches is unequal. Hence, the predict polynomial branch of a conventional 2-band dyadic design cannot be used. For example, from Fig.5, we note that in a 2-band rational wavelet system, higher (lower) rate branch samples used in prediction should be downsampled (upsampled) by a factor of kp to equal the rate to the lower (higher) predicted branch samples, defined as: q1 (20) kp = , q2 where q1 /M is the rate of ‘predicting’ branch samples and q2 /M is the rate of ‘to be predicted’ branch samples. We propose to predict the lower branch samples with the help of upper branch samples. Here, input signal x[n] is divided into two disjoint sets, x[kM +i], i = 0, 1, ..., q1 −1 and x[kM + j], j = q1 , q1 + 1, ..., M − 1. We label these outputs as a and d, respectively (Fig.7). Here, k = 0, 1, ..., L − 1, where N = LM is the length of input signal x[n] and, without loss of generality, is assumed to be a multiple of M . Thus, a given input signal is divided into L blocks of M size each at the subband output of 2-channel rational Lazy wavelet system. Here, first q1 samples of every block move to the upper branch as a block of a and next q2 samples (such that q1 + q2 = M ) move to the lower branch as a block of d. Or in other words, the rate of upper branch output is q1 samples per block and the rate of lower branch output is q2 samples per block. Fig.7 shows these blocks of outputs a and d explicitly. This motivates us to introduce the concept of rate converter that equals the output sample rate of the upper predicting branch to that of the lower predicted branch to enable predict branch design. In other words, the output of upper branch a[n] is upsampled by q2 and downsampled by q1 to match the rate of lower branch output d[n]. It should be noted that the above downsampler and upsampler cannot be connected consecutively. Since an upsampler introduces spectral images in the frequency domain, it is generally preceded by a filter, also known as interpolator or anti-imaging filter. On the other hand, a downsampler stretches the frequency spectrum of the signal, that’s why it is followed by a filter known as anti-aliasing filter. Both these conditions imply that if a rate converter is required with downsampling factor of kp = qq12 , a polynomial Rp (z) should be incorporated which acts as the anti-imaging filter for the upsampler and at the same time as the anti-aliasing filter for the downsampler. The structure of this polynomial is Rp (z) = 1 + z −1 + ... + z −(q2 −1) . The placement of this polynomial is shown in Fig. 8. We call the complete branch with rate converter kp and polynomial Rp (z) as the predict rate converter. Definition 1. Predict Rate Converter: It is a combination of the polynomial Rp (z) = 1 + z −1 + ... + z −(q2 −1) preceded by a q2 -fold upsampler and followed by a q1 -fold downsampler as shown in Fig.8. q1 Gl ( z ) a[n] M q2 Rp ( z ) 𝑥[𝑛] q2 Gh ( z ) M d [ n] Predict rate converter q1 a4 [n]  …   d new[n] Fig. 8: Illustration of Predict Rate Converter The predict rate converter with this choice of Rp (z) will lead to an appropriate repetition or drop of samples of a[n] such that the total number of samples in a4 [n] and d[n] in any k th block contains an equal number of samples. Next, we present three subsections: 1) on the structure of predict stage filter T (z), 2) how to learn T (z) from a given signal, and 3) how to update all the filters of the corresponding 2-channel rational filterbank using the learned T (z). 1) Structure of predict stage filter T (z): Once the outputs of two analysis filterbank branches are equal, lower branch samples are predicted from the upper branch samples with the help of predict stage filter T (z). This filter is introduced after the polynomial Rp (z) as shown in Fig.9. For good prediction, the current sample of input signal x contained in d[n] should be predicted from its past and future samples. The k th block of a[n] has preceding samples and (k + 1)th block of a[n] has future samples of the k th block of d[n] as is evident from Fig.7. Thus, the structure of the predict polynomial T (z) should be chosen appropriately. This is presented with a theorem for 2-tap T (z) as below. 6 q1 Gl ( z ) q2 a1[n] a2 [n] 𝑥[𝑛] a3[n] Rp ( z ) T ( z) Gh ( z ) M d [ n] ak4 [m] P( z ) Composite Predict branch q1 a4 [n] q2 a4 [n] as: a[n] M    d new[n] where m = 0, 1, ..., q2 − 1. The block size or the rate of a4 [n] is same as that of d[n] and hence, it predicts d[n] providing the k th block prediction error given by: Fig. 9: Predict Rate Converter Theorem 1. A 2-tap predict stage filter T (z) = z q2 q1 (t[0]z −q2 + t[1]) ensures that every sample in the k th block of d[n] branch is predicted from the k th and/or the (k + 1)th block samples of a[n]. Proof. As discussed in section III-B, on passing the input signal x[n] through the Lazy filterbank of analysis side, we obtain approximate and detail coefficients (a and d, respectively) in the form of blocks as shown in Fig.7. The k th block of approximate and detail signals are given by: ak [m] =a[(k − 1)q1 + m] = x[(k − 1)M + m], m = 0, 1, ..., q1 − 1, k d [m] =d[(k − 1)q2 + m] = x[(k − 1)M + q1 + m], m = 0, 1, ...q2 − 1.     q1 m =t[0]a (k − 1)q1 + + (q1 − 1) q2     q1 m + t[1]a (k − 1)q1 + + q1 , q  2      q1 m q1 m + t[1]a kq1 + , =t[0]a kq1 − 1 + q2 q2 (24) (21) For prediction, first the upper branch signal a[n] is passed through q2 -fold upsampler and polynomial Rp (z). This leads to signal a2 [n], as shown in Fig.9, that contains every element of a[n] repeated q2 number of times. Mathematically, the k th block of signal a2 [n] is given by:    m , m = 0, 1, ..., q1 q2 − 1, ak2 [m] = a (k − 1)q1 + q2 (22) where b.c denotes the floor function. Predict filter T (z) is defined as a product of two polynomials z q2 q1 and t[0]z −q2 + t[1] and is applied to this signal a2 [n]. Intuitively, the first polynomial z q2 q1 will position the (k+1)th block of a[n] over the k th block of d[n] since every block of a2 [n] contains q1 q2 number of elements. The second polynomial (t[0]z −q2 + t[1]) chooses two consecutive samples of a[n] because every sample of a[n] is repeated q2 times as explained earlier. Thus, the second term will help in choosing either both the elements of (k + 1)th block of a[n] that are future samples of d[n] or in choosing one element of (k + 1)th block and one of k th block of a[n]. Mathematically, this is seen as below. On passing a2 [n] through the predict filter T (z), the k th block of signal a3 [n] is obtained as:         m m ak3 [m] = t[0]a (k − 1)q1 + + (q1 − 1) + t[1]a (k − 1)q1 + + q1 , q2 q2 (23) where m = 0, 1, ..., q1 q2 − 1. This signal is passed through the q1 -fold downsampler resulting in the k th block of signal dnew,k [m] = dk [m] − ak4 [m], m = 0, 1, ..., q2 − 1. (25) From (21), (24), (25), and Fig.7, it can be noted that (i) The first element of k th block of d[n], i.e., dk [0] for m = 0 is x[(k − 1)M + q1 ] and this sample is predicted from a[kq1 − 1] = x[(k − 1)M + q1 − 1] and a[kq1 ] = x[kM ] that are the elements of the k th and (k + 1)th blocks of a[n], respectively. Also, these are the past and future samples of x[(k − 1)M + q1 ]. (ii) The last element of k th block, i.e., d[kq2 −1] = x[kM  −1]  q1 for m = q2 −1 is predicted from x[kM +q1 −2− q2   q1 and x[kM + q1 − 1 − that are the elements of the q2 (k + 1)th block of a[n]. (iii) From (i) and (ii) above, it is clear that in between elements of the k th block of d[n] will be predicted from only the k th and (k + 1)th block elements of a[n]. In fact, d qq12 e elements of a block of d[n] (where d.e denotes the ceil function) are predicted using the past and future samples, i.e., from the k th and (k + 1)th block elements of a[n], while rest of the elements are predicted from the elements of (k + 1)th block of a[n]. This proves Theorem-1. We name the modified rate converter branch incorporating polynomial T (z), shown in Fig.9, as the ‘Composite Predict Branch’. This is to note that the above choice of polynomial provides the best possible generic solution for prediction using nearest neighbors for different values of q1 and q2 . For example, if the the first polynomial z q2 q1 of T (z) is omitted, one may note that samples in d[n] will be predicted from far away past samples or far away future samples depending on the second polynomial. Thus, although T (z) can be chosen in many ways, we choose to use the polynomial provided in Theorem-1. The above theorem of 2-tap predict filter can be easily extended to obtain an Np -tap filter with even Np . For example, a 4-tap T (z) can be given by T (z) = z q2 q1 (t[0]z −2q2 + t[1]z −q2 + t[2] + t[3]z q2 ), (26) that will choose four consecutive samples of a[n] and are from the immediate neighboring blocks of d[n]. In general, an even 7 length Np -tap filter T (z) is given by T (z) = z q2 q1 − z relation: Np −1 Np 2 q2 X t[k]z kq2 . (27) Gnew (z) = Gh (z) − h k=0 2) Estimation of predict stage filter T (z) from a given signal: In order to estimate an Np -length predict stage filter T (z), we consider the prediction error dnew [n] shown in Fig.9 as below: dnew [n] =d[n] − a4 [n] Np −1 q2 −1 =d[n] − X X t[j]a1 [q1 n − (j + k)]), (28) j=0 k=0 and minimize it using the Least Squares (LS) criterion yielding the solution of T (z) as below: t̂ = min t NX d −1 (dnew [n])2 , (29) n=0 where Nd is the length of the difference signal dnew and t = [t[0]t[1] . . . t[Np − 1]]0 is the column vector of elements of the polynomial T (z). The above equation can be solved to learn the predict stage filter T (z). 3) Update of RFB filters using learned T (z): For the dyadic wavelet design using lifting as discussed in Section II.A, predict polynomial is used to update analysis highpass and synthesis lowpass filters using (1) and (2). However, we require to derive similar equations for a rational filterbank. Before we present this work, let us look at a Lemma that will be helpful in defining these equations. Lemma 1. A structure containing a filter H(z) followed by an M -fold downsampler and preceded by an M -fold upsampler (Fig. 10(a)) can be replaced by an equivalent filter H̃(z) (Fig. 10(b)), which is given by: M −1 1 1 X r H̃(z) = H(z M WM ), M r=0 (30) Proof. Refer section 4.3.5 of [34]. M where P (z) = Rp (z)T (z). Proof. Refer to Fig.11. Since M and q2 are relatively prime, the corresponding downsampler and upsampler can be swapped to simplify the structure in Fig.11(b). Using noble identities [34] and further simplification, we obtain the structure in Fig.11(c). As two downsampler or upsampler can swap each other, we obtain the structure in Fig.11(d). The part in red dotted rectangle in the figure can be replaced by an equivalent filter, H̃p (z) using (30) of Lemma-1 and is given by: H̃p (z) = q1 −1 q2 M 1 X r Gl (z q1 Wqq12 r )P (z q1 WqM ). 1 q1 r=0 (32) Considering the structure in Fig.11(d), signals a4 [n], d[n], and dnew [n] can be written in Z-domain as: A4 (Z) = M −1 q2 1 1 X q2 k k X(z M WM )H̃p (z M WM ), M (33) k=0 D(Z) = M −1 q2 1 1 X q2 k k X(z M WM )Gh (z M WM ), M (34) k=0 Dnew (z) = Xl (z) − Xu (z) = M −1  q2 1 1 1 X q2 k  k k X(z M WM ) Gh (z M WM ) − H̃p (z M WM ) . M k=0 (35) The above relation is equivalent to applying a new filter Gnew (z) to the lower branch of the rational wavelet system h (Fig.11(e)) and is given by Gnew (z) =Gh (z) − H̃p (z) h where WM = exp(−j 2π M ). v1[n] q1 −1 q2 M 1 X k ), Gl (z q1 Wqq12 k )P (z q1 WqM 1 q1 k=0 (31) H ( z) =Gh (z) − M v2 [n] q1 −1 q2 M 1 X k Gl (z q1 Wqq12 k )P (z q1 WqM ) 1 q1 k=0 (36) This proves Theorem-2. (a) v1[n] H ( z) v2 [n] (b) Fig. 10: Filter structure in (a) is equivalent to filter in (b) Next, we present Theorem-2 that provides the structure of updated analysis highpass filter Gnew (z) using T (z) for an h RFB. Theorem 2. The analysis highpass filter of a 2-channel rational filterbank can be updated using the predict polynomial T (z) used in the ‘composite predict branch’ via the following Next, the synthesis lowpass filter Fl (z) is updated to Flnew (z) as follows. First, the analysis RFB containing Gl (z) and Gnew (z) is converted into an equivalent M -band analysis h filterbank structure as shown in Fig.4 using (6) and (8), where lowpass filter Gl (z) is transformed to q1 upper filters of M band analysis filters and highpass filter Gnew (z) of rational h wavelet transforms to M − q1 (= q2 ) lower filters of the M band analysis filterbank. Next, the polyphase matrix Rnew (z) is obtained by using (10), (12), and (14). On using Rnew (z) from (13) in (11), we obtain all M updated synthesis filters of uniformly decimated filterbank. Out of these M synthesis filters, lower M − q1 filters are unchanged, while the upper 8 Swapping q1 Gl ( z ) q1 M Gl ( z ) a4 [n] q2 q2 Gh ( z ) M   new d q2 [ n] Gh ( z ) (a) M    M Gl ( z ) P( z ) 𝑞1 M d [ n] 𝑞2 q1  q2 𝑥[𝑛] P( z ) 𝑥[𝑛] q1 q2 𝑞1 M P( z ) 𝑥[𝑛] 𝑞2 d new Gh ( z ) M  [ n] (b) Gl ( z ) M a[n] 𝑞2 𝑥[𝑛] 𝑞1 Gl ( z q2 ) q2 G ( z) M d new[n] a4 [n] P( z M ) 𝑞1 𝑥[𝑛] new h  (c) H p ( z) q1  d new M d [ n] 𝑞2 [ n] Gh ( z ) M    d new[n] (d) (e) Fig. 11: Illustration for the proof of theorem-2. Structures in (a)-(e) are equivalent of each other. q1 filters are updated because of the predict branch. On using these filters in (7), we obtain Flnew (z). q1 C. Update Stage before adding this signal to a[n] as shown in Fig.13. a new [n] a[n] q1 Gl ( z )  M   M Fl ( z ) q1 q2 𝑥[𝑛] 𝑥 [𝑛] Ru ( z ) Update rate converter q1 q2 Gh ( z ) Gl ( z )  M d 4 [ n] d 3 [ n] In this subsection, we present the structure of update branch to be used in lifting structure of an RFB, present the estimation of the update branch polynomial from the given signal, and the theorem for the update of corresponding filters of RFB. 1) Structure of update branch: In the predict stage, we used upper branch samples to predict the lower branch samples. In the update stage, we update the upper branch samples using the lower branch samples. Since the output sample rate of two branches is unequal, we require to downsample the lower branch samples by a factor of ku given by: q2 ku = , (37) q1 M M Fh ( z ) q2 d [ n] Fig. 12: Illustration of Update Rate Converter As explained earlier, the upsampler is required to be followed by an anti-imaging filter and downsampler should be preceded by an anti-aliasing filter. We use filter Ru (z) = 1 + z −1 + ... + z −(q1 −1) that accomplishes the same and its placement is shown in Fig.12. Similar to the predict stage, we call the update branch of Fig.13 as the composite update branch that includes update rate converter, update polynomial S(z), and a summer. a new [n] v[n] a[n]  S ( z) Ru ( z ) d1[n] q1 q2 Gh ( z ) M 𝑥𝑟𝑙 [𝑛] Fl ( z ) q1 q2 d 2 [ n] 𝑥[𝑛]  U ( z) 𝑥 [𝑛] Composite Update branch M M Fh ( z ) q2 d [ n] Fig. 13: Update Stage Definition 2. Update rate converter: It is a combination of the polynomial Ru (z) = 1 + z −1 + ... + z −(q1 −1) preceded by a q1 -fold upsampler and followed by a q2 -fold downsampler as shown in Fig.12. Similar to the structure of predict stage filter T (z), the structure of update stage filter S(z) should also be chosen carefully, so that the elements in the upper branch samples are updated only with the nearest neighbors. Below we present the theorem on the structure of a 2-tap update stage filter that ensures this. Theorem 3. A 2-tap update stage filter S(z) = s[0]+s[1]z −q1 ensures that every sample in the k th block of a[n] branch is updated from the (k − 1)th and/or k th block samples of d[n]. Proof. After passing the signal through the Lazy wavelet, blocks of output signal can be formed as described in section III-B. The k th block of approximation and detail coefficients is given by ak [m] =a[(k − 1)q1 + m] = x[(k − 1)M + m], m = 0, 1, ..., q1 − 1, k d [m] =d[(k − 1)q2 + m] = x[(k − 1)M + q1 + m], m = 0, 1, ...q2 − 1, (38) where superscript denoted the k th block. For update, first the detail coefficients are passed through a q1 -fold upsampler and 9 the polynomial Ru (z). This leads to signal d2 [n] that contain every element of d[n] repeated q1 times. Mathematically, k th block of signal d2 [n] is given by:    m k d2 [m] = d (k − 1)q2 + , m = 0, 1, ..., q1 q2 − 1, q1 (39) Update filter S(z) = s[0] + s[1]z −q1 is applied on this signal d2 [n]. Unlike the predict stage, we do not require advancement of any block of d2 [n] because a2 [n] requires to be updated from the past and future samples of its block that are contained in the (k − 1)th and the k th blocks of d2 [n] (Fig. 7). Hence, S(z) requires only one polynomial (s[0] + s[1]z −q1 ) that chooses two consecutive samples of d[n] because every sample of d[n] is repeated q1 times as explained earlier. On passing d2 [n] through the update stage filter S(z), we obtain       m m − q1 dk3 [m] = s[0]d (k − 1)q2 + + s[1]d (k − 1)q2 + , q1 q1 (40) where m = 0, 1, ..., q1 q2 − 1. This signal is downsampled by a factor of q2 resulting in the k th block of d4 [n] given by       q2 m q2 m − q1 dk4 [m] = s[0]d (k − 1)q2 + + s[1]d (k − 1)q2 + , q1 q1 (41) where m = 0, 1, ..., q1 − 1. Signal d4 [n] helps with the update of signal in the upper branch. From (41), we note that (i) The first term of the ak [m], i.e., ak [0] = x[(k − 1)M ] for m = 0 is updated with d[(k − 1)q2 ] = x[(k − 1)M + q1 ] and d[(k − 1)q2 − 1] = x[(k − 1)M − 1] that are the elements of the k th and (k − 1)th blocks of d[n], respectively. Also, these are the past and future samples of x[(k − 1)M ]. (ii) The last term of ak [m], i.e., a[kq1 − 1]= x[(k− 1)M  +  q2 q1 −1] for m = q1 −1 is updated with x kM − −1 q1     q2 and x kM − − 2 that are the elements in the k th q1 block of detail signal. (iii) From (i) and (ii) above, it is clear that in between elements of the k th block of a[n] will be updated from only the (k − 1)th and k th block elements of d[n]. This proves Theorem-3. Similar to the predict stage polynomial, the above choice of polynomial provides the best possible generic solution for update using nearest neighbors for different values of q1 and q2 . Thus, although S(z) can be chosen in many ways, we choose to use the polynomial provided in Theorem-3. The above theorem of 2-tap update filter can be easily extended to obtain an even length Ns -tap filter. For example, a 4-tap S(z) can be given by S(z) = s[0]z q1 + s[1] + s[2]z −q1 + s[3]z −2q1 , (42) that will choose four consecutive samples of d[n] and are from the immediate neighboring blocks of a[n]. In general, an even length Ns -tap filter S(z) can be defined by the following relation: S(z) = z ( Ns 2 −1)q1 NX s −1 s[k]z −kq1 , (43) k=0 where Ns is the length of the update stage filter, S(z) and is assumed to be even. 2) Learning of update stage filter S(z) from a given signal: In order to estimate an Ns -length update stage filter S(z), we consider the updated signal anew [n] shown in Fig.13 as below: anew [n] =a[n] + d4 [n] =a[n] + NX 1 −1 s −1 qX s[j]d1 [q2 n − (j + k)] (44) j=0 k=0 On passing these approximate coefficients through an M -fold upsampler (Fig.13), we obtain (  n if n is a multiple of M anew M (45) v[n] = 0 otherwise. This signal v[n] is passed through the synthesis lowpass filter Fl (z) followed by a q1 -fold downsampler as shown in Fig.13 to obtain xrl [n] = Lfl X fl [k]v[q1 n − k], (46) k=0 where xrl is the reconstructed signal at the lowpass branch of synthesis side with the same length as that of the input signal x. Lfl is the length of the synthesis lowpass filter. Assuming input signals to be rich in low frequency, most of the energy of the input signal moves to lowpass branch. Hence, signal xrl [n] is assumed to be the close approximation of the input signal x[n]. Correspondingly, the following optimization problem is solved to learn the update stage filter S(z): ŝ = min s N −1 X (xrl [n] − x[n])2 , (47) n=0 where s = [s[0]s[1] . . . s[Ns − 1]]0 is the column vector of the coefficients of polynomial S(z). From equation (44), (45) and (46), it can be observed that signal xrl [n] can be written in terms of the update stage filter S(z). Thus, equation (47) is solved using LS criterion to learn the update stage filter S(z). 3) Update of RFB filters using estimated S(z): Similar to the predict stage discussed earlier, we propose equation for the update of analysis lowpass filter using the update stage polynomial S(z) for a rational filterbank in Theorem-4 below. Theorem 4. The analysis lowpass filter of a 2-channel rational filterbank can be updated using the update polynomial S(z) used in the ‘composite update branch’ via the following relation: Gnew (z) = Gl (z) + l qX 2 −1 r=0 q1 M 1 r Gh (z q2 Wqq21 r )U (z q2 WqM ), 2 q2 (48) 10 q1 Gl ( z ) M    a new [n] q1 Gl ( z ) M   q1 Gh ( z ) q1 q2 M q2 M Gl ( z ) M  q1 Gh ( z q1 ) U (zM ) a new [n] q2 M q1 Gh ( z )   x[n] U ( z) x[n] U ( z) 𝑞2 a[n] a new [n] 𝑞2 q2 x[n]  d 4 [ n] Swapping (a) (c) (b) q1 Glnew ( z ) M a new [n] 𝑞1 𝑥[𝑛] a[n] Gl ( z ) M    a new [n] x[n] 𝑞1 q2 Gh ( z ) M 𝑞2 Gh ( z q1 ) U (zM ) 𝑞2 d [ n] M d 4 [ n] Hu ( z) (d) (e) Fig. 14: Illustration for the proof of theorem-4. Structures in (a)-(e) are equivalent of each other where U (z) = Ru (z)S(z). Proof. Refer to Fig.14. Following the similar procedure as in Theorem 2, the update structure of Fig.14(a) can be equivalently converted to Fig.14(d) with the filter H̃u (z) given by: H̃u (z) = q2 −1 q1 M 1 X r Gh (z q2 Wqq21 r )U (z q2 WqM ). 2 q2 r=0 (49) Considering the structure in Fig.14(d), signals d4 [n], a[n], and anew [n] can be written in Z-domain as: D4 (z) = M −1 q1 1 1 X q1 k k X(z M WM )H̃u (z M WM ), M (50) k=0 A(Z) = M −1 q1 1 1 X q1 k k X(z M WM )Gl (z M WM ), M (51) k=0 Anew (z) = D4 (z) + A(z) = M −1  q1 1 1 1 X q1 k  k k X(z M WM ) H̃u (z M WM ) + Gl (z M WM ) . M k=0 (52) The above signal is equivalent to passing the signal X(z) through an equivalent low pass filter Gnew (z) of the RFB l (Fig.14(e)) and is given by: Gnew (z) =Gl (z) + H̃u (z) l =Gl (z) + q2 −1 q1 M 1 X r Gh (z q2 Wqq21 r )U (z q2 WqM ) 2 q2 r=0 (53) This proves the above theorem. Similar to the predict stage, the synthesis highpass filter Fhnew (z) is learned using the updated analysis lowpass filter Gnew (z), polyphase matrices, and equations (6), (8), (9), (10)l (14). Although we presented  a specific case of this work with decimation ratio of 32 , 13 in [35], the proposed method is very general and presents rational wavelet design theory for generalized rational factors. Note that the resultant filters of the rational filterbank are learned from the given signal because these are updated based on the predict and update stage filters T (z) and S(z) learned from the signal using (29) and (47), respectively. Also, the proposed method has closed-form solution because (29) and (47) can be directly solved using the least squares criterion. D. Design Examples We present some examples of M-RWTL and the corresponding rational filterbank learned with the proposed method. Table-I presents parameters used for learning M-RWTL  1 with  2 1 1 the following sampling rates in two branches: , 3 3 , 2, 2 ,   3 1 2 3 4 , 4 , 5 , 5 . The first row in table provides values of q1 , q2 , and M . Second to fifth row presents Lazy wavelet of the corresponding rational filterbank structure from which the learning is initialized. Sixth and eighth row represents polynomial Rp (z) and Ru (z) respectively. Seventh and ninth row represents the structure of 2-tap predict and update polynomial T (z) and S(z) respectively. With these parameters and structure, we learn the M-RWTL matched to given signals of interest. We consider four signals of different types: 1) ECG signal, 2) speech signal, and 3) two music signals (named as music-1 and music-2). These signals are shown We  1 in1 Fig.15.  learn  M-RWTL with sampling rates 2 1 3 1 2 3 , , , , , , , 3 3 2 2 4 4 5 5 in the two branches matched to music-1, music-2, ECG, and speech signals, respectively. Since we learn M-RWTL in the lifting framework that always satisfies PR condition, our learned wavelet system achieves PR with NMSE (normalized mean square error) of the order of 10−19 . Table-II presents coefficients of predict polynomial, update polynomial, and synthesis filters learned with the proposed method. Fig.16 shows the frequency response of filters associated with the learned M-RWTL. IV. A PPLICATION IN C OMPRESSED S ENSING BASED RECONSTRUCTION In this Section, we explore the performance of the learned M-RWTL in the applications of Compressed Sensing (CS)- 0.5 0.5 0 1.5 0 −0.5 −0.5 −1 −1 1 1 0.5 Magnitude 1 Magitude 1 Magnitude Magnitude 11 0.5 0 −0.5 −0.5 0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000 Sample (n) Sample (n) (a) Music signal-1 −1 0 0 500 1000 1500 2000 2500 3000 3500 Sample (n) (b) Music signal-2 (c) ECG signal Fig. 15: Signals used in the experiments −1 0 500 1000 1500 2000 2500 Sample (n) (d) Speech signal TABLE I: Illustration of M-RWTL learned based on the theory developed in section-III     Sampling rate 2 1 1 1 3 1 2 3 , , , , 3 3 2 2 4 4 5 5 in two branches (q1 , q2 , M ) (2, 1, 3) (1, 1, 2) (3, 1, 4) (2, 3, 5) Lazy Gl (z) 1 + z −1 1 1 + z −1 + z −2 1 + z −3 Lazy Gh (z) z2 z z3 z2 + z4 + z6 2 Lazy Fl (z) 1+z 1 1+z+z 1 + z3 Lazy Fh (z) z −2 z −1 z −3 z −2 + z −4 + z −6 Rp (z) 1 1 1 1 + z −1 + z −2 2 −1 −1 3 −1 6 T (z) z (t[0]z + t[1]) z(t[0]z + t[1]) z (t[0]z + t[1]) z (t[0]z −3 + t[1]) Ru (z) 1 + z −1 1 1 + z −1 + z −2 1 + z −1 −2 −1 −3 S(z) s[0] + s[1]z s[0] + s[1]z s[0] + s[1]z s[0] + s[1]z −2 (a) ( 23 , 13 ) LPF with Music-2 (e) ( 34 , 14 ) LPF with Music-1 (b) ( 23 , 31 ) HPF with Music-2 (c) ( 21 , 12 ) LPF with ECG (f) ( 34 , 14 ) HPF with Music-1 (g) ( 25 , 35 ) LPF with Speech Fig. 16: Frequency response of synthesis filters presented in Table-II. (d) ( 12 , 12 ) HPF with ECG (h) ( 25 , 52 ) HPF with Speech TABLE II: Coefficients of predict polynomial, update polynomial, and synthesis filters of M-RWTL learned with different sampling rates. fs is the sampling frequency of the signal in kHz and N is the number of samples of the signal used in experiments. Sampling rate Signal Predict/Update Filter coefficients in two branches (fs , N ) polynomial   Music-2 t[n] =[0.4777 0.5101] fl [n] =[ 0.1707 0.3347 0.3347 0.1599] 2 1 , 3 3 (11.025, 10000) fh [n] =[ -0.0748 -0.1466 -0.1466 0.6867 s[n] =[0.1751 0.1893] -0.1586 -0.1586 -0.0758]   ECG t[n] =[0.5119 0.5143] fl [n] =[ 0.2538 0.4935 0.2526] 1 1 , 2 2 (0.36,3600) s[n] =[0.2818 0.2834] fh [n] =[-0.1449 -0.2818 0.7100 -0.2834 -0.1451]   Music-1 t[n] =[0.6100 0.6109] fl [n] =[ 0.1447 0.2369 0.2369 0.2369 0.1445] 3 1 , 4 4 (11.025, 10000) fh [n] =[-0.0755 -0.1236 -0.1236 -0.1236 0.6751 s[n] =[0.1496 0.1499] -0.1239 -0.1239 -0.1239 -0.0756] fl [n] =[ 0.0652 0.0652 0 0.0652 0.1955   Speech t[n] =[0.7048 0.3336] 2 3 0.1378 0.1378 0.1955 0.1378] , 5 5 (11.025,2700) fh [n] =[ -0.0137 -0.0137 0 -0.0137 0 -0.0427 -0.0715 -0.0409 -0.0427 -0.0409 -0.0288 0.1959 s[n] =[ 0.1430 0.4474] -0.1280 0.2572 -0.1280 0.1959 0 0 -0.0902] based reconstruction of 1-D signals. Compressed Sensing (CS) problem aims to recover a full signal from a small number of its linear measurement [36], [37]. Mathematically, the problem is modeled as: y = Ax, (54) where x is the original signal of size N × 1, which is compressively measured as y of size M × 1 by the measurement matrix A of size M × N . Full signal x is reconstructed from compressive measurements by solving an optimization problem with signal sparsity in some transform domain as prior. Wavelets are extensively 12 used as the sparsifying transforms [37]. l1 regularized linear least square is solved for signal reconstruction as: α̃ = min ||α||1 subject to y = AWα, α (55) where x = Wα. Here, W represents the wavelet transform and α is wavelet transform of x. The above problem is known as basis pursuit (BP) [38] and we used SPGL1 solver to solve the above problem. Full signal is reconstructed as x̃ = Wα̃. Table-III presents CS-based reconstruction performance of M-RWTL with different sampling rates on the four signals shown in Fig.15. Reconstruction performance is measured via PSNR (peak signal to noise ratio) given by   (max(x))2 , (56) PSNR = 10 log10 M SE PN −1 where M SE = N1 n=0 |x[n] − x̃[n]|2 with x̃ as the reconstructed signal. Measurement matrix A is Gaussian and sampling ratio (SR) is varied from 10% to 90% with a difference of 10%, where sampling ratio= b M N c × 100. Threelevel wavelet transform decomposition has been used for all the experiments. Results are averaged over 50 independent trials. M-RWTL the following sampling rate are considered:  1 2with    2 1 3 1 , , , and , and are represented as NG 23 , 31 , 3 3 3 3 4 4  NG 13 , 23 and NG 34 , 14 , respectively. The original signal x is not available in the compressed sensing application, while the proposed method requires the signal for learning matched rational wavelet. Thus, we propose to sample one-third of the data fully (at 100% sampling ratio) to learn matched wavelet. Next, we apply the learned matched wavelet for the reconstruction of the rest of the data sampled at lower compressive sensing ratio. One may also use another approach of [39], [40] to learn matched wavelet in CS application. However, so far [39] and [40] are limited to the special case of dyadic matched wavelet and can be explored for extension in the CS application with rational wavelet transform learning as a future work. The reconstruction performance is compared with standard orthogonal Daubechies wavelets db2 and db4, and standard bi-orthogonal wavelets, bior5/3 and bior9/7 (labeled as Bi 5/3 and Bi 9/7, respectively). The performance is also compared with overcomplete rational wavelets designed in [25]. For fair comparison, we consider the same sampling rate in the low frequency branch for these overcomplete rational wavelets as used in proposed M-RWTL in Table-III, i.e., we use overcom 1  2 plete rational wavelets [25] with sampling rates: 3 , 1 , 3 , 1    and 34 , 1 , represented as S 32 , 1 , S 13 , 1 and S 43 , 1 . From Table III, it is observed that Bi 9/7 performs best among the existing wavelets used. Also, the overcomplete rational wavelet 13 , 1 performs better than the existing wavelets on ECG signal at all sampling ratios less than 60%, while these perform comparable or inferior in performance to the existing wavelets on music and speech signals. On the other hand, the proposed M-RWTL perform better, with an improvement of upto 1.8 dB of PSNR, in comparison to existing wavelets for music signals. Particularly, M-RWTL with 23 , 13 performs better on music-1 signal for all the  sampling ratios. M-RWTL with 23 , 13 performs better on music-2 signal  from 90% to 70% sampling ratio beyond which 31 , 23 performs better. On ECG signal, existing wavelet Bi 9/7 performs better than all other wavelets at higher sampling ratios from 90% to 60%. At 50% sampling ratio, overcomplete rational wavelet 13 , 1 outperforms all existing and proposed rational wavelets. Below 50% sampling ratio, M-RWTL with 13 , 23 performs better than all existing as well as overcomplete rational wavelets with an improvement of upto 10 dB than existing wavelets and upto 4 dB than overcomplete rational wavelets. Similarly, in case of speech signal, Bi 9/7 performs better than all other wavelets from sampling ratio 90% to 70%. Below 70% sampling ratio, M-RWTL with 23 , 31 outperforms all existing as well as overcomplete rational with an improvement of upto 1.7 dB. Further, this is to note that at higher sampling ratios, PSNR of the reconstructed ECG and speech signals is high at around 40 dB and 30 dB, respectively, with different wavelets. Hence, the reconstructed signal appears almost similar to the original signal. The quality of the reconstructed signal deteriorates with decreasing sampling ratios, where the proposed M-RWTL performs best with as much as 10 dB improvement. Overall, the performance of rational matched wavelets is superior at lower sampling ratios in compressive sensing application. In this paper, we have not explored the problem of choosing the optimal sampling rate of learned wavelet for a particular signal or in a particular application. This remains an open problem and can be explored in the future. V. C ONCLUSION Theory of learning rational wavelet transform in the lifting framework, namely, the M-RWTL method, has been presented. The existing theory of lifting framework is extended from dyadic to rational wavelets, where critically sampled rational matched wavelet filterbank can be designed for any general rational sampling ratios. The concept of rate converters is introduced to handle variable data rate of subbands. The learned signal-matched rational filterbank inherits all the advantages of lifting framework. The learned analysis and synthesis filters are FIR and are easily implementable in hardware, thus making RWT easily usable in applications. Closed form solution is presented for learning rational wavelet and thus, no greedy solution is required making M-RWTL computationally efficient. The proposed M-RWTL transform can be learned from a short snapshot of a single signal and hence, extends the use of transform learning from the requirement of large training data to small data snapshots. As a proof of concept, the learned MRWTL is applied in CS-based reconstruction of signals and is observed to perform better compared to the existing wavelets. Although the learned M-RWTL performs better in the above application, it is not known apriori as to what sampling ratios in the two branches of the rational filterbank are optimal for a given signal in a particular application. We leave it as an open problem for the future work. 13 TABLE III: Performance of rational filterbank in CS-based reconstruction of signals. The rational wavelet is learned from one-third of the data samples. Results are averaged over 50 independent trials. Signal Music-1 Music-2 ECG Speech Sampling ratio (in %) db2 db4 Bi 5/3 Bi 9/7 90 80 70 60 50 40 30 20 10 90 80 70 60 50 40 30 20 10 90 80 70 60 50 40 30 20 10 90 80 70 60 50 40 30 20 10 28.3 23.9 21.0 18.5 16.6 15.1 13.8 12.7 11.9 29.7 26.1 23.9 22.0 20.4 18.9 17.5 16.0 14.7 49.7 45.9 42.9 39.5 34.6 27.0 21.2 18.5 16.6 41.7 36.2 31.9 28.5 25.4 22.4 20.1 18.0 16.0 31.2 26.2 22.5 19.8 17.6 15.8 14.2 12.9 12.0 30.0 26.4 24.1 22.3 20.7 19.1 17.6 16.1 14.7 50.1 46.4 43.6 40.6 36.1 27.4 21.1 18.3 16.3 44.7 39.2 34.3 30.1 26.4 23.2 20.6 18.1 15.9 28.7 23.6 20.1 17.4 15.5 14.0 12.9 12.1 11.5 29.4 25.9 23.5 21.6 19.9 18.4 16.9 15.5 14.2 49.8 45.9 42.8 39.4 34.1 25.9 20.7 17.7 15.5 42.2 36.6 32.1 28.1 24.6 21.5 19.2 17.0 15.4 33.2 28.0 24.1 20.8 18.2 16.1 14.4 13.0 12.1 30.3 26.8 24.5 22.6 21.0 19.4 17.8 16.2 14.8 50.9 47.1 44.2 41.2 37.3 29.4 21.8 18.7 16.5 46.4 41.0 36.4 32.1 28.0 24.0 21.2 18.6 16.2 (in  PSNR   dB)  2 S 3, 1 S 13 , 1 R EFERENCES [1] S. Ravishankar and Y. Bresler, “Learning sparsifying transforms,” IEEE Transactions on Signal Processing, vol. 61, no. 5, pp. 1072–1086, 2013. [2] B. Wen, S. Ravishankar, and Y. Bresler, “Video denoising by online 3d sparsifying transform learning,” in Image Processing (ICIP), 2015 IEEE International Conference on. IEEE, 2015, pp. 118–122. [3] S. Ravishankar and Y. Bresler, “Efficient blind compressed sensing using sparsifying transforms with convergence guarantees and application to magnetic resonance imaging,” SIAM Journal on Imaging Sciences, vol. 8, no. 4, pp. 2519–2557, 2015. [4] J. Xie, L. Xu, and E. Chen, “Image denoising and inpainting with deep neural networks,” in Advances in Neural Information Processing Systems, 2012, pp. 341–349. [5] K. Kulkarni, S. Lohit, P. Turaga, R. Kerviche, and A. Ashok, “Reconnet: Non-iterative reconstruction of images from compressively sensed measurements,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 449–458. [6] S. Mallat, A wavelet tour of signal processing. Academic press, 1999. 29.6 24.4 20.6 17.6 15.4 13.8 12.5 11.7 11.4 29.4 26.0 23.7 21.8 20.2 18.6 17.1 15.7 14.4 46.8 42.9 39.6 35.0 28.2 23.3 20.4 18.5 16.7 43.4 37.6 32.5 28.4 25.1 22.1 19.9 17.8 15.9 24.8 20.8 18.2 16.3 14.8 13.7 12.8 12.1 11.6 29.2 25.7 23.4 21.6 20.0 18.5 17.2 15.9 14.7 49.0 45.6 43.3 41.1 38.6 35.3 30.7 23.8 16.6 38.3 33.5 30.2 27.4 25.2 22.8 20.6 18.0 15.5   S 34 , 1 29.2 24.0 20.6 17.9 15.8 14.2 12.9 12.1 11.6 29.4 25.9 23.5 21.6 19.9 18.2 16.8 15.4 14.3 47.5 42.1 36.2 29.5 24.4 21.5 19.6 18.2 16.6 43.6 38.0 32.8 28.4 25.2 22.4 20.2 18.1 16.1 NG  2 1 3, 3 33.3 28.7 25.4 22.6 19.9 17.7 15.6 13.8 12.3 30.6 27.1 24.8 23.1 21.5 20.0 18.5 16.9 15.3 50.1 45.9 42.6 38.6 33.1 26.7 22.9 20.3 17.9 43.6 39.1 35.7 32.4 29.1 25.7 22.9 20.2 17.6  NG  1 2 3, 3 27.6 23.6 20.8 18.6 16.7 15.0 13.6 12.3 10.9 30.2 26.9 24.8 23.3 22.0 20.8 19.6 18.3 16.6 47.9 44.5 42.4 40.5 38.3 35.6 32.1 27.0 20.1 39.0 35.1 32.2 29.5 27.2 25.0 22.7 20.2 16.9  NG  3 1 4, 4 32.7 27.8 24.3 21.3 18.9 16.8 15.1 13.6 12.3 30.1 26.5 24.1 22.2 20.5 18.9 17.5 16.1 14.8 48.3 43.1 37.2 30.1 25.5 22.7 20.7 18.9 17.0 42.2 37.1 32.8 29.1 25.9 23.3 20.9 18.8 16.8 [7] S. G. Chang, B. Yu, and M. Vetterli, “Adaptive wavelet thresholding for image denoising and compression,” IEEE transactions on image processing, vol. 9, no. 9, pp. 1532–1546, 2000. [8] T. Gagie, G. Navarro, and S. J. Puglisi, “New algorithms on wavelet trees and applications to information retrieval,” Theoretical Computer Science, vol. 426, pp. 25–41, 2012. [9] R. C. Guido, J. F. W. Slaets, R. Köberle, L. O. B. Almeida, and J. C. Pereira, “A new technique to construct a wavelet transform matching a specified signal with applications to digital, real time, spike, and overlap pattern recognition,” Digital Signal Processing, vol. 16, no. 1, pp. 24–44, 2006. [10] K. Najarian and R. Splinter, Biomedical signal and image processing. CRC press, 2005. [11] A. Depeursinge, A. Foncubierta-Rodriguez, D. Van de Ville, and H. Müller, “Rotation–covariant texture learning using steerable riesz wavelets,” IEEE Transactions on Image Processing, vol. 23, no. 2, pp. 898–908, 2014. [12] T. Blu, “A new design algorithm for two-band orthonormal rational filter banks and orthonormal rational wavelets,” Signal Processing, IEEE Transactions on, vol. 46, no. 6, pp. 1494–1504, 1998.  14 [13] G. F. Choueiter and J. R. Glass, “An implementation of rational wavelets and filter design for phonetic classification,” Audio, Speech, and Language Processing, IEEE Transactions on, vol. 15, no. 3, pp. 939–948, 2007. [14] R. Bregović and T. Saramäki, “Design of two-channels fir filterbanks with rational sampling factors using the frm technique,” in Circuits and Systems, 2005. ISCAS 2005. IEEE International Symposium on. IEEE, 2005, pp. 1098–1101. [15] J. Kovačević and M. Vetterli, “Perfect reconstruction filter banks with rational sampling factors,” Signal Processing, IEEE Transactions on, vol. 41, no. 6, pp. 2047–2066, 1993. [16] P. Blanc, T. Blu, T. Ranchin, L. Wald, and R. Aloisi, “Using iterated rational filter banks within the arsis concept for producing 10m landsat multispectral images,” International journal of remote sensing, vol. 19, no. 12, pp. 2331–2343, 1998. [17] A. Baussard, F. Nicolier, and F. Truchetet, “Rational multiresolution analysis and fast wavelet transform: application to wavelet shrinkage denoising,” Signal Processing, vol. 84, no. 10, pp. 1735–1747, 2004. [18] P. Auscher, “Wavelet bases for l2 (r) with rational dilation factor,” Wavelets and their applications, pp. 439–451, 1992. [19] T.-T. Le, M. Ziebarth, T. Greiner, and M. Heizmann, “Optimized sizeadaptive feature extraction based on content-matched rational wavelet filters,” in Signal Processing Conference (EUSIPCO), 2014 Proceedings of the 22nd European. IEEE, 2014, pp. 1672–1676. [20] O. Chertov, V. Malchykov, and D. Pavlov, “Non-dyadic wavelets for detection of some click-fraud attacks,” in Signals and Electronic Systems (ICSES), 2010 International Conference on. IEEE, 2010, pp. 401–404. [21] L. Yu and L. B. White, “Optimum receiver design for broadband doppler compensation in multipath/doppler channels with rational orthogonal wavelet signaling,” Signal Processing, IEEE Transactions on, vol. 55, no. 8, pp. 4091–4103, 2007. [22] ——, “Broadband passive sonar detection using rational orthogonal wavelet filter banks,” in Intelligent Sensors, Sensor Networks and Information Processing (ISSNIP), 2011 Seventh International Conference on. IEEE, 2011, pp. 461–466. [23] I. Bayram and I. W. Selesnick, “Design of orthonormal and overcomplete wavelet transforms based on rational sampling factors,” in Optics East 2007. International Society for Optics and Photonics, 2007, pp. 67 630H–67 630H. [24] ——, “Overcomplete discrete wavelet transforms with rational dilation factors,” Signal Processing, IEEE Transactions on, vol. 57, no. 1, pp. 131–145, 2009. [25] ——, “Frequency-domain design of overcomplete rational-dilation wavelet transforms,” Signal Processing, IEEE Transactions on, vol. 57, no. 8, pp. 2957–2972, 2009. [26] S. T. N. Nguyen and B. W.-H. Ng, “Design of two-band critically sampled rational rate filter banks with multiple regularity orders and associated discrete wavelet transforms,” Signal Processing, IEEE Transactions on, vol. 60, no. 7, pp. 3863–3868, 2012. [27] ——, “Bi-orthogonal rational discrete wavelet transform with multiple regularity orders and application experiments,” Signal Processing, vol. 93, no. 11, pp. 3014–3026, 2013. [28] L. Yu and L. B. White, “Complex rational orthogonal wavelet and its application in communications,” Signal Processing Letters, IEEE, vol. 13, no. 8, pp. 477–480, 2006. [29] W. Sweldens, “The lifting scheme: A custom-design construction of biorthogonal wavelets,” Applied and computational harmonic analysis, vol. 3, no. 2, pp. 186–200, 1996. [30] W. Dong, G. Shi, and J. Xu, “Signal-adapted directional lifting scheme for image compression,” in Circuits and Systems, 2008. ISCAS 2008. IEEE International Symposium on. IEEE, 2008, pp. 1392–1395. [31] J. Blackburn and M. N. Do, “Two-dimensional geometric lifting,” in Image Processing (ICIP), 2009 16th IEEE International Conference on. IEEE, 2009, pp. 3817–3820. [32] M. C. Kale and O. N. Gerek, “Lifting wavelet design by block wavelet transform inversion,” in Acoustics, Speech and Signal Processing (ICASSP), 2014 IEEE International Conference on. IEEE, 2014, pp. 2619–2623. [33] J. Kovačević and W. Sweldens, “Wavelet families of increasing order in arbitrary dimensions,” Image Processing, IEEE Transactions on, vol. 9, no. 3, pp. 480–496, 2000. [34] P. P. Vaidyanathan, Multirate systems and filter banks. Pearson Education India, 1993. [35] N. Ansari and A. Gupta, “Lifting-based rational wavelet design from a given signal,” in 2015 IEEE International Conference on Digital Signal Processing (DSP). IEEE, 2015, pp. 853–857. [36] E. J. Candès, J. Romberg, and T. Tao, “Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information,” IEEE Transactions on information theory, vol. 52, no. 2, pp. 489– 509, 2006. [37] D. L. Donoho, “Compressed sensing,” IEEE Transactions on information theory, vol. 52, no. 4, pp. 1289–1306, 2006. [38] S. S. Chen, D. L. Donoho, and M. A. Saunders, “Atomic decomposition by basis pursuit,” SIAM review, vol. 43, no. 1, pp. 129–159, 2001. [39] N. Ansari and A. Gupta, “Joint framework for signal reconstruction using matched wavelet estimated from compressively sensed data,” in Data Compression Conference (DCC), 2016. IEEE, 2016, pp. 580–580. [40] ——, “Image reconstruction using matched wavelet estimated from data sensed compressively using partial canonical identity matrix,” IEEE Transactions on Image Processing, 2017.
3cs.SY
Link Prediction using Top-k Shortest Distances Andrei Lebedev, JooYoung Lee, Victor Rivera, and Manuel Mazzara arXiv:1705.02936v1 [cs.SI] 4 Apr 2017 Innnopolis University, Russia {a.lebedev, j.lee, v.rivera, m.mazzara}@innopolis.ru, Abstract. In this paper, we apply an efficient top-k shortest distance routing algorithm to the link prediction problem and test its efficacy. We compare the results with other base line and state-of-the-art methods as well as with the shortest path. Our results show that using top-k distances as a similarity measure outperforms classical similarity measures such as Jaccard and Adamic/Adar. Keywords: Graph databases, Shortest paths, link prediction, graph matching, similarity 1 Introduction In a connected world, emphasis is on relationships more than isolated pieces of information. Relational databases may compute relationships at query time [11,10], but this would result in computationally expensive solution. Graph databases [4] store connections as first class citizens, allowing access to persistent connections in almost constant-time [12]. One of the fundamental topological feature in the context of graph theory and graph databases, with implications in Artificial Intelligence and Web communities [7], is the computation of the shortest-path distance between vertices [8]. Many efficient methods for searching shortest path were proposed. On the other hand, top-k distance query handling methods are not well developed and spread. They have many advantages over traditional shortest path as they reveal much more information over a simple shortest path [2]. To extract top-k distances from graph databases, an efficient indexing algorithm is needed, such as pruned landmark labeling scheme, presented in [3]. We utilize this algorithm to obtain the distances and then develop a similarity metric based on them to predict links on graphs. Figure 1 is an example to demonstrate the superiority of using top-k distance over a simple distance metric to measure relationship between two vertices. Table 1 summarizes the connections represented in Figure 1. Based on (top-1) distance alone, we can conclude that similarity (the shortest distance) between black nodes ({a, b}, {b, c}) is the same in the graph. However, it is clear that {b, c} are connected more tightly through a greater number of shortest paths. The framework described in this paper is based on previous work which is called Pruned Landmark Labeling [3]. We evaluate and test the proposed algorithm extensively to prove the efficacy of the algorithm. Then we apply the algorithm to the link prediction problem to compare with existing solutions. a b c Fig. 1: An example of representing connection between pairs of vertices using top-k shortest distance. Table 1: Distances and top-k distances between pairs of black vertices in Figure 1. Vertex Pair (Top-1) Distance Top-k Distance (a,b) 2 [2, 4, 6. . . ] 2 [2, 2, 4 . . . ] (b,c) 2 Related Work In this section, we introduce some of comparable methods for computing top-k shortest paths and link prediction as we utilize top-k distances to predict links in social graphs. 2.1 Top-k shortest paths One of the attempts to find k shortest paths is presented in [6] which achieves O(m + n log n + k) complexity. [6] also covers programming problems, including the knapsack problem, sequence alignment, maximum inscribed polygons, and genealogical relationship discovery. We adopt the algorithm presented in [2] to discover k shortest paths since it achieves six orders of magnitude faster computation given very large graphs with millions of vertices and tens of millions of edges. 2.2 Link prediction Link prediction in social network is a well known problem and extensively studied. Links are predicted using either semantic or topological information of a given network. The main idea in link prediction problem is to measure similarity between two vertices which are not yet linked to each other. If the measured similarity is high enough then the future link is predicted. [9] attempts to infer which new interactions among members of a social network are likely to occur in the near future. Authors develop approaches to link prediction based on measures for analyzing the “proximity” of nodes in a network. 3 Prediction Methods Given a graph G = (V, E), where V is the set of vertices and E is the set of edges. Then let m and n be |E| and |V | respectively. We also assume that vertices are represented by unique integers to enable the comparison of two vertices. Furthermore, let us denote Pab as the set of paths from a to b, a ∈ V and b ∈ V , and dith (s, t) as the i-th shortest path between s and t in Pst . In the following, we introduce the structure of querying and indexing algorithm from [2]. – Distance label L(v): a set of triplets (u, δ, c) of a vertex, a path length and a number of paths with length δ. – Loop label C(v): a sequence of k integers (δ1 , δ2 , ..., δk ). – Index I = (L, C): L and C sets of distance labels and loop labels. – Ordering: vertices in order of decreasing degrees. – Number of paths cw,δ′ : number of paths in the L(v) between vertex v and w, of length not exceeding δ ′ using loop label C(v). – Query(I, s, t): smallest k elements in the ∆(I, s, t). Then we can compute the multiset as follows. ∆(I, s, t) = {δsv + δvv + δvt | (v, δsv ) ∈ L(s), δvv ∈ C(v), (v, δvt ) ∈ L(t)} Referring to the original work by [2], we have measured the performance in two ways: index construction speed and the final index size. Our implementation which considers unweighted and undirected graphs have achieved a reduction in index size compared with [2]. Proposed Method: First, we implemented the algorithm presented in [2] to compute top-k shortest paths between two vertices. Then we use the sum of top-k shortest paths as the similarity measure to predict future links. This naive approach shows better results compared with other commonly used link prediction methods. k−1 Sk = Σi=0 KSP (s, t, k)[i] (1) Equation 1 shows the similarity measure based on top-k distances where KSP (s, t, k) is the list of top-k shortest paths between vertices s and t. 4 Experimental Results Setup: In these experiments, all networks were treated as undirected unweighted graphs without self-loops and multiple edges. For testing purposes, we randomly sample 60% edges for prediction and 40% for evaluation. The sampling, prediction and evaluation tasks were performed 10 times on each dataset. We use AUROC (Area Under the ROC curve) as an evaluation metric. We used five different datasets from [2] to ease comparisons. Results: The performance of our proposed method is summarized in Table 2. The best performance for each dataset is emphasized. As we can see, Top-4 shortest distance consistently performs better than others except CondMat in which the difference is negligible. Intuitively, one might think bigger k should be a better predictor but our results suggests that small k is enough. Even with such a small k value, 4, we can predict future links more effectively. From the experiments, we can demonstrate that top-k distances capture the structural similarity between vertices better than commonly used measures, namely, Common Neighbors (CN), Jaccard, Adamic/Adar and Preferential Attachment. Table 2: Performance evaluation of predictions on 5 datasets. Statistics of each graph are described in [2]. Facebook-1 Last.fm GrQc HepTh CondMat 5 Top1 0.878481 0.863925 0.853479 0.826561 0.911328 Top4 0.909458 0.88316 0.853527 0.82677 0.911099 Top16 0.899959 0.881586 0.851746 0.824997 0.90812 Top64 0.886813 0.87527 0.84664 0.819686 0.901543 CN 0.834086 0.736351 0.784875 0.730995 0.815252 Jaccard 0.833845 0.733341 0.784865 0.730984 0.815241 Adamic Preferential 0.799192 0.693485 0.721282 0.782929 0.726173 0.720467 0.671017 0.690331 0.75657 0.716568 Conclusions and Future work In this paper, we defined a new similarity metric between two users of social networks based on top-k shortest distances. We also found out through experiments that our simple metric outperforms other common metrics and also a small k suffices to accurately predict future links. Since top-k distances capture important topological properties between vertices, we plan to apply the metric in gene regulation networks to discover unknown relationships among genes which are difficult to infer using other methods. Furthermore, graph database is a growing technology these days and, in some cases, the shortest path implementations is already at their core (for example, Neo4j [1]). It is natural therefore to investigate the results in the context of this development, in order to identify possible improvements in performance gaps. References 1. Neo4j, https://neo4j.com/ 2. Akiba, T., Hayashi, T., Nori, N., Iwata, Y., Yoshida, Y.: Efficient top-k shortestpath distance queries on large networks by pruned landmark labeling (2015), http://www.aaai.org/ocs/index.php/AAAI/AAAI15/paper/view/9320 3. Akiba, T., Iwata, Y., Yoshida, Y.: Fast exact shortest-path distance queries on large networks by pruned landmark labeling. In: Proceedings of the 2013 ACM SIGMOD International Conference on Management of Data. pp. 349–360. SIGMOD ’13, ACM, New York, NY, USA (2013), http://doi.acm.org/10.1145/2463676.2465315 4. Angles, R., Gutierrez, C.: Survey of graph database models. ACM Comput. Surv. 40(1), 1:1–1:39 (Feb 2008), http://doi.acm.org/10.1145/1322432.1322433 5. Cohen, E., Halperin, E., Kaplan, H., Zwick, U.: Reachability and distance queries via 2-hop labels. SIAM Journal on Computing 32(5), 1338–1355 (2003), http://dx.doi.org/10.1137/S0097539702403098 6. Eppstein, D.: Finding the k shortest paths 28(2), 652–673 (Apr 1998), http://epubs.siam.org/sam-bin/dbq/article/29047 7. Goldberg, A.V., Harrelson, C.: Computing the shortest path: A search meets graph theory. In: Proceedings of the Sixteenth Annual ACMSIAM Symposium on Discrete Algorithms. pp. 156–165. SODA ’05, Society for Industrial and Applied Mathematics, Philadelphia, PA, USA (2005), http://dl.acm.org/citation.cfm?id=1070432.1070455 8. Lee, J., Oh, J.C.: Estimating the degrees of neighboring nodes in online social networks. In: International Conference on Principles and Practice of Multi-Agent Systems. pp. 42–56. Springer International Publishing (2014) 9. Liben-Nowell, D., Kleinberg, J.: The link-prediction problem for social networks. Journal of the American Society for Information Science and Technology 58(7), 1019–1031 (2007), http://dx.doi.org/10.1002/asi.20591 10. Qu, Q., Chen, C., Jensen, C.S., Skovsgaard, A.: Space-time aware behavioral topic modeling for microblog posts. IEEE Data Eng. Bull. 38(2), 58–67 (2015), http://sites.computer.org/debull/A15june/p58.pdf 11. Qu, Q., Liu, S., Yang, B., Jensen, C.S.: Integrating non-spatial preferences into spatial location queries. In: Proceedings of the 26th International Conference on Scientific and Statistical Database Management. pp. 8:1–8:12. SSDBM ’14, ACM, New York, NY, USA (2014), http://doi.acm.org/10.1145/2618243.2618247 12. Robinson, I., Webber, J., Eifrem, E.: Graph Databases. O’Reilly Media, Inc. (2013)
8cs.DS
This article will be published in IEEE SMC Magazine, vol. 4, no. 2, 2018 1 Human and Smart Machine Co-Learning with Brain Computer Interface Chang-Shing Lee, Mei-Hui Wang, Li-Wei Ko, Naoyuki Kubota Lu-An Lin, Shinya Kitaoka, Yu-Te Wang, and Shun-Feng Su Abstract Machine learning has become a very popular approach for cybernetics systems, and it has always been considered important research in the Computational Intelligence area. Nevertheless, when it comes to smart machines, it is not just about the methodologies. We need to consider systems and cybernetics as well as include human in the loop. The purpose of this article is as follows: (1) To integrate the open source Facebook AI Research (FAIR) DarkForest program of Facebook with Item Response Theory (IRT), to the new open learning system, namely, DDF learning system; (2) To integrate DDF Go with Robot namely Robotic DDF Go system; (3) To invite the professional Go players to attend the activity to play Go games on site with a smart machine. The research team will apply this technology to education, such as, playing games to enhance the children concentration on learning mathematics, languages, and other topics. With the detected brainwaves, the robot will be able to speak some words that are very much to the point for the students and to assist the teachers in classroom in the future. I. Introduction Rémi Coulom, a freelance developer of Go programs, said “Online games are usually played at a faster pace, which favors the computer over humans,” and he still expected a strong correlation with performance in serious tournament games [1]. Hence, the held special event Human and Smart Machine CoLearning @ IEEE SMC 2017 (http://oase.nutn.edu.tw/IEEESMC2017/, Banff, Canada, Oct. 5, 2017) still has Go games on site, not just playing Go on the Internet. The purpose of the activity in IEEE SMC 2017 is as follows: 1) To integrate the open source Facebook AI Research (FAIR) DarkForest (DF) program of Facebook (USA) [2] with Item Response Theory (IRT) [8] of NUTN, Taiwan, to the new open learning system, namely, Dynamic DF (DDF, Dynamic DarkForest) learning system [3]; 2) To integrate DDF Go with FUJISOFT Robot (led by Kubota Lab., TMU, Japan) namely Robotic DDF Go system; 3) To invite the professional Go players to attend the activity to play Go games on site with a smart machine. Chun-Hsun Chou (9P, Taiwan), Ping-Chiang Chou (6P, Taiwan), and Kai-Hsin Chang (5P, Taiwan) were invited to play Go games with DeepZenGo (Japan). Lu-An Lin (6D, Taiwan), Daisuke Horie (4D, Japan), and Shuji Takemura (1D, Japan) played games with Dynamic Darkforest (DDF, Taiwan) embedded FAIR Darkforest Open Go AI Engine [2]. In addition, the research collaborative team from National Chiao Tung University (NCTU), National University of Tainan (NUTN), University of California San Diego (UCSD), and National Center for High-Performance Computing (NCHC) jointly integrated the Brain Computer Interface (BCI) with the current Dynamic-Darkforest (called BCI-DDF) Go system, which was also firstly demonstrated, to attract more scholars in the brain machine interaction (BMI) area in IEEE SMC conference and then join the SMC society. II. Past held events in the world from 2008 to 2017 Owing to the maturity of deep learning technologies and computer hardware, Google combined them together with Monte Carlo Tree to beat many top professional Go players without handicaps in 2016 and 2017 [4-5]. This year is the first year to hold Human & Smart Machines Co-Learning @ IEEE SMC 2017. However, we have carried out the events of humans playing Go with the computer Go programs for almost a decade [6-7]. Fig. 1 shows the past held events of Human vs. Computer Go Competitions from 2008 to 2017 (https://www.youtube.com/watch?v=UkSOVnbC2Y8) funded by IEEE CIS, IEEE SMC, Taiwanese government, NUTN, and Taiwanese Association for Artificial Intelligence (TAAI). In 1998, the handicaps for the human vs. computer 19×19 game were 29 stones [6]. However, the power of computer Go programs has increased from seven-stone to zero-stone handicap against top professional Go players from 2008 to 2017. III. BCI-DDF Go System In the special event of IEEE SMC 2017, we combined the theory of deep learning with the technology of BCI [9, 10] to demonstrate playing Go. The brainwave technology has been developed for a long time; however, applying it to play Go is the world’s first case in an IEEE conference. The world latest mobile and wireless EEG system is fully utilized in the innovation of the developed BCI-DDF Go system. The wireless system, developed by the research team from Brain Research Center in NCTU, is designed to extract the Go player’s brainwaves when they play and compete with the DDF Go system directly. Fig. 2 shows the two-mode (a competitive learning mode and a predictive learning mode) scenario of Human & Smart Machine Co-Learning @ IEEE SMC 2017, including the invited Go players, computer Go programs, robot Palro, and the developed BCI-DDF Go system. Fig. 3 shows the BCI-DDF Go system diagram that Lu-An Lin (6D) played Go with DDF without using her hand and the robot Palro reported her the next moves suggested by DDF, which were demonstrated in IEEE SMC 2017. This article will be published in IEEE SMC Magazine, vol. 4, no. 2, 2018 2 IEEE SMC 2017, Canada Figure 1. Past held events in the world from 2008 to 2017. Competitive Learning BCI -DDF vs. 9P / Taiwan Japan 5P / Taiwan Predictive Learning Mindo Palro / Japan 6D / Taiwan Figure 2. Two-mode scenario of Human & Smart Machine Co-Learning @ IEEE SMC 2017, including the invited Go players, computer Go programs, robot Palro, and the developed BCI-DDF Go system. We adopted the steady-state visual evoked potential (SSVEP) technology to collect the brain signals from the visual cortex (O1 and O2 channels) and performed the real-time signal processing in the cloud server. Five pilot players have tested the developed BCI-DDF Go system and it can reach about 90% accuracy on a five-class classification task before we held this special event. The demonstration of the BCI-DDF Go system is expressive of the breakthrough in human brain and interaction Figure 3. BCI-DDF Go system diagram that Lu-An Lin (6D, right 1) played Go with DDF without using her hand and the robot Palro reported her the next moves suggested by DDF. with the artificial intelligence. Humans wore a wireless EEG headset and were instructed to gaze at the coded visual stimulus on the screen. The BCI-DDF GO system continuously decoded the on-going EEG to send the move command to the game server. While the Go player is wearing the wireless EEG headset, the technician must check the impedance between the scalp and the EEG electrodes first for collecting the good quality of EEG signals. This article will be published in IEEE SMC Magazine, vol. 4, no. 2, 2018 Fig. 4(a) shows the impedance map of O1 and O2 channels (~100 KOhm). Fig. 4(b) shows the Go player’s EEG signals collected from O1 and O2 channels in which he / she was successfully perceiving the 8Hz visual stimulus for controlling the up Arrow. Then, the DDF communicates with BCI via the WebSocket protocol. The decoded EEGs are sent to the NCHC server in Taiwan, and then to FAIR Darkforest AI Go Engine to control the stone to move. Humans do not need to use their hands to play Go anymore, and even they can play Go with the robot together because the robot is able to predict and report the next suggested moves to them for a reference before playing next move. Epoch Times Calgary also reported this special event by the topic of “The world’s first playing Go using brainwaves in Banff” in 2017 (https://www.youtube.com/watch?v=okMA0Snhj-s). 3 with Chou (9P) and Chou (6P) / Chang (5P), DeepZenGo runs (1 node / E5-2643 v4 x2 / NVIDIA Titan X (Pascal) RAM 128GB / Storage SSD480GB) and (1 node / E5-2623 v3 x2 / NVIDIA Titan X RAM 128GB / Storage SSD480GB), respectively. DeepZenGo won all of the games with three invited professional Go players. Figs. 5(a)-(b) are the games where Chou (9P) as White / DeepZenGo as Black (B + Resign), as well as Chang (5P) as White / DeepZenGo as Black (B + Resign), respectively. (a) (a) Time Series 35 O1 30 Amplitude (mV) 25 20 15 10 5 0 -5 2 2.25 2.125 2.375 2.5 2.625 2.875 2.75 3 Time (s) 50 Amplitude (mV) Amplitude(mV) 40 O2 30 20 10 0 -10 -20 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 Time (s) (b) Figure 4. (a) Impedance check of the EEG channels and (b) Go player’s EEG signals extracted from O1 and O2 channels for controlling the up arrow of the BCI-DDF Go system. IV. Game results One of the top computer Go programs, DeepZenGo, from Japan was used to play with the invited professional Go players, including Chun-Hsun Chou (9P), Ping-Chiang Chou (6P), and Kai-Hsin Chang (5P) for seven games without handicaps. This mode is a learning mode between humans and machines. In order to make DeepZenGo have different strength while it plays (b) Figure 5. (a) Chou (9P) as White, DeepZenGo as Black, and Black won the game with resignation and (b) Chang (5P) as White, DeepZenGo as Black, and Black won the game with resignation.2017. Chou (9P) talked of his experience after the event: “This is my first experience to play Go games with DeepZenGo in public and I am obviously at a disadvantage after the layout of the three games. The rise of AI challenges professional Go players’ domain knowledge of the game layout which has been learned and taught for hundreds of years. I have joined the human vs. computer Go competitions hosted by NUTN and other academic organizations since 2008. In the past, I advised computer Go programs; however, I am advised now instead.” Chang (5P) said “It is an indisputable fact that top-level This article will be published in IEEE SMC Magazine, vol. 4, no. 2, 2018 computer Go programs, like AlphaGo and DeepZenGo, can completely defeat the professional Go players. I hadn’t yet personally experienced the strength of the artificial intelligence. But, I played with DeepZenGo in the special event of IEEE SMC 2017. In the past, the layout and the situation judgment of Go were two of the most difficult problems to solve for computer Go programs; however, now, the top-level machine does much better than humans, not to mention its ability to handle life and death. AI overturns domain knowledge that humans have been learned and taught. In a sense, humans are given a different picture of the world. AI has simulated the rich variety of playing-Go models and led Go to another level.” Lin (6D) was invited to play Go with DDF by brainwaves and cooperated with the robot Palro under the predictive learning mode. Take the game (Lin + Palro as Black, DDF as White) for example. Fig. 6(a) shows the curves of the predictive winning rate and numbers of simulations. Fig. 6(b) shows the inferred game-situation results of the developed FML-based dynamic assessment agent, where B++, B+, U, W+, and W++ represent “Black is obvious at an advantage,” “Black is possible at an advantage,” “Both are in an uncertain situation,” “White is possible at an advantage,” and “White is obvious at an advantage,” respectively [3]. After the event, Lin said that “This is my first time to attend the international conference. Indeed, it is such an amazing experience, in which wearing a Mindo to play Go was a real eye-opener for me. When the sensors in the headgear measured my brainwaves continually, the sparkling points on the monitor caused my eyes a little bit of discomfort. But, this did not dampen my interest in this biometric gadget and enthusiasm for playing Go using brainwaves. This developed technology can be applied in a wide variety of fields; for example, the physically challenged can use it to express their inner thoughts. Glad to have this great opportunity to wear this biometric headgear to play Go and hope to see its further development, https://www.youtube.com/watch?v=6SuZ92z_PBU&feature=y outu.be).” 4 (b) Figure 6. (a) Predictive winning rate and numbers of simulations curves and (b) inferred game-situation results of FML-based dynamic assessment agent for the game that (Lin + Palro) as Black, DDF as White, and Black won the game. Figure 7. Special event venue, where General Chair (Anup Basu, foreground, left 1), Program Chair (Irene Cheng, the left of General Chair), YoPing Huang (the left of Program Chair), and Shinya Kitaoka (background, right 1). (a) Figure 8. Li-Wei Ko, Chang-Shing Lee, Shun-Feng Su, Naoyuki Kubota, Takenori Obo (background, from left to right), Lu-An Lin (6D, foreground) playing Go with DDF, and the robot Palro reported real-time suggested next move to Lin. This article will be published in IEEE SMC Magazine, vol. 4, no. 2, 2018 V. Conclusion Smart machine is one of the main themes of IEEE SMC 2017. A study on Go players competing with and learning from smart machine is sure to attract more attention from worldwide scholars in SMC conferences. There are 15 games played at this special event where the amateur Go players won 8 games against DDF with the small computational resource, but DeepZenGo won all the 7 games against the invited professional Go players. Therefore, DeepZenGo could be the coach of the professional Go players while DDF with small machine could be the student of the amateur Go players as long as we properly and adaptively adjust the computational resources of the intelligent robot. Figs. 7 and 8 show some of the selected photos for this special event. In the future, the research team will apply this technology to education, such as, playing games to enhance children’s concentration on learning mathematics, languages, and other topics. With the detected brainwaves, the robot will be able to speak some words that are very much to the point for the students and to assist the teachers in class in the future. Acknowledgement The event in IEEE SMC 2017 was a great success and we would like to express our heartfelt thanks to everyone who offered any help, joined, as well as watched the games and the BCI-DDF demonstration. We would also like to sincerely thank the IEEE SMC Society, IEEE SMC President Dimitar Filev, Sr. IEEE SMC Past President Philip Chen, the organizing committee of the IEEE SMC 2017, especially General Chair Anup Basu. The authors would like to thank the Ministry of Science and Technology of Taiwan (MOST 106-3114-E-024001) and IEEE SMC’s outreach project for their financial support. Finally, we would like to thank three professional Go players (Chun-Hsun Chou, Kai-Hsin Chang, and Ping-Chiang Chou), Sheng-Chi Yang, Taiwanese government, NCHC of Taiwan, and Osaka Prefecture University (OPU / Japan) for supporting this special event. Li-Wei Ko ([email protected]) received the Ph.D. degree in electrical engineering from National Chiao Tung University, Hsinchu, Taiwan in 2007. Currently, he is an associate professor with the Institute of Bioinformatics and Systems Biology, National Chiao Tung University, Taiwan. Dr. Ko is an AE of IEEE Transactions on Neural Networks and Learning Systems. Naoyuki Kubota ([email protected]) received the D. E. degree from Nagoya University, Nagoya, Japan in 1997. Currently, he is a professor in the Department of System Design, Tokyo Metropolitan University, Tokyo, Japan. Lu-An Lin ([email protected]), a 12-year-old 7th grader in Taiwan, is a 6D amateur Go player. Ever since she was in 6th grade, she has been cooperating with OASE Lab., NUTN, Taiwan to carry out research into smart machine learning of Go and language. Shinya Kitaoka ([email protected]) is one of the DeepZenGo team members and currently, he is an engineer with Dwango Media Village, DWANGO Co., Ltd., Japan. Yu-Te Wang ([email protected]) received the Ph.D. degree in computer science and engineering at the University of California San Diego (UCSD), La Jolla, in 2015. Currently, he is a Staff Research Associate III at the Swartz Center for Computational Neuroscience, UCSD, USA. Shun-Feng Su ([email protected]) received the Ph.D. degrees in electrical engineering from Purdue University, West Lafayette, IN, USA, in 1991. He is currently a Chair Professor with the Department of Electrical Engineering, National Taiwan University of Science and Technology, Taiwan. Dr. Su currently serves as an AE for the IEEE Transactions on Cybernetics, the IEEE Transactions on Fuzzy Systems, and the IEEE Access, a Subject Editor (Electrical Engineering) for the Journal of the Chinese Institute of Engineers, and the Editor-in-Chief for the International Journal of Fuzzy Systems. References [1] About the authors Chang-Shing Lee ([email protected]) received the Ph.D. degree in Computer Science and Information Engineering from the National Cheng Kung University, Tainan, Taiwan, in 1998. He is currently a Professor with the Department of Computer Science and Information Engineering, National University of Tainan. He was awarded Certificate of Appreciation for outstanding contributions to the development of IEEE Standard 1855TM-2016 (IEEE Standard for Fuzzy Markup Language). He was the Emergent Technologies Technical Committee (ETTC) Chair of the IEEE Computational Intelligence Society (CIS) from 2009 to 2010 and serves as the AE of IEEE TCIAIG. Mei-Hui Wang ([email protected]) received M.S. degree in electrical engineering from the Yuan Ze University, Chung-Li, in 1995. She is currently a Researcher with the Ontology Application and Software Engineering (OASE) Laboratory, Department of Computer Science and Information Engineering, National University of Tainan (NUTN), Taiwan. 5 [2] [3] [4] [5] [6] E. Gibney, “Google secretly tested AI bot: updated version of Google DeepMind’s AlphaGo program revealed as mystery online player,” Nature, vol. 541, pp. 142, Jan. 2017. Y. Tian and Y. Zhu, “Better computer Go player with neural network and long-term prediction,” 2016 International Conference on Learning Representations (ICLR 2016), San Juan, Puerto Rico, May 2-4, 2016 (https://arxiv.org/pdf/1511.06410.pdf). C. S. Lee, M. H. Wang, S. C. Yang, P. H. Hung, S. W. Lin, N. Shuo, and N. Kubota, “FML-based dynamic assessment agent for human-machine cooperative system on game of Go,” International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, vol. 25, no. 5, pp. 677-705, 2017. D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. van den Driessche, J. Schrittwieser, I. Antonoglou, V. Panneershelvam, M. Lanctot, S. Dieleman, D. Grewe, J. Nham, N. Kalchbrenner, I. Sutskever, T. Lillicrap, M. Leach, K. Kavukcuoglu, T. Graepel, and D. Hassabis, Mastering the game of Go with deep neural networks and tree search, Nature, vol. 529, pp. 484489, Jan. 2016. D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton, Y. Chen, T. Lillicrap, F. Hui, L. Sifre, G. van den Driessche, T. Graepel, and D. Hassabis, “Mastering the game of Go without human knowledge,” Nature, vol. 550, pp. 354–359, Oct. 2017. C. S. Lee, M. H. Wang, S. J. Yen, T. H. Wei, I. C. Wu, P. C. Chou, C. H. Chou, M. W. Wang, and T. H. Yang, “Human vs. computer Go: review and prospect,” IEEE Computational Intelligence Magazine, vol. 11, no. 3, This article will be published in IEEE SMC Magazine, vol. 4, no. 2, 2018 pp. 67-72, Aug. 2016. C. S. Lee, M. H. Wang, C Chaslot, J. B. Hoock, A. Rimmel, O. Teytaud, S. R. Tsai, S. C. Hsu, and T. P. Hong, “The computational intelligence of MoGo revealed in Taiwan’s computer Go tournaments,” IEEE Transactions on Computational Intelligence and AI in Games, vol. 1, no. 1, pp. 73-89, Mar. 2009. [8] C. S. Lee, M. H. Wang, M. J. Wu, O. Teytaud, and S. J. Yen, “T2FS-based adaptive linguistic assessment system for semantic analysis and human performance evaluation on game of Go,” IEEE Transactions on Fuzzy Systems, vol. 23, no. 2, pp. 400-420, Apr. 2015. [9] C. T. Lin, Y. T. Liu, S. L. Wu, Z. Cao, Y. K. Wang, C. S. Huang, J. T. King, S. A. Chen, S. W. Lu, and C. H. Chuang, “EEG-based braincomputer interfaces,” IEEE Systems, Man, and Cybernetics Magazine, vol. 3, no. 1, pp. 16-26. Oct. 2017. [10] K. P. Thomas and A. P. Vinod, “Toward EEG-based biometric systems,” IEEE Systems, Man, and Cybernetics Magazine, vol. 3, no. 1, pp. 6-15. Oct. 2017. [7] 6
2cs.AI
1 Indistinguishability and Energy Sensitivity of Asymptotically Gaussian Compressed Encryption arXiv:1709.05744v1 [cs.IT] 18 Sep 2017 Nam Yul Yu, Member, IEEE Abstract—The principle of compressed sensing (CS) can be applied in a cryptosystem by providing the notion of security. In information-theoretic sense, it is known that a CS-based cryptosystem can be perfectly secure if it employs a random Gaussian sensing matrix updated at each encryption and its plaintext has constant energy. In this paper, we propose a new CS-based cryptosystem that employs a secret bipolar keystream and a public unitary matrix, which can be suitable for practical implementation by generating and renewing the keystream in a fast and efficient manner. We demonstrate that the sensing matrix is asymptotically Gaussian for a sufficiently large plaintext length, which guarantees a reliable CS decryption for a legitimate recipient. By means of probability metrics, we also show that the new CS-based cryptosystem can have the indistinguishability against an adversary, as long as the keystream is updated at each encryption and each plaintext has constant energy. Finally, we investigate how much the security of the new CS-based cryptosystem is sensitive to energy variation of plaintexts. Index Terms—Compressed encryption, Hellinger distance, indistinguishability, linear feedback shift register (LFSR), probability metrics, self-shrinking generators, total variation distance. I. I NTRODUCTION Compressed sensing (CS) [1]−[3] is to recover a sparse signal from the measurements that are believed to be incomplete. A signal x ∈ RN is called K-sparse if it has at most K-nonzero entries, where K ≪ N . A sparse signal is linearly measured by y = Φx ∈ RM , where Φ is an M × N sensing matrix with M ≪ N . In CS theory, if Φ obeys the restricted isometry property (RIP) [1][3], a stable and robust reconstruction of x can be guaranteed from the incomplete measurement y. The CS reconstruction is accomplished by solving an l1 -minimization problem with convex optimization or greedy algorithms [4]. With efficient measurement and stable reconstruction, the CS technique has been of interest in a variety of research fields, e.g., communications [5]−[7], sensor networks [8]−[10], image processing [11]−[13], radar [14], etc. The CS principle can be applied in a cryptosystem for information security. A CS-based cryptosystem encrypts a plaintext through a CS measurement process, where the sensing matrix is kept secret. The ciphertext can then be decrypted through a CS reconstruction process by a legitimate recipient with the knowledge of the sensing matrix. In [15], Rachlin and Baron proved that a CS-based cryptosystem cannot be perfectly secure in itself, but might be computationally secure. Orsdemir et al. [16] showed that it is computationally secure The author is with the School of of Electrical Engineering and Computer Science (EECS), Gwangju Institute of Science and Technology (GIST), Korea. (e-mail: [email protected]). against a key search technique via an algebraic approach. In [17], Bianchi et al. analyzed the security of a CS-based cryptosystem employing a random Gaussian sensing matrix updated at each encryption. Precisely, they showed that the cryptosystem with the one-time sensing random Gaussian matrix can be perfectly secure, as long as each plaintext has constant energy. A similar analysis has been made for a CS-based cryptosystem having a circulant sensing matrix for efficient CS processes [18][19]. In [20] and [21], wireless channel characteristics could be exploited for wireless security of CS-based cryptosystems. The CS technique can also be applied in database systems [22], where random noise has been intentionally added to CS measurements for differential privacy. In practice, a variety of CS-based cryptosystems concerning the security of multimedia, imaging, and smart grid data have been suggested in [23]−[29]. In this paper, we propose a new CS-based cryptosystem that employs a secret bipolar keystream and a public unitary matrix, which can be suitable for practical implementation by generating and renewing the keystream at each encryption in a fast and efficient manner. The keystream generator, based on a linear feedback shift register (LFSR), plays a crucial role in the efficient implementation. We demonstrate that the entries of the sensing matrix are asymptotically Gaussian distributed if the plaintext length is sufficiently large. With the sensing matrix, it is obvious that the new CS-based cryptosystem, named as the asymptotically Gaussian one-time sensing (AG-OTS) cryptosystem, theoretically guarantees a stable and robust CS decryption for a legitimate recipient. For security analysis, we study the indistinguishability [30] of the AG-OTS cryptosystem. The total variation (TV) distance [31][32] between probability distributions of ciphertexts conditioned on a pair of plaintexts is examined as a security measure for the indistinguishability, where the upper and lower bounds on the TV distance are developed by the Hellinger distance [31][32]. With the probability metrics, we examine the success probability of an adversary to distinguish a pair of potential plaintexts from a given ciphertext. By proving that the success probability of any kind of attack is at most that of a random guess, we demonstrate that the AG-OTS cryptosystem can have the indistinguishability, as long as each plaintext has constant energy. Therefore, the AG-OTS cryptosystem, if it has a normalization step before encryption for equalizing the plaintext energy, can be computationally secure. Finally, we investigate how much the security of the AGOTS cryptosystem is sensitive to energy variation of plaintexts. It is worth studying the energy sensitivity, since one might need to assign unequal energy for plaintexts in the presence of noise, depending on the reliability demands. As a consequence, 2 we develop sufficient conditions on the minimum energy ratio, the plaintext length, and the maximum plaintext-tonoise power ratio, respectively, to achieve the asymptotic indistinguishability of the AG-OTS cryptosystem having unequal plaintext energy. Since the analysis relies on the Gaussianity of the sensing matrix, the results of the energy sensitivity can also be applicable to the Gaussian one-time sensing (G-OTS) cryptosystem in [17]. This paper is organized as follows. In Section II, we propose a new CS-based cryptosystem employing a secret bipolar keystream, where the sensing matrix turns out to be asymptotically Gaussian. Also, we discuss an LFSR-based efficient keystream generation for the cryptosystem. Section III introduces the indistinguishability along with the probability metrics of total variation (TV) and Hellinger distances. For security analysis, Section IV studies the indistinguishability and the energy sensitivity of the new CS-based cryptosystem in the presence of noise. Section V presents numerical results to demonstrate the security of the new CS-based cryptosystem. Finally, concluding remarks will be given in Section VI. Notations: A matrix (or a vector) is represented by a boldface upper (or lower) case letter. UT and |U| denote the transpose and the determinant of a matrix U, respectively. U(k, t) is an entry of an M × N matrix U in the kth row and the tth column, where 0 ≤ k ≤ M − 1 and 0 ≤ t ≤ N − 1. Also, U(k, :) denotes the kth row vector of U, while U(:, t) is the tth column vector of U. diag(s) is a diagonal matrix whose diagonal entries are from a vector s. An identity matrix is denoted by I, where the dimension is determined in the context. D denotes an N ×N discrete-cosine transform (DCT) matrix, where DDT = DT D = N I. For an N -dimensional vector x = (x1 , · · · , xN )T ∈ RN , the lp -norm of x is denoted P 1 N p p , where 1 ≤ p < ∞. If the by ||x||p = |x | k k=1 context is clear, ||x|| denotes the l2 -norm of x. A vector n ∼ N (0, σ 2 I) is a Gaussian random vector with mean 0 = (0, · · · , 0)T and covariance σ 2 I. Finally, E[·] denotes the average of a random vector or a random matrix. II. S YSTEM M ODEL In [17], the authors presented the Gaussian one-time sensing (G-OTS) cryptosystem, where a random Gaussian sensing matrix is used only once for each encryption, and renewed for the next. In information-theoretic sense, they showed that if each plaintext has constant energy, the G-OTS cryptosystem can be perfectly secure, which implies the indistinguishability [30] that will be discussed in next section. In practice, generating the Gaussian entries at each encryption may require high complexity and large memory for CS encryption and decryption. For efficient implementation, this section proposes a new CS-based cryptosystem in which the sensing matrix employs a bipolar keystream. A. Asymptotically Gaussian Sensing Matrices Definition 1: Let U ∈ RN ×N be a public unitary matrix, i.e., UT U = UUT = N I, where each element of U has the magnitude of O(1). Let S be a secret M × N matrix, where we assume that each element takes ±1 independently and uniformly at random. Then, a new CS-based cryptosystem has the sensing matrix of 1 SU. Φ= √ MN Theoretically, each element of S can be taken from the random Bernoulli distribution. In practice, however, we consider a keystream generator of stream ciphers to generate it in a fast and efficient manner. Employing an efficient keystream generator allows us to construct and update S at each encryption with low complexity and small memory. Since a keystream for a stream cipher is designed to have nice pseudorandomness properties [33], such as balance, large period, low autocorrelation, large linear complexity, etc., we assume that each entry of S from the keystream takes ±1 independently and uniformly at random, which facilitates the reliability and security analysis of the new CS-based cryptosystem. Theorem 1: In Definition 1, the elements of Φ follow the Gaussian distribution asymptotically for a sufficiently large N . Proof : Each row of Φ is represented by 1 S(k, :)U Φ(k, :) = √ rM N N 1 U = · √ diag(S(k, :)) √ , M N N k = 1, · · · , M (1) where S(k, :) is the kth row vector of S, and 1 = (1, · · · , 1) is all one row vector of length N , respectively. In (1), √1N is a row of a unit-norm   row matrix with absolute magnitude of all 1 √ entries of O . Also, √UN is a unit-norm column matrix N with the maximum absolute magnitude of entries of o(1). With the structure, Theorem III.1 of [34] shows that the elements of Φ(k, :) are asymptotically Gaussian if N is sufficiently large, which completes the proof. ✷ The asymptotic Gaussianity of Theorem 1 also holds if the elements of S are generated by an efficient keystream generator, under the assumption that each one takes ±1 independently and uniformly at random. The assumption will be validated by the numerical results of Section V. B. Keystream Generation For the secret matrix S of Definition 1, we employ a keystream generator based on a linear feedback shift register (LFSR), to generate the elements in a fast and efficient manner. As an example, we introduce the self-shrinking generator (SSG) [35]. Definition 2: Assume that an L-stage LFSR generates a binary m-sequence of a = (a0 , a1 , · · · ). With a clock-controlled operation, the self-shrinking generator outputs bt = a2i+1 if a2i = 1, and discards a2i+1 if a2i = 0. Then, we obtain a bipolar keystream of s = (s0 , s1 , · · · ), where st = (−1)bt for t = 0, 1, · · · , which will be arranged as the elements of S. The SSG keystream generation requires a simple structure of an L-stage LFSR along with a clock-controlled operator. 3 Moreover, the SSG keystream possesses nice pseudorandomness properties [33], such as balance, large period, and large linear complexity. Meier and Staffelbach [35] showed that the SSG keystream is balanced, and has the period of at least 2⌊L/2⌋ and the linear complexity of at least 2⌊L/2⌋−1 , respectively. Although the SSG keystream generator is considered in this paper, any other LFSR-based keystream generator can also be applied for the new CS-based cryptosystem. When each element of S is obtained by a keystream generator, the initial seed (or state) of the generator is essentially the key of the new CS-based cryptosystem. Therefore, the key should be kept secret between a sender and a legitimate recipient, while the structure of the keystream generator can be publicly known. C. AG-OTS Cryptosystem From Definition 1, the new CS-based cryptosystem encrypts a K-sparse plaintext x1 by producing a ciphertext y = Φx = √ 1 SUx, where S is updated at each encryption. Under the MN presence of noise, a legitimate recipient and an adversary have a noisy ciphertext r = Φx+n, where n ∼ N (0, σ 2 I). As Φ is asymptotically Gaussian and S is updated at each encryption, the new CS-based cryptosystem will be called the asymptotically Gaussian one-time sensing (AG-OTS) cryptosystem throughout this paper. Table I summarizes the symmetric-key AG-OTS cryptosystem proposed in this paper. The reliability and stability of the AG-OTS cryptosystem for a legitimate recipient is straightforward from the RIP result [36] of a random Gaussian matrix, under the fact that Φ is Gaussian for a sufficiently large N . Proposition 1: [36] For a legitimate recipient, if N is sufficiently large, the AG-OTS cryptosystem theoretically guarantees a stable and robust CS decryption with bounded errors of a K-sparse plaintext, as long as M = O (K log(N/K)). III. S ECURITY M EASURE This section introduces a security measure of the indistinguishability of a CS-based cryptosystem. To examine the indistinguishability, we also discuss the probability metrics of total variation (TV) and Hellinger distances. B. Total Variation (TV) and Hellinger Distances In this paper, we make use of the total variation (TV) distance [31] to evaluate the performance of an adversary in the indistinguishability experiment of Table II. In the experiment, let dTV (p1 , p2 ) be the TV distance between the probability distributions p1 = Pr(r|x1 ) and p2 = Pr(r|x2 ). Then, it is readily checked from [32] that the probability that an adversary can successfully distinguish the plaintexts by any kind of test D is bounded by 1 dTV (p1 , p2 ) + (2) 2 2 where dTV (p1 , p2 ) ∈ [0, 1]. Therefore, if dTV (p1 , p2 ) is zero, the probability of success is at most that of a random guess, which leads to the indistinguishability [30]. Since computing dTV (p1 , p2 ) directly is difficult [37], we may employ an alternative distance metric to bound the TV distance. In particular, the Hellinger distance [31], denoted by dH (p1 , p2 ), is useful by giving both upper and lower bounds on the TV distance [38], i.e., q d2H (p1 , p2 ) ≤ dTV (p1 , p2 ) ≤ dH (p1 , p2 ) 2 − d2H (p1 , p2 ) (3) where dH (p1 , p2 ) ∈ [0, 1]. Moreover, if a ciphertext r conditioned on xh , is a jointly Gaussian random vector with zero mean and the covariance matrix Ch , where h = 1 and 2, the Hellinger distance between the multivariate Gaussian distributions p1 and p2 is given by [39][40] s 1 1 |C1 | 4 |C2 | 4 dH (p1 , p2 ) = 1 − (4) 1 |C3 | 2 pd ≤ 2 . For the formal definitions and properties where C3 = C1 +C 2 of the TV and the Hellinger distances, readers are referred to [31], [32], and [37]. Throughout this paper, we use (2) − (4) to examine the success probability of the indistinguishability experiment for the AG-OTS cryptosystem, by taking the Gaussian distributed ciphertexts into account. IV. S ECURITY A NALYSIS A. Indistinguishability Assume that a cryptosystem produces a ciphertext by encrypting one of two possible plaintexts of the same length. Then, the cryptosystem is said to have the indistinguishability [30], if no adversary can determine in polynomial time which of the two plaintexts corresponds to the ciphertext, with probability significantly better than that of a random guess. In other words, if a cryptosystem has the indistinguishability, an adversary is unable to learn any partial information of the plaintext in polynomial time from a given ciphertext. Table II describes the indistinguishability experiment [30] in the presence of an eavesdropper, which will be used to investigate the indistinguishability of a CS-based cryptosystem in this paper. 1 In general, x can be K-sparse in an arbitrary orthonormal basis Ψ, i.e., x = Ψθ with ||θ||0 ≤ K, where Ψ 6= √1 UT . For simplicity, we assume N Ψ = I in this paper. In this section, we show that the AG-OTS cryptosystem can be indistinguishable, as long as each plaintext has constant energy. Moreover, we study how much the security of the AGOTS cryptosystem is sensitive to energy variation of plaintexts. A. Indistinguishability Recall the indistinguishability experiment of Table II. Given 1 a plaintext xh , E[r|xh ] = E[Φ]xh + E[n] = √MN E[S] · Uxh = 0 from E[S] = 0, where h = 1 and 2. In the following, Lemma 1 derives the covariance matrix of r conditioned on xh , by exploiting the independency and the uniformity of the entries of S. Lemma 1: In the AG-OTS cryptosystem, the covariance matrix of r conditioned on xh is given by   ||xh ||2 Ch = E[rrT |xh ] = (5) + σ2 I M 4 TABLE I S YMMETRIC - KEY AG-OTS C RYPTOSYSTEM Public: Secret: Keystream generation: CS encryption: CS decryption: Unitary matrix U, Structure of a keystream generator Initial seed k ∈ {0, 1}L of a keystream generator With the initial seed k, a keystream generator creates a bipolar keystream of length M N . A secret matrix S ∈ {−1, +1}M ×N is constructed by arranging the keystream, and updated at each encryption by a new keystream. 1 SU. With a plaintext x ∈ RN , a ciphertext is produced by y = Φx ∈ RM , where Φ = √M N Given a noisy ciphertext r = Φx + n, the plaintext x is reconstructed by a CS recovery algorithm with the knowledge of S. TABLE II I NDISTINGUISHABILITY E XPERIMENT FOR A CS- BASED C RYPTOSYSTEM An adversary creates a pair of plaintexts x1 and x2 of the same length, and submits them to a CS-based cryptosystem. The CS-based cryptosystem encrypts a plaintext xh by randomly selecting h ∈ {1, 2}, and gives a noisy ciphertext r = Φxh + n back to the adversary. Given the ciphertext r, the adversary carries out a polynomial time test D : r → h′ ∈ {1, 2}, to figure out the corresponding plaintext. The adversary passes the experiment if h′ = h, or fails otherwise. Step 1: Step 2: Step 3: Decision: where h = 1 and 2. From (5), it is obvious that   C1 + C2 ||x1 ||2 + ||x2 ||2 = + σ 2 I. C3 = 2 2M bh = Uxh = (b bh,N )T for h = 1 and 2, Proof : Let x xh,1 , · · · , x 2 respectively, where ||b xh || = N ||xh ||2 . Also, let sk = S(:, k) and sl = S(:, l) are the kth and the lth column vectors of S, respectively. Since the elements of S and n are independent to each other,    1 T T T Ch = E SUxh · xh U S xh + E[nnT ] MN " # N N X  1 X T =E x bh,k sk · x bh,l sl xh + σ 2 I (6) MN = 1 MN where E k=1 N N XX k=1 l=1  sk sTl  l=1   x bh,k x bh,l E sk sTl + σ 2 I =  I, 0, if k = l, if k = 6 l as the entries of sk and sl take ±1 independently and uniformly at random. Thus, (6) yields ! N X 1 2 Ch = x bh,k · I + σ 2 I MN k=1   ||xh ||2 2 = +σ I M which completes the proof. ✷ In Lemma 1, note that the derivation of covariance matrices does not rely on the asymptotic Gaussianity of Φ. Instead, the covariance matrices are non-asymptotic results, obtained by exploiting the independency and the uniformity of the elements of S. Using the covariance matrices of Lemma 1, we can develop upper and lower bounds on the TV distance in the AG-OTS cryptosystem, which is the main contribution of this paper. Theorem 2: In the AG-OTS cryptosystem, assume that the plaintext length N is sufficiently large such that Φ can be asymptotically Gaussian by Theorem 1. In the indistinguishability experiment, let dTV (p1 , p2 ) be the TV distance between probability distributions of ciphertexts conditioned on a pair of plaintexts in the AG-OTS cryptosystem. Let xmin and xmax be the plaintexts that have the minimum and ||xmin ||2 maximum possible energies, respectively, where γ = ||x 2 max || 2 max || is the is the minimum energy ratio and PNRmax = ||xMσ 2 maximum plaintext-to-noise power ratio, respectively, of the cryptosystem. Then, the worst-case lower and upper bounds on dTV (p1 , p2 ) are given by  M4  4γe , dTV,low = 1 − (γe + 1)2 s (7)   M2 4γe dTV,up = 1 − , (γe + 1)2 respectively, where γe = 1 + γ · PNRmax . 1 + PNRmax (8) Proof : In the indistinguishability experiment of Table II, let us 2 h || consider a pair of plaintexts x1 and x2 , where PNRh = ||x Mσ2 for h = 1 and 2. From the covariance matrices of Lemma 1, M  ||xh ||2 + σ2 = σ 2M · (PNRh + 1)M |Ch | = M for each h. Obviously, |C3 | = σ 2M ·  PNR1 + PNR2 +1 2 M . 5 In (4), 1 |C3 | 2 ! M4 where 1 + γ · PNR2 1 + PNR1 = γe = 1 + PNR2 1 + PNR2 √ 2 1 || and γ = ||x 1 − Γ, (3) yields the ||x2 ||2 . With dH (p1 , p2 ) = lower and upper bounds of the form of (7). Without loss of generality, we may assume ||x1 ||2 ≤ ||x2 ||2 , which yields 0 ≤ γ ≤ 1. As the lower and upper bounds turn out to be monotonically decreasing over γ ∈ [0, 1], we can redefine γ = ||xmin ||2 1+γ·PNRmax ||xmax ||2 and γe = 1+PNRmax , 0 ≤ γe ≤ 1 with x1 = xmin and x2 = xmax , to obtain the worst-case bounds, which completes the proof. ✷ In (8), γe is a general definition of the energy ratio covering noisy cases, which will be called the effective energy ratio in this paper. For security analysis, we assume that both a legitimate recipient and an adversary have the same energy ratio γ and the same PNRmax in the AG-OTS cryptosystem. Theorem 2 shows that the indistinguishability of the AGOTS cryptosystem depends on the ciphertext length M , the minimum energy ratio γ, and the maximum plaintext-to-noise ratio PNRmax , irrespective of the plaintext length N and the sparsity K. In particular, if γ = γe = 1, the indistinguishability can be guaranteed for the AG-OTS cryptosystem, regardless of M and PNRmax . Corollary 1: If each plaintext has constant energy or γ = 1, the AG-OTS cryptosystem has the indistinguishability, since the success probability of the indistinguishability experiment is at most 0.5 from (2), thanks to dTV (p1 , p2 ) = 0 for dTV,low = dTV,up = 0. In the AG-OTS cryptosystem, Corollary 1 ensures that no adversary can learn any partial information about the plaintext from a given ciphertext, as long as each plaintext has constant energy, which is also the case in the G-OTS cryptosystem of [17]. To achieve the indistinguishability, therefore, a normalization step for equalizing the plaintext energy is implicitly required before CS encryption in the AG-OTS cryptosystem of Table I. Since it also offers a practical benefit from the efficient keystream generation, the AG-OTS cryptosystem can be a promising option for information security, by guaranteeing the indistinguishability, reliability, and efficiency in a CS framework. B. Energy Sensitivity Theorem 2 implies that the indistinguishability of the AGOTS cryptosystem can be sensitive to the minimum energy ratio γ. Figure 1 sketches the upper and lower bounds of (7) over γ at PNRmax = ∞ in the noiseless AG-OTS cryptosystem. It indicates that the TV distance increases as γ gets away from 1. In particular, if M gets larger, the TV distance approaches to 1 more quickly as γ decreases. Such a behavior of the TV distance suggests that if γ is far less than 1, an adversary may 1 0.9 0.8 0.7 0.6 TV distance Γ= (PNR1 + 1)(PNR2 + 1) =  2 PNR1 +PNR2 +1 2  M4  4γe = (γe + 1)2 0.5 0.4 0.3 M = 32: dTV,up M = 32: dTV,low 0.2 M = 64: d TV,up M = 64: dTV,low 0.1 M = 128: dTV,up M = 128: dTV,low 0 0 0.1 0.2 0.3 0.4 0.5 γ 0.6 0.7 0.8 0.9 1 Fig. 1. Upper and lower bounds on the TV distance over γ in the noiseless AG-OTS cryptosystem. 1 0.9 0.8 Upper bound on TV distance 1 1 |C1 | 4 · |C2 | 4 0.7 0.6 0.5 Noiseless PNR = 10 dB max PNR = 0 dB PNR = −10 dB max max 0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 0.4 0.5 γ 0.6 0.7 0.8 0.9 1 Fig. 2. Upper bounds on the TV distance over γ in the noisy AG-OTS cryptosystem, where M = 64. be able to detect a correct plaintext in the indistinguishability experiment with a significantly high probability of success, which implies that the AG-OTS cryptosystem may not be indistinguishable. In addition, Figure 2 shows the upper bounds of (7) over γ for various PNRmax in the noisy AG-OTS cryptosystem, where M = 64. In the figure, the bounds are sensitive to γ for each PNRmax , as in the noiseless case of Figure 1. Moreover, the bound itself is smaller at less PNRmax , which implies that an adversary may have a difficulty in distinguishing plaintexts at low PNRmax , due to the low TV distance. As a result, it appears that the security of the AG-OTS cryptosystem would 6 1 0.35 ρmax (γ = 0.95) ρmax (γ = 0.96) ρmax (γ = 0.97) 0.3 0.95 ρmax (γ = 0.98) ρmin 0.25 Compression ratio γmin 0.9 0.85 Achievable region (theory) for K = 8 at γ = 0.98 0.2 0.15 0.8 0.1 ǫN ǫN ǫN ǫN 0.75 0.7 0 32 64 96 128 M 160 192 = = = = 1 N log N N √1 N 1 log N 224 0.05 0 500 256 1500 2000 N Fig. 3. Minimum energy ratio required for the asymptotic indistinguishability of the AG-OTS cryptosystem at PNRmax = 20 dB, where N = 512. be more sensitive to the energy ratio γ at higher PNRmax . In summary, the AG-OTS cryptosystem may not be able to achieve the indistinguishability, unless each plaintext has constant energy. In what follows, we study how much the security of the AG-OTS cryptosystem is sensitive to energy variation of plaintexts. It is worth studying the energy sensitivity, since one might need to assign unequal energy for each plaintext in the presence of noise, depending on the reliability demands. Theorems 3 − 5 present sufficient conditions for the minimum energy ratio γ, the plaintext length M , and the maximum plaintext-to-noise ratio PNRmax , respectively, to guarantee the asymptotic indistinguishability for the AG-OTS cryptosystem. Theorem 3: When M and PNRp max are given, let ϕ = (1 − 2 4ǫ2N )− M and γe,min = 2ϕ − 1 − 2 ϕ(ϕ − 1), respectively. If the minimum energy ratio γ satisfies γ ≥ γe,min − (1 − γe,min ) · PNR−1 max , γmin , 1000 (9) the success probability of the indistinguishability experiment is pd ≤ 0.5 + ǫN , where ǫN vanishes as the plaintext length N increases. In other words, the AG-OTS cryptosystem is asymptotically indistinguishable for a sufficiently large N , as long as γ ≥ γmin for given M and PNRmax . 2 2 M e Proof : When M is given, we have (γe4γ +1)2 ≥ 1 − 4ǫN from dTV,up ≤ 2ǫN . The inequality turns into γe2 − 2(2ϕ − 1)γe + 1 ≤ 0, which holds if γe ≥ γe,min , or equivalently if γ ≥ γmin in (9). Consequently, if the sufficient condition of (9) is met, the success probability of the indistinguishability experiment is pd ≤ 0.5 + ǫN by (2), which completes the proof. ✷ In Theorem 3, γmin is the minimum energy ratio required for the asymptotic indistinguishability of the AG-OTS cryptosystem. Figure 3 displays γmin over M in the AG-OTS cryptosystem at PNRmax = 20 dB, where N = 512. In the figure, γmin is sketched for various ǫN = log1N , √1N , logNN , Fig. 4. Compression ratios for the asymptotic indistinguishability and the reliability of the √ AG-OTS cryptosystem at PNRmax = 20 dB, where K = 8 and ǫN = 1/ N . and N1 . The figure reveals that the minimum energy ratio required for the asymptotic indistinguishability approaches to 1 as the ciphertext length M increases. In particular, if the AG-OTS cryptosystem allows larger energy variation for plaintexts, the asymptotic indistinguishability can be achieved at a lower rate over N . Theorem 4: When γ and PNRmax are given, recall γe = 1+γ·PNRmax 4γe 1+PNRmax . Let Cγe = log (γe +1)2 ≤ 0, where the equality holds if and only if γ = 1. If the ciphertext length M satisfies M≤ 2 log(1 − 4ǫ2N ) , Mmax , Cγe then pd ≤ 0.5 + ǫN , which implies that the AG-OTS cryptosystem is asymptotically indistinguishable for a sufficiently large N , as long as M ≤ Mmax for given γ and PNRmax . Proof : When γe is given from γ and PNRmax , the proof is similar to that of Theorem 3 from dTV,up ≤ 2ǫN . ✷ Figure 4 depicts the maximum compression ratio ρmax = Mmax over N for the AG-OTS cryptosystem to be asympN totically indistinguishable at PNRmax = 20 dB with pd ≤ 0.5+ǫN , where ǫN = √1N . We also sketch the minimum com- pression ratio ρmin = 2K log(N/K) for reliable CS decryption2 N from a random Gaussian sensing [41] with K = 8, to compare the requirements for the asymptotic indistinguishability and the reliability. Note that if γ = 1 or the plaintexts have constant energy, the indistinguishability can be achieved at any compression ratio. Meanwhile, if γ < 1, the compression ratio of the AG-OTS cryptosystem must be at most ρmax for the asymptotic indistinguishability. In particular, if ρmax < ρmin , the cryptosystem may not be valid at least in theory for the corresponding N , since the indistinguishability cannot be compatible with the reliability. Thus, Figure 4 shows that if 2 This is a theoretical ratio in noiseless recovery. 7 15 γ = 0.99*γe,min γ = 0.98*γe,min γ = 0.95*γe,min γ = 0.9*γ Upper bound on PNRmax (dB) 10 e,min 5 0 −5 −10 0 32 64 96 128 M 160 192 224 256 Fig. 5. Upper bounds on PNRmax for the asymptotic indistinguishability of √ the AG-OTS cryptosystem, where N = 512 and ǫN = 1/ N . γ = 0.98, the AG-OTS cryptosystem at PNRmax = 20 dB can achieve both reliability and security for the plaintexts of at most K = 8 nonzero entries only at the compression ratios of the achievable (shaded) region. It also shows that if γ ≤ 0.96, the AG-OTS cryptosystem has no theoretically achievable region for N > 500, where the reliability and the indistinguishability cannot be guaranteed simultaneously. In (8), note that 1−γ γe = γ + 1 + PNRmax (10) where γe ≥ γ for γ ∈ [0, 1]. Since the upper bound of (7) is monotonically decreasing over γe ∈ [0, 1], (10) implies that the upper bound on the TV distance is lower in noisy case (PNRmax < ∞) than in noiseless case (PNRmax = ∞). Ultimately, it points out that the presence of noise improves the security of the AG-OTS cryptosystem by lowering the success probability of an adversary in the indistinguishability experiment. Moreover, one can increase γe by reducing PNRmax in (10) for a given γ, which indicates that the AGOTS cryptosystem will be more secure for less PNRmax . With given γ and M , Theorem 5 presents the largest possible PNRmax to guarantee the asymptotic indistinguishability for the AG-OTS cryptosystem, where the proof is straightforward from γe ≥ γe,min in (10). Theorem 5: In the AG-OTS cryptosystem, assume that the minimum energy ratio is given as γ < γe,min for a given M , where γe,min is the minimum effective energy ratio defined in Theorem 3. Then, the asymptotic indistinguishability can be achieved for a sufficiently large N , if PNRmax ≤ 1 − γe,min . γe,min − γ Note that if γ ≥ γe,min , the AG-OTS cryptosystem is asymptotically indistinguishable, regardless of PNRmax , due to γe ≥ γ ≥ γe,min . Figure 5 displays the upper bounds on PNRmax of Theorem 5 for various γ < γe,min , where N = 512 and ǫN = √1N . From (10), it is clear that if PNRmax is sufficiently high, γe ≈ γ < γe,min from which the asymptotic indistinguishability cannot be achieved from Theorem 3. Figure 5 points out that we need to increase γe by reducing PNRmax below the upper bound for each given γ, to achieve the asymptotic indistinguishability of the AG-OTS cryptosystem. However, it appears that the largest possible PNRmax is relatively low for a reliable CS decryption. For the AG-OTS cryptosystem, therefore, it is an important issue to keep the energy variation of plaintexts as low as possible. In conclusion, it turned out that the security of the AGOTS cryptosystem is highly sensitive to the energy ratio of plaintexts. The indistinguishability can be achieved only if all the plaintexts have equal and constant energy. Therefore, if the AG-OTS cryptosystem is to be indistinguishable nonasymptotically, it is essential that each plaintext should be normalized before CS encryption to have constant energy. By analyzing the energy sensitivity, we presented the sufficient conditions of Theorems 3 − 5 for the asymptotic indistinguishability of the AG-OTS cryptosystem with unequal plaintext energy. However, we found that even the asymptotic indistinguishability can be achieved only if the plaintexts have low energy variation for most M , N , and PNRmax . As the analysis technique utilizes the result of Theorem 2 based on the Gaussianity of the sensing matrix, the energy sensitivity of this paper can also be valid for the G-OTS cryptosystem, which has never been discussed in [17]. V. N UMERICAL R ESULTS This section presents numerical results to demonstrate the indistinguishability and the energy sensitivity of the AG-OTS cryptosystem. In numerical experiments, each plaintext x has at most K nonzero entries, where the positions are chosen uniformly at random and the coefficients are taken from the 1 SU, Gaussian distribution. In CS encryption, Φ = √MN where U = D is the discrete cosine transform (DCT) matrix. Each element of the secret matrix S is taken from a bipolar keystream obtained by the self-shrinking generator (SSG) with a 128-stage LFSR. For comparison, we test with S whose elements are taken from the random Bernoulli distribution. We assume that a ciphertext is available for both an adversary 2 and a legitimate recipient with the same PNR = ||x|| Mσ2 . For CS decryption, the CoSaMP recovery algorithm [42] is employed for a legitimate recipient to decrypt each ciphertext with the knowledge of S. Meanwhile, we assume that an adversary can attempt any kind of detection in polynomial time, to pass the indistinguishability experiment by distinguishing a pair of plaintexts from a given ciphertext. Figure 6 displays the quantile-quantile (QQ) plots of the √ entries of total 100 matrices of M Φ in the AG-OTS cryptosystem, where N = 512 and M = 64. In Figure 6(a), each entry of S is taken from the random Bernoulli distribution taking ±1 independently and uniformly at random, while S of Figure 6(b) is from the bipolar SSG keystream. Since both QQ-plots are linear with slope 1, it appears that the entries 8 (b) S is a matrix from the SSG keystream 6 4 4 Quantiles of M1/2Φ Quantiles of M1/2Φ (a) S is a random Bernoulli matrix 6 2 0 −2 −4 2 0 −2 −4 −6 −6 −4 −2 0 2 Standard Normal Quantiles 4 −6 −6 6 Fig. 6. Quantile-Quantile plots (QQ-plots) of the entries of total 100 matrices of √ −4 −2 0 2 Standard Normal Quantiles 4 6 M Φ’s in the AG-OTS cryptosystem, where N = 512 and M = 64. (b) S is a matrix from the SSG keystream (a) S is a random Bernoulli matrix 70 70 60 60 50 50 1 column index column index 0.8 40 30 0.6 40 30 20 20 10 10 0.4 0.2 0 0 10 20 30 40 row index 50 60 70 0 0 10 20 30 40 row index 50 60 70 0 Fig. 7. Covariance matrices of M · E[rrT |x] in the AG-OTS cryptosystem at PNR = 20 dB, where N = 512, M = 64, and K = 8. For a given x with ||x||2 = 1, total 10000 matrices of S have been tested for the average. √ of M Φ follow the normal distribution in both cases of S. The figure gives a numerical evidence that Φ of the AGOTS cryptosystem is asymptotically Gaussian for a sufficiently large N , even if S is generated in a pseudorandom fashion by the SSG. Figure 7 illustrates the covariance matrices of M C = M · E[rrT |x] in the AG-OTS cryptosystem at PNR = 20 dB, where N = 512, M = 64, and K = 8. In the experiment, total 10000 matrices of S have been tested for the average with a given x of ||x||2 = 1. In the figure, the dark areas indicate the off-diagonal entries of each covariance matrix having very small magnitudes less than 0.04, whereas the white cells represent the diagonal components of significant values, determined by the plaintext energy and the noise variance. Figure 7 numerically confirms that the covariance analysis of Lemma 1 is valid for the AG-OTS cryptosystem, whether S is a random Bernoulli matrix or a matrix from the SSG keystream. Figure 8 displays the upper and lower bounds of Theorem 2 on the TV distance over γ in the AG-OTS cryptosystem at PNRmax = 20 dB, where N = 512, M = 64, and K = 8. In the experiment, we computed the bounds of (3) using the covariance matrices obtained by testing total 106 matrices of S, where each entry of S is taken from the random Bernoulli distribution or the SSG keystream. In both cases of S, the figure shows that the bounds from the experiment are well matched to the theoretical results of Theorem 2. In summary, Figures 6−8 validate our assumption of the independency and the uniformity of the elements of S from the SSG keystream through the numerical experiments. Figure 9 displays the success probabilities over the ciphertext length M in the AG-OTS cryptosystem at PNRmax = 20 dB, where N = 512. For an adversary, it sketches the upper bounds on the success probability of the indistinguishability experiment, obtained by (2) from the upper bound of Theorem 2. For comparison, we also sketch the empirical success probabilities of a legitimate recipient, where we tested total 10000 plaintexts each of which has at most K = 8 nonzero entries and the energy ||x||2 = α||xmax ||2 with α uniformly distributed in [γ, 1]. In CS encryption, each entry 9 1 0.9 0.9 0.8 0.8 0.7 0.7 Success probability 1 TV distance 0.6 0.5 0.4 0.3 0.2 0.1 0 0.1 Upper bound (experiment): random Bernoulli Lower bound (experiment): random Bernoulli Upper bound (experiment): SSG Lower bound (experiment): SSG Upper bound (Theorem 2) Lower bound (Theorem 2) 0.4 0.2 0.3 0.4 0.5 γ 0.6 0.1 0.7 0.8 0.9 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 Legitimate recipient (K = 8): γ = 0.9 Legitimate recipient (K = 8): γ = 0.95 Legitimate recipient (K = 8): γ = 1.0 Adversary (Theorem 2): γ = 0.9 Adversary (Theorem 2): γ = 0.95 Adversary (Theorem 2): γ = 1.0 0.2 0.1 32 48 64 80 Legitimate recipient (K = 8): γ = 0.9 Legitimate recipient (K = 8): γ = 0.95 Legitimate recipient (K = 8): γ = 1.0 Adversary (Theorem 2): γ = 0.9 Adversary (Theorem 2): γ = 0.95 Adversary (Theorem 2): γ = 1.0 0.2 1 Success probability 0.5 0.3 Fig. 8. Upper and lower bounds on TV distance over γ in the AG-OTS cryptosystem at PNRmax = 20 dB, where N = 512, M = 64, and K = 8. For a given x, total 106 matrices of S have been tested. 0 16 0.6 96 112 128 M Fig. 9. Success probabilities over M in the AG-OTS cryptosystem at PNRmax = 20 dB, where N = 512. For an adversary, the upper bounds on the success probability of the indistinguishability experiment are sketched. of the secret matrix S is from the SSG keystream, where we observed that the decryption performance is similar to that of S from the random Bernoulli distribution. The CS decryption b achieves is declared as a success if a decrypted plaintext x ||x−b x||2 −2 < 10 . The figure shows that a legitimate recipient ||x||2 enjoys a reliable and stable CS decryption for a sufficiently large M at each γ. Meanwhile, the upper bounds on the success probability of an adversary indicate that no detection test can be successful in the indistinguishability experiment with the probability more than the bounds. In particular, if γ = 1, no adversary can learn any information about the 0 −10 −5 0 5 10 15 20 PNRmax (dB) 25 30 35 40 Fig. 10. Success probabilities over PNRmax in the AG-OTS cryptosystem, where N = 512 and M = 64. For an adversary, the upper bounds on the success probability of the indistinguishability experiment are sketched. plaintext with the success probability higher than 0.5, which leads to the indistinguishability. However, if energy variation occurs in plaintexts with γ < 1, the figure reveals that an adversary may be able to distinguish the plaintexts in the experiment, with the probability higher than 0.5. It also shows that the success probability of an adversary becomes more significant as the minimum energy ratio γ decreases and the plaintext length M increases. Figure 10 depicts the success probabilities over PNRmax in the AG-OTS cryptosystem, where N = 512 and M = 64. The simulation environment is identical to that of Figure 9. As can be seen from the figure, the CS decryption performance of a legitimate recipient improves over PNRmax . However, the detection performance of an adversary is saturated at high PNRmax , where the highest possible success probability is determined by the minimum energy ratio γ. The figure also shows that if PNRmax is low, the highest possible success probability of an adversary is close to 0.5 for any γ, which implies that the AG-OTS cryptosystem can be indistinguishable at sufficiently low PNRmax , regardless of energy variation. In this case, however, a legitimate recipient also fails in CS decryption due to high noise level. VI. C ONCLUSIONS This paper has proposed a new CS-based cryptosystem, named as the AG-OTS cryptosystem, by employing a secret bipolar keystream and a public unitary matrix for efficient implementation in practice. We demonstrated that the elements of the sensing matrix are asymptotically Gaussian for a sufficiently large plaintext length, which guarantees a stable and robust CS decryption for a legitimate recipient. By means of the total variation (TV) and the Hellinger distances, we showed that the AG-OTS cryptosystem can have the indistinguishability against an adversary, as long as each plaintext 10 has constant energy. Therefore, it is essential that the AGOTS cryptosystem should have a normalization step before CS encryption for equalizing the plaintext energy, which guarantees the computational security against any kind of polynomial time attack from an adversary. Finally, we found that the indistinguishability of the AG-OTS cryptosystem is highly sensitive to energy variation of plaintexts. To support the AGOTS cryptosystem with unequal plaintext energy, we developed sufficient conditions on the minimum energy ratio, the plaintext length, and the maximum plaintext-to-noise power ratio, respectively, for the asymptotic indistinguishability. The results of the energy sensitivity can be directly applicable to the G-OTS cryptosystem of [17]. R EFERENCES [1] D. L. Donoho, “Compressed sensing,” IEEE Trans. Inf. Theory, vol. 52, no. 4, pp. 1289-1306, Apr. 2006. [2] E. J. Candes, J. Romberg, and T. Tao, “Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information,” IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 489-509, Feb. 2006. [3] E. J. Candes and T. Tao, “Near-optimal signal recovery from random projections: Universal encoding strategies,” IEEE Trans. Inf. Theory, vol. 52, no. 12, pp. 5406-5425, Dec. 2006. [4] Y. C. Eldar and G. Kutyniok, Compressed Sensing - Theory and Applications, Cambridge University Press, 2012. [5] J. Tropp, J. N. Laska, M. Duarte, J. Romberg, and R. G. Baraniuk, “Beyond Nyquist: Efficient sampling of sparse bandlimited signals,” IEEE Trans. Inf. Theory, vol. 56, no. 1, pp. 520-544, Jan. 2010. [6] M. Mishali and Y. C. Eldar, “From Theory to practice: Sub-Nyquist sampling of sparse wideband analog signals,” IEEE J. Select Top. Sig. Process., vol. 4, no. 2, pp. 375-391, 2010. [7] J. Haupt, W. Bajwa, G. Raz, and R. Nowak, “Toeplitz compressed sensing matrices with applications to sparse channel estimation,” IEEE Trans. Inf. Theory, vol. 56, no. 11, pp. 5862-5875, Nov. 2010. [8] M. F. Duarte, S. Sarvotham, D. Baron, M. B. Wakin, and R. G. Baraniuk, “Distributed compressed sensing of jointly sparse signals,” Asilomar Conf. on Signals, Systems and computers, pp. 1537-1541, Pacific Grove, CA, USA, Nov. 2005. [9] J. Haupt, W. Bajwa, M. Rabbat, and R. Nowak, “Compressed sensing for networked data,” IEEE Sig. Process. Mag., vol. 25, no. 2, pp. 92-101, Mar. 2008. [10] C. Caione, D. Brunelli, and L. Benini, “Compressive sensing optimization for signal ensembles in WSNs,” IEEE Trans. Industrial Informatics, vol. 10, no. 1, pp. 382-392, Feb. 2014. [11] M. Duarte, M. Davenport, D. Takhar, J. N. Laska, T. Sun, K. F. Kelly, and R. G. Baraniuk, “Single-pixel imaging via compressive sampling,” IEEE Sig. Process. Mag., vol. 25, no. 2, pp. 83-91, Mar. 2008. [12] R. Marcia, Z. Harmany, and R. Willet, “Compressive coded aperture imaging,” Proc. IS&T/SPIE Symp. Elec. Imag.: Comp. Imag, San Jose, 2009. [13] M. Lustig, D. Donoho, and J. Pauly, “Rapid MR imaging with compressed sensing and randomly under-sampled 3DFT trajectories,” Proc. Ann. Meeting of ISMRM, Seattle, 2006. [14] S. Goginneni and A. Nehorai, “Target estimation using sparse modeling for distributed MIMO radar,” IEEE Trans. Signal Process., vol. 59, no. 11, pp. 5315-5325, Nov. 2011. [15] Y. Rachlin and D. Baron, “The secrecy of compressed sensing measurements,” in Proc. 46th Annu. Allerton Conf. Commun.Control, Comput., pp. 813-817, Sep. 2008. [16] A. Orsdemir, H. O. Altun, G. Sharma, and M. F. Bocko, “On the security and robustness of encryption via compressed sensing,” in Proc. IEEE Military Commun. Conf. (MILCOM), pp. 1-7, Nov. 2008. [17] T. Bianchi, V. Bioglio, and E. Magli, “On the security of random linear measurements,” in Proc. IEEE Int. Conf. Acoust. Speech Signal Process (ICASSP), pp. 3992-3996, May 2014. [18] T. Bianchi and E. Magli, “Analysis of the security of compressed sensing with circulant matrices,” in Proc. IEEE Workshop on Inf. Forens. Security (WIFS), pp. 1-6, Dec. 2014. [19] T. Bianchi, V. Bioglio, and E. Magli, “Analysis of one-time random projections for privacy preserving compressed sensing,” IEEE Trans. Inf. Forens. Security, vol. 11, no. 2, pp. 313-327, Feb. 2016. [20] G. Reeves, N. Goela, N. Milosavljevic, and M. Gastpar, “A compressed sensing wire-tap channel,” in Proc. IEEE Inf. Theory Workshop (ITW), pp. 548-552, Oct. 2011. [21] S. Agrawal and S. Vishwanath, “Secrecy using compressive sensing,” in Proc. IEEE Inf. Theory Workshop (ITW), pp. 563-567, Oct. 2011. [22] Y. D. Li, Z. Zhang, M. Winslett, and Y. Yang, “Compressive mechanism: Utilizing sparse representation in differentical privacy,” in Proc. 10th Annu. ACM Workshop Privacy Electron. Soc. (WPES), pp. 177-182, 2011. [23] R. Dautov and G. R. Tsouri, “Establishing secure measurement matrix for compressed sensing using wireless physical layer security,” in Proc. Int. Conf. Comput. Netw. Commun., pp. 354-358, Jan. 2013. [24] V. Cambareri, M. Mangia, F. Pareschi, R. Rovatti, and G. Setti, “Low complexity multiclass encryption by compressed sensing,” IEEE Trans. Signal Process., vol. 63, no. 9, pp. 2183-2195, May 2015. [25] S. N. George and D. P. Pattathil, “A secure LFSR based random measurement matrix for compressive sensing,” Sens. Imag., vol. 15, no. 1, pp. 1-29, 2014. [26] Y. Zhang, J. Zhou, F. Chen, L. Y. Zhang, K.-W. Wong, and X. He, “Embedding cryptographic features in compressive sensing,” Neurocomputing, vol. 205, pp. 472-480, 2016. [27] H. Li, R. Mao, L. Lai, and R. Qui, “Compressed meter reading for delay-sensitive and secure load report in smart grid,” in Proc. IEEE SmartGridComm, Oct. 2010. [28] J. Gao, X. Zhang, H. Liang, and X. Shen, “Joint encryption and compressed sensing in smart grid data transmission,” in Proc. IEEE GLOBECOM, Commun. Inf. Syst. Security Symp., pp. 662-667, Dec. 2014. [29] Y. Zhang, L. Y. Zhang, J. Zhou, L. Liu, F. Chen, and X. He, “A review of compressive sensing in information security field,” IEEE Access, Special Section on Green Communications and Networking for 5G Wireless, vol. 4, pp. 2507-2519, 2016. [30] J. Katz and Y. Lindell, Introduction to Modern Cryptography, 2nd Ed., Chapman & Hall/CRC, 2015. [31] A. L. Gibbs and F. E. Su, “On choosing and bounding probability metrics,” International Statistical Review, vol. 70, no. 3, pp. 419-435, 2002. [32] L. Le Cam, Asymptotic Methods in Statistical Decision Theory, SpringerVerlag, New York, 1986. [33] S. W. Golomb and G. Gong, Signal Design for Good Correlation - for Wireless Communication, Cryptography and Radar. Cambridge University Press, 2005. [34] T. T. Do, L. Gan, N. H. Nguyen, and T. D. Tran, “Fast and efficient compressive sensing using structurally random matrices,” IEEE Trans. Signal Process., vol. 60, no. 1, pp. 139-154, Jan. 2012. [35] W. Meier and O. Staffelbach, “The self-shrinking generator,” Advances in Cryptology-Eurocrypt’94, Lecture Notes in Computer Science (LNCS), vol. 950, pp. 205-214, Springer-Verlag, 1995. [36] M. Rudelson and R. Vershynin, “On sparse reconstruction from Fourier and Gaussian measurements,” Comm. Pure Appl. Math., vol. 61, no. 8, pp. 1025-1045, Aug. 2008. [37] A. DasGupta, Asymptotic Theory of Statistics and Probability, Springer Science+Business Media, LLC 2008. [38] A. Guntuboyina, S. Saha, and G. Schiebinger, “Sharp inequalities for f -divergences,” IEEE Trans. Inf. Theory, vol. 60, no. 1, pp. 104-121, Jan. 2014. [39] T. Kailath, “The divergence and Bhattacharyya distance measures in signal selection,” IEEE Trans. Commun. Technol., vol. COM-15, no. 1, pp. 52-60, Feb. 1967. [40] K. T. Abou-Moustafa and F. P. Ferrie, “A note on metric properties for some divergence measures: The Gaussian case,” JMLR: Asian Conference on Machine Learning, vol. 25, pp. 1-15, 2012. [41] S. Foucart and H. Rauhut, A Mathematical Introduction to Compressive Sensing, Springer Science+Business Media, New York, 2013. [42] D. Needell and J. A. Tropp, “CoSaMP: Iterative signal recovery from incomplete and inaccurate samples,” Appl. and Comput. Harmon. Anal., vol. 26, pp. 301-321, 2009.
7cs.IT
An Integrated Design of Optimization and Physical Dynamics for Energy Efficient Buildings: A Passivity Approach arXiv:1701.05577v1 [cs.SY] 19 Jan 2017 Takeshi Hatanaka, Xuan Zhang, Wenbo Shi, Minghui Zhu and Na Li Abstract— In this paper, we address energy management for heating, ventilation, and air-conditioning (HVAC) systems in buildings, and present a novel combined optimization and control approach. We first formulate a thermal dynamics and an associated optimization problem. An optimization dynamics is then designed based on a standard primal-dual algorithm, and its strict passivity is proved. We then design a local controller and prove that the physical dynamics with the controller is ensured to be passivity-short. Based on these passivity results, we interconnect the optimization and physical dynamics, and prove convergence of the room temperatures to the optimal ones defined for unmeasurable disturbances. Finally, we demonstrate the present algorithms through simulation. I. I NTRODUCTION Stimulated by strong needs for reducing energy consumption of buildings, smart building energy management algorithms have been developed both in industry and academia. In particular, about half of the current consumption is known to be occupied by heating, ventilation, and air-conditioning (HVAC) systems, and a great deal of works have been devoted to HVAC optimization and control [1]. In this paper, we address the issue based on a novel approach combining optimization and physical dynamics. Interplays between optimization and physical dynamics have been most actively studied in the field of Model Predictive Control (MPC), which has also been applied to building HVAC control [1]–[6]. While the MPC approach regards the optimization process as a static map from physical states to optimal inputs, another approach to integrating optimization and physical dynamics is presented in [7]–[11] mainly motivated by power grid control. There, the solution process of the optimization is viewed as a dynamical system, and the combination of optimization and physical dynamics is regarded as an interconnection of dynamical systems. The benefits of the approach relative to MPC are as follows. First, the approach allows one to avoid complicated modeling and prediction of factors hard to know in advance, while MPC needs their models to predict future system evolutions. Second, since the entire system is a dynamical system, its stability and performance are analyzed based on unifying dynamical system theory. In this line of works, Shiltz et al. [7] addresses smart grid control, and interconnects a dynamic optimization process and a locally controlled grid dynamics. The entire process is T. Hatanaka is with School of Engineering, Tokyo Institute of Technology, Tokyo 152-8552, JAPAN ([email protected]). X. Zhang, W. Shi and N. Li are with Electrical Engineering and Applied Mathematics of the School of Engineering and Applied Sciences, Harvard Univ., 33 Oxford St, Cambridge, MA 02138, USA. M. Zhu is with Department of Electrical Engineering, Pennsylvania State Univ., University Park, PA 16802, USA. then demonstrated through simulation. The authors of [8], [9] incorporate the grid dynamics into the optimization process by identifying the physical dynamics with a subprocess of seeking the optimal solution. A scheme to eliminate structural constraints required in [8], [9] is presented by Zhang et al. [10] while instead assuming measurements of disturbances. A similar approach is also taken for power grid control in Stegink et al. [11]. In this paper, we address integrated design of optimization and physical dynamics for HVAC control based on passivity, where we regard the optimization process as a dynamical system similarly to [7]–[11]. Interconnections of such dynamic HVAC optimization with a building dynamics are partially studied in [12], where temperature data for all zones and their derivatives in the physics side are fed back to the dynamic optimization process to recover the disturbance terms. However, such data are not always available in practical systems. We thus present an architecture relying only on temperature data of a subgroup of zones with HVAC systems. The contents of this paper are as follows. A thermal dynamic model with unmeasurable disturbances and an associated optimization problem are first presented. We then formulate an optimization dynamics based on the primaldual gradient algorithm [13]. The designed dynamics is then proved to be strictly passive from a transformed disturbance estimate to an estimated optimal room temperature. We next design a controller so that the actual room temperature tracks a given reference, and produces a disturbance estimate. Then, the physical dynamics is proved to be passivity-short from the reference to the disturbance estimate. From these two passivity-related results, we then interconnect the optimization and physical dynamics, and prove convergence of the actual room temperature to the optimal solution defined for the unmeasurable actual disturbance. Finally, the presented algorithm is demonstrated through simulation. II. P ROBLEM S ETTINGS A. Preliminary In this section, we introduce the concept of passivity. Consider a system with a state-space representation ẋ = φ(x, u), y = ϕ(x, u), (1) where x(t) ∈ Rn is the state, u(t) ∈ Rp is the input and y(t) ∈ Rp is the output. Then, passivity is defined as below. Definition 1 The system (1) is said to be passive if there exists a positive semi-definite function S : Rn → R+ := [0, ∞), called storage function, such that Z t S(x(t)) − S(x(0)) ≤ y T (τ )u(τ )dτ (2) 0 holds for all inputs u : [0, t] → Rp , all initial states x(0) ∈ Rn and all t ∈ R+ . In the case of the static system y = ϕ(u), it is passive if y T u = ϕT (u)u ≥ 0 for all u ∈ Rp . The system (1) is also said to be output feedback passive with index ε > 0 if (2) is replaced by Z t S(x(t)) − S(x(0)) ≤ y T (τ )u(τ ) − εky(τ )k2 dτ (3) 0 If the right-hand side of (3) is changed as Z t S(x(t)) − S(x(0)) ≤ y T (τ )u(τ ) + εku(τ )k2 dτ Ti ) (i = 1, 2, . . . , n1 ), 1 is the n-dimensional real vector whose elements are all 1, and B = [In1 0]⊤ ∈ Rn×n1 . We next linearize the model at around an equilibrium as ˙ = RδT a 1 − RδT − LδT + BG(T̄ )δm − Ū δT + Bδq, C δT where δT , δm, δT a and δq describe the errors from the equilibrium states and inputs and Ū ∈ Rn×n is a diagonal matrix whose diagonal elements are m̄1 , . . . , m̄n1 , 0, . . . , 0, where m̄i is the i-th element of the equilibrium input m̄. Using the variable transformations x := C 1/2 δT, u := B ⊤ C −1/2 BG(T̄ )δm, wa := C −1/2 RδT a 1, wq := B ⊤ C −1/2 Bδq, (4) (6) is rewritten as 0 with ε > 0, the system is said to be passivity-short, and then ε is called impact coefficient [18]. (8) ẋ = −Ax + Bu + Bwq + wa , x :=  x1 x2  (9) where A := C −1/2 (R + L + Ū )C −1/2 , x1 ∈ Rn1 and x2 ∈ Rn2 . Remark that the matrix A is positive definite [19]. From control engineering point of view, x is the system state, u is the control input, and wq and wa are disturbances. We suppose that wa is measurable as well as x1 . Meanwhile, it is in general hard to measure the heat gain wq . B. System Description In this paper, we consider a building with multiple zones i = 1, 2, . . . , n. The zones i = 1, 2, . . . , n are divided into two groups: The first group consists of zones equipped with VAV (Variable Air Volume) HVAC systems whose thermal dynamics is assumed to be modeled by the RC circuit model [2]–[6], [12] as C. Optimization Problem X Tj − Ti T a − Ti Regarding the above system, we formulate the optimizaCi Ṫi = + + ai (Tis − Ti )mi + qi , (5) tion problem to be solved as follows: Ri Rij j∈Ni where Ti is the temperature of zone i, mi is the mass flow rate at zone i, T a is the ambient temperature, Tis is the air temperature supplied to zone i, which is treated as a constant throughout this paper, qi is the heat gain at zone i from external sources like occupants, Ci is the thermal capacitance, Ri is the thermal resistance of the wall/window, Rij is the thermal resistance between zone i and j and ai is the specific heat of the air. The second group is composed of other spaces such as walls and windows whose dynamics is modeled as X Tj − Ti T a − Ti + . (6) Ci Ṫi = Ri Rij j∈Ni Rooms not in use can be categorized into this group. Without loss of generality, we assume that i = 1, 2, . . . , n1 belong to the first group and i = n1 + 1, . . . , n (n2 := n − n1 ) belong to the second. Remark that the system parameters Ci , Ri , and Rij can be identified using the toolbox in [14]. The collective dynamics of (5) and (6) is described as C Ṫ = RT a 1 − RT − LT + BG(T )m + Bq, (7) where T , q and m are collections of Ti (i = 1, 2, . . . , n), qi (i = 1, 2, . . . , n1 ), and mi (i = 1, 2, . . . , n1 ) respectively. The matrices C and R are diagonal matrices with diagonal elements Ci (i = 1, 2, . . . , n) and R1i (i = 1, 2, . . . , n), respectively. The matrix L describes the weighted graph Laplacian with elements R1ij , G(T ) ∈ Rn1 ×n1 is a block diagonal matrix with diagonal elements equal to ai (Tis − min ⊤ z ⊤ ]⊤ ∈Rn+n1 z=[zx u kzx1 − hk2 + f (zu ) (10a) subject to: g(zu ) ≤ 0 (10b) −Azx + Bzu + Bdq + da = 0 (10c) ⊤ ⊤ ⊤ The variables zx := [zx1 zx2 ] (zx1 ∈ Rn1 , zx2 ∈ Rn2 ), and n1 zu ∈ R correspond to the zone temperature x and mass flow rate u after the transformation (8). The parameters dq ∈ Rn1 and da ∈ Rn are DC components of the disturbances wq and wa , respectively. These variables are coupled by (10c) which describes the stationary equation of (9). Throughout this paper, we assume the following assumption. Assumption 1 The problem (10) satisfies the following properties: (i) f : Rn1 → R is convex and its gradient is locally Lipschitz, (ii) every element of the constraint function g : Rn1 → Rc is convex and its gradient is locally Lipschitz, and (iii) there exists zu ∈ Rn1 such that g(zu ) < 0. The first term of (10a) evaluates the human comfort, where h ∈ Rn1 is the collection of the most comfortable temperatures for occupants in each zone, which might be determined directly by occupants in the same way as the current systems, or computed using human comfort metrics like PMV (Predicted Mean Vote). The quadratic function for the error is commonly employed in the MPC papers [3], [22]–[24] and [12]. Note that it is common to put weights on each element of zx1 − h to give priority to each zone, but this can be done by appropriately scaling each element of zu in the function f . This is why we take (10a). The function f is introduced to reduce power consumption. The papers [3], [22]–[24] simply take a linear or quadratic function of control efforts as such a function and then Assumption 1(i) is trivially satisfied. Also, as mentioned in [5], [12], the power consumption of supply fans is approximated by the cube of the sum of the mass flow rates, which also satisfies Assumption 1(i). A simple model of consumption at the cooling coil is given by the product of the mass flow rate mi and |Tis −T a | [25], which also belongs to the intended class 1 . The constraint function g : Rn1 → Rc reflects hardware constraints and/or an upper bound of the power consumption. For example, the constraints in [12] are reduced to the above form. The objective of this paper is to design a controller so as to ensure convergence of the actual room temperature x1 ∗ to the optimal room temperature zx1 , the solution to (10), without direct measurements of wq . A. Optimization Dynamics In this subsection, we present a dynamics to solve the above optimization problem. Before that, we eliminate zx from (10) using (10c). Then, the problem is rewritten as min kB ⊤ A−1 (Bzu + Bdq + da − h̄)k2 + f (zu ) (11a) zu ∈Rn1 (11b) with h̄ = ABh. It is easy to confirm from positive definiteness of A that the cost function of (11) is strongly convex. In this case, (11) has the unique optimal solution, denoted by zu∗ , and it satisfies the following KKT conditions [26]. M 2 (zu∗ + dq ) + N (da − h̄) + ∇f (zu∗ ) + ∇g(zu∗ )λ∗ = 0, (12a) ∗ ∗ ∗ ∗ g(zu ) ≤ 0, λ ≥ 0, λ ◦ g(zu ) = 0, (12b) where M = B ⊤ A−1 B, N := B ⊤ A−1 BB ⊤ A−1 , and ◦ represents the Hadamard product. Since (11) is essentially equivalent to (10), the solution to (11) also provides a solution for the problem (10). Precisely, if we define zx∗ ∈ Rn as zx∗ := A−1 (Bzu∗ + Bdq + da ), the pair z ∗ := [(zx∗ )⊤ (zu∗ )⊤ ]⊤ is a solution to (10). In the sequel, we also use the notation ∗ ∗ ⊤ ∗ ⊤ ⊤ ∗ ∈ Rn2 ). It is zx∗ := [(zx1 ) (zx2 ) ] (zx1 ∈ Rn1 and zx2 then easy to confirm ∗ zx1 = M (zu∗ + dq ) + B ⊤ A−1 da . primal-dual gradient algorithm [13] as one of such solutions2 . However, it is hard to obtain dq since wq is not measurable. We thus need to estimate dq from the measurements of physical quantities. This motivates us to interconnect the physical dynamics with the optimization dynamics. Taking account of the above issues, we present ẑ˙u = −α{M 2 (ẑu + dˆq ) + N (wa − h̄) + ∇f (ẑu ) + p},(14a) ˙ λ̂ = [g(ẑu )]+ , p = ∇g(ẑu )λ̂, (14b) λ̂ III. O PTIMIZATION DYNAMICS subject to: g(zu ) ≤ 0 Fig. 1. Block diagram of optimization dynamics, which is passive from ∗ (Lemma 2). ṽo = vo − vo∗ ỹo = ŷo − yo∗ with vo∗ := M dq and yo∗ := zx1 (13) Given dq and da , it would be easy to solve (12). However, in the practical applications, it is desired that dq and da are updated in real time according to the changes of disturbances. In this regard, it is convenient to take a dynamic solution process of optimization since it trivially allows one to update the parameters in real time. In particular, we employ the 1 For simplicity, we skip the dependence of d on f , but subsequent results a are easily extended to the case that f depends on da . where ẑu and λ̂ are estimates of zu∗ and λ∗ respectively, and α > 0. The notation [b]+ a for real vectors a, b with the same dimension provides a vector whose l-th element, denoted by ([b]+ a )l , is given by  0, if al = 0 and bl < 0 ([b]+ ) = , (15) l a bl , otherwise where al , bl are the l-th element of a, b, respectively. Note that (14) is different from the primal-dual gradient algorithm for (12) in that the term da is replaced by the measurement wa , and dq is replaced by its estimate dˆq whose production will be mentioned later. The system is illustrated in Fig. 1. B. Passivity Analysis for Optimization Dynamics Hereafter, we analyze passivity of the above optimization process assuming that wa is constant. In this case, wa ≡ da holds. In practice, the disturbance wa , namely the ambient temperature T a , is time-varying but the following results are applied to the practical case if wa is approximated by a piecewise constant signal, which is fully expected since T a usually varies slowly. Under the above assumption, we define the output ν := −M 2 (ẑu + dˆq ) for (14). We then have the following lemma. Lemma 1 Consider the system (14) with wa ≡ da and λ̂(0) ≥ 0. Then, under Assumption 1, it is passive from d˜q := dˆq − dq to −ν̃, where ν̃ := ν − ν ∗ and ν ∗ := −M 2 (zu∗ + dq ). Proof: See Appendix I. 2 Another benefit of using the dynamic solution is that it provides a distributed solution when the present results are extended to a more global problem, although it exceeds the scope of this paper. Please refer to [20] for more details on the issue. Let us next transform the output ν to yo := −M −1 ν + B ⊤ A−1 da = M (ẑu + dˆq ) + B ⊤ A−1 da , yo∗ := −M −1 ν ∗ + B ⊤ A−1 da . Comparing (13) and the above definition of yo , the signal yo ∗ is regarded as an estimate of zx1 . We also define vo := M dˆq ∗ and vo := M dq . Then, we can prove the following lemma. Lemma 2 Consider the system (14) with wa ≡ da and λ̂(0) ≥ 0. Then, under Assumption 1, it is output feedback passive from ṽo := vo − vo∗ to ỹo := yo − yo∗ with index 1. Proof: It is easy to see from ν ∗ = −M 2 (zu∗ + dq ) and ∗ (13) that yo∗ = zx1 . From (32), ỹo = −M −1 ν̃ and ṽo = ˜ M dq , we have the following inequality. D+ So ≤ ỹo⊤ ṽo − kỹo k2 Fig. 2. Block diagram of the physical dynamics with the local controller, ∗ which is passivity-short from ṽp = vp −vp∗ to ỹp = yp −yp∗ with vp∗ := zx1 and yp∗ := −M dq (Lemma 5). Substituting (17) into (9) yields ẋ = −Ax + kP B(r − x1 ) + κBr + Bξ +Bwq + (BF + In )wa , ˙ξ = kI (r − x1 ). (16) Integrating this in time completes the proof. In this section, we design a physical dynamics and prove its passivity. A passivity-based design for the model (9) is presented in [21], but we modify the control architecture in order to interconnect it with the optimization dynamics. Lemma 3 The steady states x∗ and ξ ∗ of (19) for r ≡ r∗ , wa ≡ da and wq ≡ dq are given as follows.     0 0 x∗ = −F ⊤ r∗ + −1 ⊤ da , Bc := ∈ Rn×n2 In2 A3 Bc ξ ∗ = (M −1 − κIn2 )r∗ − dq . A. Controller Design In this subsection, we design a controller to determine the input u so that x1 tracks a reference signal r. Here we assume the following assumption, where   A1 A⊤ 2 , A1 ∈ Rn1 ×n1 , A3 ∈ Rn2 ×n2 . A= A2 A3 Assumption 2 The matrix M A1 +A1 M is positive definite. This property does not always hold for any positive definite matrices A1 and M , but it is expected to be true in many practical cases since the diagonal elements tend to be dom−1 −1 inant both for A1 and M = (A1 − A⊤ in this 2 A3 A2 ) application [12]. Note that this assumption holds in the full actuation case (A1 = A, M = A−1 ). Inspired by the fact that many existing systems employ Proportional-Integral (PI) controllers (with logics) as the local controller, we design the following controller adding reference and disturbance feedforward terms. ξ̇ = kI (r − x1 ) (17a) u = kP (r − x1 ) + ξ + κr + F wa (17b) −1 where kP > 0, kI > 0 and F := [−In1 A⊤ 2 A3 ] and In1 is the n1 -by-n1 identity matrix. The feedforward gain κ > 0 is selected so that (18) (20) Equation (19) is now rewritten as ẋ = −Āx + BM −1 ζ + Bwq + F̄ wa ξ˙ = kI (r − x1 ), ζ = k̄P M (r − x1 ) + M ξ, (21a) (21b) where  A3−1 −A⊤ 2 Bc⊤ . Ā := A − κBB , k̄P := kP + κ, F̄ := In2 ⊤ Such a κ exists under Assumption 2. (19b) For the system, we have the following lemma. IV. P HYSICAL DYNAMICS P := M A1 + A1 M − 2κM > 0. (19a)  Remark that, at the steady state, the variable ζ is equal to ζ ∗ := M ξ ∗ = Kr∗ − M dq , K := In1 − κM. (22) B. Passivity Analysis for Physical Dynamics In this subsection, we analyze passivity of the system (21) assuming that wq and wa are constant. The case of the time varying wq will be treated in the end of the next section. Choose ζ as the output and prove passivity as follows. Lemma 4 Consider the system (21) with r ≡ r∗ , wa ≡ da and wq ≡ dq . Then, under Assumption 2, the system is passive from r̃ := r − r∗ to ζ̃ := ζ − ζ ∗ . Proof: See Appendix II. Remark that this lemma holds regardless of the value of r∗ . To extract the term M dq from (22), we define the output yp := ζ − Kr, yp∗ := ζ ∗ − Kr∗ = −M dq , (23) ∗ and vp := r and vp∗ := zx1 . Then, we have the following. Lemma 5 Consider the system (21) with r ≡ r∗ , wa ≡ da and wq ≡ dq . Then, under Assumption 2, the system from ṽp := vp − vp∗ to ỹp := yp − yp∗ is passivity-short with the impact coefficient 1 − κσ, where σ > 0 is the minimal eigenvalue of M . ∗ Proof: If we take r∗ = zx1 , we have ỹp = ζ̃ − K ṽp . Substituting this into (39) yields Ṡp ≤ ỹp⊤ ṽp + ṽp⊤ K ṽp − kP (ṽp − x̃1 )⊤ M (ṽp − x̃1 ) (24) ≤ ỹp⊤ ṽp + (1 − κσ)kṽp k2 − kP σkṽp − x̃1 k2 (25) This completes the proof. V. I NTERCONNECTION OF O PTIMIZATION AND P HYSICAL DYNAMICS Let us interconnect the optimization dynamics (14) and physical dynamics (21). Remark that the stationary value of ∗ ∗ yo , yo∗ = zx1 , is equivalent to that of vp , vp∗ = zx1 . Also, ∗ ∗ vo = −yp holds. Inspired by these facts, we interconnect these systems via the negative feedback as vo = −yp , vp = yo . We then have the following main result of this paper. Theorem 1 Suppose that λ̂(0) ≥ 0, wa ≡ da and wq ≡ dq . Then, if Assumptions 1 and 2 hold, the interconnection of ∗ (14) and (21) via vo = −yp , vp = yo ensures that x1 → zx1 . Proof: Define S := So + Sp . Then, combining (16), (25) and vo = −yp , vp = yo yields D+ S ≤ −κσkỹo k2 − kP σkṽp − x̃1 k2 . (26) ∗ This means that both of ỹo = yo − zx1 and ỹo − x̃1 = yo − x1 belong to class L2 . Since S is positive definite, all of the state variables ẑu , λ̂, x and ξ belong to L∞ . From (14), ỹ˙ o = −M −1 ν̇ = M ẑ˙u is bounded. Also, (19) means that x̃˙ = ẋ is bounded and hence ẏo − ẋ1 is bounded. Thus, invoking ∗ Barbalat’s lemma, we can prove yo −zx1 → 0, yo −x1 → 0, ∗ which means x1 → zx1 . This completes the proof. Lyapunov stability of the desirable equilibrium, tuple of zu∗ , λ∗ , x∗ and ξ ∗ , is also proved in the above proof. It is to be emphasized that the optimal solution is dependent on the unmeasurable disturbance. Nevertheless, convergence to the solution is guaranteed owing to the feedback path from physics to optimization. The above results are obtained assuming that both of wq and wa are constant. This is likely valid for wa since the ambient temperature is in general slowly varying. However, the heat gain wq may contain high frequency components. To address the issue, we decompose the signal wq into the DC components dq and others w̃q as wq = dq + w̃q . We also assume that w̃q belongs to an extended L2 space [16]. The following corollary then holds, which is proved following the proof procedure of the well-known passivity theorem [16], [17] and using the fact that the right-hand side of (26) is κkP σ kx̃1 k2 . upper bounded by − κ+k P Corollary 1 Suppose that λ̂(0) ≥ 0, wa ≡ da and wq = dq + w̃q . Then, if Assumptions 1 and 2 hold, the interconnected system (14), (21) and vo = −yp , vp = yo from w̃q ∗ to x̃1 = B ⊤ x̃ = x1 − zx1 has a finite L2 gain. Fig. 3. Hierarchical control architecture. We give some remarks on the present architecture. Figs. 1 and 2 are oriented by theoretical analysis, but the implementation does not need to follow the information processing in the figures. Indeed, the interconnected system is equivalently transformed into Fig. 3. If we let the operations shaded by dark gray be executed in the high-level controller, the low-level controller can be implemented in a decentralized fashion similarly to the existing systems. In Fig. 3, both of the high-level and low-level controller with the physical dynamics are biproper and hence a problem of algebraic loops can occur. This however does not matter in practice since the information transmissions between highand low-level processes usually suffer from possibly small delays. Although the high-level controller itself contains an algebraic loop, it is easily confirmed that the loop can be solved by direct calculations of the algebraic constraint. The transfer function from wq to the disturbance estimate dˆq , roughly speaking, is almost the same as the complementary transfer function and hence only the low frequency components are provided by the physical dynamics. This is why dˆq is regarded as an estimate of the DC component of wq . The cutoff frequency of the disturbance can be in principle tuned by kP and kI , but, once a closed-loop system is designed, the cutoff is also automatically decided. It is however not always a drawback at least qualitatively. Actually, even if optimal solutions reflecting much faster disturbance variations are provided, the physical states cannot respond to variations faster than the bandwidth. It is a consequence of the internal model control and constant disturbances that the actual disturbance wq is correctly estimated. A control architecture based on a similar concept is presented in Section VII of Stegink et al. [11]. However, it is clear that the problem (10) does not meet the structural constraints assumed in [11] and hence the architecture in [11] cannot be directly applied to our problem. Zhang et al. [12] present another kind of interconnection between physical and optimization dynamics based on a quasi-disturbance feedforward, where the disturbance is computed by state measurements and their derivatives, and VI. S IMULATION In this section, we demonstrate the presented control architecture through simulation. For this purpose, we build a building on 3D modeling software SketchUp (Trimble Inc.), which contains three rooms (n1 = 3) and other 46 zones (n2 = 46) including walls, ceilings, and windows. The building model is then installed into EnergyPlus [27] in order to simulate the evolution of zone temperatures. Then, using the acquired data, we identify the model parameters in (5) and (6) via BRCM toolbox [14]. We next specify the optimization problem (10). All the elements of h are set to 22C 1/2◦ C and we take f (zu ) = 150kzuk2 . The constraints are also selected as |zui | ≤ P 0.61 i = 1, 2, 3 and 3i=1 |zui | ≤ 1.25, where zui is the i-th element of zu . Collecting these constraints, we define the function g. However, since it turns out that directly using g(zu ) ≤ 0 has a response speed problem in penalizing the constraint violation in the primal-dual algorithm, we instead take the constraint θg(zu ) ≤ 0 with θ = 15, which does not essentially change the optimization problem. In the simulation, we take the feedback gains kP = 6.0 × 10−2 and kI = 1.0 × 10−3 , and κ = 1.0 × 10−3 , which are tuned so that the peak gain of σ-plot from r to x1 is smaller than the well-known criterion. It is then confirmed that Assumption 2 is satisfied. In the following simulation, we use the disturbance data shown in Fig. 4. Here, we compare the results with the ideal case that the disturbance wq is directly measurable. In this case, the feedback path from the physical dynamics to optimization is not needed and hence we take the cascade 0.7 0.6 30 heat[kW] ambient temp.[c-degree] 35 25 0.5 0.4 0.3 20 0 100 200 300 400 0.2 0 500 100 time[min] Fig. 4. 200 300 400 500 time[min] Ambient temperature (left) and external heats (right). 0.7 0.6 heat[kW] heat[kW] 0.6 0.5 0.4 0.2 0 0.58 0.56 0.54 0.3 0.52 100 200 300 400 140 500 room temp. [c-degree] 28 26 24 22 0 100 200 300 time[min] 160 180 200 220 400 500 time[min] time[min] reference temp. [c-degree] then fed back to the optimization dynamics. The differences of the present scheme from [12] are listed as follows: The approach of [12] requires the measurements of state variables. If the states include temperatures of windows and/or walls, its technological feasibility may be problematic or at least increases the system cost. On the other hand, our approach needs only x1 which is usually measurable. Since there is no sensor to measure ẋ, it has to be computed using the difference approximation, which provides approximation errors. Meanwhile, the present approach does not need such an approximation. In [12], the difference approximation errors together with sensor noises and high frequency components of the disturbances are directly sent to the optimization dynamics, which may cause fluctuations for the output and internal variables in the optimization process unless it is carefully designed in the sense of the noise reduction. Adding a low-pass filter to the computed disturbance might eliminate these undesirable factors. However, the filter is not designed independently of stability of the entire system in the presence of uncertainties in ẋ and the system model since, in this case, the quasi-feedforward system becomes a feedback system and the filter is included into the loop. Meanwhile, the noises are automatically rejected by the physical dynamics in our algorithm. 400 500 28 26 24 22 0 100 200 300 time[min] Fig. 5. Time responses of the estimated heat gain (top 2 figures), the estimated optimal room temperatures (bottom-left) and actual room temperatures (bottom-right). In all figures, the solid curves show the responses delivered by the proposed method, and the dotted ones with light colors are those by the disturbance feedforward scheme. In the top figures, the dotted lines coincide with the actual heat gain. connection from optimization to physics. It is to be noted that it is hard to implement this in practice. The trajectories of the estimated heat gains, the estimated optimal room temperatures and temperatures T1 , T2 , T3 are illustrated by solid curves in Fig. 5. The dotted lines with light colors show the trajectories using the above disturbance feedforward scheme. We see from the top left figures that the presented algorithm almost correctly estimate the disturbance wq . It is also observed from the right fine-scale figure that high frequency signal components are filtered out in the case of our algorithm. The trajectories in the bottom-left figure sometimes get far from 22◦ C since the constraint P 3 i=1 |zui | ≤ 1.25 gets active during the periods due to the high ambient temperature and heat gains. We see from the bottom figures that the response of the present method to the constraint violations is slower than the disturbance feedforward scheme because high frequency components of the disturbance are filtered out by the physical dynamics. Let us next show that the high sensitivity of disturbance feedforward can cause another problem. Here, small noises are added to the heat gain wq at every 20s, whose absolute value is upper bounded by 1.0 × 10−3 . We then run the above two algorithms. The resulting responses are illustrated in Fig. 6. It is observed from the top-left figure that the 0.7 heat[kW] heat[kW] 0.5 0.4 0.58 0.56 0.54 0.3 0.2 0 signed a primal-dual algorithm-based optimization dynamics and a local physical control system, and proved the system properties related to passivity. We then interconnected the optimization and physical dynamics, and proved convergence of the room temperatures to the optimal ones. We finally demonstrated the present algorithms through simulation. 0.6 0.6 0.52 100 200 300 400 140 500 room temp. [c-degree] reference temp. [c-degree] 30 28 26 24 22 20 18 16 0 100 200 300 160 180 200 220 time[min] time[min] 400 28 26 24 22 0 500 R EFERENCES 30 100 200 300 400 500 time[min] time[min] room temp. [c-degree] reference temp. [c-degree] Fig. 6. Time responses in the presence of noises on the heat gain wq , where every line has the same meaning as Fig. 5. 28 26 24 22 0 100 200 300 400 500 28 26 24 22 0 100 200 300 400 500 time[min] time[min] Fig. 7. Time responses of the reference (left) and room temperatures (right) with the scaling factor θ = 50. noise is filtered out in the present algorithm, and its effects do not appear on the estimates. Accordingly, the trajectories of the estimated optimal and actual room temperatures are almost the same as Fig. 5. Meanwhile, the disturbance feedforward approach suffers significant effects from the noise. The trajectories of the bottom-left get smaller than 22◦ C. Namely, the fluctuations are caused by the undesirable over- and undershoots. Although the trajectories of the actual temperatures get smooth, this behavior of the reference is not desirable from an engineering point of view. Adding a low-pass filter or reducing the gain of the optimization dynamics would eliminate the oscillations but it spoils the advantage, namely response speed. It is to be noted that if the disturbance feedforward is implemented using the recovery technique in [12], the low-pass filter is not designed independently of system stability as stated in Section V. If the response speed of the present algorithm in Fig. 5 is still problematic, it can be accelerated by tuning the scaling factor θ. The results for θ = 50 are shown in Fig. 7, where it is observed that almost the same speed as the disturbance feedforward in Fig. 5 is achieved by the present algorithm. Simulation for a larger-scale system with more practical settings is left as a future work of this paper. VII. C ONCLUSION In this paper, we presented a novel combined optimization and control algorithm for HVAC control of buildings. We de- [1] A. Afram, F. Janabi-Sharif, “Theory and applications of HVAC control systems e A review of model predictive control (MPC),” Building and Environment, vol. 72, pp. 343–355, 2014. [2] F. Oldewurtel, A. Parisio, C.N. Jones, D. Gyalistras, M. Gwerder, V. Stauch, B. Lehmann and M. Morari, “Use of model predictive control and weather forecasts for energy efficient building climate control,” Energy and Buildings, vol. 45, pp. 15–27, 2012. [3] A. Aswani, N. Master, J. Taneja, D. Culler and C. Tomlin, “Reducing transient and steady state electricity consumption in HVAC using learning-based model-predictive control,” Proc. IEEE, vol. 100, no. 1, pp. 240–253, 2012. [4] Y. Ma, A. Kelman, A. Daly and F. Borrelli, “Predictive control for energy efficient buildings with thermal storage: modeling, stimulation, and experiments,” IEEE Control Syst. vol. 32, no. 1, pp. 44–64, 2012. [5] Y. Ma, J. Matusko and F. Borrelli, “Stochastic model predictive control for building HVAC systems: Complexity and conservatism,” IEEE Trans. Control Systems Technology, vol. 23, no. 1, pp. 101–116, 2015. [6] S. Goyal, H. Ingley and P. Barooah, “Zone-level control algorithms based on occupancy information for energy efficient buildings,” Proc. American Control Conf., pp. 3063–3068, 2012. [7] D.J. Shiltz, M. Cvetkovic and A.M. Annaswamy, “An integrated dynamic market mechanism for real-time markets and frequency regulation,” IEEE Trans. Sust. Ener., vol. 7, no. 2, pp. 875–885, 2016. [8] C. Zhao, U. Topcu, N. Li and S. Low “Design and stability of load-side primary frequency control in power systems,” IEEE Trans. Automatic Control, vol. 59, no. 5, pp. 1177–1189, 2014. [9] E. Mallada, C. Zhao and S. Low, “Optimal load-side control for frequency regulation in smart grids,” Proc. 52nd Annual Allerton Conf. Communication, Control, and Computing, pp. 731–738, 2014. [10] X. Zhang, A. Papachristodoulou and N. Li, “Distributed optimal steady-state control using reverse- and forward-engineering,” Proc. 54th IEEE Conf. Decision and Control, pp. 5257–5264, 2015. [11] T. Stegink, C.D. Persis and A. van de Shaft, “A unifying energy-based approach to optimal frequency and market regulation in power grids,” arXiv:1510.05420v1, 2015 (downloadable at https://arxiv.org/abs/1510.05420v1). [12] X. Zhang, W. Shi, X. Li, B. Yan, A. Malkawi and N. Li, “Decentralized and distributed temperature control via HVAC systems in energy efficient buildings,” Automatica, submitted, 2016. [13] A. Cherukuri, E. Mallada, and J. Cortés, “Asymptotic convergence of constrained primal-dual dynamics,” System and Control Letters, vol. 87, pp. 10–15, 2016. [14] D. Sturzenegger, D. Gyalistras, V. Semeraro, M. Morari and R.S. Smith, “BRCM Matlab toolbox: Model generation for model predictive building control,” Proc. 2014 American Control Conf., pp. 1063– 1069, 2014. [15] T. Hatanaka, N. Chopra, T. Ishizaki and N. Li, “Passivity-based distributed optimization with communication delays using PI consensus estimator,” IEEE Trans. Automatic Control, submitted, 2016 (downloadable at https://arxiv.org/abs/1609.04666). [16] A.J. van der Schaft, L2-Gain and Passivity Techniques in Nonlinear Control, 2nd edn. Communications and Control Engineering Series. Springer, London, 2000. [17] T. Hatanaka, N. Chopra, M. Fujita and M.W. Spong, Passivity-Based Control and Estimation in Networked Robotics, Communications and Control Engineering Series, Springer-Verlag, 2015. [18] Z. Qu and M.A. Simaan, “Modularized design for cooperative control and plug-and-play operation of networked heterogeneous systems,” Automatica, vol. 50, no. 9, pp. 2405–2414, 2014. [19] Y. Hong, J. Hu, and L. Gao, “Tracking control for multi-agent consensus with an active leader and variable topology,” Automatica, vol. 42, no. 7, pp. 1177–1182, 2006. [20] T. Hatanaka, X. Zhang, W. Shi, M. Zhu, and N. Li, “Physics Integrated Hierarchical/Distributed HVAC Optimization for Multiple Buildings with Robustness against Time Delays,” arXiv, 2017 (downloadable at https://arxiv.org/abs/1510.05420v1). [21] J.T. Wen, S. Mishra S. Mukherjee, N. Tantisujjatham and N. Minakais, “Building Temperature Control with Adaptive Feedforward,” Proc. 52nd IEEE Conf. Decision and Control, pp. 4827–4832, 2013. [22] S. Privara, J. Siroky, L. Ferkl, and J. Cigler, “Model predictive control of a building heating system: the first experience,” Energy and Buildings, vol. 43, pp. 564–572, 2011. [23] P.-D. Morosan, R. Bourdais, D. Dumur, and J. Buisson, “Building temperature regulation using a distributed model predictive control,” Energy and Buildings, vol. 42, pp. 1445–1452, 2010. [24] S. Yuan and R. Perez, “Multiple-zone ventilation and temperature control of a single-duct VAV system using model predictive strategy,” Energy and Buildings, vol. 38, pp. 1248-1261, 2006. [25] M. Maasoumya, M. Razmara, M. Shahbakhti, and A.S. Vincentelli, “Handling model uncertainty in model predictive control for energy efficient buildings,” Energy and Buildings, vol. 77, pp. 377–392, 2014. [26] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004. [27] US Dept. of Energy’s: EnergyPlus, https://energyplus.net/ A PPENDIX I P ROOF OF L EMMA 1 Lemma 6 Consider the system (14b) with λ̂(0) ≥ 0. Then, under Assumption 1, it is passive from z̃u = ẑu − zu∗ to p̃ = p − p∗ with p∗ := ∇g(zu∗ )λ∗ . Proof: Define the energy function U := 21 kλ̂ − λ∗ k2 . Then, following the same procedure as [15], we have D+ U ≤ (p − p∗ )⊤ (ẑu − zu∗ ) = p̃⊤ z̃u , (27) where the notation D+ represents the upper Dini derivative. Integrating this in time completes the proof. We next consider (14a). Now, replace −M 2 (ẑu + dˆq ) − p by an external input µ and consider the system ẑ˙u = −α{N (wa − h̄) + ∇f (ẑu ) − µ}. (28) Lemma 7 Suppose wa ≡ da . Then, under Assumption 1, the system (28) is passive from µ̃ := µ − µ∗ to z̃u = ẑu − zu∗ , where µ∗ := −M 2 (zu∗ + dq ) − p∗ . Proof: Subtracting (12a) from (28) under yields z̃˙u = −α(∇f (ẑu ) − ∇f (zu∗ )) + αµ̃. (29) 1 1 Now, define V := 2α kz̃u k2 = 2α kẑu − zu∗ k2 . Then, the time derivative of V along the trajectories of (29) is given by V̇ = −(ẑu − zu∗ )⊤ (∇f (ẑu ) − ∇f (zu∗ )) + z̃u⊤ µ̃. (30) ∇f (zu∗ )) From convexity of f , (ẑu − − ≥0 holds [26]. This completes the proof. The system (14) is given by interconnecting (14b) and (28) via µ = ν − p. It is then easy to confirm that µ̃ = ν̃ − p̃. We are now ready to prove Lemma 1. Define So := V +U . Then, combining (27), (30) and µ̃ = ν̃ − p̃, we have + D So ≤ −(ẑu − zu∗ )⊤ (∇f (ẑu ) − ∇f (zu∗ )) − (ẑu − zu∗ )⊤ (∇f (ẑu ) − ∇f (zu∗ )) ≤ −ν̃ ⊤ d˜q . (32) This completes the proof. A PPENDIX II P ROOF OF L EMMA 4 Lemma 8 Under Assumption 2 and (18), the system (Ā3 , A2 P −1/2 ) is stabilizable and (P −1/2 M A⊤ 2 , Ā3 ) is detectable, where Ā3 := −A3 + A2 M P −1 A⊤ . 2 Proof: Define Φs := P 1/2 M P −1 A⊤ 2 and Φd := A2 M P −1 M −1 P 1/2 . Then, Ā3 − A2 P −1/2 Φs = −A3 , Ā3 − Φd P −1/2 M A⊤ 2 = −A3 hold and −A3 is stable. This completes the proof. Using Lemma 8, we next prove the following result. Lemma 9 Consider the system (21a) with r ≡ r∗ , wa ≡ da and wq ≡ dq . Then, under Assumption 2, the system is passive from ζ̃ to x̃1 := B ⊤ x̃ with x̃ := x − x∗ . Proof: We first formulate the error system x̃˙ = −Āx̃ + BM −1 ζ̃ ˙ ξ̃ = kI (r̃ − B x̃), ζ̃ = k̄P M (r̃ − B ⊤ x̃) + M ξ̃ (33a) (33b) where ξ̃ := ξ − ξ ∗ . Take a positive definite matrix Ψ ∈ M 0 Rn2 ×n2 and define Ψ̄ := ∈ Rn×n . Then, by 0 Ψ calculation, we have   ⊤ P M A⊤ 2 + A2 Ψ Ψ̄Ā + ĀΨ̄ = . (34) ΨA2 + A2 M ΨA3 + A3 Ψ From Schur complement, under Assumption 2, Ψ̄Ā + ĀΨ̄ > 0 is equivalent to the following Riccati inequality. −1 ⊤ − Ā3 Ψ − ΨĀ⊤ A2 Ψ + A2 M P −1 M A⊤ 3 + ΨA2 P 2 < 0 (35) Then, we have the following lemma. zu∗ )⊤ (∇f (ẑu ) Since ν̃ = −M 2 (z̃u + d˜q ), it follows D+ So ≤ −ν̃ ⊤ d˜q − (z̃u + d˜q )⊤ M 2 (z̃u + d˜q ) + z̃u⊤ ν̃. (31) A positive semi-definite solution Ψ to (35) is shown to exist from Lemma 8. Now, define an energy function Sx := 1 ⊤ 2 x̃ Ψ̄x̃ for the solution Ψ to (35). Then, the time derivative of Sx along the trajectories of (33a) is given by 1 Ṡx = − x̃(Ψ̄Ā + ĀΨ̄)x̃ + x̃⊤ Ψ̄BM −1 ζ̃ 2 ≤ x̃⊤ B ζ̃ = (B ⊤ x̃)ζ̃ = x̃⊤ (36) 1 ζ̃. This completes the proof. We are now ready to prove Lemma 4. Replace r̃ − B ⊤ x̃ in (33b) by ẽ as ˜ ξ˜˙ = k ẽ, ζ̃ = k̄ M ẽ + M ξ. (37) I P 1 ˜⊤ 2kI ξ M ξ̃. Then, the time derivative of Sξ Define Sξ := along the trajectories of (37) is given as Ṡξ = ξ˜⊤ M ẽ = (ζ̃ − k̄P M ẽ)⊤ ẽ = ζ̃ ⊤ ẽ − kP ẽ⊤ M ẽ. (38) Define Sp := Sx + Sξ . Then, from (36) and (38), we have Ṡp ≤ ζ̃ ⊤ r̃ − kP (r̃ − x̃1 )⊤ M (r̃ − x̃1 ) ≤ ζ̃ ⊤ r̃. This completes the proof. (39)
3cs.SY
Language Support for Reliable Memory Regions arXiv:1611.02823v2 [cs.DC] 23 Nov 2016 Saurabh Hukerikar, Christian Engelmann Computer Science and Mathematics Division Oak Ridge National Laboratory Oak Ridge, TN, USA Email: {hukerikarsr, engelmann}@ornl.gov Abstract. The path to exascale computational capabilities in highperformance computing (HPC) systems is challenged by the inadequacy of present software technologies to adapt to the rapid evolution of architectures of supercomputing systems. The constraints of power have driven system designs to include increasingly heterogeneous architectures and diverse memory technologies and interfaces. Future systems are also expected to experience an increased rate of errors, such that the applications will no longer be able to assume correct behavior of the underlying machine. To enable the scientific community to succeed in scaling their applications, and to harness the capabilities of exascale systems, we need software strategies that enable explicit management of resilience to errors in the system, in addition to locality of reference in the complex memory hierarchies of future HPC systems. In prior work, we introduced the concept of explicitly reliable memory regions, called havens. Memory management using havens supports reliability management through a region-based approach to memory allocations. Havens enable the creation of robust memory regions, whose resilient behavior is guaranteed by software-based protection schemes. In this paper, we propose language support for havens through type annotations that make the structure of a program’s havens more explicit and convenient for HPC programmers to use. We describe how the extended haven-based memory management model is implemented, and demonstrate the use of the language-based annotations to affect the resiliency of a conjugate gradient solver application. This work was sponsored by the U.S. Department of Energy’s Office of Advanced Scientific Computing Research. This manuscript has been authored by UT-Battelle, LLC under Contract No. DE-AC05-00OR22725 with the U.S. Department of Energy. The United States Government retains and the publisher, by accepting the article for publication, acknowledges that the United States Government retains a non-exclusive, paid-up, irrevocable, world-wide license to publish or reproduce the published form of this manuscript, or allow others to do so, for United States Government purposes. The Department of Energy will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan (http://energy.gov/downloads/doe-public-access-plan). 1 Introduction The high-performance computing (HPC) community has their sights set on exascale-class computers, but there remain several challenges in designing these systems and preparing application software to harness the extreme-scale parallelism. Due to constraints of power, emerging HPC system architectures will employ radically different node and system architectures. Future architectures will emphasize increasing on-chip and node-level parallelism, in addition to scaling the number of nodes in the system, in order to drive performance while meeting the constraints of power [1]. Technology trends suggest that present memory technologies and architectures will yield much lower memory capacity and bandwidth per flop of compute performance. Therefore, emerging memory architectures will be more complex, with denser memory hierarchies and utilize more diverse memory technologies [2]. The management of resilience to the occurrence of frequent faults and errors in the system has also been identified as a critical challenge [3]. HPC applications and their algorithms will need to adapt to these evolving architectures, which will also be increasingly unreliable. These challenges have led to suggestions that our existing approaches to programming models must change to complement existing system-level approaches [4]. The demands for massive concurrency and the emergence of high fault rates require that programming model features also support the management of resilience and data locality in order to achieve high performance. Recent efforts in the HPC community have focused on improvements in the scalability of numerical libraries and implementations of Message Passing Interface (MPI) libraries for these to be useful on future extreme-scale machines. However, there is also a need to develop new abstractions and methods to support fault resilience. In prior work, we proposed a resilience-driven approach to memory management using havens [5]. Havens offer an explicit method for affecting resilience in the context of memory management decisions. In haven-based memory management, each allocated object is placed in a program-specified haven. The havens guarantee a specified level of robustness for all the program objects contained in a memory region. The objects contained in havens may not be freed individually; instead the entire haven is deallocated, leading to the deletion of all the contained objects. Each haven is protected by a detection/correction mechanism, and different havens in a program may be protected using different resilience schemes. The use of havens provides structure to resiliency management of the program memory by grouping related objects based on the objects’ individual need for robustness and the performance overhead of the resilience mechanism. This approach to memory management enables HPC applications to write their own disciplines to enhance the resilience features of arbitrary types of memory. Traditional region-based systems were designed to statically assign program objects to memory regions, based on compiler analysis, in order to eliminate the need for runtime garbage collection [6]. In contrast, the primary goal of havens is to provide a scheme for creating regions within heap-allocated memory with various resilience features. In our initial design, we defined interfaces for the creation and use of havens that were implemented by a library interface [5]. In this paper, we develop language support in order to make havens clearer and more convenient to use in HPC application programs by supporting as many C/C++ language constructs as possible. This paper makes the following contributions: – We make a realistic proposal for adding language support for havens to mainstream HPC languages. – We develop type annotations, which enable static encoding of the decisions for a program object’s allocation and deallocation into the robust regions. They also provide opportunities to optimize the trade-off between the robustness and performance overhead for protecting program objects. – We investigate how affecting the resilience of individual program objects using these static annotations affects their fault coverage and performance during application execution. 2 Havens: Reliable Memory Regions Havens are designed to support resilience-driven memory management. The runtime memory is partitioned into robust regions, called havens, into which program objects are allocated. Various object deallocation policies may be defined for each haven, but the default is to free all the objects in a haven at once by deleting the entire pool of memory. Therefore, havens enable the association of lifetime to the reliable memory regions. Each memory region is protected by a predefined robustness scheme that provides error detection and/or correction for all objects in the haven. Any robustness scheme used by a haven is intended to be agnostic to the algorithm features, and to the structure of the data objects placed in havens. The concept of havens maintains a clear separation between the memory management policies and the mechanism that provides error resilience. Different havens used by an application may be protected using different detection/correction schemes, such as software-based parity, hashing, replication, etc., each of which may carry a different level of performance overhead. Therefore, havens enable the program memory to be logically partitioned into distinct regions each of which possess a specific level of error resilience and performance overhead. From the perspective of an HPC application program, havens enable applications to exert fine-grained control on the resilience properties of individual program objects. Since different havens may have varying guarantees of resilience and performance overhead, object placement in havens may be driven by the trade-off between criticality of the object to program correctness and the associated overhead. Havens are used to create a logical grouping of objects that require similar resilience characteristics. Havens also enable improvements to the locality of dynamically allocated objects by placement and aggregation of various objects based on an application’s pattern of use. Furthermore, havens permit HPC applications to balance the locality of program objects with their resilience needs. For example, a runtime system may dynamically map a haven onto specific hardware units in the memory hierarchy in an effort to improve the locality of its program objects; such mapping may also be guided by the availability of hardware-based error detection/correction in the memory unit that cooperates with the software-based protection scheme of the haven. 3 3.1 Using Havens for Resilience-driven Memory Management Basic Operations While developing the concept of havens, we defined an interface for HPC programs to effectively use the reliable memory regions in their application codes [5]. The abstract interface is based on the notion of a haven manager, which provides a set of basic operations that must be implemented to fully support the use of havens. The operations are summarized below: 1. 2. 3. 4. 5. 6. 3.2 haven create : The request for the creation of a haven by an application returns a handle to the memory region, but no memory is allocated. The choice of the error protection scheme is specified during the haven creation operation. haven alloc : An application requests a specified block of memory within a haven using this interface. This operation results in the allocation of the memory and the initialization of state related to the protection scheme. haven delete : The interface indicates intent to delete an object within the haven, but the memory is not released until the haven is destroyed. haven read and haven write : These interfaces read and update the program objects contained in the haven; the operations are performed through these interfaces, rather than directly on the objects, to enable the haven manager to maintain updated state about the robustness mechanism. haven destroy : The interface requests that the haven be destroyed, which results in all memory blocks allocated in the region to be deallocated. Upon completion of this operation, no further operation on the haven are permitted, and the memory is available for reuse. The state related to the robustness scheme maintained by the haven manager is also destroyed. haven relax and haven robust : These interfaces enable the error protection scheme applied to a haven to be turned on and off based on the needs of the application during program execution. Haven Library Interface The implementation of the havens library is similar to the one in [5], in which the heap is divided into fixed-size pages, and each new haven creation is aligned to a page boundary. The library maintains a linked list of these pages. We provide the library API functions for each of the primitives that enable basic haven operations: the haven alloc() and haven new() implement the abstraction for the allocation of objects into the associated region. With the library-based implementation of the haven interfaces, we require no changes to the representation of pointers. Pointers may reference havens or access individual objects in the havens. Since the library implementation does not differentiate between the pointer types, any conversions between these two kinds of pointers are potentially unsafe, and may lead to incorrect behavior. We only support per-region allocation and deallocation, and therefore per-object deallocation is an illegal operation. The haven release() enables the expression of the end of object life. However, the haven destroy() operation must be invoked to release the memory, which is achieved by concatenating the haven’s page list to the global list of free pages. 3.3 Protection Schemes for Havens In our initial implementation of havens, the memory regions are guaranteed highly-reliable behavior through comprehensive protection based on a lightweight software-based parity scheme. This scheme requires the haven manager to maintain a pair of signatures for each memory region, which are of word length for error correction, and an additional word length signature for error detection. The detection signature contains one parity bit per word in the memory region. As memory is allocated for the region and initialized, the correction signature S1 retains the XOR of all words that are written to the memory region. We apply an XOR operation on every word that is updated in the memory region and the correction signature S2. Silent data corruptions or multi-bit errors are detected by checking the detection signature for parity violations. The detection signature also enables the location of the corrupted memory word to be identified. The value at the corrupted memory location may be recovered using the signatures S1 and S2. The XOR of these two signatures S1 and S2 equals the XOR of all the uncorrupted locations in the haven. Therefore, the corrupted value in the memory region is recovered by performing an XOR operation on the remaining words in the haven with the XOR of the signatures S1 and S2. The recovered value overwrites the corrupted value, and the detection signature is recomputed. This parity-based protection is an adaptation of an erasure code. Using this scheme, multibit corruptions may be recovered from unlike hardware-based ECC, which offers only single bit error correction and double bit error detection. The scheme maintains limited state for the detection and correction capabilities and therefore carries very little space overhead in comparison to other software-based schemes such as software-based ECC and checksums. Additionally, the detection/recovery operations are transparent to the application. The detection is a constant time operation while the recovery is a O(n) operation based on the size of the haven. 4 4.1 A Haven Type System Goals Havens express the intended relationships between locality and resilience requirements of various program objects. The use of havens brings structure to memory management by grouping related program objects based on their resiliency and locality needs. The initial prototype implementation of havens contains library interfaces for each of the primitive haven operations [5]. The language support for havens aims to make programming HPC applications with havens straightforward and productive by making the programs using havens clearer and easier to write and to understand. Our design of the haven language support seeks to address the following seemingly conflicting goals: – Explicit: HPC programmers control where their program objects are allocated and explicitly define their robustness characteristic and lifetime. – Convenience: A minimal set of explicit language annotations that support as many C/C++ idioms as possible in order to facilitate the use of havensbased memory management in existing HPC application codes, as well as in the development of new algorithms. – Safety: The language annotations must prevent dangling-pointer dereferences and space leaks. – Scalability: The havens must support various object types and the performance overhead of any resilience scheme scales well even with large number of objects. The language support enables HPC programmers to statically encode memory management decisions for various program objects. By making the structure of the havens and their resilience features explicit, the number of runtime checks and modifications to the haven structure and the resilience scheme are reduced. 4.2 Type Annotations for Havens In the haven-based model for memory management, the heap is divided into regions, each containing a number of program objects. Therefore, havens are abstract entities that represent an aggregation of program objects. Pointers to havens refer to these abstract entities in the heap, whose resilience scheme is defined upon creation and provides protection to all program objects that are contained within the haven. The definition of a haven pointer type provides a statically enforceable way of specifying the resilience scheme, type and size information for the encapsulated objects inside the haven. A haven type statically ensures that programs using this region-based model of memory management are memory-safe, i.e., they dont permit dangling references. The haven ptr is a new type for handles to havens. The declaration of a haven ptr typed pointer leads to the creation of a haven, but the declaration of a haven does not allocate any memory. The haven-typed pointer object is declared and the haven is subsequently deleted as shown in Listing 1.1. haven_ptr h1 ; . . . deletehaven h1 ; Listing 1.1. Type Annotations for Havens The haven ptr is smart pointer object that contains the pointer reference to a haven and also maintains bookkeeping information about the objects resident to the haven, including their sizes and a reference count. This information enables the library to optimize the resilience scheme that protects the haven. For example, in the parity-based protection scheme, the haven is protected using a pair of parity signatures. The availability of the count and sizes of the objects inside the haven enables statically creating sub-havens that are each protected by pair of signatures. We define the deletehaven operator that provides a static mechanism to reclaim the memory allocated for objects inside a haven, and also discards the bookkeeping information and any state maintained by the resilience scheme (for e.g., the signatures that provide parity protection for the haven). The library implementation of havens permits unsafe operations, since a haven h may be deleted even if the program contains accessible pointers to objects in h. With the introduction of the haven ptr type, we also address the issue of safety. When the deletehaven operator is encountered, the safety of the delete operation is guaranteed by checking the reference counts included in the haven ptr typed pointer object. The delete operation succeeds when the haven ptr contains all null object pointers, and the operation results in releasing the storage space for the haven, along with the program objects contained in the haven. When the haven ptr typed pointer object contains a non-zero count of active object references, the delete operation fails. 4.3 Subtyping Annotations A subtype annotation is used to constrain the membership of an object to a specific haven. Each object type is annotated with a region expression, which explicitly specifies the haven to which values of that type belong. The region expression is always bound to the type declaration of an object. // Declare new haven pointer h1 haven_ptr h1 ; // Declare v a r i a b l e x as member of the haven h1 int < h1 > x ; x = 4; // Delete haven r e l e a s e s memory for haven and the c o n t a i n e d v a r i a b l e x deletehaven h1 ; Listing 1.2. Subtype Annotations for Havens The type<haven ptr> defines a subtype for non-pointer variables that guarantees the allocation of the qualified object within a haven. The type annotation enables local variables and global variables in C/C++ programs to be associated with a haven. The haven membership of the annotated variable also guarantees the variable with the protection offered by the haven’s specified resilience scheme. The declaration of a single integer variable inside a haven is written as shown in Listing 1.2. The type*<haven ptr> annotation defines a subtype for pointer objects. The inclusion of the haven ptr specifies membership of the object referenced by the annotated variable to the haven. The declaration of an array inside a haven and the allocation of memory for the array is written as shown in Listing 1.3. // Declare new haven pointer h2 haven_ptr h2 ; // Declare vector pointer as member of the haven h2 double * < h2 > vector ; // A l l o c a t e memory for vector of size N vector = haven_alloc ( N * sizeof ( double ) ) ; . . . // Set vector pointer to be null ; without this d e l e t e h a v e n fails vector = null ; // Delete haven release memory for haven and the c o n t a i n e d vector deletehaven h2 ; Listing 1.3. Declaration of an array object within a haven The membership relationship between variables and havens expressed by the subtyping annotations also enables programmers to imply locality of reference for all program objects that are associated with a haven. Restrictions: With the use of the type annotations for object pointers, programmers need to differentiate between traditional C/C++ pointers and pointers that specify haven membership. Any conversion between these two kinds of pointers is potentially unsafe and may lead to incorrect program behavior. Therefore, we define a null haven, which enables traditional C/C++ pointers to be viewed as pointers to objects inside this null region. The compiler guarantees safe assignments of pointer variables through static analysis or runtime checks. 4.4 Defining Lifetimes Through language support, we also define the notion of lifetimes for havens. The basic idea is to define the scope of computation for which a haven is valid. We define the reference lifetime for a haven as shown in Listing 1.4. This syntax enables the creation of dynamic havens, whose lifetime is the execution of the statement s; the statement s may be a compound statement. The program objects that are allocated within the haven hx are guaranteed error protection through the haven’s default resilience scheme. The explicit definition of lifetimes for the havens enables programs to scope specific regions of computation that must be executed with high reliability. haven hx { // s t a t e m e n t s } Listing 1.4. Defining lifetime scope for havens 4.5 Example: Vector Addition The example in Listing 1.5 shows the skeleton of the vector addition code, in which the objective is to protect the operand vectors a and b. The example omits the details of the vector initialization and the addition routines. The declaration of the haven ptr pointer variable with identifier h3 creates the haven. Upon creation of the haven, the parity signatures are initialized, but no memory is allocated. // Create a haven for vectors haven_ptr h3 ; // Declare vectors as members of the haven h1 double * < h3 > a = haven_alloc ( N * sizeof ( double ) ) ; double * < h3 > b = haven_alloc ( N * sizeof ( double ) ) ; // Declare t r a d i t i o n a l vector pointer as member of null haven double * < null > c = malloc ( N * sizeof ( double ) ) ; // Vector a d d i t i o n c = a + b v ec to r _a dd it i on (c , a , b ) ; // Set vector p o i n t e r s to null ; without this d e l e t e h a v e n fails a = null ; b = null ; free ( c ) ; deletehaven h3 ; Listing 1.5. Example: Resilient Vector Addition using Havens Language Support The sub-type declaration of the array pointers makes the relationship between the operand vectors and the haven h3 explicit and ensures the allocation of the vectors inside the haven. When the haven alloc allocation requests are made, the library initializes the resilience scheme for the haven and allocates the vectors a and b of size N elements. The array pointer to the result vector c is a traditional pointer that is declared as a sub-type to a double* that establishes membership of the null haven. When the vector addition function returns, the operand vector pointers are set to null so that the deletehaven operator is able to release the memory associated with the haven h3 that includes vectors a and b. 5 Application-Level Resilience Models using Havens A variety of algorithm-based fault tolerance (ABFT) strategies have been extensively studied over the past decades. Many of these techniques are designed to take advantage of the unique features of an application’s algorithm or data structures. These techniques are also able to leverage the fact that different aspects of the application state have different resilience requirements, and that these needs vary during the execution of an application. However, the key barrier to the broader adoption of algorithm-based resilience techniques in the development of HPC applications is the lack of sufficient programming model support since the use of these features requires significant programming effort. We explore three generalized application-level resilience models that may be developed using havens, and whose construction is facilitated by the languagebased annotations. These models are intended to serve as guidelines for HPC application programmers to develop new algorithms as well as adapt the existing application codes to incorporate algorithm-based resilience capabilities: – Selective Reliability: Based on the insight that different variables in an HPC program exhibit different vulnerabilities to errors, havens provide specific regions of program memory with comprehensive error protection. With this model, HPC programmers use havens as mechanisms to explicitly declare specific data and compute regions to be more reliable than the default reliability of the underlying system. – Specialized Reliability: Various protection schemes that provide error/detection and correction capabilities for havens guarantee different levels of resiliency. Also, based on the placement of havens in physical memory, the software-based schemes may complement any hardware-based capabilities. Havens provide simplified abstractions to design resilience strategies that seek to complement the requirements of different program objects with the various hardware and software-based protection schemes available. – Phased Reliability: The vulnerability of various program objects and computations to errors varies during program execution. Havens may also be used to partition applications into distinct phases of computations. Since the various resilience schemes incur overheads to the application performance, the protection features of specific data regions and compute phases may be enabled or disabled in order to trade-off performance overhead and resilience. 6 Experimental Results To apply the static annotations in an HPC application, we must identify program objects that must be allocated in havens, and annotate their declarations with the type qualifiers. These experiments evaluate the use of haven-based memory management using the type qualifiers for a conjugate gradient code by including the type and subtype qualifiers on its various application objects. We use a pre-conditioned iterative CG algorithm and we validate the correctness of the Fig. 1. Performance overheads of havens with static annotations outcome of the solver with a solution produced using a direct solver. We compare the evaluation with the results from our previous implementation that required insertion of raw library interfaces. One of the important advantages of using the static annotations is that the number of lines of code changed is reduced significantly when compared to the changes required for insertion of library calls in the same application code, which improves code readability. In the CG algorithm, which solves a system of linear equations A.x = b, the algorithm allocates the matrix A, the vector b and the solution vector x. Additionally, the conjugate vectors p and the residual vector r are referenced during each iteration of the algorithm. The program objects in the CG application demonstrate different sensitivities to errors. Errors in the operand matrix A or vector b fundamentally changes the linear system being solved. For errors in these structures even if the CG solver converges to a solution, it may be significantly different from a correct solution. The preconditioner matrix M demonstrates lower sensitivity to the errors, as do the vectors x, p, r. These features of the CG algorithm form the basis for the strategic placement of the objects into havens, since the allocation of only sensitive data structures into havens provides a substantially higher resilient behavior in terms of completion rates of the CG algorithm for reasonable overheads to performance than a naive placement strategy. We present a detailed sensitivity analysis in [5]. Here, we evaluate the performance benefits gained from the use of static annotations for the various objects in the CG code. We perform two sets of experiments: (i) we allocate only one structure using the haven static annotations, while the remaining structures are allocated using the standard memory allocation interfaces; (ii) we strategically annotate the data structures of the CG to allocate structures to havens in specific combinations. We evaluate the following combinations: (i) allocation of only the static state, i.e., the matrix A and vector B, the preconditioner M into havens, while the dynamic state, i.e., all the solution vectors, are allocated using standard memory allocation functions; (ii) allocation of only matrix A and vector B into havens; (iii) only the dynamic state is provided fault coverage using havens. We compare these strategies with allocations in which havens provide complete coverage and with experimental runs which do not allocate any structure using havens. The performance overhead of using havens in terms of the time to solution of the CG solver for the above selection of program objects for allocation into havens is shown in Figure 1. The annotation of all the program variables to be allocated into havens provides higher fault coverage, but it results in higher overhead to the time to solution for the CG application. When the variables are allocated using raw library interfaces, each program object is protected by a pair of signatures, which provides monolithic protection for the entire haven. When these objects are qualified with the static annotations in the application code, the compiler and library have a better understanding of the size and structure of the program objects. Therefore, the larger program objects, notably the operand matrix A and the preconditioner matrix M, are split and protected by multiple pairs of parity signatures. This split protection is transparent to the application programmer and the application still accesses the matrix elements as a single data structure. The use of multiple signatures improves the read/write overhead for the objects and the observed overhead with static annotations for all program objects is 11% lower than the library-based allocation for the same set of objects. The operand matrix A occupies a dominant part of the solver’s memory, occupying over 50% of the active address space, whereas the solution vector x, the conjugate vectors p and the residual vector r and the preconditioner matrix M account for the remaining space. Therefore, the annotation of matrix A individually results in 9% lower overhead than with monolithic parity protection using library interfaces. The improvement in performance when smaller data objects are statically annotated is only within 2% of the version using library interfaces for the same objects. 7 Related Work Much research has been devoted to studies of algorithms for memory management, which are based on either automatic garbage collection or explicit allocation/deallocation schemes. The concept of regions was implemented in storage systems, which allowed objects to be allocated in specific zones [7]. While each zone permits a different allocation policy, the deallocation is performed on a per-object basis. The vmalloc library [8] provides programmers with an interface to allocate memory and to define policies for each allocation. Region-based systems, such as arenas [9], enable writing special-purpose memory allocators that achieve performance by creating heap memory allocation disciplines that are suited to the application’s needs. Implementations such as vmalloc place the burden of determining policy of allocation of objects to regions on the programmer [8]. Other schemes have used profiling to identify allocations that are short-lived and place such allocations in fixed-size regions [10]. Several early implementations of region-based systems were unsafe; the deletion of regions often left dangling pointers that were subsequently accessible. Such safety concerns were addressed through reference counting schemes for the regions [11]. For dynamic heap memory management through static analysis, regions provide [6] an alternative to garbage collection methods. In this approach, the assignment of program objects to regions is statically directed by the compiler in an effort to provide more predictable and lower memory space. The approach was refined by relaxing the restriction that region lifetimes must be lexical [12]. Language support for regions is available in many declarative programming languages such as ML [13], Prolog [14]. Cyclone is a language designed to be syntactically very close to C, but which provides support for regions through an explicit typing system [15]. The Rust programming language [16] also provides support for regions. Recent efforts seek provide programming model support for reliability, such as containment domains [17], which offer programming constructs that impose transactional semantics for specific computations. Our previous work on havens [5] provided a reliability-driven method for memory allocations. Rolex [18] offers language-based extensions that support various resilience semantics on application data and computations. Global View Resilience (GVR) supports reliability of application data by providing an interface for applications to maintain versionbased snapshots of the application data [19]. In support of fault tolerance of in explicit memory allocation/deallocation, the malloc failable interface is used by the application to allocate memory on the heap; callback functions are used to handle error recovery for the memory block [20]. 8 Conclusion Resilience is among the major concerns for the next generation of extreme-scale HPC systems. With the rapid evolution of HPC architectures and the emergence of increasingly complex memory hierarchies, applications running on future HPC systems must manage the locality and maintain reliability of their data. Havens provide an explicit software-based approach for HPC applications to manage the resilience and locality of their programs. In this paper, we focused on developing language support for havens with emphasis on providing structure to the haven-based memory management. Through type annotations, a programmer expresses the intended relationships between locality and resilience requirements of various objects in the application program. The type annotations enable the resilience requirements of program objects to be encoded within the heap memory-management idioms. The static typing discipline for application codes written in C/C++ also guarantees the safety of memory operations by preventing dangling-pointer dereferences and space leaks. The structured haven-based management facilitated by the language support provides the mechanisms for the development of effective application-based resilience models for HPC applications. References 1. Shalf, J., Dosanjh, S., Morrison, J.: Exascale computing technology challenges. In: Proceedings of the 9th International Conference on High Performance Computing for Computational Science. VECPAR’10, Springer-Verlag (2011) 1–25 2. Kogge, P., Bergman, K., Borkar, S., Campbell, D., Carlson, W., Dallya, W., Denneau, M., Franzon, P., Harrod, W., Hill, K., Hiller, J., Karp, S., Keckler, S., Klein, D., Lucas, R., Richards, M., Scarpelli, A., Scott, S., Snavely, A., Sterling, T., Williams, R.S., Yelick, K.: Exascale Computing Study: Technology Challenges in Achieving Exascale systems. Technical report, DARPA (September 2008) 3. DeBardeleben, N., Laros, J., Daly, J., Scott, S., Engelmann, C., Harrod, B.: HighEnd Computing Resilience: Analysis of issues facing the hec community and pathforward for research and development. Whitepaper (December 2009) 4. Amarasinghe, S., Hall, M., Lethin, R., Pingali, K., Quinlan, D., Sarkar, V., Shalf, J., Lucas, R., Yelick, K., Balaji, P., Diniz, P.C., Koniges, A., Snir, M., Sachs, S.R., Yelick, K.: Exascale Programming Challenges: Report of the 2011 workshop on exascale programming challenges. Technical report, U.S. Department of Energy, Office of Science, Office of Advanced Scientific Computing Research (ASCR) (July 2011) 5. Hukerikar, S., Engelmann, C.: Havens: Explicit reliable memory regions for hpc applications. In: IEEE High Performance Extreme Computing Conference (HPEC). (September 2016) 1–6 6. Tofte, M., Talpin, J.P.: Implementation of the typed call-by-value λ-calculus using a stack of regions. In: Proceedings of the 21st ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages. POPL ’94, New York, NY, USA, ACM (1994) 188–201 7. Ross, D.T.: The aed free storage package. Communications of ACM 10(8) (August 1967) 481–492 8. Vo, K.P.: Vmalloc: A general and efficient memory allocator. Software: Practice and Experience 26(3) (1996) 357–374 9. Hanson, D.R.: Fast allocation and deallocation of memory based on object lifetimes. Softwre Practices & Experience 20(1) (January 1990) 5–12 10. Barrett, D.A., Zorn, B.G.: Using lifetime predictors to improve memory allocation performance. In: Proceedings of the ACM SIGPLAN 1993 Conference on Programming Language Design and Implementation. PLDI ’93, New York, NY, USA (1993) 187–196 11. Gay, D., Aiken, A.: Memory management with explicit regions. In: Proceedings of the ACM SIGPLAN 1998 Conference on Programming Language Design and Implementation. PLDI ’98, New York, NY, USA, ACM (1998) 313–323 12. Aiken, A., Fähndrich, M., Levien, R.: Better static memory management: Improving region-based analysis of higher-order languages. In: Proceedings of the ACM SIGPLAN 1995 Conference on Programming Language Design and Implementation. PLDI ’95 (1995) 174–185 13. Tofte, M., Birkedal, L., Elsman, M., Hallenberg, N., Olesen, T.H., Sestoft, P., Bertelsen, P.: Programming with regions in the ml kit, technical report (diku-tr97/12). Technical report, University of Copenhagen, Denmark (April 1997) 14. Makholm, H.: A region-based memory manager for prolog. In: Proceedings of the 2nd International Symposium on Memory Management. ISMM ’00, New York, NY, USA, ACM (2000) 25–34 15. Grossman, D., Morrisett, G., Jim, T., Hicks, M., Wang, Y., Cheney, J.: Regionbased memory management in cyclone. In: Proceedings of the ACM SIGPLAN 2002 Conference on Programming Language Design and Implementation. PLDI ’02, New York, NY, USA, ACM (2002) 282–293 16. : The rust programming language http://www.rust-lang.org. 17. Chung, J., Lee, I., Sullivan, M., Ryoo, J.H., Kim, D.W., Yoon, D.H., Kaplan, L., Erez, M.: Containment domains: a scalable, efficient, and flexible resilience scheme for exascale systems. In: Proceedings of the International Conference on High Performance Computing, Networking, Storage and Analysis. (2012) 58:1–58:11 18. Hukerikar, S., Lucas, R.F.: Rolex: Resilience-oriented language extensions for extreme-scale systems. The Journal of Supercomputing (2016) 1–33 19. Chien, A., Balaji, P., Beckman, P., Dun, N., Fang, A., Fujita, H., Iskra, K., Rubenstein, Z., Zheng, Z., Schreiber, R., Hammond, J., Dinan, J., Laguna, I., Richards, D., Dubey, A., van Straalen, B., Hoemmen, M., Heroux, M., Teranishi, K., Siegel, A.: Versioned distributed arrays for resilience in scientific applications: Global view resilience. Procedia Computer Science 51 (2015) 29 – 38 20. Bridges, P.G., Hoemmen, M., Ferreira, K.B., Heroux, M.A., Soltero, P., Brightwell, R.: Cooperative application/os dram fault recovery. In: Proceedings of the 2011 International Conference on Parallel Processing - Volume 2. Euro-Par’11, SpringerVerlag (2011) 241–250
6cs.PL
Finite State Machine Synthesis for Evolutionary Hardware Andrey Bereza, Maksim Lyashov, Luis Blanco Dept. of Information Systems and Radio Engineering, Don State Technical University [email protected], [email protected], [email protected] Abstract This article considers application of genetic algorithms for finite machine synthesis. The resulting genetic finite state machines synthesis algorithm allows for creation of machines with less number of states and within shorter time. This makes it possible to use hardware-oriented genetic finite machines synthesis algorithm in autonomous systems on reconfigurable platforms. 1. Introduction Nowadays evolutionary algorithms (EA) are applied to design digital and analog devices [1]. This trend is called evolutionary electronics [2, 3]. The application of EA on hardware platforms with reconfigurable elements, which allows for rebuilding the systems in process of operation, is called evolutionary hardware [4]. Evolutionary hardware (EH) is a new type of hardware based on various probabilistic algorithms such as genetic algorithms (GA) and evolutionary programming. In EH design the reconfigurable parts are dynamically rebuilding combinatory or sequential logic circuits [4]. To dynamically rebuild digital logic circuit it is necessary for GA to be able to synthesize circuit on a gate level. Hence, the task of digital logic circuit synthesis EA development arises. Current methods of finite state machine synthesis always use the specifics of a problem, which makes it impossible to use that same state machine generation technic for any different kind of problem. The quest is to make universal state machine synthesis method applicable to a wide range of problems. The application of EA for finite machine synthesis is shown in the work [5]. However given algorithms are applied for state machine programming, where program is described with finite state machines, which doesn’t allow their usage in autonomous hardware systems or reconfigurable platforms. ______________________________________________________________ * The research is supported by Russian Foundation for Basic Research (grant #13-07-00951) 2. The problem of state machine evolutionary synthesis The problem of state machine evolutionary synthesis is defined as set: { } where H – the synthesized solution genotype, О – the { }, F – the objective genetic operators function. The synthesized solution genotype is defined as set: { }, where , S – the amount of finite machine states, x – the amount of inputs. The objective function is defined as expression: , where – the amount of states, – the amount of iterations, w – the weight coefficients for particular criteria. The task for GA is to minimize the objective function, e.g. . 3. Hardware-oriented genetic algorithm of finite state machine synthesis Schematic diagram of proposed hardware oriented genetic algorithm of finite state machine synthesis, designed for EH creation on a reconfigurable platform is shown on a figure 1. On the first step user sets requirements for finite state machine being designed. Those are the numbers of states and triggers. Also the amount of generations and mutation and crossing over probabilities must be set to organize the process of evolutionary synthesis. Then, according to the algorithm, initial set of solutions is generated and evaluated; transition correction algorithm is also executed. If during the evaluation process of initial set of solutions population has a solution that meets all the requirements, then it gets saved and algorithm ends. The terminal criteria of GA are reaching the maximum number of generations or having a solution. Primary genetic operators in FSMGA are mutation and crossing over. Chromosome coding with bit string applies some constrains on operator types. deleted from population (since after crossing over and mutation operators 3 more chromosomes are inserted into population). Since proposed GA is designed to function on autonomous EH, chromosomes are coded with bit strings. Let’s consider chromosome coding on a specific example of finite machine. Combinatory logic built on logical elements is replaced by RAM on the schematic diagram (figure 2). To do so, finite machine transition table should be converted first to be able to replace combinatory circuit with RAM. Transition graph of the finite machine is given on figure 3; it describes the behavior of some control device. The amount of triggers required to represent four states is equal to 2. Start Determine chromosome structure Generate initial set of solutions Evaluate initial set of solutions Define counter C = 1 C >= generation count OR solutionНет found? Yes No Select chromosome pair Apply crossing over and get 2 descendants Triggers Combinatory logic C ++ Input Mutate descendants Evaluate descendants and include them in population Result output Selection End Output Triggers RAM Figure 1 – Schematic diagram of hardware oriented genetic algorithm of finite state machine synthesis Mutation operator is random, e.g. it does not depend on chromosome fitness or gene residing in chromosome. As the result of mutation it randomly changes either the output value of state machine or state number which will be selected by randomly picked transition. Crossing over operator randomly exchanges genetic information between two solutions, while existing genetic information is being preserved. The GA solution quality is largely dependent on crossing over operator type selection. In proposed finite state machine synthesis algorithm one-point and two-point crossing over operators were applied, as those have the simplest hardware implementation [4]. Experimental studies have shown that two-point crossing over is preferable. Selection operation algorithm is based on bubble sort algorithm, since its hardware implementation takes the least resources among other sorting algorithms [4]. After the population has been sorted in descending order (e.g. chromosomes with a higher objective function value are moved to the top of the population), 3 chromosomes with worst objective function value are Input Output Figure 2 – Memory usage in state machine 0 / 10 1 / 11 00 0 / 01 01 1 / 11 0 / 01 11 0 / 00 10 1 / 10 1 / 11 Figure 3 – State machine sample For a state machine the number of RAM address inputs is equal to a sum of the amount of triggers and the number of inputs. RAM output number is accordingly equal to a sum of the amount of triggers and the number of outputs. For this finite machine sample RAM should contain 3 address inputs and 4 outputs; e.g. the required memory size is bits. The truth table of the RAM is equal to that one of a state machine. The schematic diagram for the state machine on a figure 3, which has been made by replacement of combinatory logics with RAM, is shown on figure 4. T2 А2 T1 А1 Input А0 Z3 8X4 RAM Z2 Z1 Z0 Output Figure 4 – Schematic diagram of state machine 4. Experimental studies The developed algorithm has been tested on two different problems: «Santa Fe Trail» problem («Smart Ant») and autopilot construction for simplified helicopter model problem. «Santa Fe Trail» – is a problem from the area of cooperative usage of GA and finite state machines [5]. The ant is on the surface of torus, which has size of 32x32 cells. The food is placed in some of the cells (on figure 5 marked as black). It is located along the broken line, but not in all cells. Broken line cells with no food are marked as gray. White cells do not belong to a broken line and contain no food. Altogether the field contains 89 food cells. Figure 5 – The «Santa Fe Trail» field Ant’s starting location is marked with “Start”. Ant occupies one cell and looks in one of four directions (up, down, left, right). Ant is able to determine if the food is directly in front of him. In one game turn ant is able to make one of three actions: − step forward, eating any food in destination; − turn left; − turn right. The food eaten by the ant won’t refill, the ant is always alive, the food isn’t vital for him. Broken line isn’t random, but strictly fixed. Ant is able to walk through any cell of field. The game is 200 moves long; each move ant performs one of the three actions. After 200 turns the amount of eaten food is calculated. That is the result of the game. The goal is to design an ant which will eat as much as possible food within 200 turns (all 89 is desirable). One of the ways to describe the behavior of the ant is Mealy machine, which has one input variable (tells if the food is in front of the ant), and a set of output actions consisting of three, described above. Schematic diagram of this machine is given on figure 6. Forward The food is in front F X Finite state machine L R Left «Ant» Right Figure 6 – Schematic diagram of «Smart Ant» finite state machine It is hard to heuristically build a machine solving problem. For instance, a heuristically built Mealy machine with five states [6] can’t solve this problem. Finite state machine describes the ant which eats only 81 food cells within 200 moves, and it takes 314 moves to eat all the food. Experimental studies of «Santa Fe Trail» problem were conducted with population size of 1200, crossing over probability of 0.4 and mutation probability of 0.25. The comparison of the results of proposed finite machine synthesis GA (FMSGA) versus the results of heuristic algorithm and GA proposed in work [6] are shown in table 1. Table 1 – Experimental results Amount Algorithm of states Heuristic 5 GA[6] 7 FMSGA 7 Moves 314 198 190 Synthesis time 269 s. 29 s. As implied by the above results, FMSGA has been able to find a machine with 7 states which solves the problem in 190 moves. It also takes 9 times less amount of time to synthesize the machine then existing analogs do. Consider the second problem of autopilot construction for simplified helicopter model. An autopilot has to be created for a simplified helicopter model which moves on a flat surface [5]. In one move helicopter model can either rotate through a certain predefined angle or change velocity. The autopilot’s task is to drive a helicopter through N markers within a limited time. The best autopilot is the one who manages to visit the highest number of markers. If two autopilots reach the same amount of markers, the one closest to a next marker at the last moment of the flight wins. The autopilot input variable receives the sight sector number (figure 7). Current target position relative to a helicopter is given as an angle between helicopter’s movement direction and the direction to a next marker (figure 7a). Helicopter always flies in the middle of the current sector. All sectors are static relative to a helicopter (figure 7b). 5 4 6 3 7 2 8 1 а) б) Figure 7 – Helicopter output data Autopilot model is a finite machine with discrete input and output actions. Machine state indirectly maps helicopter’s current position, its speed and history of state transitions. Experimental studies have been conducted with sector sizes of 4 and 6. For each parameter set 50 tests have been conducted. Experimental results are shown in table 2. The «Result» column shows amount of markers visited by autopilot designed with FMSGA. In work [5] the finite machine with 12 states is able to drive the helicopter through the first 18 out of 20 markers within given time. Table 2 – Experimental results of helicopter autopilot design with FMSGA Result Number of sectors Worst Average Best 4 11 18 20 6 12 17 20 5. Conclusion As shown by the given experimental results, developed FMSGA allows for machine synthesis within shorter time and with less number of states. This proves that developed hardware-oriented FMSGA can be effectively used in autonomous systems on reconfigurable platforms. 6. References [1] Pauline C. Haddow, Andy M. Tyrrell. Challenges of evolvable hardware: past, present and the path to a promising future. Genetic Programming and Evolvable Machines. September 2011, Volume 12, Issue 3, pp 183-215. [2] Bereza A.N., Kureichik V.M. Evolutionary methods of synthesis of circuit decisions for designing data telecommunication systems. Journal of Computer and Systems Sciences International. 2006. Т. 45. № 2. p. 234-244. [3] Zebulum, R.S., Pacheco, M.A., Vellasco, M.M. Evolutionary Electronics: Automatic Design of Electronic Circuits and Systems by Genetic Algorithms. USA, CRC Press LLC, 2002. – 320 p. [4] Garrison, W. Introduction to Evolvable Hardware: A Practical Guide for Designing Self-Adaptive Systems. IEEE Press Series on Computational Intelligence. – 2006. – 224 p. [5] Lobanov, P.G. Genetic algorithms usage in finite machine synthesis: tech. science cand. dissertation. – St. Petersburg, 2008. – 114 p. [6] Tsarev, F.N., Shalyto, A.A. About finite machine design with minimum number of states for «Santa Fe Trail» problem. Xth soft computing conference’s collection of reports. St. Petersburg Electrotechnical University. vol.2, 2007, p. 88–91.
9cs.NE
arXiv:1607.01731v2 [math.AC] 19 Aug 2017 ON THE ATOMIC STRUCTURE OF PUISEUX MONOIDS FELIX GOTTI Abstract. In this paper, we study the atomic structure of the family of Puiseux monoids, i.e, the additive submonoids of Q≥0 . Puiseux monoids are a natural generalization of numerical semigroups, which have been actively studied since midnineteenth century. Unlike numerical semigroups, the family of Puiseux monoids contains non-finitely generated representatives. Even more interesting is that there are many Puiseux monoids which are not even atomic. We delve into these situations, describing, in particular, a vast collection of commutative cancellative monoids containing no atoms. On the other hand, we find several characterization criteria which force Puiseux monoids to be atomic. Finally, we classify the atomic subfamily of strongly bounded Puiseux monoids over a finite set of primes. 1. Introduction A Puiseux monoid is an additive submonoid of the non-negative rational numbers. The family of Puiseux monoids is a natural generalization of that one comprising all numerical semigroups. In this paper, we explore the atomic structure of the former family, which is far more complex than the atomic structure of numerical semigroups. However, the controlled atomic behavior of numerical semigroups will guide our initial approach to Puiseux monoids. Numerical semigroups are atomic monoids that have been systematically studied since the mid-nineteen century; see the monograph [13] of Garcı́a-Sánchez and Rosales. In algebraic geometry, Noetherian local domains whose integral closures are finitely generated modules and discrete valuation rings pop up very often, and their associated valuations turn out to be numerical semigroups. Many properties of the previously mentioned domains can be fully characterized in terms of their valuation numerical semigroups. For more details, see [3]. Understanding the atomicity of Puiseux monoids can set the groundwork for a future exploration of the arithmetic properties and factorization invariants of their atomic subfamilies. Once we obtain good insight of the algebraic properties of Puiseux monoids, we might expect to use this family of commutative monoids to understand certain behaviors of Puiseux domains (see, e.g., [11, Sec. 13.3]) and other subdomains of power series with rational exponents. This would mirror the way numerical semigroups have been used to understand many attractive properties of subdomains of power series with natural exponents (see [3]). Date: March 6, 2018. 1 2 F. GOTTI Although significant effort has been put in exploring the arithmetic invariants of many families of atomic monoids (see, for instance, [6, 7, 9, 18]), very little work has gone into an attempt to classify them. In this paper, we find an entirely new family of atomic monoids hidden inside the realm of Puiseux monoids, increasing the current spectrum of atomic monoids up to isomorphism and, therefore, contributing to a classification of the aforementioned family. In addition, Puiseux monoids provide a source of examples of both atomic and non-atomic monoids. This new arsenal of examples might help to test several existence conjectures concerning commutative semigroups and factorization theory. The family of Puiseux monoids contains a vast collection of non-atomic representatives, monoids containing non-unit elements with no factorizations into irreducibles. Even more surprising, Theorem 5.2 identifies a subfamily of antimatter representatives, Puiseux monoids possessing no irreducible elements. In contrast, there are various subfamilies of Puiseux monoids whose members are atomic even when they are not isomorphic to numerical semigroups. We devote this paper to introduce and study the fascinating atomic structure of Puiseux monoids. In Section 2, we establish the terminology we will be using throughout this paper. In Section 3, after pointing out how Puiseux monoids naturally appear in commutative ring theory, we introduce some members of the targeted family, illustrating how much Puiseux monoids differ from numerical semigroups in terms of atomic configuration. Once we have highlighted the wildness of the atomic structure of the family being investigated, we show that its atomic members are precisely those containing a minimal set of generators (Theorem 3.6). We then present two sufficient conditions for atomicity (Proposition 3.8 and Theorem 3.10). In Section 4, we introduce the subfamily of strongly bounded Puiseux monoids, presenting simultaneously atomic and antimatter subfamilies failing to be strongly bounded. In the last section, we study the atomic configuration of strongly bounded Puiseux monoids. We present a sufficient condition for strongly bounded Puiseux monoids to be antimatter (Theorem 5.2). To conclude, we dedicate the second part of the last section to the classification of the atomic subfamily of strongly bounded Puiseux monoids over a finite set of primes (defined in Section 3). 2. Preliminary We begin by presenting some of the terminology related to the atomicity of commutative cancellative monoids. Then we briefly mention a few basic properties of numerical semigroups, the objects we generalize in this work. Our goal in this section is not to formally introduce the elementary concepts and results of commutative semigroups and factorization theory, but rather to fix notation and establish ON THE ATOMIC STRUCTURE OF PUISEUX MONOIDS 3 the nomenclature we will use later. For extensive background information on commutative semigroups and non-unique factorization theory, we refer readers to the monographs [15] of Grillet and [14] of Geroldinger and Halter-Koch, respectively. We use the double-struck symbols N and N0 to denote the sets of positive integers and non-negative integers, respectively. Moreover, if r is a real number, we will denote the set {z ∈ Z | z ≥ r} simply by Z≥r ; with a similar intention, we will use the notations Z>r , Q≥r , and Q>0 . If S ⊆ Q, we often write S • instead of S \ {0}. For r ∈ Q>0 , we denote the unique a, b ∈ N such that r = a/b and gcd(a, b) = 1 by n(r) and d(r), respectively. If R ⊆ Q>0 , we call the sets n(R) = {n(r) | r ∈ R} and d(R) = {d(r) | r ∈ R} the numerator and denominator set of R, respectively. Unless otherwise specified, the word monoid in this paper means commutative cancellative monoid. Let M be a monoid. Because every monoid is assumed to be commutative, unless we state otherwise, we will always use additive notation; in particular, “ + ” denotes the operation of M, while 0 denotes the identity element. The invertible elements of a monoid are called units, and the set of all units of M is denoted by M × . The monoid M is said to be reduced if M × = {0}. If M is generated by a subset S, we write M = hSi. The monoid M is finitely generated if M = hSi for some finite set S. For a brief but precise exposition of finitely generated commutative monoids, readers might find [12] very useful. An element a ∈ M \M × is irreducible or an atom if a = x + y implies either x or y is a unit. The set of atoms of M is denoted by A(M). The monoid M is atomic if every non-unit element of M can be expressed as a sum of atoms, i.e., M = hA(M)i. We briefly comment on general properties of numerical semigroups. A numerical semigroup N is a submonoid of the additive monoid N0 such that N0 \N is finite. Every numerical semigroup has a unique minimal set of generators, which happens to be finite. For n ∈ N, if N = ha1 , . . . , an i is minimally generated by a1 , . . . , an ∈ N, then gcd(a1 , . . . , an ) = 1 and A(N) = {a1 , . . . , an }. Consequently, every numerical semigroup is atomic and has finitely many atoms. The family of numerical semigroups has been intensely studied for more than three decades. For an entry point to the realm of numerical semigroups, readers might consider [13] to be a valuable resource. Let N = ha1 , . . . , an i be a minimally generated numerical semigroup. The Frobenius number of N, denoted by F (N), is the greatest natural number not contained in N, i.e., the smallest integer F (N) such that for all b ∈ N with b > F (N) the Diophantine equation a1 x1 + · · · + an xn = b has a solution in Nn0 . We will need later the following result (taken from [4]), which gives an upper bound for the Frobenius number F (N) in terms of the minimal set of generators a1 , . . . , an . Theorem 2.1. Let N = ha1 , . . . , an i be a minimally generated numerical semigroup, where a1 < · · · < an . Then F (N) < (a1 − 1)(an − 1). 4 F. GOTTI Numerical semigroups not only have a canonical set of generators, but also exhibit a very convenient atomic structure. In the next section, we explore how these desirable properties behave in the more general setting of Puiseux monoids. 3. Atomic Characterization of Puiseux Monoids We start this section with a brief discussion on how Puiseux monoids show up naturally in commutative ring theory. We explain their connection to the field of Puiseux series, justifying our choice of the name Puiseux. Then we move to study the atomicity of the family of Puiseux monoids; we find an atomic characterization and two sufficient conditions for atomicity. Let F be a field. A valuation on F is a map val : F → R ∪ {∞} satisfying the following three axioms: (1) val(r) = ∞ if and only if r = 0; (2) val(rs) = val(r) + val(s) for all r, s ∈ F× ; (3) val(r + s) ≥ min{val(r), val(s)} for all r, s ∈ F× . Example 3.1. Take F to be the field of Laurent series C((T )) in the formal variable T . Consider the map valL : C((T )) → R ∪ {∞} defined by X  n valL = min{n ∈ Z≥N | cn 6= 0} cn T n≥N P n if n≥N cn T 6= 0, and val(0) = ∞. It is not difficult to verify that the function valL is a valuation on C((T )); this is a standard result that is explained in many introductory textbook in algebra. The algebraic closure of the field of Laurent series C((T )) in the formal variable T is denoted by C{{T }} and called the field of Puiseux series; it was first studied by Puiseux in [17]. We can write the field of Puiseux series as [ 1  C{{T }} = C Td , d∈N where C((T 1/d )) is the field of Laurent series in the formal variable T 1/d . The nonzero elements in C{{T }} are formal power series of the form c(T ) = c1 T n1 d + c2 T n2 d + ..., where c1 , c2 , . . . are complex numbers such that c1 6= 0, the denominator d is a natural number, and n1 < n2 < . . . are integers. Also, the function valP : C{{T }} → R∪{∞} mapping c(T ) to n1 /d and 0 to ∞ is a valuation on C{{T }}. Let R be a subring of C{{T }}. By the second axiom in the definition of valuation, the image of R under valP is closed under addition. Since valP (1) = 0, it follows ON THE ATOMIC STRUCTURE OF PUISEUX MONOIDS 5 that valP (R) is an additive submonoid of Q. In particular, Puiseux monoids arise naturally in commutative ring theory as images under valP of subdomains of C{{T }} of positive valuations. Given this connection, the monoids investigated in this paper are named Puiseux, honoring the French mathematician Victor A. Puiseux (18201883). We now proceed to study the atomic structure of Puiseux monoids. The family of Puiseux monoids is a natural generalization of that of numerical semigroups. The following proposition, whose proof follows immediately, characterizes those Puiseux monoids isomorphic to numerical semigroups. Proposition 3.2. A Puiseux monoid is isomorphic to a numerical semigroup if and only if it is finitely generated. Let M be a Puiseux monoid. If S = {s1 , s2 , . . . } is a set of rational numbers generating M, instead of M = hSi sometimes we write M = hs1 , s2 , . . . i, omitting the brackets in the description of S. We say that M is minimally generated by S if no proper subset of S generates M. In clear contrast to numerical semigroups, there are Puiseux monoids that are neither atomic nor finitely generated; see examples below. In addition, unlike numerical semigroups, not every Puiseux monoid is atomic. In fact, there are nontrivial Puiseux monoids containing no atoms at all. On the other hand, there are atomic and non-atomic Puiseux monoids having infinitely many atoms. The following examples illustrate the facts just mentioned. Example 3.3. Fix a prime number p. Let M be the Puiseux monoid generated by the set S = {1/pn | n ∈ N}. Although M is not finitely generated, its set of atoms is empty. This is because A(M) ⊆ S and 1/pn is the sum of p copies of 1/pn+1 for every positive integer n. Example 3.4. Let P be the set comprising all prime numbers, and consider the Puiseux monoid M = h1/p | p ∈ P i. We shall check that 1/p is an atom for every p ∈ P . For a prime p, suppose 1 1 1 +···+ , = p p1 pn where n is a natural number and the pk are not necessarily distinct primes. Setting m = p1 . . . pn and mk = m/pk for k = 1, . . . , n, we obtain m/p = m1 + · · · + mn ∈ N. Therefore p divides m, and so pk = p for some k. Thus, n = 1, which means that 1/p ∈ A(M). Since M is generated by atoms, it is atomic. Finally, it follows that M, albeit atomic, is not isomorphic to a numerical semigroup; to confirm this, note that M contains infinitely many atoms. Example 3.5. Let p1 , p2 , . . . be an enumeration of the odd prime numbers. Let M be the Puiseux monoid generated by the set S ∪ T , where S = {1/2n | n ∈ N} and 6 F. GOTTI T = {1/pn | n ∈ N}. It is easy to check, in the same way we did in Example 3.4, that 1/pn is an atom of M for each n ∈ N. On the other hand, it follows immediately that 1/2n ∈ / A(M) for any n ∈ N. Since T ⊆ A(M) ⊆ S ∪ T and S ∩ A(M) is empty, one has A(M) = T . We verify now that 1/2n cannot be written as a sum of atoms for any n ∈ N. Suppose, by way of contradiction, that for some n ∈ N there exist a positive integer k and non-negative coefficients c1 , . . . , ck satisfying k X 1 1 = ci . 2n pi i=1 (3.1) Multiplying (3.1) by m = p1 . . . pk , one gets k m X = ci mi ∈ N, 2n i=1 where mi = m/pi . This implies that 2n divides m. Since m is odd, we get a contradiction. Hence 1/2n cannot be written as a sum of atoms for any n ∈ N. Consequently, M is a non-atomic monoid with infinitely many atoms. Like numerical semigroups, Puiseux monoids are reduced. Therefore the set of atoms of a Puiseux monoid is contained in every set of generators. As we mentioned before, a numerical semigroup has a unique minimal set of generators, namely its set of atoms. Theorem 3.6 shows that having a (unique) minimal set of generators characterizes the family of atomic Puiseux monoids. Theorem 3.6. If M is a Puiseux monoid, the following conditions are equivalent: (1) M contains a minimal set of generators; (2) M contains a unique minimal set of generators; (3) M is atomic. Proof. First, we show that conditions (1) and (2) are equivalent. Since (1) follows immediately from (2), it suffices to prove (1) implies (2). Suppose S and S ′ are two minimal sets of generators of M. Take an arbitrary s ∈ S. The fact that M = hS ′i leads to the existence of n ∈ N and s′1 , . . . , s′n ∈ S ′ such that s = s′1 + · · · + s′n . Because S also generates M, for each i = 1, . . . , n, we have s′i = si1 + · · · + sini for some ni ∈ N and sij ∈ S for j ∈ {1, . . . , ni }. As a result, s= n X i=1 s′i = ni n X X sij . i=1 j=1 The minimality of S implies n = 1 and, therefore, one gets s = s′1 ∈ S ′ . Then S ⊆ S ′ and, using a similar argument, we can check that S ′ ⊆ S. Hence, if a minimal set of generators exists, then it must be unique. ON THE ATOMIC STRUCTURE OF PUISEUX MONOIDS 7 Now we prove that (1) and (3) are equivalent. First, assume condition (1) holds. Let S be a minimal set of generators of M. Let us show that every element in S is an atom. Suppose, by way of contradiction, that a ∈ S is not an atom. So a = x + y for some x, y ∈ M • . Since x and y are both strictly less than a, each of them can be written as a sum of elements in S\{a}. As a result, a = x+y ∈ hS\{a}i, contradicting the minimality of S. Therefore S ⊆ A(M). As A(M) must be contained in any set of generators, S = A(M). Thus, M is atomic, which is condition (3). Finally, we check (3) implies (1). Assume M is atomic, i.e., M = hA(M)i. As M is reduced, no atom can be written as a sum of positive elements of M. Hence A(M) is a minimal set of generators.  In contrast with numerical semigroups, there are Puiseux monoids containing no minimal sets of generators. When M is not atomic we still have A(M) ⊆ S for every minimal set of generators S. Nevertheless, A(M) might not generate M, as Example 3.5 shows. In fact, M can fail to be finitely generated and still have finitely many atoms; Example 3.3 sheds light upon this situation. Let p be a prime. For a nonzero integer a, define vp (a) to be the exponent of the maximal power of p dividing a, and set vp (0) = ∞. In addition, for b ∈ Z\{0}, set vp (a/b) = vp (a) − vp (b). It follows immediately that the map vp : Q → R ∪ {∞}, which is called the p-adic valuation, is an actual valuation on Q. In particular, it satisfies the third condition in the definition of valuation given before, that is (3.2) vp (r + s) ≥ min{vp (r), vp (s)} for all r, s ∈ Q× . Definition 3.7. Let P be a set of primes. A Puiseux monoid M over P is a Puiseux monoid such that vp (m) ≥ 0 for every m ∈ M and p ∈ / P. If P is finite, then we say that M is a finite Puiseux monoid over P . The Puiseux monoid M is said to be finite if there exists a finite set of primes P such that M is finite over P . To simplify the notation, if P contains only one prime p, we write Puiseux monoid over p instead of Puiseux monoid over {p}. We will see that the p-adic valuation maps play an important role in describing the atomic configuration of Puiseux monoids over P . For example, in Proposition 3.8, we check that a finite Puiseux monoid over P is atomic if for each p ∈ P the sequence of p-adic valuations of its generators is bounded from below. The remainder of this section is devoted to finding characterization criteria for atomicity of Puiseux monoids. Proposition 3.8 and Theorem 3.10 identify two atomic subfamilies of Puiseux monoids. Proposition 3.8. For a Puiseux monoid M the following conditions are equivalent: (1) M is finite and {vp (M)} is bounded from below for every prime p; (2) M is finite, and M = hRi implies that {vp (R)} is bounded from below for every prime p; 8 F. GOTTI (3) The denominator set d(M • ) is bounded; (4) If M = hRi, then the denominator set d(R• ) is bounded. If one (and so all) of the conditions above holds, then M is atomic. Proof. Condition (1) trivially implies condition (2). Assume condition (2), and let P be a finite set of primes over which M is finite. For all p ∈ P one has {vp (R)} is bounded from below; therefore d(R• ) is finite. Taking m to be the product of all elements in d(R• ), one has d | m for all d ∈ d(M • ). Hence d(M • ) is finite and (3) holds. Condition (3) implies condition (4) trivially. Finally, assume condition (4). Since d(R• ) is bounded, so is d(M • ). As a consequence, only finitely many primes divide elements in d(M • ). The boundedness of d(M • ) also implies that {vp (M)} is bounded from below for every prime p, which is condition (1). Now we will check that condition (2) implies that M is atomic. Suppose that M is finite over P = {p1 , . . . , pn } for some n ∈ N. Let R be a subset of rationals such that M = hRi. Set  mi = min 0, min• {vp i (r)} r∈R −m1 p1 . . . n . The function ϕ : M → mM for each i ∈ {1, . . . , n}, and take m = p−m n defined by ϕ(x) = mx is an isomorphism. An arbitrary x ∈ M • can be written as x = c1 r1 + · · · + ck rk , where k ∈ N while ci ∈ N and ri ∈ R• for every i ∈ {1, . . . , k}. By inequality (3.2), one has  X  k ci ri ≥ min {vp j (mci ri )} ≥ min {vp j (ci )} ≥ 0 vp j (mx) = vp j m i=1 1≤i≤k 1≤i≤k for j = 1, . . . , n. Since vp (mM • ) ⊆ N0 for each p ∈ P , it follows that mM is isomorphic to a numerical semigroup, and so it is atomic. Hence M is also atomic, as expected.  If one of the four conditions of Proposition 3.8 fails, namely that {vp (r) | r ∈ R} is not bounded from below for some p ∈ P , then M might not be atomic. This is illustrated in Example 3.3. Besides, if we allowed |P | = ∞, Proposition 3.8 would not hold, as we can see in the next example. Example 3.9. Let P = {p1 , p2 , . . . } be an infinite set of primes. Then we define the Puiseux monoid over P   1 1 M= , , . . . , where dn = p1 . . . pn d1 d2 for every n ∈ N. Observe that 1/dn ∈ / A(M) for any n ∈ N; this is because 1/dn is the sum of pn+1 copies of 1/dn+1. Since A(M) is contained in any set of generators, the fact that 1/dn ∈ / A(M) for every natural n implies that A(M) is empty. Therefore M is not atomic. In Section 4 we will give a special name to the Puiseux monoids whose set of atoms is empty. ON THE ATOMIC STRUCTURE OF PUISEUX MONOIDS 9 Theorem 3.10. Let M be a Puiseux monoid. If 0 is not a limit point of M, then M is atomic. Proof. Assume, by way of contradiction, that M is not atomic. Let N be the nonempty subset of M • comprising all the elements that cannot be written as a sum of atoms. For r1 ∈ N there exist positive elements r2 and r2′ in M such that r1 = r2 + r2′ . Either r2 or r2′ must be contained in N; otherwise r1 would not belong to N. Let us suppose then, without loss of generality, that r2 ∈ N. We have r1 > r2 and r1 − r2 = r2′ ∈ M. Because r2 ∈ N, there exist r3 , r3′ ∈ M • such that r2 = r3 + r3′ and either r3 ∈ N or r3′ ∈ N. Assume r3 ∈ N. Again, one obtains r2 > r3 and r2 − r3 = r3′ ∈ M. Continuing in this fashion, we can build two sequences {rn } and ′ {rn′ } of elements of M such that {rn } is decreasing and rn − rn+1 = rn+1 ∈ M for every n ∈ N. Since {rn } is a decreasing sequence of positive terms, it converges, and so it is a Cauchy sequence. This implies that the sequence {rn′ } converges to zero. But it contradicts the fact that 0 is not a limit point of M. Thus, M is atomic, which establishes the theorem.  The converse of Theorem 3.10 does not hold. The next example not only illustrates the failure of its converse, but also indicates that Proposition 3.8 and Theorem 3.10 are not enough to fully characterize the family of atomic Puiseux monoids. Example 3.11. Let p1 , p2 , . . . be an enumeration of the odd prime numbers. Define the sequence of positive integers {kn } as follows. Take k1 ∈ N to be arbitrary, and once kn has been chosen, take kn+1 ∈ N such that both inequalities kn+1 > kn and 2kn+1 p1 . . . pn > 2kn +1 p1 . . . pn+1 hold. Now define the Puiseux monoid p1 . . . pn M = hr1 , r2 , . . . i, where rn = . 2kn We verify that 0 is a limit point of M and that M is atomic. The way we defined the sequence {kn } ensures that rn+1 < rn /2 for every n ∈ N. As a result, the sequence {rn } converges to 0 and, hence, 0 is a limit point of M. Additionally, for j ∈ N, suppose m m X X p1 . . . pi (3.3) rj = ci ri = ci , ki 2 i=1 i=1 for some m ∈ N and coefficients c1 , . . . , cm ∈ N0 . As ri > rj when i < j, we get ci = 0 for i < j and cj ∈ {0, 1}. If cj = 0, then (3.3) can be written as (3.4) 2km −kj p1 . . . pj = m X 2km −ki ci p1 . . . pi . i=j+1 Every summand in the right-hand side of (3.4) is divisible by pj+1 , which contradicts that the left-hand side of (3.4) is not divisible by pj+1. Therefore cj = 1, and so rj is an atom. Since M is generated by atoms, it is atomic. 10 F. GOTTI Clearly, if M is a Puiseux monoid as in Theorem 3.10 (i.e., 0 is not a limit point of M), then every submonoid of M must also be atomic as a result of Theorem 3.10. However, in general, it is not true that every submonoid of an atomic monoid is atomic. The next example illustrates this observation. Example 3.12. Let {pn } be the sequence comprising the odd prime numbers in strictly increasing order. Then consider the Puiseux monoid M = hSi, where   1 n∈N . S= 2 n pn Since each odd prime divides exactly one element of the set d(S), it follows that A(M) = S. Hence M is atomic. On the other hand, the element 1/2n is the sum of pn copies of the atom 1/(2n pn ) for every n ∈ N. Thus, the monoid N = h1/2n | n ∈ Ni contains no atoms, which immediately implies that N is not atomic. Therefore N is a submonoid of the atomic monoid M that fails to be atomic. Proposition 3.8 and Theorem 3.10 give rise to large families of atomic monoids. Theorem 3.10 applies, in particular, when a Puiseux monoid is generated by an eventually increasing sequence. Many factorization invariants of numerical semigroups have been investigated during the last two decades. For example, the set of lengths, elasticity, delta set, and catenary/tame degree have been actively studied in terms of minimal sets of generators (see [1, 2, 6, 8, 16] and references therein). Studying these factorization invariants on the atomic Puiseux monoids provided by Theorem 3.10 would contribute significantly to understanding their algebraic and combinatorial structure. 4. Strongly Bounded Puiseux Monoids In this section, we restrict attention to the atomic structure of those Puiseux monoids that can be generated by a subset S ⊂ Q satisfying that n(S) is bounded. To be more precise, we say that a subset S of rational numbers is strongly bounded if its numerator set, n(S), is bounded. Definition 4.1. A Puiseux monoid is bounded (resp., strongly bounded ) if it can be generated by a bounded (resp., strongly bounded) set of rational numbers. Although, for the remainder of this paper, we focus on studying the subfamily of strongly bounded Puiseux monoids, they are by no means the only subfamily containing atomic representatives. The following proposition explains a way of constructing a family of atomic Puiseux monoids whose members are not strongly bounded. ON THE ATOMIC STRUCTURE OF PUISEUX MONOIDS 11 Proposition 4.2. There exist infinitely many atomic Puiseux monoids that are not strongly bounded. Proof. Let p be a prime. Suppose {an } is the sequence recurrently defined as follows. Choose a1 ∈ N such that a1 > p. Suppose a1 , . . . , an have been selected. Take an+1 ∈ N so that gcd(an+1 , p) = 1 and an+1 /pn+1 > an /pn . Consider the Puiseux monoid over p   an n∈N . M = hSi, where S = pn Let us check that an /pn ∈ A(M) for every n ∈ N. Since a1 /p is the smallest element in M • , it is an atom. By inequality (3.2), for n > 1, the p-adic valuation of every element of the monoid   a1 an−1 Mn−1 = , . . . , n−1 p p n is at least −n + 1. As a consequence, an /p ∈ / Mn−1 . From the fact that an /pn is the smallest element in M \Mn−1 , we deduce that it is an atom. Hence S ⊆ A(M) and, therefore, A(M) = S. Because every generating set of M contains S, and an > pn for n ∈ N, it follows that M is not strongly bounded. Yet, the monoid M is atomic because it is generated by atoms. Note also that A(M) = ∞. For each prime p, we have found a Puiseux monoid over p that is atomic but not strongly bounded; thus, there are infinitely many atomic Puiseux monoids failing to be strongly bounded. In fact, from the way we constructed the Puiseux monoid M over p, we can infer that for each prime p, there are infinitely many atomic Puiseux monoids over p that are not strongly bounded.  We have just found a subfamily of atomic Puiseux monoids that are not strongly bounded. By contrast, it is natural to ask whether the family of strongly bounded Puiseux monoids comprises all Puiseux monoids containing no atoms. We postpone the answer to this question until we prove Proposition 4.4. Let us introduce some terminology for those monoids containing no atoms. An integral domain is called antimatter domain if it contains no irreducible elements. Antimatter domains have been in-depth studied by Coykendall et al. [10]. However, no relevant investigation has been carried out concerning monoids containing no atoms. Definition 4.3. Let M be a monoid. If A(M) is empty, we say that M is an antimatter monoid. We should point out that, in general, the concepts of antimatter and atomic monoids are independent. Abelian groups are atomic and antimatter. The additive monoid N0 is atomic, but it is not antimatter. Also, the additive monoid Q≥0 is antimatter; however, it is not atomic. Finally, the set of polynomials M = { p(x) ∈ Q[x] | p(0) ∈ Z } 12 F. GOTTI endowed with the standard multiplication of polynomials is not atomic; this is proved in [5]. In addition, since every prime, seen as a constant polynomial, is an atom of M, one finds that M is not antimatter. Having indicated the independence of antimatter and atomic monoids in a general setting, we should notice that, in the particular case of Puiseux monoids, a nontrivial atomic monoid automatically fails to be antimatter; more generally, this is actually true for every nontrivial reduced monoid. At this point we know there are infinitely many atomic Puiseux monoids failing to be strongly bounded. For the sake of completeness, we will also construct in Proposition 4.6 an infinite subfamily of antimatter Puiseux monoids whose members fail to be strongly bounded. We know that every generating set of a Puiseux monoid M contains A(M). In particular, if M is atomic, then every generating set of M contains a generating subset consisting of atoms, namely A(M). This suggests the question of whether every generating set of a bounded (resp., strongly bounded) Puiseux monoid contains a bounded (resp., strongly bounded) generating subset. As we show now, we can reduce any generating set of a bounded Puiseux monoid to a bounded generating subset. Proposition 4.4. If M is a bounded Puiseux monoid, then every generating set of M contains a bounded generating subset. Proof. Let R be a set of generators of M. Take B to be a bounded subset of rational numbers such that M = hBi. For each b ∈ B define [ S= Sb , where Sb = {r ∈ R | r divides b in M}. b∈B Since b is an upper bound of Sb for each b, the fact that B is bounded implies that S is also bounded. So S is a bounded subset of R. We verify now that S is a generating set of M. It is enough to check that M ⊆ hSi. Take an arbitrary r ∈ R. Since B generates M, there exist k ∈ N and b1 , . . . , bk ∈ B such that r = b1 +· · ·+bk . Because M is generated by R, for each i ∈ {1, . . . , k} there exist ni ∈ N and ri1 , . . . , rini ∈ R such that bi = ri1 + · · · + rini . Consequently, we have (4.1) r= k X i=1 bi = ni k X X rij . i=1 j=1 Notice that for every i = {1, . . . , k} and j ∈ {1, . . . , ni }, the element rij divides bi in M. Thus, equality (4.1) forces r ∈ hSi and, therefore, M = hRi ⊆ hSi. Hence S is a bounded subset of R generating M.  Proposition 4.4 naturally suggests the question of whether every generating set of a strongly bounded Puiseux monoid contains a strongly bounded generating subset. Unlike its parallel statement for boundedness, this desirable claim does not hold for strongly bounded Puiseux monoids. ON THE ATOMIC STRUCTURE OF PUISEUX MONOIDS 13 Example 4.5. Let p be an odd prime, and let us consider the following two sets of rational numbers:    2n  n 2 p − 1 p2 + 1 S= n ∈ N and T = , 2n+1 n∈N . p2n p2n+1 p To verify that S and T generate the same Puiseux monoid, it suffices to notice that n n n n p2 ± 1 p2 ± 1 2 2 p2 − 1 p2 + 1 + and = . = p2n 2 p2n+1 p2n+1 p2n+1 p2n+1 Let M be the Puiseux monoid generated by any of the sets S or T . Since S is strongly bounded then so is M. On the other hand, every strongly bounded subset of T must contain only finitely many elements; this is because the sequences of numerators of n n+1 n n+1 {(p2 − 1)/p2 } and {(p2 + 1)/p2 } both increase to infinite. In addition, as M is antimatter, and so non-finitely generated, any subset of T generating M must contain infinitely many elements. Hence we can conclude that T does not contain any strongly bounded subset generating M. Let us resume now our search for a family of antimatter Puiseux monoids failing to be strongly bounded. To accomplish this goal, we make use of Proposition 4.4. Proposition 4.6. There exist infinitely many antimatter Puiseux monoids that are not strongly bounded. Proof. Since every strongly bounded Puiseux monoid is also bounded, it is enough to find a family failing to be bounded. Let p be an odd prime, and consider the sets  2    p +1 1 1 Sp = n∈N . + n n ∈ N and T = p 2 2n Let P be the collection of all infinite subsets of odd prime numbers. Let P ∈ P and take MP to be the Puiseux monoid generated by the set [  X =T ∪ Sp . p∈P We claim that MP is antimatter but not bounded. Let us verify first that MP is antimatter. Notice that for every p ∈ P and n ∈ N,  2  1 p +1 1 1 p2 + 1 + n = + n+1 + n+1 , p 2 p 2 2 which means that Sp ⊆ X + X. Additionally, 1/2n = 2(1/2n+1) and so T ⊆ X + X. Therefore X ⊆ X + X, which implies that X does not contain any atoms of MP . Since MP = hXi, it follows immediately that MP is antimatter. Now we show that MP is not bounded. Suppose, by way of contradiction, that MP is bounded. By Proposition 4.4, the set X must contain a bounded subset Y generating MP . Observe that, for every prime p ∈ P , the set Sp is bounded from 14 F. GOTTI below by p. Therefore there exists a natural N such that Y ∩ Sp is empty for all prime p > N. If q ∈ P is a prime greater than N, then q does not divide 2q 2 + q + 2 and so q2 + 1 1 2q 2 + q + 2 + = ∈ Sq \ hY i, q 2 2q which contradicts the fact that Y generates MP . Hence, for each P ∈ P, the Puiseux monoid MP is not bounded. Since MP 6= MP ′ when P and P ′ are distinct elements of P, one gets that {MP | P ∈ P} is an infinite family of antimatter Puiseux monoids that are not bounded.  As shown in the above example, not every generating set of a strongly bounded Puiseux monoid M can be reduced to a strongly bounded subset generating M. However, if M is not only strongly bounded but also atomic, then any set of generators of M can certainly be reduced to a strongly bounded generating set. We record this observation in Proposition 4.7, whose proof follows straightforwardly from the fact that the set of atoms of a reduced monoid must be contained in every generating set. Proposition 4.7. Let M be a strongly bounded Puiseux monoid. If M is atomic, then every generating set of M contains a strongly bounded generating subset. 5. Atomic Structure of Strongly Bounded Puiseux Monoids In this section, we restrict attention to the atomic structure of the subfamily of Puiseux monoids that happen to be strongly bounded. First, we find a condition under which members of this subfamily are antimatter; presenting Theorem 5.2 as the first main result. Then we move our focus to the classification of the atomic subfamily of strongly bounded Puiseux monoids over a finite set of primes P , which is stated in our second main result, Theorem 5.8. Let us introduce some definitions. We say that a sequence {an } of integers stabilizes at a positive integer d if there exists N ∈ N such that d divides an for every n ≥ N. The spectrum of a sequence {an }, denoted by Spec({an }), is the set of primes p for which {an } stabilizes at p. Lemma 5.1. Let {an } be a sequence of positive integers having an upper bound B. If the spectrum of {an } is empty, then for each N ∈ N there exist k ≤ B + 1 and n1 , . . . , nk ∈ N such that N < n1 < · · · < nk and gcd(an1 , . . . , ank ) = 1. Proof. Fix N ∈ N. Since the sequence {an } is bounded, there are only finitely many primes dividing at least one of the terms of {an }. Let P be the set comprising such primes. If P is empty, then an = 1 for all n ∈ N, and we can take k = 2 and both n1 and n2 to be two distinct integers greater than N such that N < n1 < n2 . In such a case, k = 2 ≤ B + 1 for every upper bound B of {an }, and one has N < n1 < n2 and gcd(an1 , an2 ) = gcd(1, 1) = 1. Assume, therefore, that {an } is not the constant ON THE ATOMIC STRUCTURE OF PUISEUX MONOIDS 15 sequence whose terms are all ones. Thus, P is not empty; let P = {p1 , . . . , pk }. The fact that 2 ≤ pn ≤ B when n ∈ {1, . . . , k} implies k ≤ B + 1. As the spectrum of {an } is empty, there exists n1 > N such that p1 does not divide an1 . Similarly, there exists n2 > n1 for which p2 does not divide an2 . In general, if for i < k one has chosen n1 , . . . , ni ∈ N so that N < n1 < · · · < ni and pi ∤ ani , then there exists ni+1 ∈ N satisfying ni+1 > ni and pi+1 ∤ ani+1 . After following the described procedure finitely many times, we will obtain n1 , . . . , nk ∈ N such that N < n1 < · · · < nk and pi ∤ ani . Now it follows immediately that gcd(an1 , . . . , ank ) = 1.  The next theorem gives a sufficient condition for a Puiseux monoid to be antimatter. Theorem 5.2. Let {rn | n ∈ N} be a strongly bounded subset of rationals generating M. If d(rn ) divides d(rn+1 ), the sequence {d(rn )} is unbounded, and the spectrum of {n(rn )} is empty, then M is antimatter. Proof. For every n ∈ N, let us denote n(rn ) and d(rn ) by an and bn , respectively. Let B be an upper bound for the sequence {an }. Fix an arbitrary positive integer N. We will show that b−1 N is contained in M. By Lemma 5.1, there exist k ∈ N and n1 , . . . , nk ∈ N such that n1 < · · · < nk and gcd(an1 , . . . , ank ) = 1, where n1 is large enough to satisfy B 2 < bn1 b−1 N (we are using here the unboundedness of {bn }). On the other hand, gcd(ank , bnk b−1 ni ) = 1; this makes sense because bni divides bnk for i = 1, . . . , k. Since gcd(an1 , . . . , ank ) = 1 and gcd(ank , bnk b−1 ni ) = 1 for every −1 −1 i ≤ k, one has gcd(bnk bn1 an1 , . . . , bnk bnk−1 ank−1 , ank ) = 1. Let F (S) be the Frobenius −1 number of the numerical semigroup S = hbnk b−1 n1 an1 , . . . , bnk bnk−1 ank−1 , ank i. Taking −1 j ∈ {1, . . . , k} so that bnk b−1 nj anj = max{bnk bni ani | 1 ≤ i ≤ k} and using Theorem 2.1, one can see that −1 2 −1 −1 F (S) < (ank − 1)(bnk b−1 nj anj − 1) < ank anj bnk bnj ≤ B bnk bn1 < bnk bN , −1 where the last inequality follows from B 2 < bn1 b−1 N . As bnk bN > F (S), there exist c1 , . . . , ck ∈ N0 such that k X −1 bnk bN = cj bnk b−1 nj anj j=1 and, accordingly, k X anj 1 ∈ M. = cj bN b n j j=1 Therefore 1/bn ∈ M for every n ∈ N. Since 1/bn = (bn+1 /bn )1/bn+1 for every n ∈ N, none of the elements 1/bn is an atom of M. Moreover, each generator an /bn can be written as the sum of an copies of 1/bn ; hence an /bn ∈ / A(M) for all n ∈ N. Having checked that none of the generators of M is an atom, we can conclude that M is antimatter.  16 F. GOTTI There is some additional information about M in the proof of Theorem 5.2. We list it in the following corollary. Corollary 5.3. If M is a Puiseux monoid satisfying the conditions in Theorem 5.2, then   m m ∈ N0 and n ∈ N . M= bn Let us verify that Theorem 5.2 is sharp, meaning that none of its hypotheses are redundant. First, if one drops the condition of M being strongly bounded, then it might not be antimatter; see Proposition 4.2. Further, Example 3.4 illustrates that the condition d(rn ) | d(rn+1 ) for every n ∈ N is also required. Numerical semigroups are evidence that the sequence {d(rn )} has to be necessarily unbounded. Finally, the family of strongly bounded Puiseux monoids constructed in Proposition 5.4 shows that the emptiness of the spectrum of {n(rn )} also needs to be imposed to guarantee M is antimatter. Proposition 5.4. For each m ∈ N there exists a non-finitely generated strongly bounded Puiseux monoid having exactly m atoms. Proof. Take p and q to be prime numbers satisfying p 6= q and q > m. Consider the Puiseux monoid over p   q q M = m, . . . , 2m − 1, m+1 , m+2 , . . . . p p We check that A(M) = {m, . . . , 2m − 1}. Suppose a ∈ Z such that m ≤ a ≤ 2m − 1. Since a ∈ M, it can be written as X q a = a′ + cn m+n , p n≥1 for a′ ∈ {0} ∪ {m, . . . , 2m − 1} and for a suitable set of non-negative coefficients cn , all but finitely many of them being zero. Then (a − a′ )pm X cn = . (5.1) n q p n≥1 Because the q-adic valuations of the right-hand side of (5.1) are at least zero, the lefthand side of this equation must be an integer. As a result, a − a′ = 0 and so cn = 0 for every n ∈ N. This implies that a ∈ A(M). Thus, {m, . . . , 2m − 1} ⊆ A(M). On the other hand, no generator of the form q/pn for n > m can be an atom of M, for q/pn is the sum of p copies of q/pn+1 . Hence A(M) = {m, . . . , 2m − 1}.  Proposition 5.4 tells us that there are infinitely many Puiseux monoids (by varying our choice of p) with any fixed finite number of atoms that are not finitely generated and, therefore, non-atomic. This fact, along with Example 3.3 and Example 3.5, ON THE ATOMIC STRUCTURE OF PUISEUX MONOIDS 17 gives evidence of the complexity of the atomic structure of Puiseux monoids. We conclude our discussion about the atomic structure of Puiseux monoids with a classification of the atomic subfamily of strongly bounded Puiseux monoids over a finite set of primes. First, let us introduce some terminology. The spectrum of a natural n, which is denoted by Spec(n), is the set of all prime divisors of n. In addition, given a finite set of primes P = {p1 , . . . , pk } with p1 < · · · < pk , the support of n ∈ N with respect to P , denoted by SuppP (n), is the set of indices i such that pi | n. The next lemmas will be used in the proof of Theorem 5.8. Lemma 5.5. Let P be a finite set of primes, and let M be a Puiseux monoid over P . If there is a sequence {bn } ⊆ N such that 1/bn ∈ M for every n ∈ N and {vp (bn )} is strictly increasing for each p ∈ P , then M is antimatter. Proof. Fix a sequence {bn } of positive integers such that 1/bn ∈ M and {vp (bn )} is strictly increasing for each p ∈ P . Take an element q in M • , and let a, b ∈ N such that q = a/b and gcd(a, b) = 1. Set mq = min{vp (q)}. p∈P Since {vp (bn )} is strictly increasing for each p ∈ P and the set P is finite, there exists N ∈ N such that −vp (bn ) < mq for every p ∈ P and n ∈ Z>N . Therefore we obtain   Y   Y  Y 1 vp (q)−vp (1/bn ) vp (q)−vp (1/bn ) vp (1/bn ) = d p (5.2) q= d p p , bn p∈P p∈P p∈P where d is the greatest natural number dividing a such that p ∤ d for every p ∈ P . Because vp (1/bn ) = −vp (bn ) < mq ≤ vp (q) for each p ∈ P , it follows that the exponents vp (q) − vp (1/bn ) in (5.2) are all positive. Thus, q is the sum of more than one copy of 1/bn , whence we find that q is not an atom. Since q was taken arbitrarily in M • , we conclude that A(M) is empty.  Lemma 5.6. Let k ∈ N and P = {p1 , . . . , pk } be a finite set of primes. Let {sn } be a sequence so that Spec(sn ) ⊆ P for every n ∈ N. Then there exists N ∈ N satisfying the following property: if there exist n ∈ N and I ⊆ {1, . . . , k} such that vp i (sn ) > N for each i ∈ I, then there is a subsequence {s′n } of {sn } for which {vp i (s′n )} is strictly increasing for each i ∈ I. Proof. Let J be the set of all subsets of indices J of {1, . . . , k} for which {sn } does not contain any subsequence {s′n } such that {vp j (s′n )} is strictly increasing for each j ∈ J. For each J ∈ J there must exist NJ ∈ N satisfying that, for every n ∈ N, the inequality vp j (sn ) ≤ NJ holds for at least an index j ∈ J. Take N ∈ N to be max{NJ | J ⊆ {1, . . . , k}}. Suppose now that n is a natural number and I is a subset of {1, . . . , k} such that vp i (sn ) > N for each i ∈ I. If I ∈ / J , then we are done. Suppose, by way of contradiction, that I ∈ J . Then N ≥ NI and so vp i (sn ) > NI 18 F. GOTTI for every i ∈ I. This means that the inequality vp i (sn ) ≤ NI does not hold for any i ∈ I, contradicting the fact that I ∈ J .  Lemma 5.7. Let M be a Puiseux monoid generated by a set S. Suppose also that S = S1 ∪ · · · ∪ Sn , where n ∈ N and Si is a nonempty subset of S for i = 1, . . . , n. Then the next set inclusion holds: n [ (5.3) A(M) ⊆ A(hSi i). i=1 Proof. If A(M) is empty, then (5.3) follows trivially. So assume A(M) is not empty, and take a to be an atom of M. Then a ∈ S, and therefore there exists i ∈ {1, . . . , n} such that a ∈ Si ⊆ hSi i. Because A(M) ∩ hSi i ⊆ A(hSi i), one gets n [ a ∈ A(hSi i) ⊆ A(hSi i). i=1 Since a was taken arbitrarily in A(M), the inclusion (5.3) holds, as desired.  We are now in a position to prove our last main result. Theorem 5.8. Let M be a strongly bounded finite Puiseux monoid. Then M is atomic if and only if M is isomorphic to a numerical semigroup. Proof. Let P be a set of primes such that M is finite over P . First, we will prove that M has only finitely many atoms. We proceed by induction on the cardinality of P . If |P | = 0, then M is isomorphic to a numerical semigroup and, therefore, A(M) is finite. Suppose k is a positive integer such that the statement of the theorem is true when |P | < k. We shall prove that every strongly bounded Puiseux monoid M over P has finitely many atoms when |P | = k. Set P = {p1 , . . . , pk }. In addition, let {an } and {bn } be two sequences of natural numbers such that {an } is bounded and gcd(an , bn ) = 1 for every n ∈ N. Let   an M = hSi, where S = n∈N . bn To show that M has only finitely many atoms, we will distribute the generators an /bn of M into finitely many submonoids of M and then we will apply Lemma 5.7. Since the sequence {an } is bounded, it has a maximum, namely m. Set   an Sj,I = n ∈ N, an = j, and SuppP (bn ) = I bn for each j ∈ {1, . . . , m} and I ⊆ {1, . . . , k}. Let J be the set of pairs (j, I) such that Sj,I is not empty. Using Lemma 5.7, we obtain [ (5.4) A(M) ⊆ S ⊆ A(hSj,I i). (j,I)∈J ON THE ATOMIC STRUCTURE OF PUISEUX MONOIDS 19 For each (j, I) ∈ J , let Mj,I be the Puiseux monoid over P generated by Sj,I . By the inclusion (5.4), we are done once we show that Mj,I contains only finitely many atoms for every pair (j, I) ∈ J . We fix an arbitrary pair (j, I) ∈ J and prove that A(Mj,I ) is finite. Since Mj,I is a strongly bounded finite Puiseux monoid over P , if I is strictly contained in {1, . . . , k}, then A(Mj,I ) is finite (induction hypothesis). So it just remains to check that Mj,I contains only finitely many atoms when I = {1, . . . , k}. If Sj,I is finite, then by Proposition 3.2 the monoid Mj,I is isomorphic to a numerical semigroup; in this case, A(Mj,I ) is finite. So we will assume Sj,I is not finite. Let {sn } be a subsequence of the sequence {bn } such that Sj,I = {j/sn | n ∈ N}. If {sn } contains a subsequence {s′n } such that vp i (s′n ) is strictly increasing for each i ∈ {1, . . . , k}, then Lemma 5.5 implies that Mj,I is antimatter and, therefore, contains no atoms (notice that Mj,I is isomorphic to h1/sn | n ∈ Ni via x 7→ j −1 x). So we assume such a subsequence of {sn } does not exist. Using Lemma 5.6, we can find N ∈ N so that if for I ′ ⊆ {1, . . . , k} there is n ∈ N satisfying vp i (sn ) > N for each i ∈ I ′ , then there exists a subsequence {s′n } of {sn } such that {vp i (s′n )} is strictly increasing for each N i ∈ I ′ (note that now I ′ must be a proper subset of {1, . . . , k}). Set µ = pN 1 . . . pk ′ ′ and M = µMj,I . Thus, one has that Mj,I is isomorphic to M via multiplication by µ, meaning x 7→ µx. So it suffices to check that M ′ has finitely many atoms. To prove that M ′ has only finitely many atoms, consider its proper generating set S ′ = {a′n /b′n | n ∈ N}, where a′n /b′n results from reducing the fraction µj/sn to lowest terms. For every n ∈ N and I ⊆ {1, . . . , k}, it follows that pi divides b′n for each i ∈ I if and only if vp i (sn ) > N for each i ∈ I, which implies the existence of a subsequence {s′n } of {sn } such that {vp i (s′n )} is strictly increasing for each i ∈ I (by Lemma 5.6). Since {sn } contains no subsequence {s′n } such that {vp i (s′n )} is strictly increasing for each i ∈ {1, . . . , k}, for every n ∈ N there is at least an index i ∈ {1, . . . , k} such that vp i (sn ) ≤ N, i.e., SuppP (b′n ) is a proper subset of {1, . . . , k} for every n ∈ N. Set  ′  an ′ SI = n ∈ N and SuppP (bn ) = I b′n for each I ⊂ {1, . . . , k}. Let I be the collection of all subsets I of {1, . . . , k} such that SI is not empty. Notice that every element of I is a proper subset of {1, . . . , k}. By Lemma 5.7, [ (5.5) A(M ′ ) ⊆ S ′ ⊆ A(hSI i). I∈I Since each set of indices I ∈ I is strictly contained in P , by the induction hypothesis we get that hSI i contains only finitely many atoms for every I ∈ I. Therefore |A(M ′ )| < ∞ follows from (5.5). At this point, we have proved that every strongly bounded Puiseux monoid over a finite set of primes P has finitely many atoms. Suppose M is atomic. By Theorem 20 F. GOTTI 3.6, M has a minimal set of generators, which must be A(M). Hence M is finitely generated and, by Proposition 3.2, it is isomorphic to a numerical semigroup. Conversely, suppose M is isomorphic to a numerical semigroup. Since every numerical semigroup is atomic, M must be atomic. This completes the proof.  Example 3.4 can be used as evidence that Theorem 5.8 does not hold if we do not require P to be finite. Besides, the strongly boundedness of the Puiseux monoid M over P is not superfluous, as one can see in Proposition 4.2, which guarantees the existence of an atomic Puiseux monoid over a prime p with infinitely many atoms that fails to be strongly bounded. 6. Acknowledgments While working on this paper, the author was supported by the UC Berkeley Chancellor Fellowship. The author is grateful to Scott Chapman for valuable feedback and encouragement, to Marly Gotti for her dedicated final review, and to Chris O’Neill for helpful discussions on early drafts of this paper. Also, the author thanks an anonymous referee for suggesting recommendations that improved this paper. References [1] J. Amos, S. T. Chapman, N. Hine, and J. Paixao: Sets of lengths do not characterize numerical monoids, Integers 7 (2007) Paper A50, 8pp. [2] T. Barron, C. O’Neill, and R. Pelayo: On the set of elasticities in numerical monoids, to appear, Semigroup Forum. [arXiv:1409.3425] [3] V. Barucci, D. E. Dobbs, and M. Fontana: Maximality Properties in Numerical Semigroups and Applications to One-Dimensional Analytically Irreducible Local Domains, vol. 125, Memoirs of the Amer. Math. Soc., 1997. [4] A. Brauer: On a problem of partitions, Amer. J. Math. 64 (1942) 299–312. [5] S. T. Chapman: A tale of two monoids: A friendly introduction to nonunique factorizations, Math. Mag. 87 (2014) 163–173. [6] S. T. Chapman, M. Corrales, A. Miller, C. Miller, and D. Patel: The catenary and tame degrees on a numerical monoid are eventually periodic, J. Aust. Math. Soc. 97 (2014) 289–300. [7] S. T. Chapman , F. Gotti, and R. Pelayo: On delta sets and their realizable subsets in Krull monoids with cyclic class groups, Colloq. Math. 137 (2014) 137–146. [8] S. T. Chapman, R. Hoyer, and N. Kaplan: Delta sets of numerical monoids are eventually periodic, Aequationes Math. 77 (2009) 273–279. [9] S. T. Chapman and D. Steinberg: On the elasticity of generalized arithmetical congruence monoids, Results Math. 58 (2010) 221–231. [10] J. Coykendall, D. E. Dobbs, and B. Mullins: On integral domains with no atoms, Comm. Algebra 27 (1999) 5813–5831. [11] D. Eisenbud: Commutative Algebra With a View Toward Algebraic Geometry, Grad. Texts in Math., vol. 150, Springer-Verlag, New York, 1995. [12] P. A. Garcı́a-Sánchez and J. C. Rosales: Finitely Generated Commutative Monoids, Nova Science Publishers Inc., New York, 1999. ON THE ATOMIC STRUCTURE OF PUISEUX MONOIDS 21 [13] P. A. Garcı́a-Sánchez and J. C. Rosales: Numerical Semigroups, Developments in Mathematics, vol. 20, Springer-Verlag, New York, 2009. [14] A. Geroldinger and F. Halter-Koch: Non-Unique Factorizations: Algebraic, Combinatorial and Analytic Theory, Pure and Applied Mathematics, vol. 278, Chapman & Hall/CRC, Boca Raton, 2006. [15] P. A. Grillet: Commutative Semigroups, Advances in Mathematics, vol. 2, Kluwer Academic Publishers, Boston, 2001. [16] M. Omidali: The catenary and tame degree of numerical monoids generated by generalized arithmetic sequences, Forum Math. 24 (2012) 627–640. [17] V. A. Puiseux: Recherches sur les fonctions algébriques, Jour. de Math. 15 (1850) 365–480. [18] W. Schmid: Differences in sets of lengths of Krull monoids with finite class group, J. Théor. Nombres Bordeaux 17 (2005) 323–345. Mathematics Department, UC Berkeley, Berkeley, CA 94720 E-mail address: [email protected]
0math.AC
String Cadences Amihood Amir1,2,⋆ , Alberto Apostolico3 , Travis Gagie4 , and Gad M. Landau5,6,⋆ arXiv:1610.03337v1 [cs.DS] 11 Oct 2016 1 Department of Computer Science, Bar-Ilan University, Ramat-Gan 52900, Israel. E-mail: [email protected] 2 Department of Computer Science, Johns Hopkins University, Baltimore, MD 21218. 3 School of Computational Science and Engineering, College of Computing, Georgia Institute of Technology, Klaus Advanced Computing Building, 266 Ferst Drive, Atlanta, GA 30332-0765. Email: [email protected] 4 School of Computer Science and Telecommunications, Diego Portales University, Santiago, Chile. E-mail: [email protected] 5 Department of Computer Science, University of Haifa, Mount Carmel, Haifa 31905, Israel. Email: [email protected] 6 Department of Computer Science and Engineering, NYU Polytechnic School of Engineering, 2 MetroTech Center, Brooklyn, NY 11201. Abstract. We say a string has a cadence if a certain character is repeated at regular intervals, possibly with intervening occurrences of that character. We call the cadence anchored if the first interval must be the same length as the others. We give a sub-quadratic algorithm for determining whether a string has any cadence consisting of at least three occurrences of a character, and a nearly linear algorithm for finding all anchored cadences. 1 Introduction Finding interesting patterns in strings is an important problem in many fields, with an extensive literature (see, e.g., [1] and references therein). One might therefore expect all obvious kinds to have already been considered; however, as far as we are aware, no one has previously investigated how best to determine whether a string contains a certain character repeated at regular intervals, possibly with intervening occurrences of that character. To initiate the study of this natural problem, we introduce the following notions: Definition 1. A cadence of a string S[1..n] is a pair (i, d) of natural numbers such that i ≤ d ≤ n and S[j] = S[i] for every j between 1 and n with j ≡ i mod d. A cadence (i, d) with ⌊(n − i)/d⌋ + 1 = k is a k-cadence. Definition 2. An anchored cadence of a string S[1..n] is a natural number i such that (i, i) is a cadence of S: i.e., such that S[j] = S[i] for every j between 1 and n with j ≡ 0 mod i. ⋆ Partly supported by ISF grant 571/14. Informally, if the first interval must be the same length as the others then the cadence is anchored. For example, if S = ALABARALAALABARDA then (3, 7) is a 3-cadence, because S[3] = S[10] = S[17] = A, and 7 is an anchored cadence, because S[7] = S[14] = A. If d ≥ max(i, n − i + 1) then (i, d) is trivially a 1-cadence; otherwise, we can check that (i, d) is a cadence by comparing S[i], S[2i], . . . , S[i + ⌊(n − i)/d⌋d]. Therefore, we can find all cadences of S in time at most proportional to n X n X n/d = n i=1 d=i n X d X 1/d = n2 . d=1 i=1 Since there can still be Θ(n2 ) k-cadences for k ≥ 2, this bound is worst-case optimal. In Section 2, however, we give a sub-quadratic algorithm for determining whether a string has a 3-cadence. In Section 3 we give an O(n log log n) time algorithm for finding all anchored cadences, of which there can be at most n. We leave as open problems finding an output-sensitive algorithm for reporting all the k-cadences for k ≥ 2, or a subquadratic algorithm for determining whether there is a k-cadence for a given k ≥ 4, or a linear algorithm for finding all anchored cadences. We are also curious about how to define properly and find efficiently approximate cadences, and whether there exists, e.g., a subquadraticspace data structure that, given the endpoints of a substring, quickly reports all the cadences of that substring. 2 Detecting 3-Cadences A string has a 2-cadence if, for some character a, the positions i and j of the leftmost and rightmost occurrences of a satisfy 2i ≤ j and 2j − i > n. We can easily check this in O(n log n) time. In this section we show that when the string is binary, we can also check if it has a 3-cadence in O(n log n) time. It follows that we still need only O(n logn) time when the string is over any constant-size alphabet, and O n3/2 log1/2 n time in general. Specifically, we show how to convert a string S[1..n] ∈ {0, 1}n into a set W of n integer weights in [−2n, . . . , 2n] such that S has a 3-cadence (i, d) with S[i] = 1 if and only if W is in 3Sum (i.e., three of its weights sum to 0), which we can check in O(n log n) time via the Fast-Fourier Transform (see, e.g., [3, Exercise 30.1–7]). Since our reduction is essentially reversible, we suspect that improving this O(n log n) bound will be challenging. Without loss of generality, assume we are interested only in detecting 3cadences (i, d) with S[i] = 1; we can detect 3-cadences (i, d) with i = 0 symmetrically. Let L1 = {j : S[j] = 1, j ≤ n/3} L2 = {j : S[j] = 1, n/3 < j ≤ 2n/3} L3 = {j : S[j] = 1, 2n/3 < j} . By definition, for any 3-cadence (i, d) we have i ∈ L1 , i + d ∈ L2 and i + 2d ∈ L3 . Therefore, there is a 3-cadence if and only if the average of some element in L1 and some element in L3 is an element in L2 . It follows that S has a 3-cadence if and only if L1 ∪ L3 ∪ {−2j : j ∈ L2 } ∈ 3Sum . If S is over any constant-size alphabet then we can create a binary string of length n for each character in the alphabet, with the 1s marking the occurrences of that character, and apply our reduction to each one in O(n log n) total time. If S is over an arbitrary alphabet then we can perform this partitioning and detect 3-cadences in the binary string for each character a using time at most proportional to min n2a , n log n , where na is the number of occurrences of a in  S. This takes a total of at most O n3/2 log1/2 n time. Theorem 1. We   can determine whether a string of length n has a 3-cadence in 1/2 3/2 O n log n time. If the alphabet has constant size then we use O(n log n) time. To see that our reduction is essentially reversible, suppose we want to determine whether a set W of n integer weights in [−n, . . . , n] is in 3Sum. Without loss of generality, we need check only whether there exist two positive weights and one negative weight that sum to 0: the case when there are two negative weights and one positive weight that sum to 0 is symmetric. For each positive weight w we set S[2w] to 1; for each negative weight w we set S[−w] to 1; and for each remaining i ≤ n we set S[i] to 0 if i ≤ n/2 and to 2 otherwise. There are two positive weights w1 and w3 and one negative weight w2 in W with w1 + w2 + w3 = 0 if and only if −w2 is the average of 2w1 and 2w3 , in which case S has a 3-cadence. 3 Finding Anchored Cadences We can check whether i ≤ n is an anchored cadence of S[1..n] by comparing S[i], S[2i], . . . , S[⌊n/i⌋i], which takes O(n/i) time. Since X 1/i = O(log n) , i≤n obviously we can find all anchored cadences in O(n log n) time. In this section we use the following lemma to improve this bound to O(n log log n). Lemma 1. If a natural number i ≤ n is not an anchored cadence of a string S[1..n], then for some prime p with p i ≤ n, either S[i] 6= S[p i] or p i is not an anchored cadence of S. Proof. Let p i be the smallest multiple of i greater than i itself such that either S[i] 6= S[p i] or p i is not an anchored cadence; p i must exist or i would be an anchored cadence. To see why p must be prime, assume it has a prime factor r < p: then S[i] = S[p i/r] and p i/r is an anchored cadence, by our choice of p, meaning S[i] = S[p i/r] = S[p i] and S[p i] is an anchored cadence, contradicting our choice of p. ⊓ ⊔ We start by computing the sorted set Primes(n) of primes between 1 and n, which takes O(n/ log log n) time [4]. We then build a Boolean array B[1..n] and set B[i] to true for ⌈n/2⌉ ≤ i ≤ n. For i from ⌈n/2⌉ − 1 down to 1, we check whether S[i] = S[p i] and B[p i] = true for each prime p with p i ≤ n, in increasing order; if so, we set B[i] to true and otherwise we set it to false. By Lemma 1, i is an anchored cadence if and only if we eventually set B[i] to true. We check each cell B[j] at most once for each of the distinct prime factors of j. Since each prime p ≤ n divides ⌊n/p⌋ numbers between 1 and n and X 1/p = ln ln n + O(1) p∈Primes(n) (see, e.g., [2]), we use O(n log log n) time. If m is the smallest anchored cadence of S then, for i < m/2, we set B[i] to false immediately after checking B[2i]. Therefore we check each cell B[j] at most once for each prime factor p of j such that p = 2 or j/p ≥ m/2. For each prime p ≤ 2n/m there are n/p − m/2 + O(1) choices of j between 1 and n such that p divides j and j/p ≥ m/2; there is no such choice of j for any larger prime. It follows that we use time O(n log log(n/m)) time. Moreover, if S is chosen uniformly at random from over a non-unary alphabet then in the expected case, for i ≤ n, we check at most 2 primes before finding one p such that S[i] 6= S[p i]. In the average case, therefore, we use O(n) total time. Theorem 2. We can find all the anchored cadences in a string of length n in O(n log log(n/m)) time, where m is the smallest cadence. If the alphabet is nonunary then we use O(n) time on average. Acknowledgments Many thanks to the organizers and other participants of the AxA workshop. References 1. Alberto Apostolico. Pattern discovery and the algorithmics of surprise. In Artificial Intelligence and Heuristic Methods for Bioinformatics, NATO Advanced Science Institutes Series, pages 111–127. 2003. 2. Eric Bach and Jeffrey Shallit. Algorithmic Number Theory: Efficient Algorithms. MIT Press, 1996. 3. Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Cliff Stein. Introduction to Algorithms. MIT Press and McGraw-Hill, 2nd edition, 2001. 4. Paul Pritchard. Improved incremental prime number sieves. In Proceedings of the First International Symposium on Algorithmic Number Theory (ANTS), volume 877 of Lecture Notes in Computer Science, pages 280–288. Springer, 1994.
8cs.DS
Stress-constrained continuum topology optimization: a new approach based on elasto-plasticity arXiv:1608.06654v1 [cs.CE] 23 Aug 2016 Oded Amir Faculty of Civil and Environmental Engineering Technion – Israel Institute of Technology [email protected] August 25, 2016 Abstract A new approach for generating stress-constrained topological designs in continua is presented. The main novelty is in the use of elasto-plastic modeling and in optimizing the design such that it will exhibit a linear-elastic response. This is achieved by imposing a single global constraint on the total sum of equivalent plastic strains, providing accurate control over all local stress violations. The single constraint essentially replaces a large number of local stress constraints or an approximate aggregation of them–two common approaches in the literature. A classical rate-independent plasticity model is utilized, for which analytical adjoint sensitivity analysis is derived and verified. Several examples demonstrate the capability of the computational procedure to generate designs that challenge results from the literature, in terms of the obtained stiffness-strength-weight trade-offs. A full elasto-plastic analysis of the optimized designs shows that prior to the initial yielding, these designs can sustain significantly higher loads than minimum compliance topological layouts, with only a minor compromise on stiffness. keywords: Topology Optimization, Stress Constraints, Elasto-plasticity 1 Introduction Topology optimization of continua is a computational method aimed at optimizing the distribution of one or several materials in a given design domain. The purpose is typically to achieve a minimumweight structural design with a constraint on displacements, or vice-versa: minimize compliance (i.e. maximize stiffness) using a given amount of available material. For extensive reviews see for example [Eschenauer and Olhoff, 2001], [Bendsøe and Sigmund, 2003] and recently [Sigmund and Maute, 2013] and [Deaton and Grandhi, 2014]. One of the most challenging aspects in developing computational topology optimization procedures is the consideration of stress constraints. From an engineering standpoint, limiting the stresses of an optimized design is a fundamental requirement: load-bearing components are typically designed to remain in the linear-elastic regime throughout their service life, meaning that the yield stress should not be exceeded. In this article, a new approach to satisfying stress constraints is proposed. The central idea is to consider the nonlinear, inelastic material behavior and via optimization to drive the design towards a linear-elastic response. The incorporation of stress constraints imposes several challenges. First and foremost is the local nature of stress constraints. Most applications of topology optimization involve an objective functional 1 of global nature and only a few global constraints that control volume, weight, displacements or compliance. If stress constraints are incorporated it is considerably more difficult to tackle the corresponding optimization problem. In principle, stress constraints should be imposed on every material point in the design domain, meaning that the number of constraints is comparable to the number of design variables– both related to the resolution of the underlying finite element mesh. Therefore it is expected that the solution time will be significantly longer than for standard topology optimization problems, and the vulnerability of numerical algorithms to arrive at local minima will be aggravated. Review of the existing literature highlights two dominating strategies for formulating and solving the optimization problem: 1) All local stress constraints are considered in the problem formulation, whereas in the actual solution only a subset of “active” constraints are included; and 2) Local stress constraints are aggregated into a single or into a few global constraints. The former strategy was implemented in one of the earliest publications on stress-constrained topology optimization of continua by [Duysinx and Bendsøe, 1998]. Throughout most of the optimization process, roughly one third of the local constraints are considered for sensitivity analysis and optimization. In the final optimization steps, 180 local constraints (corresponding to 15% of the design variables) are actually active. Local constraints were also considered by [Bruggi and Venini, 2008] and similar tendencies are reported in a later, detailed study of various problem formulations with local stress constraints [Bruggi and Duysinx, 2012]. Therefore the efficiency of imposing local constraints in large-scale problems is questionable. A similar problem formulation but with a different numerical treatment was presented by [Pereira et al., 2004]. All stress constraints are considered but an Augmented Lagrangian technique is utilized for solving the optimization problem, facilitating a reduction in computational effort invested in sensitivity analysis. Results appear promising as they exhibit layouts that circumvent regions with potentially high stress concentrations. On the other hand, the authors report computational times of up to 10 times higher than for standard minimum compliance formulations. The Augmented Lagrangian approach was followed also by [Fancello, 2006] who presented layouts that avoid stress concentrations. The author reported difficulties regarding the numerical implementation and the number of function evaluations indicates that the procedure may not be suitable for large-scale applications. The second, widely adopted strategy for dealing with the large number of stress constraints involves various forms of constraint aggregation, i.e. collecting the constraints into a global stress function. In an early study, [Yang and Chen, 1996] examine the use of both Kreisselmeier-Steinhauser (KS) functions and p-norm functions (referring to [Park, 1995]). Their problem formulations aim at minimizing either the global stress or a weighted combination of global stress and compliance, subject to a volume constraint. A step further in the direction of utilizing global stress measures was proposed by [Duysinx and Sigmund, 1998]. Two global stress functions were suggested, namely the p-mean and the p-norm. It was shown that for any given p, the maximum local stress is bounded from above by the p-norm and from below by the p-mean. Due to ill-conditioning and oscillatory behavior, the maximum value of p is limited to 4 which is not large enough for identifying the actual peak stress. Several recent studies demonstrate that constraint aggregation can in fact lead to satisfactory results–in particular for the classical L-bracket case. [Le et al., 2010] provide an extensive critical review and propose regional stress measures where local stress constraints are grouped in interlacing regions according to their stress level. A similar approach of “block aggregation” was suggested also by [Parı́s et al., 2007, Parı́s et al., 2010]. The regional stress measures are based on p-norms and a normalization with respect to the actual maximum stress is proposed in order to improve the approximation of the maximum stress. In the optimization of the L-bracket, a layout that avoids the re-entrant corner is generated, demonstrating the potential of the approach. On the other hand, the numerical implementation suffers from several drawbacks: First, the normalization is non-differentiable as it changes discontinuously between optimization cycles; second, also the regional constraints can change every optimization cycle according to the sorting of local stresses, thus introducing some inconsistency in the optimization process; and third, it is shown that increasing the number of regions does not always improve the optimized design as one might expect 2 due to the tighter control over local stresses. Level-set methods combined with topological derivatives have also succeeded in generating designs free of stress concentrations. [Allaire and Jouve, 2008] minimize an integral measure of a power-law penalty of stress. This approach is shown to provide smooth designs in re-entrant corners but there is no direct control over the actual stress nor the compliance. Another set of positive results was presented by [Amstutz and Novotny, 2010]. The large number of stress constraints is replaced by an external penalty functional that mimics the point-wise constraint. The layouts avoid stress concentrations at re-entrant corners. Only slight violations of the stress constraints occur in the final designs, e.g. 1%-2% above the target stress. The main drawbacks of this approach are the reliance on penalization parameters that may be problem-specific, and the lack of direct control over local stresses. The significance of the latter depends on the degree of localization of high stresses–which may differ considerably from one problem to another. An interesting novel approach was proposed recently by [Verbart et al., 2013]. Instead of imposing a large number of constraints, material is penalized if the stress exceeds the allowable stress. Numerical results demonstrate the potential of this efficient approach. However, because penalization is utilized it is hard to satisfy the admissible stress criterion accurately. Furthermore, the penalized material law is somewhat artificial so it may be difficult to generalize the method. Another major difficulty in computational stress-constrained topology optimization is the so-called “singularity” problem, originally demonstrated in the context of truss topological design. It was shown that the optimal topology might correspond to a singular point in the design space, therefore making it difficult or in some cases impossible to arrive at the true optimum by numerical search algorithms [Sved and Ginos, 1968, Kirsch, 1990, Cheng and Jiang, 1992]. Such singular points are encountered in cases where removal of a certain truss bar (or a material point in the continuum case) results in a feasible design space with better optimum due to the removal of the corresponding constraint. This article does not target the difficulties related to the singularity phenomenon. In fact, an appropriate relaxation scheme is an essential ingredient in the suggested computational approach. Possibly the most widespread remedy for dealing with the singularity problem is the so-called ε-relaxation [Cheng and Guo, 1997], where the actual stress constraints are relaxed so that the resulting feasible domain does not possess degenerate branches. A similar relaxation scheme involving smooth envelope functions was suggested in the context of local buckling constraints by [Rozvany, 1996]. The ε-relaxation approach was first integrated into continuum topology optimization by [Duysinx and Bendsøe, 1998] and by [Duysinx and Sigmund, 1998], who implemented a continuation scheme for gradually reducing ε hence approaching the actual constraints. The ε-relaxation approach was successfully applied to various test cases in later studies, see for example [Pereira et al., 2004, Fancello, 2006, Le et al., 2010]. An alternative relaxation for avoiding the singularity phenomenon was introduced by [Bruggi, 2008]. In the SIMP rule, the penalization power q for the yield stress was chosen to be lower than the penalization power p for the stiffness. It is noted that separate penalization exponents for stiffness and for yield stress have been suggested much earlier in an extension of SIMP-based topology optimization to elasto-plastic structures [Maute et al., 1998]. In practice, the so-called qp-relaxation appears to provide similar results to those of the ε-relaxation. Both are highly dependent on the continuation scheme and it was shown that a sequence of solutions to the relaxed problem may not converge to the global optimum [Stolpe and Svanberg, 2001]. Despite these shortcomings it seems necessary to apply some form of continuous relaxation in order to arrive at practical structural designs that satisfy stress requirements. The central idea of the approach proposed in this article is to optimize the inelastic structural response for the particular purpose of satisfying stress constraints in linear elasticity. Up to date, applications of topology optimization that considered inelastic response were concerned with objectives other than the one pursued herein. Material nonlinearities in topology optimization were initially considered by [Yuge and Kikuchi, 1995]. Layout optimization of frame structures undergoing plastic deformation was presented, based on homogenization of porous material. [Swan and Kosaka, 1997] sug- 3 gested a framework for topology optimization of structures with material nonlinearity based on Voigt and Reuss mixing rules. The SIMP (Solid Isotropic Material with Penalization) interpolation scheme, originally proposed for linear elastic material [Bendsøe, 1989], was extended for elasto-plastic behavior by [Maute et al., 1998]. Although several other articles on the subject were published over the last two decades, topology optimization involving elasto-plasticity is still not well established. One difficulty lies in obtaining accurate design sensitivities. In some cases, several derivative terms are neglected [Maute et al., 1998, Schwarz et al., 2001]. Apparently this has a minor effect on the outcome of the optimization but in general these terms are not negligible. Moreover, when comparing analytical design sensitivities to finite difference calculations, errors in the order of 10−2 are observed [Swan and Kosaka, 1997, Yoon and Kim, 2007]. Recent studies have incorporated analytical adjoint sensitivity analysis for rateindependent elasto-plasticity, based on the framework by [Michaleris et al., 1994]. Accurate sensitivities have been reported for problems involving reinforced concrete design [Bogomolny and Amir, 2012] and effective energy management under dynamic loading [Nakshatrala and Tortorelli, 2015]. Another analytical sensitivity analysis scheme for topology optimization of elasto-plastic structures was recently presented by [Kato et al., 2015]. Highly accurate derivatives are obtained, however the formulation is limited to cases in which the load is applied only to the nodes whose displacements are controlled. Adjoint sensitivity analysis has been applied also for topology optimization with viscoelastic material [James and Waisman, 2015]; viscoplastic micro-heterogeneous materials in a multiscale approach [Fritzen et al., 2015]; and continuum damage models [Amir and Sigmund, 2013, Amir, 2013, James and Waisman, 2014]. The latter study in fact targeted a similar goal as in this article–mitigating failure, i.e. imposing stress constraints–but was based on a different constitutive model and on a different problem formulation which involved constraint aggregation. As mentioned in the beginning of the introduction, the proposed approach relies on modeling the inelastic behavior and driving the design towards a linear-elastic response. This is achieved by constraining the total sum of equivalent plastic strains. A single global constraint is added to the standard stiffness vs. volume problem, inherently providing accurate control over all local stress violations. Consequently stress limits can be implicitly satisfied, without imposing a large number of local constraints. The corresponding computational procedure can alleviate one of the major obstacles in stress-constrained topology optimization–the need to solve a nonlinear, non-convex optimization problem with a large number of design variables and an equally large number of constraints. The remainder of the article is organized as follows. In Section 2 we briefly review the elasto-plastic material model and the nonlinear finite element analysis formulation. The topology optimization problem formulation and the design parametrization are then introduced in Section 3, followed by a derivation and verification of the adjoint sensitivity analysis in Section 4. Several examples demonstrate the applicability of the proposed approach in Section 5. Finally, a discussion of the results and of necessary future investigations is given in Section 6. 2 Elasto-plastic model and finite element analysis In this section we briefly review the material model and the subsequent nonlinear finite element analysis. The purpose is to provide the necessary background for the optimization problem formulation presented in Section 3, which involves state variables related to the elasto-plastic material model, as well as for the adjoint sensitivity analysis, presented in Section 4. 2.1 Classical rate-independent plasticity The derivation of the governing equations herein follows the textbooks by [Simo and Hughes, 2006] and by [Zienkiewicz and Taylor, 2000]. The model is essentially composed of the following assumptions and 4 rules: elastic stress-strain relationships; a yield condition, defining the elastic domain; a flow rule and hardening law; Kuhn-Tucker complementarity conditions; and a consistency condition. We first assume that the total strain tensor ε can be split into its elastic and plastic parts, ε el and ε pl respectively, ε = ε el + ε pl . (1) Furthermore, we relate the stress tensor σ to the elastic strains using the elastic constitutive tensor D, σ = Dε el . (2) The yield criterion f is a function that defines the admissible stress states f (σ , q) ≤ 0 (3) where q are internal variables related to the plastic strains and to the hardening parameters. The elastic domain is defined by the interior of the yield criterion where f < 0; the yield surface is defined by f = 0; and the stress state corresponding to f > 0 is considered non-admissible. The irreversible plastic flow is governed by the evolution of plastic strains and internal variables ε̇ pl = λ̇ r(σ , q) q̇ = −λ̇ h(σ , q) (4) (5) where r and h are functions defining the direction of plastic flow and the hardening of the material. The parameter λ̇ is typically called the consistency parameter or plastic multiplier. Together with the yield criterion, λ̇ must satisfy the Kuhn-Tucker complementarity conditions λ̇ ≥ 0 f (σ , q) ≤ 0 λ̇ f (σ , q) = 0 (6) λ̇ f˙(σ , q) = 0. (7) as well as the consistency requirement The consistency requirement means that during plastic loading, the stress state must remain on the yield surface, meaning f˙ = 0 if λ̇ > 0. A widely accepted model of rate-independent plasticity in metals is usually known as J2 flow theory or simply J2 -plasticity. It is based on the von Mises yield criterion [von Mises, 1928] that relates the yielding of the material to the deviatoric stresses, measured by the second deviatoric stress invariant J2 . The model is hereby presented as a particular case of rate-independent plasticity. The yield criterion is the von Mises yield function expressed as p (8) f (σ , κ) = 3J2 − σy (κ) ≤ 0 √ where the expression 3J2 is usually named the von Mises stress or equivalent stress. σy is the yield stress in uniaxial tension, which depends on a single internal parameter κ according to an isotropic hardening function. Kinematic hardening is not considered in the current work. A popular choice for the hardening rule is the bi-linear function (9) σy (κ) = σy0 + HEκ where σy0 is the initial yield stress, H is a scalar (usually in the order of 10−2 ) and E is Young’s modulus. An associative flow rule is assumed, meaning that the flow of plastic strains is in a direction normal to the yield surface ∂f . (10) ε̇ pl = λ̇ ∂σ 5 Finally, the internal variable governing the hardening is the equivalent plastic strain, evolving according to the rule r 2 pl . (11) ε̇ κ̇ = 3 2 q The factor 23 is introduced so that for the particular one-dimensional case (involving uniaxial plastic deformation), the obvious relation will be obtained, i.e. κ̇ = ε˙pl . 2.2 Finite element implementation For finite element analysis, the process of rate-independent plasticity is conveniently represented as a flow evolving in time, where each time step corresponds to an increment of load or displacement. In the current work, a standard Newton-Raphson incremental-iterative scheme with displacement control is employed. For the purpose of sensitivity analysis in optimal design, the finite element equations are cast into the framework for transient, coupled and nonlinear systems suggested by [Michaleris et al., 1994]. In the coupled approach, for every ‘time’ increment n in the transient analysis, we determine the unknowns un , vn and θn that satisfy the residual equations Rn (un , un−1 , vn , vn−1 , θn ) = 0 (12) Hn (un , un−1 , vn , vn−1 ) = 0 (13) where un is the displacements vector, θn is the load factor and vn are the internal variables–all corresponding to the time tn . Rn = 0 is satisfied at the global level and Hn = 0 is satisfied at each Gauss integration point. The transient, coupled and nonlinear system of equations is uncoupled by treating the response v as a function of the response u. When solving the residual equations for the n-th increment, the responses un−1 and vn−1 are known from the previous converged increment. The independent response un is found by an iterative prediction-correction procedure in the global level, while for each iterative step the dependent response vn (un ) is found by an inner iterative loop. The responses un and its dependant vn are corrected until Eqs. (12) and (13) are satisfied to sufficient accuracy. This procedure is repeated for all N increments. Neglecting body forces, Rn is defined in the current study as the difference between external and internal forces and depends explicitly on vn and θn only Rn (vn , θn ) = θn f̂ − Z V BT σ n dV (14) where f̂ is a constant reference load vector with non-zero entries only at loaded degrees of freedom and B is the standard strain-displacement matrix in the context of finite element procedures. For the particular material model used in this study, the vector vn is given by  pl  εn  κn   vn =  (15)  σn . λn For solving the local nonlinear constitutive problem, an implicit backward-Euler scheme is employed. The central feature of this scheme is the introduction of a trial elastic state. For any given incremental displacement field, it is first assumed that there is no plastic flow between time tn and the next time step tn+1 , meaning the incremental elastic strains are the incremental total strains. It can be shown that the loading/unloading situation which is governed by the Kuhn-Tucker conditions can be identified using the 6 trial elastic state [Simo and Hughes, 2006]. Once a plastic increment occurs, the new state variables can be found by solving a nonlinear equation system resulting from the time discretization of the governing equations. This results in the nonlinear system Hn which is derived specifically for any given elastoplastic model. For the particular model used in the current study, Hn is defined as the collection of four incremental residuals, resulting from the time linearization of the governing constitutive equations 1 Hn 2 Hn 3 Hn 4 Hn ∂f T pl + (λn − λn−1 )( ) − ε npl = ε n−1 ∂σn s 2 ∂f T ∂f ( = κn−1 + (λn − λn−1 ) ) ( ) − κn 3 ∂σn ∂σn i h pl ) −σn = σ n−1 + D Bun − Bun−1 − (ε npl − ε n−1 1 = J2n − (σy (κn ))2 . 3 (16) The equation 1 Hn represents the associative flow rule; 2 Hn represents the evolution of the isotropic hardening parameter; 3 Hn relates stresses to elastic strains;; and 4 Hn is the yield criterion in squared form, with J2n representing the second deviatoric stress invariant evaluated using σ n . It is worth noting that the local nonlinear problem of Eq. (16) can be solved efficiently as a scalar equation by a return-mapping algorithm, for example as derived by [Simo and Taylor, 1986] for plane stress situations. Nevertheless, for the purpose of sensitivity analysis we find it convenient to use the full representation as suggested by [Michaleris et al., 1994]. 3 Topology optimization approach The central goal of the proposed formulation is to generate optimized structural layouts that can sustain a certain load under a prescribed range of displacements, while not exceeding the allowable stress limitations. In the majority of studies so far, this design problem was formulated as an optimization of a linear-elastic structure, aimed at minimizing either compliance or volume. Stress limitations were imposed as constraints, either locally on each material point or in a global, aggregated manner. In the suggested formulation, we approach the same design goal in a completely different way. Essentially, we seek the best trade-off between three quantities: 1) The weight of the structure, coinciding with the volume for single-material layouts; 2) The load-bearing capacity, represented by the end-compliance–the product of loads and displacements at the final (time-wise) equilibrium point; and 3) The overall sum of plastic strains, representing the violation of allowable stress limits. In the numerical experiments, two variants of the optimization problem are examined. These arise from assigning each of the above quantities 1 or 2 as an objective, while constraining the other one and quantity 3. It will be shown that both variants lead to satisfactory results. In the remainder of this section, quantity 2 is considered in the objective. Other variants can be derived in a very similar manner. 3.1 Problem formulation For the purpose of optimizing the topological layout of a continuum, we follow the material distribution approach [Bendsøe and Kikuchi, 1988] together with the SIMP interpolation scheme [Bendsøe, 1989] and its extension to multiple phases, usually known as Modified SIMP [Sigmund and Torquato, 1997]. This implies that the design variables x are densities at discrete material points, assigned at the centroid of each finite element in the design domain and varying between zero (void) and 1 (solid). The optimization 7 problem can be stated as follows min g0 = −θN f̂T uN s.t.: g1 = x Ne ∑ ve xe − g?1 ≤ 0 e=1 Ne NGP g2 = ∑ ∑ κNek − g?2 ≤ 0 e=1 k=1 0 ≤ xe ≤ 1, with: e = 1, ..., Ne Rn (vn , θn , x) = 0 n = 1, ..., N Hn (un , un−1 , vn , vn−1 , x) = 0 n = 1, ..., N. (17) The objective is to maximize the end-compliance for a given prescribed displacement, i.e. maximize the load-bearing capacity for a given magnitude of deformation. This quantity is evaluated using the terminal values of the load factor θN and of the displacements uN . The constraint g1 ensures that no more than a certain prescribed volume g?1 is utilized. The design volume is measured according to the physical material density xe of each finite element. The physical densities x are related to the mathematical variables x via widely used filtering and projection techniques which will be presented explicitly in the next section. The constraint g2 ensures that the overall spatial sum of the plastic strains does not exceed a certain small threshold g?2 , which can in theory be zero. The sum of plastic strains is evaluated based on the quantity κN , i.e. the plastic strain measured at each Gauss point in the finite element mesh, at the terminal equilibrium point. Finally, the nonlinear residuals Rn and Hn are as defined in Section 2.2 according to the respective elasto-plastic model. 3.2 Design parametrization The correspondence between the mathematical optimization variables x and the nonlinear finite element analysis is as follows. First a standard density filter is applied [Bruns and Tortorelli, 2001, Bourdin, 2001] with a simple linear weighting function to obtain e x. The purpose of applying a density filter is to overcome the well-known difficulty of artificial checkerboard patterns as well as to introduce a length scale in the design, thus avoiding results with very thin features that are difficult to manufacture. Then, a Heaviside projection function [Guest et al., 2004, Xu et al., 2010] is utilized in order to ‘push’ the design towards a distinct 0-1 (or void-solid) layout. This yields the physical density distribution,  h i  η e−βHS (1−exe /η) − (1 − xee /η)e−βHS    h xe = (1 − η) 1 − e−βHS (exe −η)/(1−η) +  i    (e xe − η)/(1 − η)e−βHS + η 0 ≤ xee ≤ η (18) η < xee ≤ 1 where η is a threshold value and βHS is a parameter determining the ‘sharpness’ of the smooth projection function. In the current study we use η = 0.5, meaning that any filtered density above 0.5 is projected to 1 and any value below 0.5 is projected to 0. The initial value of βHS is usually set to 1 and it is increased gradually as the optimization progresses. Heaviside projections are typically introduced in order to achieve crisp void-solid layouts which are necessary in some design problems due to manufacturing requirements. For the cases addressed in this article, it is not absolutely necessary to utilize such projections, which increase the degree of nonlinearity and may cause difficulties in convergence. Nevertheless, it is useful to apply the Heaviside projection, even with rather mild βHS values, in order to minimize material transition regions. In regions where the density is between zero and one, also known as gray regions in topology optimization, the elasto-plastic material law is artificial. This is due to the choice of penalization scheme as will be explained below. Therefore, the true stress in the actual (manufactured) 8 (a) (b) Figure 1: Normalized stress-strain curves for various densities ρ and separate penalty exponents pE and pσy . Left: pE = 1.0, pσy = 0.5; Right: pE = 3.0, pσy = 2.5. For densities smaller than 1, the yield strain is relatively delayed. structure may differ from the computed stress within the optimization. This motivates the minimization of gray transition regions. The constitutive model corresponding to J2 flow theory involves three material parameters: Young’s modulus E, the hardening fraction H and the initial yield stress σy0 . As mentioned above, an extension to the SIMP approach for interpolating the three parameters was originally presented by [Maute et al., 1998]. For evaluating the tangent stiffness matrix and the internal forces vector, Young’s modulus is interpolated in each finite element as follows E(xe ) = Emin + (Emax − Emin )xe pE . (19) In general, Emin and Emax are the values of Young’s modulus of two candidate materials which are distributed in the design domain. For the case of distributing a single material and void, Emin is set to be several orders of magnitude smaller than Emax . Finally, pE is a penalization factor required to drive the design toward a 0-1 layout. The initial yield stress is penalized similarly, 0 0 0 + (σy,max − σy,min )xe pσy σy0 (xe ) = σy,min (20) 0 0 where σy,min and σy,max are the initial yield stresses for the two candidate materials, corresponding to x = 0 and x = 1 respectively. From a physical point of view, the penalization factor pσy should be equal to pE so that the yield strain does not depend on the density. However, in many cases it is necessary to set pσy < pE in order to avoid numerical difficulties arising when low density elements reach their yield limit. The physical consequence is that in intermediate densities, the yield strain is artificially higher than that of the full material, as shown in Figure ?? Separate exponents in elasto-plastic topology optimization were already introduced by [Maute et al., 1998]. This approach is used also with stress constraints (namely the qp-relaxation, [Bruggi, 2008]) and is similar to ε-relaxation approaches [Cheng and Guo, 1997]. In this study, we keep H independent of the design variables because the post-yield stiffness is already penalized via Eq. (19). Furthermore, the essence of the design problem is to find designs that do not yield or that have a very short post-yield response. For such cases, it is not necessary to consider an accurate post-yield response, especially for intermediate material densities. For solid material, a constant H is the same as having a SIMP-type interpolation for H. 9 4 Sensitivity Analysis Considering the optimization problem in Eq. (17), the derivatives of the volume constraint g1 are straightforward. The objective g0 and the constraint g2 involve state variables, therefore an adjoint sensitivity analysis procedure is necessary. As mentioned earlier, the design sensitivities are computed following the framework for transient, nonlinear coupled problems described by [Michaleris et al., 1994]. In the following we focus on computing the derivatives of a general functional with respect to the physical densities x whereas the derivatives with respect to x can then be computed by the chain rule. 4.1 Backwards-incremental adjoint procedure We begin by forming the augmented functional ĝ(x) N ĝ(x) = g− ∑ λ Tn Rn (vn (x), θn (x)) n=1 N − γ Tn Hn (un (x), un−1 (x), vn (x), vn−1 (x), x) n=1 ∑ (21) where for clarity, the dependency of g on state and design variables was omitted. From here on, λ n represents an adjoint vector corresponding to increment n, not to be confused with the scalar λn which is used for the time discretization of the plastic multiplier λ̇ . Furthermore, λ n is a global adjoint vector whereas γ n is a local (Gauss-point) adjoint vector. In principle, g can be a function of all state variables throughout all time steps, in addition to its dependency on design variables. For the particular functionals in Eq. (17), we see that g0 (x) = g0 (uN (x), θN (x)) and g2 (x) = g2 (vN (x)). These relations are utilized in the particular implementation of the adjoint procedure for each functional. The purpose of the adjoint procedure is to eliminate all terms involving derivatives of state variables with respect to design variables, which cannot be computed explicitly. It can be seen that the only explicit dependency upon design variables is contained in Hn , yielding the expression for the explicit sensitivity with respect to an element physical density N ∂ ĝexp ∂ (Hn ) = − ∑ γ Tn . ∂ xe ∂ xe n=1 (22) The adjoint vectors γ n (n = 1, ..., N) are computed in each Gauss integration point by a backwardsincremental procedure, which is required due to path dependency of the elasto-plastic response. The backwards procedure consists of the collection of equation systems resulting from the requirement that all implicit derivatives with respect to the design variables will vanish. Complete details regarding the derivation of the adjoint procedure can be found in [Michaleris et al., 1994], whereas specific implementations are described by [Amir, 2011], [Bogomolny and Amir, 2012] and [Nakshatrala and Tortorelli, 2015]. Implementations with other nonlinear material models were mentioned in the introduction. The adjoint procedure begins with a coupled system to be solved for λ N , " λN = " #T ∂g T ∂ g ∂ (HN ) −1 ∂ (HN ) − ∂ (uN ) ∂ (vN ) ∂ (vN ) ∂ (uN ) ∂ (RN ) T λN ∂ (θN ) = ∂g ∂ (θN ) ∂ (RN ) ∂ (HN ) −1 ∂ (HN ) − ∂ (vN ) ∂ (vN ) ∂ (uN )  where ∂ (RN ) ∂ (HN ) −1 ∂ (HN ) ∂ (vN ) ∂ (vN ) ∂ (uN ) #T (23)  is the tangent stiffness matrix corresponding to the converged state at incre- 10 ment N [Michaleris et al., 1994]. γ N is then determined on a Gauss-point level by solving ∂ (RN ) T ∂g T ∂ (HN ) T γN = − λN + . ∂ (vN ) ∂ (vN ) ∂ (vN ) (24) Proceeding incrementally backwards in time, in the n-th increment the coupled adjoint equations are solved to determine λ n #T " #T " ∂g T ∂ g ∂ (Hn ) −1 ∂ (Hn ) ∂ (Rn ) ∂ (Hn ) −1 ∂ (Hn ) λn = − − ∂ (vn ) ∂ (vn ) ∂ (un ) ∂ (un ) ∂ (vn ) ∂ (vn ) ∂ (un ) " #T ∂ (Hn+1 ) ∂ (Hn+1 ) ∂ (Hn ) −1 ∂ (Hn ) − γ n+1 − ∂ (un ) ∂ (vn ) ∂ (vn ) ∂ (un ) ∂ (Rn ) T λn ∂ (θn ) = ∂g ∂ (θn ) (25) followed by the solution of the local adjoint vector γ n on a Gauss-point level ∂ (Rn ) T ∂ (Hn+1 ) T ∂g T ∂ (Hn ) T γn = − λn − γ n+1 + . ∂ (vn ) ∂ (vn ) ∂ (vn ) ∂ (vn ) (26) Once γ n is determined, its contribution to the design sensitivities is computed. Then the procedure continues to the previous increment denoted by n − 1. This is repeated until all contributions are collected to obtain the required design sensitivities. The partial derivatives of the objective, the constraints, the global residuals and the local residuals with n) respect to the state variables are required for implementing the adjoint procedure. The derivatives ∂∂(R (vn ) and ∂ (Rn ) ∂ (θn ) ∂ (Hn ) ∂ xe can be easily obtained from Eq. (14) whereas the derivatives ∂ (Hn ) ∂ (Hn+1 ) ∂ (Hn ) ∂ (Hn+1 ) ∂ (un ) , ∂ (un ) , ∂ (vn ) , ∂ (vn ) and are related to the particular elasto-plastic model and to the choice of the internal variables v. For the model considered herein based on classical J2 flow theory, they can be derived by differentiation of Eq. (16). An explicit example of these derivatives was given in [Amir, 2011]. Finally, the partial ∂g ∂g derivatives ∂ (u , ∂ g , ∂ g , ∂ g , ∂ g and ∂ (θ can be derived explicitly for each functional to be n) N ) ∂ (vN ) ∂ (θN ) ∂ (un ) ∂ (vn ) considered in the problem formulation of Eq. (17). It should be noted that when implementing the adjoint procedure, the derivatives of the local residuals Hn and Hn+1 should maintain consistency with respect to the analysis. In essence, four situations are possible at a certain sequence of increments {n, n + 1}: 1) Continuous elastic response; 2) Elastic-plastic transition; 3) Continuous plastic response; and 4) Plastic-elastic transition (during unloading). The actual situation encountered affects the computation of the derivatives of the respective residuals. In general, the derivatives of the local residual are matrices of varying sizes, depending on the situation which is determined exclusively by the elastic trial state. The final component required for performing the sensitivity analysis is the derivative of the residual Hn with respect to the physical material density. Combining Eqs. (16), (19) and (20), we obtain   0  0 ∂ (Hn )    ∂E el − ε el ) = (27)  D (ε n−1 ∂ xe 0 n ∂ xe   0 ∂σ − 23 (σy0 + HEκn )( ∂ xey + H ∂∂xEe κn ) ∂σ0 where ∂∂xEe and ∂ xey are derived from Eqs. (19) and (20), and D0 is the elastic constitutive tensor for Young’s modulus equal to 1. 11 Figure 2: Problem setup for topology optimization of an L-bracket. The load is distributed over the top 10 nodes in order to avoid artificial stress concentrations at the loading point. 5 Examples In this section we present several numerical examples that demonstrate the applicability of the proposed approach for solving stress-constrained topology optimization. Different variants of the optimization problem of Eq. (17) are considered. All optimization problems are solved by the method of moving asymptotes - MMA [Svanberg, 1987]. Specific parameters required for reproducing the results are given within the text. 5.1 Example 1: Stress-constrained L-bracket design In the first example we consider the classical case of an L-bracket, which is often used to evaluate new procedures for topology optimization under stress constraints (e.g. [Le et al., 2010] and references therein), see Figure 2 for the problem setup. In particular, the case of the L-shaped domain with a point load at the position {1.0, 0.4} is thoroughly examined, as by [Le et al., 2010] and [Verbart et al., 2013]. A similar problem, often appearing in articles about stress constraints, is the case in which the point load is applied at the position {1.0, 0.2}. In the author’s opinion, the latter is somewhat easier to deal with because the optimized layout for compliance only has a wider angle in the re-entrant corner, see Figures 3a and 6a. Therefore the former case is in the center of the following examination, in order to fairly evaluate the proposed approach. The latter case is presented subsequently for the sake of completeness. 5.1.1 Reference design: no stress limitation First, a maximization of end-compliance subject to a volume constraint only is performed, given a certain prescribed displacement. This is necessary in order to identify the “stiffest design” achievable without any limitation on stresses. The load is distributed over the top 10 nodes in order to avoid artificial stress concentrations at the loading point. Assuming the 10 adjacent nodes will have almost identical vertical displacements, it is sufficient to measure the end-compliance based on a single DOF where the displacement is prescribed, instead of measuring the complete end-compliance. This somewhat simplifies the computational implementation of the adjoint equations, though the derivation above is general and applicable to any loading situation. With reference to the formulation in Eq. (17), the objective now includes 12 Emin Emax ν 0 σy,min 0 σy,max H 1.0 · 10−3 1.0 · 103 0.3 0.0 2.0 0.01 Table 1: Material parameters used for all examples. only the product of force and displacement at the prescribed DOF and the constraint g2 is omitted. The resulting layout and performance coincide with those that can be obtained by a linear-elastic minimum compliance topology optimization procedure. This is expected for an elasto-plastic single-material optimization with strain hardening. If multiple materials are considered, with distinct yield stresses and hardening behaviors, the optimized layouts may differ from linear-elastic minimum compliance layouts, see for example [Kato et al., 2015]. The model is discretized with a 150×150 mesh resolution consisting of 14,400 square, bi-linear elements. The available volume is set to 35% of the total volume of the L-shaped domain and the filter radius is 0.02. The prescribed displacement the position {1.0, 0.4} is set to u p = 0.01 and automatic displacement incrementation is applied, where the increment size is adapted based on the convergence of Newton-Raphson iterations in the previous increment. The material parameters are given in Table 1 and are essentially constant for all examples. According to the numerical experiments, a continuation scheme involving both the penalty exponents and the Heaviside sharpness yields the best results. The parameters pE and pσy are increased gradually throughout the optimization process. The initial values are set to pE = 1.0 and pσy = 0.5 and they are increased by 0.1 every 10 design cycles, up to the values of 5.0 and 4.5 respectively. The parameter βHS is initialized at 1.0 and multiplied by 1.1 every 10 design cycles, but only when pE ≥ 3.0. The upper limit for βHS is set to 10.0 in order to avoid highly nonlinear projection functions. In the call to MMA, the derivatives of the compliance objective are multiplied by 105 in order to obtain good scaling and consequently fast convergence of the MMA sub-problems. It is known that the performance of MMA can be affected by this scaling parameter, which should be chosen according to the values of the actual quantities–in this case, the magnitude of the end-compliance is in the order of 10−5 . According to the author’s experience, if the problem is badly scaled then it can slow down convergence and in some cases lead the overall optimization process to inferior local minima. An external move limit of 0.2 on the MMA update is enforced. In all examples presented in this section, the optimization is terminated after 500 design cycles. The stopping criterion that was imposed, requiring that the maximum change in an element’s density is below 10−3 , was never achieved. The optimized topology is presented in Figure 3a. The end-compliance, the sum of equivalent plastic strains and the volume are presented in the second column of Table 2. The stress distribution in terms of von Mises stresses is presented in Figure 3d and the distribution of equivalent plastic strains in Figure 3g. It can be seen that there is a significant stress concentration in the vicinity of the re-entrant corner. 5.1.2 Constraining the plastic straining The exact same problem setup is used to demonstrate the capability of the proposed approach to capture local stress concentrations and to consider them when seeking the optimized topology. We add a single, global constraint on the total sum of the equivalent plastic strains at the final equilibrium state, as in Eq. (17). It will be shown that the stress distribution is improved and stress concentrations are avoided. All material parameters remain the same, so does the continuation scheme for pE , pσy and βHS as well as the scaling and move limit for MMA. A small threshold of g?2 = 10−4 is set for the constraint on plastic 13 End-compliance θN fˆp uNp NGP ek Ne Plastic strains ∑e=1 ∑k=1 κN Ne v ∑e=1 e xe Volume 0.35·Ne ·ve − 1 Figures Max. end-comp. s.t. vol. 4.3013 · 10−5 2.0455 · 10−1 Max. end-comp. s.t. vol. and plastic strains 3.8163 · 10−5 1.3353 · 10−3 Min. vol. s.t. end-comp. and plastic strains 3.7908 · 10−5 1.1369 · 10−2 −5.6769 · 10−7 3a, 3d, 3g −2.7461 · 10−4 3b, 3e, 3h −2.2727 · 10−2 3c, 3f, 3i Table 2: Results of the topology optimization of an L-bracket with a load at the top right corner. For the same volume and under the same prescribed displacement, constraining the sum of equivalent plastic strains leads to nearly zero plastic strains while compromising the end-compliance by 13%. strains, providing some slack and improving the convergence to a design with almost zero plastic strains. The optimized topology is presented in Figure 3b. The end-compliance, the sum of equivalent plastic strains and the volume are presented in the third column of Table 2. The stress distribution in terms of von Mises stresses is presented in Figure 3e and the distribution of equivalent plastic strains in Figure 3h. Examining the optimized layout, it can be seen that the proposed approach can indeed generate designs that circumvent stress concentrations. The modification of the design, compared to the one optimized for end-compliance only, is quite subtle–a rounding of the re-entrant corner and a stiffening of the bars meeting at the corner in order to compensate for the reduced stiffness of the rounded corner. This result is slightly different from those achieved in previous studies referenced above, and appears to be the closest one to the layout obtained without stress considerations. It resembles a result achieved by [Bruggi and Venini, 2008] with a mixed-FEM approach and local stress constraints. It also resembles a result achieved by [Le et al., 2010] but for a much higher volume fraction (see Figure 7(b) in the referenced article). It can be argued that the primary design change is a shape modification, that may have been generated by a shape optimization procedure following topology optimization without stress constraints. Nevertheless, it can be seen that topological changes are indeed introduced, for example the stiffening bars orthogonal to the main bars. This means that performing topology optimization without stress considerations, followed by shape optimization with stress considerations, may not be sufficient for finding the best layout in terms of both topology and shape. As for the stress distribution, it is evident that adding a constraint on plastic strains leads to a more uniform distribution of extreme stresses. Hence stress constraint violation is implicitly avoided without actually imposing local stress constraints on each material point. Nevertheless, a slight violation of the global constraint on plastic strains is observed–plastic straining is present in the first element near the re-entrant corner. This violation can be attributed to several factors: 1) The non-differentiability of plastic strains at the yield point, causing difficulties in satisfying the constraint precisely; 2) The inherent approximation due to the use of a sequential convex programming method for solving a non-convex problem; and 3) The quality of the optimization algorithm itself. Another possibility to achieve the same design goal is by interchanging end-compliance and volume in Eq. (17). This corresponds to minimizing the volume of the optimized design while requiring a certain load-bearing capacity for a given prescribed displacement. The end-compliance is constrained to 3.8 · 10−5 in order to obtain a good comparison with the result of maximization of end-compliance s.t. volume and plastic strains. The initial volume fraction is set to 100% of the domain. All parameters used in the solution of the previous case retain the same values, except for the continuation on the penalization exponents. For effectively constraining the end-compliance, it is necessary to begin the process with some penalty in Eqs. (19) and (20). Therefore the initial values are chosen as pE = 3.0 and pσy = 2.5 and they are kept constant for the first 200 iterations. Then, the same continuation scheme is applied as for the previous cases. The optimized topology is presented in Figure 3c. The end-compliance, the sum of equivalent plastic strains and the volume are presented in the fourth column of Table 2. The stress distribution in terms of von Mises stresses is presented in Figure 3f and the distribution of equivalent 14 (a) (b) (c) (d) (e) (f) (g) (h) (i) Figure 3: Topology optimization of an L-bracket with a load and a prescribed displacement at the top right corner. From the left: maximizing the end-compliance s.t. a volume constraint; maximizing the end-compliance s.t. constraints on volume and on the total sum of equivalent plastic strains; minimizing volume s.t. constraints on end-compliance and on the total sum of equivalent plastic strains. From top: optimized layouts; von Mises stress distributions; equivalent plastic strains in the vicinity of the re-entrant corner. Note the two orders of magnitude difference between the scale of sub-figure (g) compared to (h) and (i). 15 plastic strains in Figure 3i. 5.1.3 Elasto-plastic performance of the optimized design For examining the actual benefit from the proposed formulation, it is interesting to examine the elastoplastic response of the three designs obtained in the numerical experiments. The responses are directly compared to a result by [Le et al., 2010], in particular the layout in Fig. 7(e) of the referenced article whose setting is the closest to the current problem definition. A simple post-processing of the layouts obtained in the current study is performed, consisting of a rounding of all intermediate design variables to 10−6 and 1, with 0.5 as the threshold value. This hardly affects the layout and volume (thanks to the Heaviside projection utilized within the optimization process) and facilitates a more accurate elasto-plastic analysis. The only visible difference is that a thin bar in the minimum volume design is partially deleted, hence it will not contribute to the transfer of forces in the comparative case. An image of the layout obtained by [Le et al., 2010] was imported and processed to obtain a design as similar as possible to the original, while adapting the mesh resolution to that of the current study. The same projection scheme was applied as described above but with a threshold of 0.26 for achieving roughly the same volume fraction as the other designs. The four post-processed layouts are presented in Figure 4. For convenience, the layouts are tagged #1, #2, #3 and #4, corresponding to maximum end-compliance s.t. volume, maximum end-compliance s.t. volume and plastic strains, minimum volume s.t. end-compliance and plastic strains and the design by a p-norm approach [Le et al., 2010]. The computed volume fractions are 34.82%, 35.11%, 34.21% and 34.84% for designs #1, #2, #3 and #4, respectively. The solution parameters are the same as for the optimization runs, except for the prescribed displacement: it is increased to 0.02 in order to guarantee that all designs enter the plastic regime so that their actual elasto-plastic response can be compared. The displacement is applied within 40 equal increments. Furthermore, pE = 1.0 and pσy = 0 so that for the solid parts the true elasto-plastic law is obtained, whereas for void there is effectively no yield. The load-displacement curves at the prescribed DOF are presented in Figure 5. It can be seen that the responses of the designs obtained with a constraint on plastic strains exhibit a significant delay of the initial yield. The magnitude of load at the initial yield is increased by approximately 42% and 54% compared to the reference design #1, for designs #2 and #3, respectively. This comes with a certain compromise on stiffness, meaning that the displacement level for a given load is slightly higher–this is expected because layout #1 is “the stiffest design”. The elasto-plastic optimization also provides superior trade-offs of volume-stiffness-strength, compared to those obtained with a p-norm approach [Le et al., 2010]. These results demonstrate the capability of the proposed approach to provide topological designs that account for stress constraints in an early design stage. The topological layouts obtained provide a significant delay in the initial yield alongside a minor compromise on stiffness. This coincides with the common design goal of finding a stiff design that does not fail prematurely. 5.1.4 L-bracket with a load at the mid-point As mentioned above, many previous studies on stress-constrained topology optimization treated the Lbracket problem with a point load at the middle of the right side edge. Therefore it is interesting to apply the proposed approach also for this setup. A direct comparison of the final elasto-plastic response to a published result is sought also for this case. For this purpose, a layout obtained using a level-set approach by [Amstutz and Novotny, 2010] is chosen–in particular, the design in the center of Fig. 8 in the referenced article. The volume fraction for the current optimization runs is modified accordingly, to 48.25%. All other parameters remain the same as for the L-bracket with a load at the top, except for the local distribution of the point load–5 nodes are sufficient in this case for avoiding a stress concentration. 16 (a) (b) (c) (d) Figure 4: Post-processed 0/1 designs of the L-bracket for comparing the elasto-plastic responses of the optimized layouts. Only minor changes are observed compared to the results of optimization in Figure 3. The layout (d) is based on an image imported from [Le et al., 2010] for comparing the responses. 17 Figure 5: Elasto-plastic response of the three optimized designs of an L-bracket loaded at the top of the right side. The designs optimized with a constraint on plastic strains exhibit an initial yield delayed by 42% and 54% in terms of forces, compared to the stiffest design without stress considerations. Performance is improved in comparison to the p-norm approach [Le et al., 2010], in terms of stiffness and strength for a given volume. End-compliance θN fˆp uNp NGP ek e Plastic strains ∑N e=1 ∑k=1 κN Volume Figures Ne ve xe ∑e=1 0.4825×Ne ×ve −1 Max. end-comp. s.t. vol. 1.0979 × 10−4 1.1400 × 100 Max. end-comp. s.t. vol. and plastic strains 8.4328 × 10−5 5.9734 × 10−4 Min. vol. s.t. end-comp. and plastic strains 8.4309 × 10−5 2.1170 × 10−3 −5.1457 × 10−7 6a −3.9286 × 10−4 6b −2.0116 × 10−1 6c Table 3: Results of the topology optimization of an L-bracket with a load at the mid point of the right side. For the same volume and under the same prescribed displacement, constraining the sum of equivalent plastic strains leads to nearly zero plastic strains while compromising the end-compliance by 23%. The minimum volume procedure reaches a volume fraction of 39.45% but exhibits higher plastic strains. The results for the three formulations are presented in Table 3. The observed trends are the same as for the first case: with the same volume fraction, nearly zero plastic strains can be achieved with a compromise of roughly 23% on end-compliance. The minimum volume procedure achieves a volume fraction far below 48.25% but also exhibits slightly higher plastic strains. The optimized layouts are presented in Figure 6. It can be seen again, that rather subtle shape and topological changes are introduced in order to avoid the stress concentrations. This facilitates a relatively minor reduction in stiffness compared to the “stiffest” design. The generated layouts resemble those achieved by other approaches, e.g. by [Allaire and Jouve, 2008], by [Amstutz and Novotny, 2010] and by [James and Waisman, 2014]. For examining the elasto-plastic response, the post-processing described above is repeated. The computed volume fractions are 48.05%, 48.35%, 38.54% and 48.74% for designs #1, #2,#3 and #4, respectively. Design #4 corresponds to the interpretation from [Amstutz and Novotny, 2010], Fig. 8. The loaddisplacement curves at the prescribed DOF for a displacement incrementation up to 0.02 are presented in Figure 7. Again, the responses of the designs obtained with a constraint on plastic strains exhibit a significant delay of the initial yield. Quite remarkably, the magnitude of load at the initial yield is increased by 18 (a) (b) (c) (d) Figure 6: Topology optimization of an L-bracket with a load and a prescribed displacement at the mid right. The presented layouts are obtained after a 0/1 projection as described in the text, for further examination of the elasto-plastic response. (a) maximizing the end-compliance s.t. a volume constraint; (b) maximizing the end-compliance s.t. constraints on volume and on the total sum of equivalent plastic strains; (c) minimizing the volume s.t. constraints on the end-compliance and on the total sum of equivalent plastic strains. (d) The layout based on an image imported from [Amstutz and Novotny, 2010] for comparing the responses. 19 Figure 7: Elasto-plastic response of the three optimized designs of an L-bracket loaded at the middle of the right side. The designs optimized with a constraint on plastic strains exhibit an initial yield delayed by 72% in terms of forces, compared to the stiffest designs without stress considerations. The attained stiffness is slightly lower than in [Amstutz and Novotny, 2010], but the initial yield is postponed. approximately 72% compared to the reference design #1, for designs #2 and #3. The yielding of designs #2 and #3 is postponed also in comparison with the layout taken from [Amstutz and Novotny, 2010]. This highlights the capability of the proposed approach to generate designs with high quality trade-offs of volume-stiffness-strength. 5.2 Example 2: Stress-constrained U-bracket design The second example demonstrates the topological design of a U-bracket with a horizontal load, see Figure 8 for the problem setup. Here, two regions of stress concentrations are expected because the load path should pass via both re-entrant corners. The model is discretized with a 200×100 mesh resolution consisting of 17,500 square, bi-linear elements. The available volume of material is set to 40% of the total volume and the filter radius is 0.03. The prescribed displacement is set to u p = 0.01 and automatic displacement incrementation is applied. The material parameters are the same as for the previous examples. The continuation scheme is slightly modified in order to examine the capability of beginning the optimization with some penalization, namely pE = 3.0 and pσy = 2.5. This is in contrast to initial values of pE = 1.0 and pσy = 0.5 that were used in the previous example. Again, for the minimum volume case the penalty exponents pE = 3.0 and pσy = 2.5 are constant for the first 200 design cycles in order to enable feasibility of the compliance constraint. Otherwise, the continuation scheme and MMA parameters are identical to the previous example. As before, we examine three optimization problems: maximizing the end-compliance s.t. a volume constraint; maximizing the end-compliance s.t. constraints on volume and on the total sum of equivalent plastic strains; and minimizing volume s.t. constraints on end-compliance and on the total sum of equivalent plastic strains. The optimized topologies, the von Mises stress distributions and the distributions of equivalent plastic strains are presented in Figure 9. The end-compliances, the sums of equivalent plastic strains and the volumes are presented in Table 4. It can be clearly seen that the suggested approach generates designs that circumvent stress concentrations in the vicinity of the re-entrant corners. The two latter solutions provide different trade-offs of 20 Figure 8: Problem setup for topology optimization of a U-bracket. The load is distributed over the top 10 nodes in order to avoid artificial stress concentrations at the loading point. End-compliance θN fˆp uNp NGP ek Ne Plastic strains ∑e=1 ∑k=1 κN Volume Figures Ne ve xe ∑e=1 0.40·Ne ·ve −1 Max. end-comp. s.t. vol. 8.6786 · 10−5 5.0729 · 10−1 Max. end-comp. s.t. vol. and plastic strains 6.7559 · 10−5 8.9298 · 10−4 Min. vol. s.t. end-comp. and plastic strains 6.7756 · 10−5 2.9604 · 10−3 −4.4559 · 10−7 9a, 9d, 9g −2.0678 · 10−3 9b, 9e, 9h −9.0041 · 10−2 9c, 9f, 9i Table 4: Results of the topology optimization of an U-bracket with a load at the top right corner. For the same volume and under the same prescribed displacement, constraining the sum of equivalent plastic strains leads to nearly zero plastic strains while compromising the end-compliance by 22%. stiffness to weight while keeping plastic strains at a very low level. It is interesting to see that the maximization of end-compliance subject to a constraint on plastic strains suggests an alternative load path to that of the reference design. Considerable forces are transferred via a vertical bar in the right side edge, enabling the reduction of stresses near the re-entrant corners. This force transfer appears also in the minimum volume design but to a lesser extent. In both designs, the compliance is compromised by roughly 22% in comparison to the reference solution which exhibits two significant stress concentrations. The minimum volume procedure appears to deliver slightly better results as it provides the same compliance, but uses only 36.4% of the design domain. Finally, a comparison of the initial yield level of the optimized designs reveals an increase of approximately 89% in the applied force prior to yielding, compared to the optimized design achieved without stress considerations. This demonstrates the capability of the proposed approach to deal with multiple stress concentrations without any added complexity. 6 Conclusion A new approach for achieving stress-constrained topological designs in continua was presented. The main novelty is in the use of material nonlinearity, in the form of classical elasto-plasticity models, in order to avoid imposing a very large number of local stress constraints. Stress constraints are implicitly satisfied by imposing a single global constraint on the spatial sum of the equivalent plastic strains. Incorporating this constraint into formulations for maximizing the end-compliance (for a given prescribed 21 (a) (b) (c) (d) (e) (f) (g) (h) (i) Figure 9: Topology optimization of a U-bracket with a load and a prescribed displacement at the top right corner. From the left: maximizing the end-compliance s.t. a volume constraint; maximizing the end-compliance s.t. constraints on volume and on the total sum of equivalent plastic strains; minimizing volume s.t. constraints on end-compliance and on the total sum of equivalent plastic strains. From top: optimized layouts; von Mises stress distributions; equivalent plastic strains in the vicinity of the re-entrant corners. Note the two orders of magnitude difference between the scale of sub-figure (g) compared to (h) and (i). 22 displacement) subject to a volume constraint, or vice-versa, leads to optimized designs with practically no stress violations. A critical comparison of the proposed approach to existing techniques, in terms of accuracy and efficiency, is hereby presented. 6.1 Designed trade-off In view of real-world applications, stress-constrained topology optimization should aim at finding the best trade-off between three competing quantities: volume, compliance and stress. The results presented in this article demonstrate the capability of the plasticity-based formulation to attain high-quality layouts in terms of this trade-off. Two types of optimized L-brackets were favorably compared to results obtained with constraint aggregation, using either p-norms or an external penalty. This highlights one advantage of the current formulation–stress constraints are captured accurately without actually imposing a large number of constraints on local stress values. There is an evident computational price tag for the improvement in design quantities, as will be discussed in the following. Nevertheless, the results achieved in this study provide another, unexplored view on stress-constrained topology optimization. This can motivate and fertilize further development of more efficient, approximate techniques. 6.2 Computational cost In its present form, the proposed formulation is not as efficient as linear-elastic approaches. The results reported in this paper were achieved with a constant number of 500 design iterations for enabling continuation on penalty parameters and on the sharpness of the smoothed Heaviside. Each design iteration requires a full nonlinear finite element analysis, which typically uses 20-40 solutions of linear equation systems, depending on the automatic incrementation and on the convergence of Newton-Raphson iterations. The compliance and plastic strains functionals require an adjoint procedure in each design iteration, which typically uses 5-10 solutions of linear equation systems, depending on the automatic incrementation. Stress-constrained procedures based on linear elasticity and a certain constraint aggregation technique will require only a few solutions of linear equation systems per design iteration, depending on the specific aggregation scheme. Therefore, the current implementation of the plasticity-based approach is expected to be slower than existing approaches. This includes p-norm techniques (e.g. [Le et al., 2010]), external penalty (e.g. [Amstutz and Novotny, 2010]) and to some extent also formulations that introduce local constraints (e.g. [Bruggi and Venini, 2008]). Nevertheless, the relatively good design trade-offs encourage further exploration of this approach, focusing on reducing the computational burden. It may be possible to utilize other material models that may not be suitable for capturing the full elasto-plastic response accurately, but may suffice for the purpose of achieving a no-yield design. Another path to be explored is the reduction of displacement increments and Newton-Raphson iterations, by exploiting the fact that in most design cycles the response is either linear-elastic or very close to linear. Then the computational cost of a single design cycle can be reduced to a level similar to that of a linear-elastic analysis. Another option is to utilize reanalysis techniques, which in fact motivated the investigation of this formulation in the first place [Amir, 2011]. These possibilities will be investigated thoroughly in future work. 6.3 Further considerations In many publications on stress constraints, the local constraints are aggregated into a single, or a few, global approximations of the maximum stress. This often requires specific tuning of parameters and does not ensure that the actual local stresses will not exceed allowable values. This highlights another advantage of the proposed approach, which only requires a well-known elasto-plastic material model that is already incorporated in standard FEA packages. In the examples it can be observed that the optimized 23 designs involve both topological as well as shape changes. This means that simply optimizing the shape of an optimized topology generated without stress considerations, may not suffice for achieving the best possible design. Furthermore, incorporating stress considerations in the topological design phase can eliminate the need to post-process the optimized topology, create a CAD model, generate a new mesh and then optimize the shape–a process that can be very time consuming in the industrial context, particularly in 3-D. A potential disadvantage of the proposed formulation lies in the non-differentiability of the plastic strains. According to the adopted formulation of the elasto-plastic model, the yield stress limit represents a non-differentiable point. This does not affect the end-compliance functional because the product of forces and displacements (or stresses and strains) retains smoothness when passing the yield point. Plastic strains however are strictly zero up to this point and instantaneously increase when passing this point. As can be seen in the numerical experiments above, the non-differentiability did not appear to hamper the convergence towards a design which does not violate the stress limit. This is despite the fact that in the optimized designs, several material points are indeed very close to their yield limit. A differentiable approximation was introduced in several numerical experiments in order to examine its affect on the quality of the attained solutions. The results did not show any improvement in comparison to the original implementation, but this important issue will be thoroughly examined in a continuation of the research. 7 Acknowledgements This research was supported by the Israel Science Foundation (grant No. 750/15). 8 Appendix In this appendix we present a numerical verification of the adjoint sensitivity analysis procedure. Implementing this procedure can be a somewhat cumbersome task, so we believe this verification can prove useful for readers who are not well-acquainted with such procedures. Furthermore, accurate and efficient sensitivity analysis for elasto-plastic response is still a rather open issue, as discussed in a recent publication [Kato et al., 2015]. In the following, results of the adjoint computations are compared to numerical derivatives computed by forward finite differences. We consider a small problem of a symmetric clamped beam, where the symmetric half is modeled with a finite element mesh of 2×2 square bi-linear elements. A downwards vertical displacement is prescribed at the top right corner. Two separate loading situations are considered, see Figure 10 for the problem setup: 1) A point load at the top right corner; and 2) A distributed load at the right edge. The first case is easier to implement because the equations for the global adjoint vectors in Eqs. (23) and (25) take a simple form when the force is applied only at the prescribed DOF. However, the second case is much more useful, especially in the particular application considered in this article: It is necessary to distribute the applied load over several adjacent nodes because the numerical solution with a point load will inherently include stress concentrations. The material and optimization parameters are given in Table 5. The density xe in all four elements is set to 0.8. The prescribed displacement of 0.01 is applied within 10 equal increments. Convergence of each increment is assumed when the relative norm of the residual forces is below 10−6 . For the finite difference check, the perturbation value is set to ∆xe = 10−6 . We compare the design sensitivities of two critical quantities in the context of the current application: 1) The end-compliance at the prescribed DOF, gec = −θN fˆp uNp , where the superscript p denotes the prescribed DOF; and 2) The sum of plastic strains NGP ek e in the whole domain at the final equilibrium state, g ps = ∑Ne=1 κN . ∑k=1 The comparisons between the derivatives computed by the adjoint procedure to those obtained by finite differences are presented in Tables 6 and 7 for the point load and distributed load, respectively. 24 Figure 10: Problem setup for verification of the adjoint sensitivity analysis. Left: point load; right: distributed load. Emin Emax ν 0 σy,min 0 σy,max pE pσy H 1.0 · 10−3 1.0 · 103 0.3 0.0 2.0 3.0 3.0 0.01 Table 5: Material and optimization parameters used in the validation of the adjoint sensitivity analysis. 25 Element 1 2 3 4 ∂ gec ∂ xe −1.6222 · 10−4 −1.4139 · 10−4 −1.4260 · 10−2 −2.3066 · 10−4 ∆gec ∆xe −1.6222 · 10−4 −1.4139 · 10−4 −1.4260 · 10−2 −2.3066 · 10−4 Rel. error 1.3347 · 10−6 1.7011 · 10−6 1.0736 · 10−6 2.2037 · 10−6 ∂ g ps ∂ xe 9.0330 · 10−3 1.5416 · 10−2 −3.3840 · 10−2 9.3906 · 10−3 ∆g ps ∆xe 9.0330 · 10−3 1.5416 · 10−2 −3.3840 · 10−2 9.3906 · 10−3 Rel. error 4.8305 · 10−7 9.9510 · 10−7 6.7529 · 10−7 1.2330 · 10−6 Table 6: Verification of the sensitivity analysis, 4 element domain with a point load. Element 1 2 3 4 ∂ gec ∂ xe ∆gec ∆xe −1.8691 · 10−3 −1.8691 · 10−3 −1.8841 · 10−3 −1.9104 · 10−3 −1.8544 · 10−3 −1.8841 · 10−3 −1.9104 · 10−3 −1.8544 · 10−3 Rel. error 1.5426 · 10−6 1.5246 · 10−6 1.8419 · 10−6 1.7176 · 10−6 ∂ g ps ∂ xe ∆g ps ∆xe −1.4077 · 10−2 −1.4076 · 10−2 8.3362 · 10−3 4.4694 · 10−2 −3.8954 · 10−2 8.3364 · 10−3 4.4694 · 10−2 −3.8953 · 10−2 Rel. error 1.1225 · 10−5 1.8334 · 10−5 2.9863 · 10−6 1.1161 · 10−5 Table 7: Verification of the sensitivity analysis, 4 element domain with a distributed load. It can be seen that the design sensitivities are practically identical, thus verifying the derivation and the implementation of the adjoint procedure. The nonlinear response of both test cases is presented in Figure 11, in terms of load-displacement curves at the prescribed DOF and equivalent plastic strain. From the tables it can be seen that even elements that are in the elastic regime contribute to the sum of plastic strains, in two opposite modes–i.e. the addition of material can either increase or decrease the overall plastic strain, whereas it always has a stiffening effect on compliance. Finally, the analysis and sensitivity analysis were repeated with 30 and 50 displacement increments. Practically identical results were obtained for the nonlinear respones as well as their design sensitivities. References [Allaire and Jouve, 2008] Allaire, G. and Jouve, F. (2008). Minimum stress optimal design with the level set method. Engineering analysis with boundary elements, 32(11):909–918. [Amir, 2011] Amir, O. (2011). Efficient Reanalysis Procedures in Structural Topology Optimization. PhD thesis, Technical University of Denmark. [Amir, 2013] Amir, O. (2013). A topology optimization procedure for reinforced concrete structures. Computers & Structures, 114-115:46–58. [Amir and Sigmund, 2013] Amir, O. and Sigmund, O. (2013). Reinforcement layout design for concrete structures based on continuum damage and truss topology optimization. Structural and Multidisciplinary Optimization, 47(2):157–174. [Amstutz and Novotny, 2010] Amstutz, S. and Novotny, A. A. (2010). Topological optimization of structures subject to von mises stress constraints. Structural and Multidisciplinary Optimization, 41(3):407– 420. [Bendsøe, 1989] Bendsøe, M. P. (1989). Optimal shape design as a material distribution problem. Structural optimization, 1(4):193–202. [Bendsøe and Kikuchi, 1988] Bendsøe, M. P. and Kikuchi, N. (1988). Generating optimal topologies in structural design using a homogenization method. Computer methods in applied mechanics and engineering, 71(2):197–224. 26 0 -0.1 -0.05 -0.2 -0.1 3 3 0 -0.3 -0.15 -0.4 -0.2 -0.5 -0.01 -0.008 -0.006 -0.004 -0.002 -0.25 -0.01 -0.008 -0.006 -0.004 -0.002 0 up 0 up (a) (b) #10-3 #10-3 5 16 14 4.5 12 4 10 8 3.5 6 3 4 2 2.5 0 2 -2 (c) (d) Figure 11: Nonlinear response of the test cases used for verification of the sensitivity analysis. (a) Load factor vs. prescribed displacement, point load; (b) Load factor vs. prescribed displacement, distributed load; (c) Equivalent plastic strains, point load; (d) Equivalent plastic strains, distributed load. 27 [Bendsøe and Sigmund, 2003] Bendsøe, M. P. and Sigmund, O. (2003). Theory, Methods and Applications. Springer, Berlin. Topology Optimization - [Bogomolny and Amir, 2012] Bogomolny, M. and Amir, O. (2012). Conceptual design of reinforced concrete structures using topology optimization with elastoplastic material modeling. International Journal for Numerical Methods in Engineering, 90(13):1578–1597. [Bourdin, 2001] Bourdin, B. (2001). Filters in topology optimization. Numerical Methods in Engineering, 50:2143–2158. International Journal for [Bruggi, 2008] Bruggi, M. (2008). On an alternative approach to stress constraints relaxation in topology optimization. Structural and multidisciplinary optimization, 36(2):125–141. [Bruggi and Duysinx, 2012] Bruggi, M. and Duysinx, P. (2012). Topology optimization for minimum weight with compliance and stress constraints. Structural and Multidisciplinary Optimization, 46(3):369–384. [Bruggi and Venini, 2008] Bruggi, M. and Venini, P. (2008). A mixed fem approach to stress-constrained topology optimization. International journal for numerical methods in engineering, 73(12):1693–1714. [Bruns and Tortorelli, 2001] Bruns, T. E. and Tortorelli, D. A. (2001). Topology optimization of nonlinear elastic structures and compliant mechanisms. Computer Methods in Applied Mechanics and Engineering, 190:3443–3459. [Cheng and Guo, 1997] Cheng, G. and Guo, X. (1997). ε-relaxed approach in structural topology optimization. Structural Optimization, 13(4):258–266. [Cheng and Jiang, 1992] Cheng, G. and Jiang, Z. (1992). Study on topology optimization with stress constraints. Engineering Optimization, 20(2):129–148. [Deaton and Grandhi, 2014] Deaton, J. and Grandhi, R. (2014). A survey of structural and multidisciplinary continuum topology optimization: post 2000. Structural and Multidisciplinary Optimization, 49(1):1–38. [Duysinx and Bendsøe, 1998] Duysinx, P. and Bendsøe, M. P. (1998). Topology optimization of continuum structures with local stress constraints. International Journal for Numerical Methods in Engineering, 43:1453–1478. [Duysinx and Sigmund, 1998] Duysinx, P. and Sigmund, O. (1998). New developments in handling stress constraints in optimal material distribution. In Proceedings of 7th AIAA/USAF/NASA/ISSMO symposium on multidisciplinary design optimization, AIAA, Saint Louis, Missouri, AIAA Paper, pages 98–4906. [Eschenauer and Olhoff, 2001] Eschenauer, H. A. and Olhoff, N. (2001). Topology optimization of continuum structures: a review. Applied Mechanics Reviews, 54(4):331–389. [Fancello, 2006] Fancello, E. (2006). Topology optimization for minimum mass design considering local failure constraints and contact boundary conditions. Structural and Multidisciplinary Optimization, 32(3):229–240. [Fritzen et al., 2015] Fritzen, F., Xia, L., Leuschner, M., and Breitkopf, P. (2015). Topology optimization of multiscale elastoviscoplastic structures. International Journal for Numerical Methods in Engineering. 28 [Guest et al., 2004] Guest, J. K., Prévost, J. H., and Belytschko, T. (2004). Achieving minimum length scale in topology optimization using nodal design variables and projection functions. International Journal for Numerical Methods in Engineering, 61:238–254. [James and Waisman, 2014] James, K. A. and Waisman, H. (2014). Failure mitigation in optimal topology design using a coupled nonlinear continuum damage model. Computer Methods in Applied Mechanics and Engineering, 268:614–631. [James and Waisman, 2015] James, K. A. and Waisman, H. (2015). Topology optimization of viscoelastic structures using a time-dependent adjoint method. Computer Methods in Applied Mechanics and Engineering, 285:166–187. [Kato et al., 2015] Kato, J., Hoshiba, H., Takase, S., Terada, K., and Kyoya, T. (2015). Analytical sensitivity in topology optimization for elastoplastic composites. Structural and Multidisciplinary Optimization, pages 1–20. [Kirsch, 1990] Kirsch, U. (1990). On singular topologies in optimum structural design. Structural Optimization, 2(3):133–142. [Le et al., 2010] Le, C., Norato, J., Bruns, T., Ha, C., and Tortorelli, D. (2010). Stress-based topology optimization for continua. Structural and Multidisciplinary Optimization, 41:605–620. [Maute et al., 1998] Maute, K., Schwarz, S., and Ramm, E. (1998). Adaptive topology optimization of elastoplastic structures. Structural Optimization, 15(2):81–91. [Michaleris et al., 1994] Michaleris, P., Tortorelli, D. A., and Vidal, C. A. (1994). Tangent operators and design sensitivity formulations for transient non-linear coupled problems with applications to elastoplasticity. International Journal for Numerical Methods in Engineering, 37(14):2471–2499. [Nakshatrala and Tortorelli, 2015] Nakshatrala, P. and Tortorelli, D. (2015). Topology optimization for effective energy propagation in rate-independent elastoplastic material systems. Computer Methods in Applied Mechanics and Engineering, 295:305 – 326. [Parı́s et al., 2007] Parı́s, J., Navarrina, F., Colominas, I., and Casteleiro, M. (2007). Block aggregation of stress constraints in topology optimization of structures. In Hernández, S. and Brebbia, C. A., editors, Computer Aided Optimum Design of Structures X. [Parı́s et al., 2010] Parı́s, J., Navarrina, F., Colominas, I., and Casteleiro, M. (2010). Block aggregation of stress constraints in topology optimization of structures. Advances in Engineering Software, 41(3):433–441. [Park, 1995] Park, Y. K. (1995). Extensions of optimal layout design using the homogenization method. PhD thesis, University of Michigan, Ann Arbor. [Pereira et al., 2004] Pereira, J. T., Fancello, E. A., and Barcellos, C. S. (2004). Topology optimization of continuum structures with material failure constraints. Structural and Multidisciplinary Optimization, 26(1-2):50–66. [Rozvany, 1996] Rozvany, G. (1996). Difficulties in truss topology optimization with stress, local buckling and system stability constraints. Structural Optimization, 11(3-4):213–217. [Schwarz et al., 2001] Schwarz, S., Maute, K., and Ramm, E. (2001). Topology and shape optimization for elastoplastic structural response. Computer Methods in Applied Mechanics and Engineering, 190(15):2135–2155. 29 [Sigmund and Maute, 2013] Sigmund, O. and Maute, K. (2013). Topology optimization approaches. Structural and Multidisciplinary Optimization, 48(6):1031–1055. [Sigmund and Torquato, 1997] Sigmund, O. and Torquato, S. (1997). Design of materials with extreme thermal expansion using a three-phase topology optimization method. Journal of the Mechanics and Physics of Solids, 45(6):1037–1067. [Simo and Taylor, 1986] Simo, J. and Taylor, R. (1986). A return mapping algorithm for plane stress elastoplasticity. International Journal for Numerical Methods in Engineering, 22:649–670. [Simo and Hughes, 2006] Simo, J. C. and Hughes, T. J. (2006). Computational inelasticity, volume 7. Springer Science & Business Media. [Stolpe and Svanberg, 2001] Stolpe, M. and Svanberg, K. (2001). On the trajectories of the epsilon-relaxation approach for stress-constrained truss topology optimization. Structural and multidisciplinary optimization, 21(2):140–151. [Svanberg, 1987] Svanberg, K. (1987). The method of moving asymptotes - a new method for structural optimization. International Journal for Numerical Methods in Engineering, 24:359–373. [Sved and Ginos, 1968] Sved, G. and Ginos, Z. (1968). Structural optimization under multiple loading. International Journal of Mechanical Sciences, 10(10):803–805. [Swan and Kosaka, 1997] Swan, C. C. and Kosaka, I. (1997). Voigt-reuss topology optimization for structures with nonlinear material behaviors. International Journal for Numerical Methods in Engineering, 40(20):3785–3814. [Verbart et al., 2013] Verbart, A., Langelaar, M., and van Keulen, F. (2013). A new approach for stressbased topology optimization: Internal stress penalization. In 10th World Congress on Structural and Multidisciplinary Optimization, Orlando, Florida, USA. [von Mises, 1928] von Mises, R. (1928). Mechanics of the ductile form changes of crystals. Zeitschrift für Angewandte Mathematik und Mechanik, 8:161–185. [Xu et al., 2010] Xu, S., Cai, Y., and Cheng, G. (2010). Volume preserving nonlinear density filter based on heaviside functions. Structural and Multidisciplinary Optimization, 41(4):495–505. [Yang and Chen, 1996] Yang, R. and Chen, C. (1996). Stress-based topology optimization. Structural Optimization, 12(2-3):98–105. [Yoon and Kim, 2007] Yoon, G. H. and Kim, Y. Y. (2007). Topology optimization of material-nonlinear continuum structures by the element connectivity parameterization. International journal for numerical methods in engineering, 69(10):2196–2218. [Yuge and Kikuchi, 1995] Yuge, K. and Kikuchi, N. (1995). Optimization of a frame structure subjected to a plastic deformation. Structural optimization, 10(3-4):197–208. [Zienkiewicz and Taylor, 2000] Zienkiewicz, O. C. and Taylor, R. L. (2000). The finite element method: Solid mechanics, volume 2. Butterworth-heinemann. 30
5cs.CE
arXiv:1608.08815v2 [math.GR] 17 Mar 2017 Invariant forms on irreducible modules of simple algebraic groups Mikko Korhonen∗†‡ February 1, 2018 Abstract Let G be a simple linear algebraic group over an algebraically closed field K of characteristic p ≥ 0 and let V be an irreducible rational Gmodule with highest weight λ. When V is self-dual, a basic question to ask is whether V has a non-degenerate G-invariant alternating bilinear form or a non-degenerate G-invariant quadratic form. If p 6= 2, the answer is well known and easily described in terms of λ. In the case where p = 2, we know that if V is self-dual, it always has a non-degenerate G-invariant alternating bilinear form. However, determining when V has a non-degenerate G-invariant quadratic form is a classical problem that still remains open. We solve the problem in the case where G is of classical type and λ is a fundamental highest weight ωi , and in the case where G is of type Al and λ = ωr + ωs for 1 ≤ r < s ≤ l. We also give a solution in some specific cases when G is of exceptional type. As an application of our results, we refine Seitz’s 1987 description of maximal subgroups of simple algebraic groups of classical type. One consequence of this is the following result. If X < Y < SL(V ) are simple algebraic groups and V ↓ X is irreducible, then one of the following holds: (1) V ↓ Y is not self-dual; (2) both or neither of the modules V ↓ Y and V ↓ X have a non-degenerate invariant quadratic form; (3) p = 2, X = SO(V ), and Y = Sp(V ). Section de mathématiques, École Polytechnique Fédérale de Lausanne, CH-1015 Lausanne, Switzerland † Email address: [email protected] ‡ The author was supported by a grant from the Swiss National Science Foundation (grant number 200021_146223). ∗ 1 Contents 1 Introduction 3 2 Invariant forms on irreducible G-modules 6 3 Fundamental representations for type Cl 9 3.1 Representation theory . . . . . . . . . . . . . . . . . . . . . . 10 3.2 Construction of V (ωr ) . . . . . . . . . . . . . . . . . . . . . . 13 3.3 Computation of a quadratic form Q on V (ωr ) . . . . . . . . . 18 4 Fundamental irreducibles for classical types 19 5 Representations L(ωr + ωs ) for type Al 21 5.1 Representation theory . . . . . . . . . . . . . . . . . . . . . . 21 5.2 Construction of V (ωr + ωn−r ) . . . . . . . . . . . . . . . . . . 23 5.3 Computation of a quadratic form Q on V (ωr + ωn−r ) . . . . . 31 6 Simple groups of exceptional type 31 7 Applications and further work 7.1 Connection with representations of the symmetric group 7.2 Reduction for Problem 1.1 . . . . . . . . . . . . . . . . . 7.3 Application to maximal subgroups of classical groups . . 7.4 Fundamental self-dual irreducible representations . . . . 7.5 Fixed point spaces of unipotent elements . . . . . . . . . 34 35 36 37 40 41 2 . . . . . . . . . . . . . . . 1 Introduction Let V be a finite-dimensional vector space over an algebraically closed field K of characteristic p ≥ 0. A fundamental problem in the study of simple linear algebraic groups over K is the determination of maximal closed connected subgroups of simple groups of classical type (SL(V ), Sp(V ) and SO(V )). Seitz [Sei87] has shown that up to a known list of examples, these are given by the images of p-restricted, tensor-indecomposable irreducible rational representations ϕ : G → GL(V ) of simple algebraic groups G over K. Then given such an irreducible representation ϕ, one should still determine which of the groups SL(V ), Sp(V ) and SO(V ) contain ϕ(G). In most cases the answer is known. • If V is not self-dual, then ϕ(G) is only contained in SL(V ). Furthermore, we know when V is self-dual (see Section 2). • If p 6= 2 and V is self-dual, then ϕ(G) is contained in Sp(V ) or SO(V ), but not both [Ste68, Lemma 78, Lemma 79]. Furthermore, we know for which irreducible representations the image is contained in Sp(V ) and for which the image is contained in SO(V ) (see Section 2). • If p = 2 and V is self-dual, then ϕ(G) is contained in Sp(V ) [Fon74, Lemma 1]. Currently what is still missing is a method for determining in characteristic two when exactly ϕ(G) is contained in SO(V ). This problem is the main subject of this paper, and we can state it equivalently as follows. Problem 1.1. Assume that p = 2 and let LG (λ) be an irreducible G-module with highest weight λ. When does LG (λ) have a non-degenerate G-invariant quadratic form? This is a nontrivial open problem. There is some literature on the subject [Wil77], [SW91], [GW95], [GW97], [GN16], but currently only partial results are known. The main result of this paper is a solution to Problem 1.1 in the following cases: • G is of classical type (Al , Bl , Cl or Dl ) and λ is a fundamental dominant weight ωr for some 1 ≤ r ≤ l (Theorem 4.2). • G is of type Al and λ = ωr + ωs for 1 ≤ r < s ≤ l (Theorem 5.1). 3 In the case where G is of exceptional type, we will give some partial results in Section 6. For G of type G2 and F4 , we are able to give a complete solution (Proposition 6.1, Proposition 6.3). For types E6 , E7 , and E8 , we give the answer for some specific λ (Table 6.1). In the final section of this paper, we will give various applications of our results and describe some open problems motivated by Problem 1.1. One particular application, given in subsection 7.3, is a refinement of Seitz’s [Sei87] description of maximal subgroups of simple algebraic groups of classical type. In [Sei87], Seitz gives a full list of all non-maximal irreducible subgroups of SL(V ), but the question of which classical groups contain the image of an irreducible representation is not considered. For example, it is possible that we have a proper inclusion X < Y of irreducible subgroups of SL(V ) such that X is a maximal subgroup of SO(V ). In subsection 7.3, we go through the list given by Seitz and describe when exactly such inclusions occur. In particular, our results have the consequence (Theorem 7.9) that if X < Y < SL(V ) are simple algebraic groups and V ↓ X is irreducible, then one of the following holds: (i) The module V ↓ Y is not self-dual; (ii) Both V ↓ X and V ↓ Y have an invariant quadratic form; (iii) Neither of V ↓ X or V ↓ Y has an invariant quadratic form; (iv) p = 2, X = SO(V ) and Y = Sp(V ). The general approach for the proofs of our main results is as follows. A basic method used throughout is Theorem 9.5. from [GN16] (recorded here in Proposition 2.2), which allows one to determine whether LG (λ) is orthogonal (when p = 2) by computing within the Weyl module VG (λ). For G of classical type and V irreducible with fundamental highest weight, we will first prove our result in the case where G is of type Cl (Proposition 3.1). From this the result for other classical types is a fairly straightforward consequence (Theorem 4.2). In the case where G is of type Cl and λ = ωr , and in the case where G is of type Al and λ = ωr + ωs , the proofs of our results are heavily based on various results from the literature on the representation theory of G. We will use results about the submodule structure of the Weyl module VG (λ) found in [PS83], [Ada84] and [Ada86]. We will also need the first cohomology groups of LG (λ) which were computed in [KS99] and [KS01, Corollary 3.6]. One more key ingredient in our proof will be the results of Baranov and Suprunenko in [BS00] and [BS05], which give the structure of the restrictions of LG (λ) to certain subgroups defined in terms of the natural module of G. 4 Notation and terminology We fix the following notation and terminology. Throughout the whole text, let K be an algebraically closed field of characteristic p ≥ 0. All groups that we consider are linear algebraic groups over K, and by a subgroup we always mean a closed subgroup. All modules and representations will be finite-dimensional and rational. Unless otherwise mentioned, G denotes a simply connected simple algebraic group over K with l = rank G, and V will be a finite-dimensional vector space over K. Throughout we will view G as its group of rational points over K, and most of the time G will studied either as a Chevalley group constructed with the usual Chevalley construction (see e.g. [Ste68]), or as a classical group with its natural module (i.e. G = SL(V ), G = Sp(V ) or G = SO(V )). We will occasionally denote G by its type, so notation such as G = Cl means that G is a simply connected simple algebraic group of type Cl . We fix the following notation, as in [Jan03]. • T : a maximal torus of G, with character group X(T ). • X(T )+ : the set of dominant weights for G, with respect to some system of positive roots. • ch V : the character of a G-module V . Here ch V is an element of Z[X(T )]. • ω1 , ω2 , . . . , ωl : the fundamental dominant weights in X(T )+ . We use the standard Bourbaki labeling of the simple roots, as given in [Hum72, 11.4, pg. 58]. • L(λ), LG (λ) : the irreducible G-module with highest weight λ ∈ X(T )+ . • V (λ), VG (λ) : the Weyl module for G with highest weight λ ∈ X(T )+ . • rad V (λ) : unique maximal submodule of V (λ). P For a dominant weight λ ∈ X(T )+ , we can write λ = li=1 mi ωi where mi ∈ Z≥0 . We say that λ is p-restricted if p = 0, or if p > 0 and 0 ≤ mi ≤ p−1 for all 1 ≤ i ≤ l. The irreducible representation LG (λ) is said to be prestricted if λ is p-restricted. A bilinear form b is non-degenerate, if its radical rad b = {v ∈ V : b(v, w) = 0 for all w ∈ V } is zero. For a quadratic form Q : V → K on a vector space V , its polarization is the bilinear form bQ defined by 5 bQ (v, w) = Q(v + w) − Q(v) − Q(w) for all v, w ∈ V . We say that Q is non-degenerate, if its radical rad Q = {v ∈ rad bQ : Q(v) = 0} is zero. For a KG-module V , a bilinear form (−, −) is G-invariant if (gv, gw) = (v, w) for all g ∈ G and v, w ∈ V . A quadratic form Q : V → K is G-invariant if Q(gv) = Q(v) for all g ∈ G and v ∈ V . We say that V is symplectic if it has a non-degenerate G-invariant alternating bilinear form, and we say that V is orthogonal if it has a non-degenerate G-invariant quadratic form. Note that if V has a G-invariant bilinear form, then for λ, µ ∈ X(T ) the weight spaces Vλ and Vµ are orthogonal if λ 6= −µ. Thus to compute the form on V it is enough to work in the zero weight space of V and Vλ ⊕ V−λ for nonzero λ ∈ X(T ). For a G-invariant quadratic form Q on V , we have Q(v) = 0 for any weight vector v ∈ V with non-zero weight. Given a morphism φ : G′ → G of algebraic groups, we can twist representations of G with φ. That is, if ρ : G → GL(V ) is a representation of G, then ρφ is a representation of G′ . We denote the corresponding G′ -module by V φ . When p > 0, we denote by F : G → G the Frobenius endomorphism induced by the field automorphism x 7→ xp of K, see for example [Ste68, Lemma 76]. When G is simply connected and λ ∈ X(T )+ , we have LG (pλ) ∼ = LG (λ)F . If a representation V of G has composition series V = V1 ⊃ V2 ⊃ · · · ⊃ Vt ⊃ Vt+1 = 0 with composition factors Wi ∼ = Vi /Vi+1 , we will occasionally denote this by V = W1 /W2 / · · · /Wt . Acknowledgements I am very grateful to Prof. Donna Testerman for suggesting the problem, and for her many helpful suggestions and comments on the earlier versions of this text. I would also like to thank Prof. Gary Seitz for providing the argument used in the proof of Lemma 4.1, and the two anonymous referees for their helpful comments for improvements. 2 Invariant forms on irreducible G-modules Let L(λ) be an irreducible representation of a simple algebraic group G with P P highest weight λ = li=1 mi ωi . Write d(λ) = α>0 hλ, α∨i, where the sum runs over the positive roots α, where α∨ is the coroot corresponding to α, and h , i is the usual dual pairing between X(T ) and the cocharacter group. We know that L(λ) is self-dual if and only if w0 (λ) = −λ, where w0 is the longest element in the Weyl group [Ste68, Lemma 78]. Furthermore, if L(λ) is self-dual and p 6= 2, then L(λ) is orthogonal if d(λ) is even and 6 Root system When is λ = −w0 (λ)? d(λ) mod 2 when λ = −w0 (λ) Al (l ≥ 1) iff mi = ml−i+1 for all i 0, when l is even l+1 · m l+1 , when l is odd 2 2 Bl (l ≥ 2) always 0, when l ≡ 0, 3 mod 4 ml , when l ≡ 1, 2 mod 4. Cl (l ≥ 2) always m1 + m3 + m5 + · · · Dl (l ≥ 4) l even: always l odd: iff ml = ml−1 0, when l 6≡ 2 mod 4 ml + ml−1 , when l ≡ 2 mod 4. G2 F4 E6 E7 E8 always always iff m1 = m6 and m3 = m5 always always 0 0 0 m2 + m5 + m7 0 Table 2.1: Values of d(λ) modulo 2 for a weight λ = Pl i=1 mi ωi symplectic if d(λ) is odd [Ste68, Lemma 79]. Hence in characteristic p 6= 2 deciding whether an irreducible module is symplectic or orthogonal is a straightforward computation with roots and weights. In Table 2.1, we give the value of d(λ) mod 2 (when λ = −w0 (λ)) for each simple type, in terms of the coefficients mi . In characteristic 2, it turns out that each nontrivial, irreducible self-dual module is symplectic, as shown by the following lemma found in [Fon74]. We include a proof for convenience. Lemma 2.1. Assume that char K = 2. Let V be a nontrivial, irreducible self-dual representation of a group G. Then V is symplectic for G. Proof. [Fon74] Since V is self-dual, there exists an isomorphism ϕ : V → V ∗ of G-modules, which induces a non-degenerate G-invariant bilinear form (−, −) defined by (v, w) = ϕ(v)(w). Since ϕt : V → V ∗ defined by ϕt (v)(w) = ϕ(w)(v) is also an isomorphism of G-modules, by Schur’s lemma there exists a scalar c such that (v, w) = c(w, v) for all v, w ∈ V . Then (v, w) = c2 (v, w), so c2 = 1 because (−, −) is nonzero. Because we are in characteristic two, it follows that c = 1, so (−, −) is a symmetric form. Now {v ∈ V : (v, v) = 0} 7 is a submodule of G. Because V is nontrivial and irreducible, this submodule must be all of V and so (−, −) is alternating. Lemma 2.1 above shows that the image of any irreducible self-dual representation lies in Sp(V ). The following general result reduces determining whether L(λ) is orthogonal (in characteristic two) to a computation within the Weyl module V (λ). Proposition 2.2. Assume that char K = 2. Let λ ∈ X(T )+ be nonzero, λ = −w0 (λ) and suppose that λ 6= ω1 if G has type Cl . Then (i) The Weyl module V (λ) has a nonzero G-invariant quadratic form Q, unique up to scalar. (ii) The unique maximal submodule of V (λ) is equal to rad bQ . (iii) The irreducible module L(λ) has a nonzero, G-invariant quadratic form if and only if rad Q = rad bQ . If this is not the case, then rad Q is a submodule of rad bQ with codimension 1, and H1 (G, L(λ)) 6= 0. (iv) If V (λ) has no trivial composition factor, then L(λ) is orthogonal. Proof. See Theorem 9.5. and Proposition 10.1. in [GN16] for (i), (ii) and (iii). The claim in (iii) about H1 (G, L(λ)) can also be deduced from [Wil77, Satz 2.5]. The claim (iv) is a consequence of (iii), since H1 (G, L(λ)) ∼ = Hom (rad V (λ), K) by [Jan03, II.2.14]. Ext1G (K, L(λ)) ∼ = G In the case where G is of type Cl and λ = ω1 , we have the following result which is well known. We include a proof for completeness. Proposition 2.3. Assume that char K = 2 and that G is of type Cl . Then V = V (ω1 ) = L(ω1 ) has no nonzero G-invariant quadratic form. Proof. ([GN16, Example 8.4]) The claim follows from a more general result that any G-invariant rational map f : V → K is constant. Indeed, for such f we have f (gv) = f (v) for all g ∈ G, v ∈ V . Because G acts transitively on nonzero vectors in V , it follows that f (w) = f (v) for all w ∈ V − {0}. Thus f (w) = f (v) for all w ∈ V since f is rational. Lemma 2.4. Let V and W be G-modules. If V and W are both symplectic for G, then V ⊗ W is orthogonal for G. Proof. See [SW91, Proposition 3.4], [GN16, Proposition 9.2], or [KL90, 4.4, pg. 126-127]. 8 Remark 2.5. Assume that char K = 2. Then lemmas 2.1 and 2.4 show that if V is a non-orthogonal irreducible G-module, then V must be tensor indecomposable. By Steinberg’s tensor product theorem, this implies that V is a Frobenius twist of LG (λ) for some 2-restricted weight λ ∈ X(T )+ . Therefore to determine which irreducible representations of G are orthogonal, it suffices to consider V = LG (λ) with λ ∈ X(T )+ a 2-restricted dominant weight. 3 Fundamental representations for type Cl Throughout this section, assume that G is simply connected of type Cl , l ≥ 2. In this section we determine when in characteristic 2 a fundamental irreducible representation L(ωr ), 1 ≤ r ≤ l, of G has a nonzero G-invariant quadratic form. The answer is given by the following proposition, which we will prove in what follows. Proposition 3.1. Assume char K = 2. Let 1 ≤ r ≤ l. Then L(ωr ) is not orthogonal if and only if r = 1, or r = 2i+1 for some i ≥ 0 and l + 1 ≡ 2i+1 + 2i + t mod 2i+2 , where 0 ≤ t < 2i . The following examples are immediate consequences of Proposition 3.1. Example 3.2. If char K = 2, then L(ω2 ) is orthogonal if and only if l 6≡ 2 mod 4. Example 3.3. If char K = 2, then L(ω4 ) is orthogonal if and only if l 6≡ 5, 6 mod 8. Example 3.4. If char K = 2, then L(ωl ) is orthogonal if and only if l ≥ 3 (this was also proven in [Gow97, Corollary 4.3]) and L(ωl−1 ) is orthogonal if and only if l = 3, l = 4 or l ≥ 6. A rough outline for the proof of Proposition 3.1 is as follows. Various results from the literature about the representation theory of G will reduce the claim to specific r which must be considered. We will then study V (ωr ) by using a standard realization of it in the exterior algebra of the natural module V of G. Here we can explicitly describe a nonzero G-invariant quadratic form Q on V (ωr ). We will then find a vector γ ∈ rad V (ωr ) such that L(ωr ) is orthogonal if and only if Q(γ) = 0. The proof is finished by computing Q(γ). 9 3.1 Representation theory The composition factors of V (ωr ) were determined in odd characteristic by Premet and Suprunenko in [PS83, Theorem 2]. Independently, the composition factors and the submodule structure of V (ωr ) were found in arbitrary characteristic by Adamovich in [Ada84], [Ada86]. Using the results of Adamovich, it was shown in [BS00, Corollary 2.9] that the result of Premet and Suprunenko also holds in characteristic two. To state the result about composition factors of PV (ωr ),i we need to Pmake ai few definitions first. Let a, b ∈ Z≥0 and write a = i≥0 ai p and b = i≥0 bi p for the expansions of a and b in base p. We say that a contains b to base p if for all i ≥ 0 we have bi = ai or bi = 0. 1 For r ≥ 1, we define Jp (r) to be the set of integers 0 ≤ j ≤ r such that j ≡ r mod 2 and l + 1 − j contains r−j to base p. The main result of [PS83], also valid in characteristic 2, can 2 be then described as follows. Here we set ω0 = 0, so that L(ω0 ) is the trivial irreducible module. Theorem 3.5. Let 1 ≤ r ≤ l. Then in the Weyl module V (ωr ), each composition factor has multiplicity 1, and the set of composition factors is {L(ωj ) : j ∈ Jp (r)}. In view of Proposition 2.2 (iii), it will also be useful to know when the first cohomology group H1 (G, L(ωr )) is nonzero. This has been determined by Kleshchev and Sheth in [KS99] [KS01, Corollary 3.6]. P Theorem 3.6. Let 1 ≤ r ≤ l and write l + 1 − r = i≥0 ai pi in base p. Then H1 (G, L(ωr )) 6= 0 if and only if r = 2(p − ai )pi for some i such that ai > 0, and either ai+1 < p − 1 or r < 2pi+1 . In characteristic 2, the result becomes the following. Corollary 3.7. Assume that char K = 2. Let 1 ≤ r ≤ l. Then H1 (G, L(ωr )) 6= 0 if and only if r = 2i+1 for some i ≥ 0, and l + 1 ≡ 2i + t mod 2i+1 for some 0 ≤ t < 2i . Throughout this section we will consider subgroups Cl′ < Cl = G, which are embedded into G as follows. Consider G = Sp(V ) and let (−, −) be the non-degenerate G-invariant alternating form (−, −) on V . Fix a symplectic basis e1 , . . . , el , e−1 , . . . , e−l of V , where (ei , e−i ) = 1 = −(e−i , ei ) and (ei , ej ) = 0 for i 6= −j. Then for 2 ≤ l′ < l, the embedding Cl′ < Cl is Note that in [PS83] there is a typo, the definition on pg. 1313, line 9 should say “for every i = 0, 1, . . . , n . . .” 1 10 Sp(V ′ ) < Sp(V ), where V ′ ⊆ V has basis e±1 , . . . , e±l′ and Sp(V ′ ) fixes the basis vectors e±(l′ +1) , . . . , e±l . The module structure of the restrictions L(ωr ) ↓ Cl−1 have been determined by Baranov and Suprunenko in [BS00, Theorem 1.1 (i)]. We will only need to know the composition factors which occur in such a restriction, and in this case the result is the following. Below we define LCl−1 (ωr ) = 0 for r < 0. Theorem 3.8. Let 1 ≤ r ≤ l and assume that l ≥ 3. Set d = νp (l + 1 − r), and ε = 0 if l + 1 − r ≡ −pd mod pd+1 and ε = 1 otherwise. Then the character of LCl (ωr ) ↓ Cl−1 is given by ! d−1 X ch LCl−1 (ωr )+2 ch LCl−1 (ωr−1 )+ 2 ch LCl−1 (ωr−2pk ) +ε ch LCl−1 (ωr−2pd ) k=0 where the sum in the brackets is zero if d = 0. Above νp denotes the p-adic valuation on Z, so for a ∈ Z+ we have νp (a) = d, where d ≥ 0 is maximal such that pd divides a. Note that if d = ν2 (l+1−r), then l+1−r ≡ 2d ≡ −2d mod 2d+1 . Therefore if char K = 2, we always have ε = 0 in Theorem 3.8. In particular, the composition factors occurring in L(ωr ) ↓ Cl−1 are LCl−1 (ωr ) and LCl−1 (ωr−2k ) for 0 ≤ k ≤ d. We will now give some applications of Theorem 3.8 and Theorem 3.5 in characteristic two, which will be needed in our proof of Proposition 3.1. Lemma 3.9. Assume that char K = 2, and let l ≥ 2i+1 , where i ≥ 0. Suppose that l + 1 ≡ 2i + t mod 2i+1 , where 0 ≤ t < 2i . Then for t + 1 ≤ j ≤ 2i+1 , the following hold: (i) All composition factors of the restriction L(ωj ) ↓ Cl−1 have the form L(ωj ′ ) for some l − 1 ≥ j ′ ≥ t. (ii) LCl (ωj ) ↓ Cl−t has no trivial composition factors. Proof. If t = 0 there is nothing to prove, so suppose that t ≥ 1. It will be enough to prove (i) as then (ii) will follow by induction on t. Let d = ν2 (l + 1 − j). Suppose first that 0 ≤ d < i + 1. Now l + 1 − j ≡ t − j mod 2i , so then ν2 (l + 1 − j) = ν2 (j − t). By Theorem 3.8, the composition factors occurring in L(ωj ) ↓ Cl−1 are L(ωj ) and L(ωj−2k ) for 0 ≤ k ≤ d, so the claim follows since ν2 (j − t) = d and thus j − 2d ≥ t. Consider then the case where d ≥ i + 1. Then l + 1 − j ≡ 2i + (t − j) ≡ 0 mod 2i+1 , so j − t ≡ 2i mod 2i+1 . On the other hand 0 < j − t < 2i+1 , so j −t = 2i . By Theorem 3.8, the composition factors occurring in L(ωj ) ↓ Cl−1 are L(ωj ) and L(ωj−2k ) for 0 ≤ k ≤ i (because j − 2k < 0 for i + 1 ≤ k ≤ d), so again the claim follows. 11 Lemma 3.10. Let x ≥ 2i+1 , where i ≥ 0. Suppose that x ≡ 2i mod 2i+1 . If 0 ≤ k ≤ 2i and x − 2k contains 2i − k to base 2, then k = 0 or k = 2i . Proof. If i = 0 there is nothing to do, so suppose that i > 0. Replacing k by 2i − k, we see that it is equivalent to prove that if x + 2k contains k to base 2, then k = 0 or k = 2i . Suppose that 0 ≤ k < 2i and that x + 2k contains k to base 2. Consider first the case where 0 ≤ k < 2i−1 . Here since x + 2k ≡ 2k mod 2i , we have that 2k contains k to base 2, which can only happen if k = 0. Consider then 2i−1 ≤ k < 2i and write k = 2i−1 + k ′ , where 0 ≤ k ′ < 2i−1 . Then x + 2k ≡ 2k ′ mod 2i , so 2k ′ contains k = 2i−1 + k ′ to base 2. But then 2k ′ must also contain k ′ to base 2, so k ′ = 0 and k = 2i−1 . In this case x + 2k ≡ 2i + 2i ≡ 0 mod 2i+1 , so x + 2k does not contain k to base 2, contradiction. Lemma 3.11. Let x ≥ 2i+1 , where i ≥ 0. Suppose that x ≡ 2i + t mod 2i+1 , where 0 ≤ t < 2i . If 0 ≤ 2j ≤ t and x − 2j contains 2i − j to base 2, then j = 0. Proof. We prove the claim by induction on i. If i = 0 or i = 1, then the claim is immediate since 0 ≤ 2j ≤ t < 2. Suppose then that i > 1. Assume that 0 < 2j ≤ t and that x − 2j contains 2i − j to base 2. Now 2j ≤ t < 2i , so 0 < j < 2i−1 . Therefore 2i−1 must occur in the binary expansion of 2i − j = 2i−1 + (2i−1 − j), so by our assumption 2i−1 occurs in the binary expansion of x − 2j. Note that this also means that x − 2j contains 2i−1 − j to base 2. Now x − 2j ≡ 2i + (t − 2j) mod 2i+1 and 0 ≤ t − 2j < 2i , so it follows that 2i−1 will occur in the binary expansion of t − 2j. Write t = 2i−1 + t′ , where 0 ≤ t′ < 2i−1 . Here t′ ≥ 2j because t − 2j ≥ 2i−1 . Finally, since x − 2j contains 2i−1 − j in base 2 and x ≡ 2i−1 + t′ mod 2i , we have j = 0 by induction. Now the following corollaries are immediate from Theorem 3.5 and lemmas 3.10 and 3.11. Corollary 3.12. Assume that char K = 2, and let l ≥ 2i+1 , where i ≥ 0. Suppose that l + 1 ≡ 2i mod 2i+1 . Then V (ω2i+1 ) = L(ω2i+1 )/L(0). Proof. For 0 ≤ j ≤ 2i+1 , by Theorem 3.5 the irreducible L(ωj ) is a composition factor of V (ω2i+1 ) if and only if j = 2j ′ and l + 1 − 2j ′ contains 2i − j ′ to base 2. By Lemma 3.10, this is equivalent to j ′ = 0 or j ′ = 2i . Corollary 3.13. Assume that char K = 2, and let l ≥ 2i+1 , where i ≥ 0. Suppose that l + 1 ≡ 2i + t mod 2i+1 , where 0 ≤ t < 2i . Then any nontrivial composition factor of V (ω2i+1 ) has the form L(ω2j ), where 2i+1 ≥ 2j ≥ t + 1. 12 Proof. For 0 ≤ j ≤ 2i+1 , by Theorem 3.5 the irreducible L(ωj ) is a composition factor of V (ω2i+1 ) if and only if j = 2j ′ and l + 1 − 2j ′ contains 2i − j ′ to base 2. If 0 ≤ j ≤ t, then by Lemma 3.11 we have j = 0. 3.2 Construction of V (ωr ) We now describe the well known construction of the Weyl modules V (ωr ) for G using the exterior algebra of the natural module. We will consider our group G as a Chevalley group constructed from a complex simple Lie algebra of type Cl . For details of the Chevalley group construction see [Ste68]. Let e1 , . . . , el , e−l , . . . , e−1 be a basis for a complex vector space VC , and let VZ be the Z-lattice spanned by this basis. We have a non-degenerate alternating form (−, −) on VC defined by (ei , e−i ) = 1 = −(e−i , ei ) and (ei , ej ) = 0 for i 6= −j. Let sp(VC ) be the Lie algebra formed by the linear endomorphisms X of VC satisfying (Xv, w) + (v, Xw) = 0 for all v, w ∈ VC . Then sp(VC ) is a simple Lie algebra of type Cl . Let h be the Cartan subalgebra formed by the diagonal matrices in sp(VC ). Then h = {diag(h1 , . . . , hl , −hl , . . . , −h1 ) : hi ∈ C}. For 1 ≤ i ≤ l, define maps εi : h → C by εi (h) = hi where h is a diagonal matrix with diagonal entries (h1 , . . . , hl , −hl , . . . , −h1 ). Now Φ = {±(εi ± εj ) : 1 ≤ i < j ≤ l} ∪ {±2εi : 1 ≤ i ≤ l} is the root system for sp(VC ), Φ+ = {εi ± εj : 1 ≤ i < j ≤ l} ∪ {2εi : 1 ≤ i ≤ l} is a system of positive roots, and ∆ = {εi − εi+1 : 1 ≤ i < l} ∪ {2εl } is a base for Φ. For any i, j let Ei,j be the linear endomorphism on VC such that Ei,j (ej ) = ei and Ei,j (ek ) = 0 for k 6= j. Then a Chevalley basis for sp(VC ) is given by Xεi −εj = Ei,j −E−j,−i for all i 6= j, by X±(εi +εj ) = E±j,∓i +E±i,∓j for all i 6= j, by X±2εi = E±i,∓i for all i, and by Hεi −εi+1 = Ei,i −E−i,−i , H2εl = El,l −E−l,−l . Let UZ be the Kostant Z-form with respect to this Chevalley basis of sp(VC ). That is, UZ is the subring of the universal enveloping algebra of k sp(VC ) generated by 1 and all Xk!α for α ∈ Φ and k ≥ 1. Now VZ is a UZ -invariant lattice in VC . We define V = VZ ⊗Z K. Note that (−, −) also defines a non-degenerate alternating form on V . Then the simply connected Chevalley group of type Cl induced by V is equal to the group G = Sp(V ) of invertible linear maps preserving (−, −) [Ree57, pg. 396-397]. By abuse of notation we identify the basis (ei ⊗ 1) of V with (ei ). Note that for all 1 ≤ k ≤ 2l, the Lie algebra sp(VC ) acts naturally on ∧k (VC ) by X · (v1 ∧ · · · ∧ vk ) = k X v1 ∧ · · · ∧ vi−1 ∧ Xvi ∧ vi+1 ∧ · · · ∧ vk i=1 for all X ∈ sp(VC ) and vi ∈ VC . With this action, the Z-lattice ∧k (VZ ) is 13 invariant under UZ and this induces an action of G on ∧k (VZ ) ⊗Z K. One can show that g · (v1 ∧ · · · ∧ vk ) = gv1 ∧ · · · ∧ gvk for all g ∈ G and vi ∈ V , so we can and will identify ∧k (VZ ) ⊗Z K and ∧k (V ) as G-modules. The diagonal matrices in G form a maximal torus T . Then a basis of weight vectors of ∧k (V ) is given by the elements ei1 ∧ · · · ∧ eik , where −l ≤ i1 < · · · < ik ≤ l. The basis vector e1 ∧ · · · ∧ ek has weight ωk . The form on V induces a form on the exterior power ∧k (V ) by hv1 ∧ · · · ∧ vk , w1 ∧ · · · ∧ wk i = det((vi , wj ))1≤i,j≤k for all vi , wj ∈ V [Bou59, §1, Définition 12, pg. 30]. This form on ∧k (V ) is invariant under the action of G since (−, −) is. Furthermore, let ei1 ∧ · · · ∧ eik and ej1 ∧ · · · ∧ ejk be two basis elements of ∧k (V ). Then ( ±1, if {i1 , · · · , ik } = {−j1 , · · · , −jk }. hei1 ∧ · · · ∧ eik , ej1 ∧ · · · ∧ ejk i = 0, otherwise. Therefore it follows that the form h−, −i on ∧k (V ) is nondegenerate if 1 ≤ k ≤ l. In precisely the same way we find a basis of weight vectors for ∧k (VZ ) and define a form h−, −iZ on ∧k (VZ ). Note that h−, −i, h−, −iZ are alternating if k is odd and symmetric if k is even. It is well known that there is a unique submodule of ∧k (V ) isomorphic to the Weyl module V (ωk ) of G, as shown by the following lemma. The following lemma is also a consequence of [AJ84, 4.9]. Lemma 3.14. Let 1 ≤ k ≤ l, and let W be the G-submodule of ∧k (V ) generated by e1 ∧ · · · ∧ ek . Then (i) W is equal to the subspace of ∧k (V ) spanned by all v1 ∧ · · · ∧ vk , where hv1 , · · · , vk i is a k-dimensional   totally isotropic subspace of V . Further2l 2l more, dim W = k − k−2 . (ii) W is isomorphic to the Weyl module V (ωk ). Proof. (i) Since G acts transitively on the set of k-dimensional totally isotropic subspaces of V , it follows that W is spanned by all v1 ∧· · ·∧vk , where hv1 , . . . , vk i is a k-dimensional totally isotropic subspace of V . Then the claim about the dimension of W follows from a result proven for example in [DB09, Theorem 1.1], [Bro92, Theorem 1.1] or (in odd characteristic) [PS83, pg. 1337]. 14 (ii) Since e1 ∧ · · · ∧ ek is a maximal vector of weight ωk for G, the submodule W generated by it is an image of V (ωk ) [Jan03, II.2.13]. Now  2l 2l dim V (ωk ) = k − k−2 [Bou75, Ch. VIII, 13.3, pg. 203], so by (i) W must be isomorphic to V (ωk ). In what follows we will identify V (ωk ) as the submodule W of ∧k (V ) given by Lemma 3.14. Set V (ωk )Z = UZ (e1 ∧ · · · ∧ ek ). Note that now we can (and will) identify V (ωk ) and V (ωk )Z ⊗Z K. We will denote yi = ei ∧ e−i for all 1 ≤ i ≤ l. Then if k = 2s is even, a basis for the zero weight space of ∧k (V ) is given by vectors of the form yi1 ∧ · · · ∧ yis , where 1 ≤ i1 < · · · < is ≤ l. There is also a description of a basis for the zero weight space of V (ωk )Z in [Jan73, Lemma 10, pg.43]. For our purposes, we will only need a convenient set of generators given by the next lemma. Lemma 3.15 ([Jan73, pg. 40, Lemma 6]). Suppose that k is even, say k = 2s, where 1 ≤ k ≤ l. Then the zero weight space of V (ωk )Z (thus also of V (ωk )) is spanned by vectors of the form (yj1 − yk1 ) ∧ · · · ∧ (yjs − yks ), where 1 ≤ kr < jr ≤ l for all r and jr , kr 6= jr′ , kr′ for all r 6= r ′ . Lemma 3.16. Suppose that k is even, say k = 2s, where 1 ≤ k ≤ l. Then the vector X y i1 ∧ · · · ∧ y is γ= 1≤i1 <···<is ≤l is fixed by the action of G on ∧k (V ). Furthermore, any G-fixed point in ∧k (V ) is a scalar multiple of γ. Proof. To see that γ is fixed by G, see for example [DB10, 3.4] where it is shown that the definition of γ does not depend on the symplectic basis chosen. For the other claim, note first that any G-fixed point must have weight zero. Recall that the zero weight space of ∧k (V ) has basis B = {yi1 ∧ · · · ∧ yis : 1 ≤ i1 < · · · < is ≤ l}. Now the group Σl of permutations of {1, 2, . . . , l} acts on V by σ · e±i = e±σ(i) for all σ ∈ Σl . Clearly this action preserves the form (−, −) on V , so this gives an embedding Σl < G. Note also that Σl acts transitively on B. Thus any P Σl -fixed point in the linear span of B must be a scalar multiple of b∈B b = γ. 15 With these preliminary steps done, we now move on to proving Proposition 3.1. For the rest of this section, we will make the following assumption. Assume that char K = 2. Let 1 ≤ r ≤ l. By Proposition 2.3, we know that L(ω1 ) is not orthogonal. Suppose then that r ≥ 2 and that L(ωr ) is not orthogonal. By Proposition 2.2 (iii) we have H1 (G, L(ωr )) 6= 0, so by Corollary 3.7 we have r = 2i+1 for some i ≥ 0 and l + 1 ≡ 2i + t mod 2i+1 for some 0 ≤ t < 2i . What remains is to determine when L(ωr ) is orthogonal for such r. With the lemma below, we reduce this to the evaluation of Q(v) for a single vector v ∈ V (ωr ), where Q is a non-zero G-invariant quadratic form on V (ωr ). Lemma 3.17. Let l ≥ 2i+1 , where i ≥ 0. Suppose that l + 1 ≡ 2i + t i+1 mod 2i+1 , where 0 ≤ t < 2i . Define the vector γ ∈ ∧2 (V ) to be equal to X yi1 ∧ · · · ∧ yi2i . 1≤i1 <···<i2i ≤l−t Then (i) γ is in V (ω2i+1 ) and is a fixed point for the subgroup Cl−t < G, (ii) γ is in rad V (ω2i+1 ), (iii) L(ω2i+1 ) is orthogonal if and only if Q(γ) = 0, where Q is a nonzero G-invariant quadratic form on V (ω2i+1 ). Proof. (i) By Lemma 3.16, γ is a fixed by the action of Cl−t . It follows from Lemma 3.14 that the Cl−t -submodule W generated by e1 ∧ · · · ∧ e2i+1 is isomorphic to the Weyl module VCl−t (ω2i+1 ). By Corollary 3.12, the module W has a trivial Cl−t -submodule and by Lemma 3.16 it is generated by γ. Since W is contained in V (ω2i+1 ), the Cl -submodule generated by e1 ∧ · · · ∧ e2i+1 , the claim follows. For a different proof, one can also prove γ ∈ V (ω2i+1 ) by showing that γ is in the kernel of certain linear maps as defined in [DB10, Theorem 3.5] or [Bro92, Theorem 3.1, Proposition 3.3]. (ii) By Proposition 2.2 (ii), it will be enough to show that γ is orthogonal to i+1 V (ω2i+1 ) with respect to the form h−, −i on ∧2 (V ). Since γ has weight 0, it is orthogonal to any vector of weight 6= 0. Therefore it suffices to show that γ is orthogonal to any vector of weight 0 in V (ω2i+1 ). By 16 Lemma 3.15, this follows once we show that γ is orthogonal to any vector δ = (yj1 + yk1 ) ∧ · · · ∧ (yj2i + yk2i ), where 1 ≤ kr < jr ≤ l for all r and kr , jr 6= jr′ , kr′ for all r 6= r ′ . Because the set {j1 , . . . , j2i } contains 2i distinct integers, we cannot have jr ≥ l − t + 1 for all r. Indeed, otherwise jr ≥ l − t + 2i ≥ l + 1 for some r, contradicting the fact that jr ≤ l. Let q > 0 be the number of jr such that jr ≤ l − t. The vector δ can be written as X yf1 ∧ · · · ∧ yf2i . fs ∈{js ,ks } Now ( 1, if {f1 , . . . , f2i } = {g1 , . . . , g2i }. hyf1 ∧ · · · ∧ yf2i , yg1 ∧ · · · ∧ yg2i i = 0, otherwise. and so hδ, γi is an integer, equal to the number of yf1 ∧ · · · ∧ yf2i in the sum such that fs ≤ l − t for all 1 ≤ s ≤ 2i . Thus if jr ≥ kr ≥ l − t + 1 for some r, then hγ, δi = 0. If kr ≤ l − t for all r, then it follows that hδ, γi = 2q = 0 since q > 0. (iii) By Lemma 3.7 we have H1 (G, L(ω2i+1 )) 6= 0 and so there exists a nonsplit extension of L(ω2i+1 ) by the trivial module K. We can find this extension as an image of the Weyl module V (ω2i+1 ) [Jan03, II.2.13, II.2.14], so rad V (ω2i+1 )/M ∼ = K for some submodule M of rad V (ω2i+1 ). Since each composition factor of V (ω2i+1 ) occurs with multiplicity one (Theorem 3.5), each composition factor of M is nontrivial. Then by Corollary 3.13 and Lemma 3.9 (ii), the restriction M ↓ Cl−t has no trivial composition factors. But by (i) γ is a fixed point for Cl−t , so it follows that γ 6∈ M and then rad V (ω2i+1 ) = hγi ⊕ M as Cl−t -modules. Now let Q be a nonzero G-invariant quadratic form on V (ω2i+1 ). Since for the polarization bQ of Q we have rad bQ = rad V (ω2i+1 ) (Proposition 2.2 (ii)), composing Q with the square root map K → K defines a morphism rad V (ω2i+1 ) → K of G-modules. Therefore Q must vanish on M, since M has no trivial composition factors. Thus for all m ∈ M and scalars c we have Q(cγ + m) = c2 Q(γ), so Q vanishes on rad V (ω2i+1 ) if and only if Q(γ) = 0. Hence by Proposition 2.2 (iii) L(ω2i+1 ) is orthogonal if and only if Q(γ) = 0. 17 3.3 Computation of a quadratic form Q on V (ωr ) To finish the proof of Proposition 3.1 we still have to compute Q(γ) for the vector γ from Lemma 3.17. We retain the notation from the previous subsection and keep the assumption that char K = 2. Let r be even, say r = 2s, where 1 ≤ r ≤ l. Now the form h−, −iZ on ∧r (VZ ) induces a quadratic form qZ on ∧r (VZ ) by qZ (x) = hx, xi. We will use this form to find a nonzero G-invariant quadratic form on V (ωr ) = V (ωr )Z ⊗Z K. Lemma 3.18. We have qZ (V (ωr )Z ) ⊆ 2Z and qZ (V (ωr )Z ) 6⊆ 4Z. Proof. Let α = e1 ∧ · · · ∧ er and β = e−1 ∧ · · · ∧ e−r . Now α, β ∈ V (ωr )Z and hα, αi = hβ, βi = 0 and hα, βi = 1, giving qZ (α + β) = 2 and so qZ (V (ωr )Z ) 6⊆ 4Z. If we had qZ (V (ωr )Z ) 6⊆ 2Z, then Q = qZ ⊗Z K defines a nonzero G-invariant quadratic form on V (ωr ). But then the polarization of Q is equal to 2h−, −i = 0, which by Proposition 2.2 (i) and (ii) is not possible. Therefore Q = 21 qZ ⊗Z K defines a nonzero G-invariant quadratic form on V (ωr ) with polarization h−, −i. A similar construction when r is odd is discussed in [GN16, Proposition 8.1]. Now if we consider a zero weight vector of the form X y i1 ∧ · · · ∧ y is {i1 ,··· ,is }∈I in V (ωr ), the value of Q for this vector is equal to |I2 | since ( 1, if {f1 , . . . , fk } = {g1 , . . . , gk }. hyf1 ∧ · · · ∧ yfk , yg1 ∧ · · · ∧ ygk i = 0, otherwise. Now we can compute the value of Q(γ) from Lemma 3.17. Since there  are l−t 1 l−t terms occurring in the sum that defines γ, we have Q(γ) = 2 2i . Thus 2i  Q(γ) = 0 if and only if l−t is divisible by 4. Now the proof of Proposition i 2 3.1 is finished with the following lemma. Lemma 3.19. Let l + 1 ≡ 2i + t mod 2i+1 , where 0 ≤ t < 2i . The integer  l−t is divisible by 4 if and only if l + 1 ≡ 2i + t mod 2i+2 . 2i Proof. According to Kummer’s theorem, if p is prime and d ≥ 0 is maximal  x d such that p divides y (x ≥ y ≥ 0), then d is the number of carries that 18 occur when adding y to x−y in base p. Now (l −t)−2i ≡ −1 ≡ 2i +· · ·+2+1 mod 2i+1 . If l − t − 2i ≡ 2i + · · · + 2 + 1 mod 2i+2 , then adding 2i to l − t − 2i in binary results in just one carry. The other possibility is that l − t − 2i ≡ 2i+1 + 2i + · · · + 2 + 1 mod 2i+2 ,  and in this case there are ≥ 2 carries. Therefore l−t is divisible by 4 if and i 2 only if l − t − 2i ≡ 2i+1 + 2i + · · · + 2 + 1 ≡ −1 mod 2i+2 , which is equivalent to l + 1 ≡ 2i + t mod 2i+2 . 4 Fundamental irreducibles for classical types With a bit more work, we can use Proposition 3.1 to determine for all classical types the fundamental irreducible representations that are orthogonal. In this section assume that char K = 2. For a groups of type Al (l ≥ 1), the only self-dual fundamental irreducible representations are those of form L(ω l+1 ), where l is odd. Furthermore, all 2 fundamental representations are minuscule, so V (ω l+1 ) = L(ω l+1 ) and thus 2 2 by Proposition 2.2 (iv) the representation L(ω l+1 ) is orthogonal. 2 Now for type Bl , there exists an exceptional isogeny ϕ : Bl → Cl between simply connected groups of type Bl and Cl [Ste68, Theorem 28]. Then irreducible representations of Cl induce irreducible representations Bl by twisting with the isogeny ϕ. For fundamental irreducible representations, we have LCl (ωr )ϕ ∼ = LBl (ωr ) if 1 ≤ r ≤ l − 1, and LCl (ωl )ϕ is a Frobenius twist of LBl (ωl ). Therefore for all 1 ≤ r ≤ l, the representation LCl (ωr ) is orthogonal if and only if LBl (ωr ) is orthogonal. Consider then type Dl (l ≥ 4). First note that the natural representation LDl (ω1 ) of Dl is orthogonal. Now since we are working in characteristic two, there is an embedding Dl < Cl as a subsystem subgroup generated by the short root subgroups. Then if 1 ≤ r ≤ l − 2, we have LCl (ωr ) ↓ Dl ∼ = LDl (ωr ) for 1 ≤ r ≤ l − 2 by [Sei87, Theorem 4.1]. By combining this fact with the lemma below, we see for 2 ≤ r ≤ l − 2 that LCl (ωr ) is orthogonal if LDl (ωr ) is orthogonal. Lemma 4.1. Let G be simple of type Cl and consider H < G of type Dl as the subsystem subgroup generated by short root subgroups. Suppose that V is a nontrivial irreducible 2-restricted representation of G and V 6= LG (ω1 ). 19 Then if V ↓ H is 2-restricted irreducible, the representation V is orthogonal for G if and only if V is orthogonal for H. Proof (G. Seitz). If V is an orthogonal G-module, it is clear that it is an orthogonal H-module as well. Suppose then that V is not orthogonal for G. Since V is not the natural module for G, by Proposition 2.2 (iii) there exists a nonsplit extension 0 → hwi → M → V → 0 of G-modules, where w ∈ M. Furthermore, there exists a nonzero G-invariant quadratic form Q on M such that Q(w) 6= 0. We claim that M ↓ H is also a nonsplit extension. If this is not the case, then M ↓ H = W ⊕ hwi for some H-submodule W of M. We will show that W is invariant under G, which is a contradiction since M is nonsplit for G. Now W is 2-restricted irreducible for H, so by a theorem of Curtis [Bor70, Theorem 6.4] the module W is also an irreducible representation of Lie(H). Since Lie(H) is an ideal of Lie(G) that is invariant under the adjoint action of G, it follows that gW is Lie(H)-invariant for all g ∈ G. But as a Lie(H)-module M is the sum of a trivial module and W , so we must have gW = W for all g ∈ G. Thus if V ↓ H = LH (λ), then there exists a surjection π : VH (λ) → M of H-modules [Jan03, II.2.13]. Now the quadratic form Q induces via π a nonzero, H-invariant quadratic form on VH (λ) which does not vanish on the radical of VH (λ). By Proposition 2.2 (iii) the representation V is not orthogonal for H. Finally, the half-spin representations of Dl are minuscule representations, so LDl (ωl ) = VDl (ωl ) and LDl (ωl−1 ) = VDl (ωl−1 ). As before, by Proposition 2.2 (iv) it follows that LDl (ωl ) and LDl (ωl−1) are orthogonal if they are selfdual. Therefore we can conclude that for l = 4 and l = 5 all self-dual LDl (ωi ) are orthogonal. Note that if l ≥ 6, then LCl (ωl ) and LCl (ωl−1 ) are also orthogonal (Example 3.4). Thus for l ≥ 6 we have for all 2 ≤ r ≤ l that LCl (ωr ) is orthogonal if LDl (ωr ) is orthogonal. Taking all of this together, Proposition 3.1 is improved to the following. Theorem 4.2. Assume that char K = 2. Let G be simple of type Al (l ≥ 1), Bl (l ≥ 2), Cl (l ≥ 2) or Dl (l ≥ 4). Suppose 1 ≤ r ≤ l and ωr = −w0 (ωr ). Then L(ωr ) is not orthogonal if and only if one of the following holds: • G is of type Bl (l ≥ 2) or Cl (l ≥ 2) and r = 1. • G is of type Bl (l ≥ 2), Cl (l ≥ 2) or Dl (l ≥ 6) and r = 2i+1 for some i ≥ 0 such that l + 1 ≡ 2i+1 + 2i + t mod 2i+2 , where 0 ≤ t < 2i . 20 Representations L(ωr + ωs ) for type Al 5 Assume that G is simply connected of type Al , l ≥ 2. Set n = l + 1. In this section, we determine when in characteristic 2 the irreducible representation L(ωr + ωs ), 1 ≤ r < s ≤ l, of G has a nonzero G-invariant quadratic form. Now L(ωr + ωs ) is not orthogonal if it is not self-dual, so it will be enough to consider L(ωr + ωn−r ), where 1 ≤ r < n − r ≤ l (see Table 2.1). In this case, the answer and the methods to prove it are very similar to those found in Section 3. The result is the following theorem, which we will prove in what follows. Theorem 5.1. Assume char K = 2. Let 1 ≤ r < n−r ≤ l. Then L(ωr +ωn−r ) is not orthogonal if and only if r = 2i for some i ≥ 0 and n + 1 ≡ 2i+1 + 2i + t mod 2i+2 , where 0 ≤ t < 2i . The following examples follow easily from Theorem 5.1 (cf. examples 3.2 and 3.3). Example 5.2. If char K = 2, then L(ω1 + ωl ) is orthogonal if and only if n 6≡ 2 mod 4. This result was also proven in [GW95, Theorem 3.4 (b)]. Example 5.3. If char K = 2, then L(ω2 + ωl−1 ) is orthogonal if and only if n 6≡ 5, 6 mod 8. 5.1 Representation theory The composition factors and the submodule structure of the Weyl modules V (ωr + ωs ), 1 ≤ r < s ≤ l, were determined by Adamovich [Ada92]. Using her result, Baranov and Suprunenko have given in [BS05, Theorem 2.3] a description of the set of composition factors, similarly to Theorem 3.5. For 1 ≤ r < s ≤ l, define Jp (r, s) be the set of pairs (r − k, s + k), where n − s, r ≥ k ≥ 0 and s − r + 1 + 2k contains k to base p. Here we will define ω0 = 0 and ωn = 0, so then L(ω0 ) = L(ωn ) = L(ω0 + ωn ) is the trivial irreducible module and L(ω0 + ωr ) = L(ωr + ωn ) = L(ωr ). Now [BS05, Theorem 2.3] gives the following2 . Theorem 5.4. Let 1 ≤ r < s ≤ l. Then in the Weyl module V (ωr + ωs ), each composition factor has multiplicity 1, and the set of composition factors is {L(ωj + ωj ′ ) : (j, j ′ ) ∈ Jp (r, s)}. Baranov and Suprunenko give the result in terms of πi,j = L(ω(j−i+1)/2 + ω(i+j−1)/2 ), but from πy−x+1,x+y = L(ωx + ωy ) we get the formulation in Theorem 5.4. 2 21 The result of Kleshchev and Sheth in [KS99] [KS01, Corollary 3.6] about the first cohomology groups for groups of type Al gives the following (cf. Corollary 3.7). Theorem 5.5. Assume that char K = 2. Let 1 ≤ r < s ≤ l. Then H1 (G, L(ωr + ωs )) 6= 0 if and only if r = 2i , s = n − 2i for some i ≥ 0, and n + 1 ≡ 2i + t mod 2i+1 for some 0 ≤ t < 2i . Throughout this section we will consider subgroups Al′ < Al = G, which are embedded into G as follows. We consider G = SL(V ), where V has basis e1 , e2 , . . . , el+1 . Then for 1 ≤ l′ < l, the embedding Al′ < Al is SL(V ′ ) < SL(V ), where V ′ ⊆ V has basis e1 , . . . , el′ +1 and SL(V ′ ) fixes the basis vectors el′ +2 , . . . , el+1 . Baranov and Suprunenko have determined the submodule structure of the restrictions L(ωr + ωs ) ↓ Al−1 for all 0 ≤ r ≤ s ≤ n in their article [BS05, Theorem 1.1]. As in Section 3, for our purposes it will be enough to know which composition factors occur in the restriction. To state the result l of Baranov and Suprunenko, we will denote πr,s = LAl (ωr + ωl+1−s ) for all l 0 ≤ r ≤ l +1−s ≤ l +1. We will define πr,s = 0 if r < 0, s < 0 or r +s > l +1. Now the main result of [BS05] gives the following3 (cf. Theorem 3.8). Theorem 5.6. Let 0 ≤ r ≤ n − s ≤ n and assume that n ≥ 3. Set d = νp (n + 1 − (r + s)), and ε = 0 if n + 1 − (r + s) ≡ −pd mod pd+1 and ε = 1 l otherwise. Then the character of πr,s ↓ Al−1 is given by l−1 l−1 l−1 ch πr,s−1 + ch πr−1,s + ch πr,s + d−1 X l−1 2 ch πr−p k ,s−pk k=0 ! l−1 + ε ch πr−p d ,s−pd where the sum in the brackets is zero if d = 0. As with Theorem 3.8, note that when char K = 2, we always have ε = 0 in Theorem 5.6. The following applications of theorems 5.4 and 5.5 in characteristic two will be needed later. Lemma 5.7. Assume that char K = 2, and let n ≥ 2i+1 , where i ≥ 0. Suppose that n + 1 ≡ 2i + t mod 2i+1 , where 0 ≤ t < 2i . Let 0 ≤ x ≤ 2i and 0 ≤ y ≤ 2i be such that 2i+1 ≥ x + y ≥ t + 1. Then l (i) All composition factors of the restriction πx,y ↓ Al−1 have the form πxl−1 ′ ,y ′ ′ ′ i ′ ′ for some 0 ≤ x , y ≤ 2 such that x + y ≥ t. Baranov and Supruneko give their result in terms of Lli,j = LAl (ωi + ωj ) for 0 ≤ i ≤ j ≤ n, but replacing j by n − j gives the formulation in Theorem 5.6. 3 22 l (ii) πx,y ↓ Al−t has no trivial composition factors. Proof. (cf. Lemma 3.9) If t = 0 there is nothing to prove, so suppose that t ≥ 1. It will be enough to prove (i) as then (ii) will follow by induction on t. Let d = ν2 (n + 1 − (x + y)). Suppose first that 0 ≤ d < i + 1. Then n + 1 − (x + y) ≡ t − (x + y) mod 2i , so ν2 (n + 1 − (x + y)) = ν2 ((x + y) − t). l−1 l By Theorem 5.6, the composition factors occurring in πx,y ↓ Al−1 are πx,y−1 , l−1 l−1 l−1 πx−1,y , πx,y , and πx−2k ,y−2k for 0 ≤ k ≤ d − 1. Therefore the claim follows since ν2 ((x + y) − t) = d and thus x + y − 2d ≥ t. Consider then the case where d ≥ i + 1. Then n + 1 − (x + y) ≡ 2i + t − (x + y) ≡ 0 mod 2i+1 , so (x + y) − t ≡ 2i mod 2i+1 . On the other hand 0 ≤ (x + y) − t < 2i+1 , so (x + y) − t = 2i . By Theorem 5.6 the composition l−1 l−1 l−1 l l−1 factors occurring in πx,y ↓ Al−1 are πx,y−1 , πx−1,y , πx,y , and πx−2 k ,y−2k for k k 0 ≤ k ≤ i − 1 (since x − 2 < 0 or y − 2 < 0 for i ≤ k ≤ d − 1), so again the claim follows. As a consequence of Theorem 5.4 and lemmas 3.10 and 3.11, we get the following (cf. corollaries 3.12 and 3.13). Corollary 5.8. Assume that char K = 2 and let n > 2i+1 , where i ≥ 0. Suppose that n + 1 ≡ 2i mod 2i+1 . Then V (ω2i + ωn−2i ) = L(ω2i + ωn−2i )/L(0). Proof. According to Theorem 5.4, the composition factors of V (ω2i + ωn−2i ) are L(ω2i −k + ωn−2i +k ), where 0 ≤ k ≤ 2i and n + 1 − 2i+1 + 2k contains k to base 2. We can replace k by 2i − k, and then the condition is equivalent to n + 1 − 2k containing 2i − k to base 2, which implies k = 0 or k = 2i by Lemma 3.10. Corollary 5.9. Assume that char K = 2 and let n > 2i+1 , where i ≥ 0. Suppose that n + 1 ≡ 2i + t mod 2i+1 , where i ≥ 0 and 0 ≤ t < 2i . Then any nontrivial composition factor of V (ω2i + ωn−2i ) has the form L(ωx + ωn−y ) for some 0 ≤ x ≤ n − y ≤ n and x + y ≥ t + 1. Proof. According to Theorem 5.4, the composition factors of V (ω2i + ωn−2i ) are L(ω2i −k + ωn−2i +k ), where 0 ≤ k ≤ 2i and n + 1 − 2i+1 + 2k contains k to base 2. Setting k ′ = 2i − k, the composition factors are L(ωk′ + ωn−k′ ), where n + 1 − 2k ′ contains 2i − k ′ to base 2. By Lemma 3.11 we have k ′ = 0 or 2k ′ ≥ t + 1, which proves the claim. 5.2 Construction of V (ωr + ωn−r ) We now describe a construction of V (ωr + ωn−r ), in many ways similar to that of VCl (ωr ) described in Section 3.2. We will consider our group G as a Chevalley group constructed from a complex simple Lie algebra of type Al . 23 Let e1 , e2 , . . . , en be a basis for a complex vector space VC , and let VZ be the Z-lattice spanned by this basis. Let sl(VC ) be the Lie algebra formed by the linear endomorphisms of VC with trace zero. Then sl(VC ) is a simple Lie algebra of type Al . Let h be the Cartan subalgebra formed by the diagonal matrices in sl(VC ) (with respect to the basis (ei )). For 1 ≤ i ≤ n, define maps εi : h → C by εi (h) = hi where h is a diagonal matrix with diagonal entries (h1 , h2 , . . . , hn ). Now Φ = {εi − εj : i 6= j} is the root system for sl(VC ) and Φ+ = {εi − εj : i < j} is a system of positive roots, and ∆ = {εi − εi+1 : 1 ≤ i ≤ l} is a base for Φ. For any i, j let Ei,j be the linear endomorphism on VC such that Ei,j (ej ) = ei and Ei,j (ek ) = 0 for k 6= j. Now a Chevalley basis for sl(VC ) is given by Xεi −εj = Ei,j for i 6= j and Hεi −εi+1 = Ei,i − Ei+1,i+1 for 1 ≤ i ≤ l. Let UZ be the Kostant Z-form with respect to this Chevalley basis of sl(VC ). That is, UZ is the subring of the universal enveloping algebra of sl(VC ) generated k by 1 and all Xk!α for α ∈ Φ and k ≥ 1. Now VZ is a UZ -invariant lattice in VC . We define V = VZ ⊗Z K. Then the simply connected Chevalley group of type Al induced by V is equal to G = SL(V ). Let e∗1 , e∗2 , . . . , e∗n be a basis for VC∗ , dual to the basis (e1 , e2 , . . . , en ) of VC (so here e∗i (ej ) = δij ). Denote the Z-lattice spanned by e∗1 , e∗2 , . . . , e∗n by VZ∗ . Then VZ∗ is UZ -invariant and we can identify VZ∗ ⊗Z K and V ∗ as G-modules. Here the action of G on V ∗ is given by (g · f )(v) = f (g −1v) for all g ∈ G, f ∈ V ∗ and v ∈ V . By abuse of notation we identify the basis (ei ⊗ 1) of V with (ei ), and the basis (e∗i ⊗ 1) of V ∗ with (e∗i ). Let 1 ≤ k < n − k ≤ l. Now the Lie algebra sl(VC ) acts naturally on ∧k (VC ) by X · (v1 ∧ · · · ∧ vk ) = k X v1 ∧ · · · ∧ vi−1 ∧ Xvi ∧ vi+1 ∧ · · · ∧ vk i=1 for all X ∈ sl(VC ) and vi ∈ VC . Similarly we have an action of sl(VC ) on ∧k (VC∗ ). Furthermore, sl(VC ) acts on ∧k (VC ) ⊗ ∧k (VC∗ ) by X · (v ⊗ w) = Xv ⊗ w + v ⊗ Xw for all X ∈ sl(VC ), v ∈ ∧k (VC ) and w ∈ ∧k (VC∗ ). Here ∧k (VZ ) ⊗ ∧k (VZ∗ ) is an UZ -invariant lattice in ∧k (VC ) ⊗ ∧k (VC∗ ), and we can and will identify ∧k (VZ ) ⊗ ∧k (VZ∗ ) ⊗Z K and ∧k (V ) ⊗ ∧k (V ∗ ) as G-modules. The diagonal matrices in G form a maximal torus T . Then a basis of weight vectors of ∧k (V ) ⊗ ∧k (V ∗ ) is given by the elements (ei1 ∧ · · · ∧ eik ) ⊗ (e∗j1 ∧ · · · ∧ e∗jk ), where 1 ≤ i1 < · · · < ik ≤ n and 1 ≤ j1 < · · · < jk ≤ n. The basis vector (e1 ∧ · · · ∧ ek ) ⊗ (e∗n ∧ e∗n−1 ∧ · · · ∧ e∗n−k+1) has weight ωk + ωn−k . 24 The natural dual pairing between ∧k (V ) and ∧k (V ∗ ) (see for example [FH91, B.3, pg.475-476]) induces a G-invariant symmetric form h−, −i on ∧k (V ) ⊗ ∧k (V ∗ ). If x = (v1 ∧ · · · ∧ vk ) ⊗ (f1 ∧ · · · ∧ fk ) and y = (w1 ∧ · · · ∧ wk ) ⊗ (g1 ∧ · · · ∧ gk ), with vi , wj ∈ V and fi , gj ∈ V ∗ , we define hx, yi = det(fi (wj ))1≤i,j≤k det(gi (vj ))1≤i,j≤k . Let b = (ei1 ∧· · ·∧eik )⊗(e∗j1 ∧· · ·∧e∗jk ) and b′ = (ei′1 ∧· · ·∧ei′k )⊗(e∗j ′ ∧· · ·∧e∗j ′ ) 1 k be two basis elements of ∧k (V ) ⊗ ∧k (V ∗ ). Then ( ±1, if {i1 , · · · , ik } = {j1′ , · · · , jk′ } and {i′1 , · · · , i′k } = {j1 , · · · , jk }. hb, b′ i = 0, otherwise. Therefore the form h−, −i on ∧k (V ) ⊗ ∧k (V ∗ ) is non-degenerate. In precisely the same way we can find a basis of weight vectors for ∧k (VZ ) ⊗ ∧k (VZ∗ ) and define a symmetric form h−, −iZ on ∧k (VZ ) ⊗ ∧k (VZ∗ ). We can find the Weyl module V (ωk + ωn−k ) as a submodule of ∧k (V ) ⊗ ∧k (V ∗ ), as shown by the following lemma (cf. Lemma 3.14). Lemma 5.10. Let 1 ≤ k < n − k ≤ l, and let W be the G-submodule of ∧k (V ) ⊗ ∧k (V ∗ ) generated by v + = (e1 ∧ · · · ∧ ek ) ⊗ (e∗n ∧ e∗n−1 ∧ · · · ∧ e∗n−k+1 ). Then W is isomorphic to the Weyl module V (ωk + ωn−k ). Proof. It is a general fact about Weyl modules that V (λ) ⊗ V (µ) always has V (λ + µ) as a submodule. For simple groups of classical type (in particular, for our G of type Al ) this follows from results proven first by Lakshmibai et. al. [LMS79, Theorem 2 (b)] or from a more general result of Wang [Wan82, Theorem B, Lemma 3.1]. For other types, the fact is a consequence of results due to Donkin [Don85] (all types except E7 and E8 in characteristic two) or Mathieu [Mat90] (in general). In any case, now the weight λ + µ occurs with multiplicity 1 in V (λ) ⊗ V (µ), so any vector of weight λ + µ in V (λ) ⊗ V (µ) will generate a submodule isomorphic to V (λ + µ). To prove our lemma, note that ∧k (V ) = L(ωk ) and ∧k (V ∗ ) = L(ωn−k ). Furthermore, ωk and ωn−k are minuscule weights, so L(ωk ) = V (ωk ) and L(ωn−k ) = V (ωn−k ). Here v + is a vector of weight ωk + ωn−k in ∧k (V ) ⊗ ∧k (V ∗ ), so the claim follows from the result in the previous paragraph. For all 1 ≤ k < n − k ≤ l, we will identify V (ωk + ωn−k ) with the submodule W from Lemma 5.10. Set V (ωk + ωn−k )Z = UZ v + where v + is as in Lemma 5.10. Then we can and will identify V (ωk + ωn−k )Z ⊗Z K and V (ωk + ωn−k ) as G-modules. Note that a basis for the zero weight space of ∧k (V ) ⊗ ∧k (V ∗ ) is given by vectors of the form (ei1 ∧· · ·∧eik )⊗(e∗i1 ∧· · ·∧e∗ik ), where 1 ≤ i1 < · · · < ik ≤ n. 25 We will need the following lemma, which gives a set of generators for the zero weight space of V (ωk + ωn−k ) (cf. Lemma 3.15). Lemma 5.11. Suppose that 1 ≤ k < n − k ≤ l. Then the zero weight space of V (ωk + ωn−k )Z (thus also of V (ωk + ωn−k )) is spanned by vectors of the form X (−1)|{s:fs =js}| (ef1 ∧ · · · ∧ efk ) ⊗ (e∗f1 ∧ · · · ∧ e∗fk ), fs ∈{js ,ks } where (k1 , . . . , kk ) and (j1 , . . . , jk ) are sequences such that 1 ≤ kr < jr ≤ n for all r, and jr , kr 6= jr′ , kr′ for all r 6= r ′ . Proof. We give a proof somewhat similar to that of Lemma 3.15 given in [Jan73, pg. 40, Lemma 6]. The zero weight space of V (ωk +ωn−k )Z is generated by elements of the form kα Y X−α v+ k ! α α∈Φ+ P where kα are non-negative integers, α∈Φ+ kα α = ωk + ωn−k and the product is taken with respect to some fixed ordering of the positive roots. For α ∈ Φ+ such that X−α v + = 0, we can assume kα = 0 by choosing a suitable ordering of Φ+ . Therefore we will assume that if kα > 0, then α is of one of the following types. (I) α = εi − εj for 1 ≤ i ≤ k and k + 1 ≤ j ≤ n − k. (II) α = εi − εj for k + 1 ≤ i ≤ n − k and n − k + 1 ≤ j ≤ n. (III) α = εi − εj for 1 ≤ i ≤ k and n − k + 1 ≤ j ≤ n. Note that the X−α with α of type (I) commute with each other. The same is also true for types (II) and (III). Writing ωk + ωn−k in terms of the simple roots, we see that ωk + ωn−k is equal to α1 +2α2 +· · ·+(k−1)αk−1 +kαk +· · ·+kαn−k +(k−1)αn−k+1 +· · ·+αn−1 (*) P Then from the fact that α∈Φ+ kα α = ωk +ωn−k we will deduce the following. (1) For any 1 ≤ i ≤ k, there exists a unique α ∈ Φ+ such that kα = 1 and α = εi − εj ′ for some k + 1 ≤ j ′ ≤ n. (2) For any n − k + 1 ≤ j ≤ n there exists a unique α ∈ Φ+ such that kα = 1 and α = εi′ − εj for some 1 ≤ i′ ≤ n − k. 26 For i = 1 and j = n these claims are clear, since α1 and αn−1 occur only once in the expression (*) of ωk +ωn−k as a sum of simple roots. For i > 1 claim (1) follows by induction, since εi −εj ′ contributes αi +αi+1 +· · ·+αk +· · ·+αj ′ −1 to the expression (*) of ωk + ωn−k as a sum of simple roots. Claim (2) follows similarly for j < n. In particular, it follows from claims (1) and (2) that kα ∈ {0, 1} for all α ∈ Φ+ . Let A1 , A2 and A3 be the sets of α ∈ Φ+ of type (I), (II) and (III) respectively such that kα = 1. It follows from claim (1) that |A1 | + |A3 | = k and from claim (2) that |A2 | + |A3 | = k, so then |A1 | = |A2 | = k ′ for some 0 ≤ k ′ ≤ k. Thus we can write A1 = {εi1 − εw1 , . . . , εik′ − εwk′ } A2 = {εz1 − εj1 , . . . , εzk′ − εjk′ } A3 = {εik′ +1 − εjk′+1 , . . . , εik − εjk } where 1 ≤ ir ≤ k and n − k + 1 ≤ jr ≤ n for all 1 ≤ r ≤ k, and k + 1 ≤ wr , zr ≤ n − k for all 1 ≤ r ≤ k ′ . Furthermore, {i1 , . . . , ik } = {1, 2, . . . , k} and {j1 , . . . , jk } = {n − k + 1, . . . , n − 1, n}. We choose the ordering of Φ+ so that kα Y X−α Y Y Y X−α . X−α X−α = k ! α + α∈A α∈A α∈A α∈Φ 2 3 1 P It is another consequence of α∈Φ+ kα α = ωk +ωn−k that {w1 , . . . , wk′ } = {z1 , . . . , zk′ }. Indeed, in the expression (*) of ωk + ωn−k as a sum of simple roots, for any k + 1 ≤ r ≤ n − k the simple root αr occurs k times. On the other hand, the α of types (I), (II), (III) that contribute to αr in the sum are precisely those of type (I) or (III) with j > r (total of k − |{r ′ : wr′ ≤ r}|), ′ ′ and those of type (II) with Pj ≤ r (total of |{r : zr ≤ r}|). Therefore in the sum α∈Φ+ kα α, the contribution to αr is equal to k − |{r ′ : wr′ ≤ r}| + |{r ′ : zr′ ≤ r}|. Since this has to be equal to k, we get |{r ′ : zr′ ≤ r}| = |{r ′ : wr′ ≤ r}| for all k + 1 ≤ r ≤ n − k, which implies {w1 , . . . , wk′ } = {z1 , . . . , zk′ }. Then since the X−α with α of type (II) commute with each other, we Q′ may assume that zr = wr for all 1 ≤ r ≤ k ′ . Denote w = kr=1 Ewr ,ir v + . A straightforward computation shows that w = (eπ(1) ∧ · · ·∧ eπ(k) ) ⊗ (e∗n ∧ e∗n−1 ∧ 27 · · · ∧ e∗n−k+1 ), where π(r) = wr′ if r = ir′ and π(r) = r otherwise. Now ′ ′ k k k kα Y Y X−α Y Y + Ejr ,ir Ejr ,wr v = Ewr ,ir v + k ! α + r=1 r=1 r=k ′ +1 α∈Φ = k Y ′ Ejr ,ir = Ejr ,wr w r=1 r=k ′ +1 k Y k Y Ejr ,kr w r=1 where (k1 , . . . , kk ) = (w1 , . . . , wk′ , ik′ +1 , . . . , ik ). In the last equality we just combine the terms, and this makes sense since X−α of type (III) commute with those of type (II). Q Computing the expression kr=1 Ejr ,kr w, we see that it is equal to a sum of 2k distinct elements of ∧k (VZ ) ⊗ ∧k (VZ∗ ), with each summand being equal to w transformed in the following way: • For all 1 ≤ s ≤ k ′ , replace ews by ejs , or replace e∗js by −e∗ws . • For all k ′ + 1 ≤ s ≤ k, replace eis by ejs , or replace e∗js by −e∗is . Q X kα + For this we conclude that up to a sign, α∈Φ+ k−α v is as in the statement α! of the lemma, with sequences (k1 , . . . , kk ) and (j1 , . . . , jk ) as defined here. Lemma 5.12. Suppose that 1 ≤ k < n − k ≤ l. Then the vector X (ei1 ∧ · · · ∧ eik ) ⊗ (e∗i1 ∧ · · · ∧ e∗ik ) 1≤i1 <···<ik ≤n is fixed by the action of G on ∧k (V ) ⊗ ∧k (V ∗ ). Furthermore, any G-fixed point in ∧k (V ) ⊗ ∧k (V ∗ ) is a scalar multiple of γ. Proof. (cf. Lemma 3.16) The fact that γ is fixed by G is an exercise in linear algebra. We will give a proof for convenience of the reader. For this we first need to introduce some notation. Let A be an n × n matrix with entries in K and denote the entry on ith row and jth column of A by Ai,j . For indices 1 ≤ i1 < · · · < ik ≤ n and 1 ≤ j1 < · · · < jk ≤ n, set I = {i1 , . . . , ik } and J = {j1 , . . . , jk }. Then the k × k minor of A defined by I and J is the determinant of the k × k matrix (Aip ,jq ). We denote this minor by [A]I,J . The following relation between minors of matrices A, B, and AB (similar to the matrix multiplication rule) is a special case of the Cauchy-Binet formula. A proof can be found in [BW89, 4.6, pg. 208-214]. 28 Proposition (Cauchy-Binet formula). Let A and B be n × n matrices. For any k-element subsets I, J of {1, . . . , n}, we have X [AB]I,J = [A]I,T [B]T,J T where the sum runs over all k-element subsets T of {1, . . . , n}. Consider A ∈ GL(V ) as a matrix with respect to the basis e1 , . . . , en of V . Now for any 1 ≤ k ≤ n, the matrix A acts on the exterior power ∧k (V ) by A · (v1 ∧ · · · ∧ vk ) = Av1 ∧ · · · ∧ Avk for all vi ∈ V . For I = {i1 , . . . , ik } with 1 ≤ i1 < · · · < ik ≤ n, denote P eI = ei1 ∧ · · · ∧ eik . A straightforward calculation shows that A · eI = J [A]J,I eJ , where the sum runs over all k-element subsets J of {1, . . . , n}. With respect to the dual basis e∗1 , . . . , e∗n of V ∗ , the action of A on V ∗ has matrix (At )−1 , where AtP is the transpose ofP A. If we denote e∗I = e∗i1 ∧ · · · ∧ e∗ik as before, then A · e∗I = J [(At )−1 ]J,I e∗J = J [A−1 ]I,J e∗J where the sum runs over all k-element subsets J of {1, . . . , n}. P We are∗ now ready to prove that A fixes the vector γ. Note that γ = I eI ⊗ eI , where the sum runs over all k-element subsets I of {1, . . . , n}. From the observations before, we see that A · γ is equal to XX XX [A]J,I [A−1 ]I,J ′ eJ ⊗ e∗J ′ = [A]J,I [A−1 ]I,J ′ eJ ⊗ e∗J ′ I J,J ′ J,J ′ I where the sums run over k-element subsets I, J, and J ′ of {1, . . . , n}. From P the Cauchy-Binet formula, we have I [A]J,I [A−1 ]I,J ′ = [1]J,J ′ = 0 if J 6= J ′ and 1 if J = J ′ . Therefore A · γ = γ. To show that γ is a unique G-fixed point up to a scalar, note first that any G-fixed point must have weight zero. Recall also that the zero weight space of ∧k (V ) has basis B = {(ei1 ∧ · · · ∧ eik ) ⊗ (e∗i1 ∧ · · · ∧ e∗ik ) : 1 ≤ i1 < · · · < ik ≤ n}. Now the group Σn of permutations of {1, 2, . . . , n} acts on V by σ · ei = eσ(i) for all σ ∈ Σn . This gives an embedding Σn < GL(V ). Note that then σ · e∗i = e∗σ(i) for all σ ∈ Σn , so it follows that Σn acts on B. For σ ∈ Σn we have det σ = 1 if and only if σ is an even permutation, so we get an embedding Alt(n) < G for the alternating group. It is well known that Alt(n) is (n − 2)-transitive, so Alt(n) acts transitively on B since k ≤ n − 2. Thus P any Alt(n)-fixed point in the linear span of B must be a scalar multiple of b∈B b = γ. 29 We now begin the proof of Theorem 5.1. For the rest of this section, we will make the following assumption. Assume that char K = 2. Let 1 ≤ k < n − k ≤ l and suppose that L(ωk + ωn−k ) is not orthogonal. Then by Proposition 2.2 (iii) we have H1 (G, L(ωk + ωn−k )) 6= 0, so by Corollary 5.5 we have k = 2i for some i ≥ 0 and n + 1 ≡ 2i + t mod 2i+1 for some 0 ≤ t < 2i . What remains is to determine when L(ωk + ωn−k ) is orthogonal for such k. The main argument is the following lemma (cf. Lemma 3.17), which reduces the question to the evaluation of the invariant quadratic form on V (ωk + ωn−k ) on a particular v ∈ V (ωk + ωn−k ). Lemma 5.13. Let n > 2i+1 , where i ≥ 0. Suppose that n + 1 ≡ 2i + t i i mod 2i+1 , where 0 ≤ t < 2i . Define the vector γ ∈ ∧2 (V ) ⊗ ∧2 (V ∗ ) to be equal to X (ei1 ∧ · · · ∧ ei2i ) ⊗ (e∗i1 ∧ · · · ∧ e∗i2i ). 1≤i1 <···<i2i ≤n−t Then (i) γ is in V (ω2i + ωn−2i ) and is a fixed point for the subgroup Al−t < G, (ii) γ is in rad V (ω2i + ωn−2i ), (iii) L(ω2i + ωn−2i ) is orthogonal if and only if Q(γ) = 0, where Q is a nonzero G-invariant quadratic form on V (ω2i + ωn−2i ). Proof. (i) Same as Lemma 3.17 (i). Apply Lemma 5.12, Lemma 5.10 and Corollary 5.8. (ii) Same as Lemma 3.17 (ii). Apply Lemma 5.11, and note that for b = (ej1 ∧· · ·∧ej2i )⊗(e∗j1 ∧· · ·∧e∗j2i ) and b′ = (ek1 ∧· · ·∧ek2i )⊗(e∗k1 ∧· · ·∧e∗k i ) 2 we have ( 1, if {j1 , . . . , j2i } = {k1 , . . . , k2i }. hb, b′ i = 0, otherwise. (iii) Same as Lemma 3.17 (iii). Apply Theorem 5.5 to find a submodule M ⊆ rad V (ω2i + ωn−2i ) such that rad V (ω2i+1 )/M ∼ = K. Each composition factor of V (ω2i + ωn−2i ) occurs with multiplicity one by Theorem 5.4, so M has no nontrivial composition factors. By Lemma 5.7 (i) the restriction M ↓ Al−t has no trivial composition factors, and by (i) the vector γ is fixed by Al−t . Thus γ 6∈ M and then rad V (ω2i + ωn−2i ) = 30 hγi ⊕ M as Al−t -modules. As in Lemma 3.17 (iii), we see that L(ω2i + ωn−2i ) is orthogonal if and only if Q(γ) = 0 for a nonzero G-invariant quadratic form Q on V (ω2i + ωn−2i ). 5.3 Computation of a quadratic form Q on V (ωr + ωn−r ) To finish the proof of Theorem 5.1 we still have to compute Q(γ) for the vector γ from Lemma 5.13. We retain the notation and assumptions from the previous subsection. Let 1 ≤ k < n − k ≤ l. Now the form h−, −iZ on ∧k (VZ ) ⊗ ∧k (VZ∗ ) induces a quadratic form qZ on ∧k (VZ ) ⊗ ∧k (VZ∗ ) by qZ (x) = hx, xi. We will use this form to find a nonzero G-invariant quadratic form on V (ωk + ωn−k ) = V (ωk + ωn−k )Z ⊗Z K. Lemma 5.14. We have qZ (V (ωk + ωn−k )Z ) ⊆ 2Z and qZ (V (ωk + ωn−k )Z ) 6⊆ 4Z. Proof. Same as Lemma 3.18, but with α = (e1 ∧ · · · ∧ ek ) ⊗ (e∗n ∧ e∗n−1 ∧ · · · ∧ e∗n−k+1 ) and β = (en ∧ en−1 ∧ · · · ∧ en−k+1 ) ⊗ (e∗1 ∧ · · · ∧ e∗k ). Therefore Q = 12 qZ ⊗Z K defines a nonzero G-invariant quadratic form on V (ωk + ωn−k ) with polarization h−, −i. As in Section 3.3, we have Q(γ) = i i 1 n−t for the vector γ ∈ ∧2 (V )⊗∧2 (V ∗ ) from Lemma 5.13. Finally applying 2 2i Lemma 3.19 completes the proof Theorem 5.1. 6 Simple groups of exceptional type In this section, let G be a simple group of exceptional type and assume that char K = 2. We will give some results about the orthogonality of irreducible representations of G. For G of type G2 or F4 we give a complete answer. For types E6 , E7 , and E8 we only have results for some specific representations, given in Table 6.1 below and proven at the end of this section. For irreducible representations occurring in the adjoint representation of G, answers were given earlier by Gow and Willems in [GW95, Section 3]. Proposition 6.1. Let G = G2 and let V be a non-trivial irreducible representation of G. Then V is not orthogonal if and only if V is a Frobenius twist of LG (ω1 ). Proof. In view of Remark 2.5, it will be enough to consider V = LG (λ) with λ ∈ X(T )+ a 2-restricted dominant weight. If λ = ω2 or λ = ω1 + ω2 , then VG (λ) = LG (λ) and so V is orthogonal by Proposition 2.2. 31 What remains is to show that V = LG (ω1 ) not orthogonal. There are several ways to see this, for example since dim V = 6 this could be done by a direct computation. Alternatively, note that the composition factors of ∧2 (V ) are LG (ω2 ) and LG (0) [LS96, Proposition 2.10], so H1 (G, ∧2 (V )) = 0. Then by [SW91, Proposition 2.7], the module V is not orthogonal. For a third proof, note that the action of a regular unipotent u ∈ G on V has a single Jordan block [Sup95, Theorem 1.9], but no such element exists in SO(V ) [LS12, Proposition 6.22]. The following lemma will be useful throughout this section to show that certain representations are orthogonal. Lemma 6.2. Let V be a nontrivial, self-dual and irreducible G-module. Suppose that one of the following holds: (i) dim V ≡ 2 mod 4, and ∧2 (V ) has exactly one trivial composition factor as a G-module. (ii) dim V ≡ 0 mod 8, and ∧2 (V ) has exactly two trivial composition factors as a G-module. Then any nontrivial composition factor of ∧2 (V ) occuring with odd multiplicity is an orthogonal G-module. Proof. Since V is nontrivial, we can assume G < Sp(V ) by Lemma 2.1. If (i) holds, then by applying results in Section 3.2 (or [McN98, Lemma 4.8.2]) we can find a vector γ ∈ ∧2 (V ) such that ∧2 (V ) = Z ⊕ hγi as an Sp(V )-module. Here Z is irreducible of highest weight ω2 for Sp(V ), so by Proposition 3.1 (see Example 3.2) the module Z is orthogonal for Sp(V ). Therefore Z is an orthogonal G-module with no trivial composition factors. From this [GW95, Lemma 1.3] shows that any composition factor of Z with odd multiplicity is an orthogonal G-module. In case (ii), the assumption on dim V implies (for example by [McN98, Lemma 4.8.2]) that there exist Sp(V )-submodules Z ′ ⊆ Z ⊆ ∧2 (V ) such that dim Z ′ = dim ∧2 (V )/Z = 1. Furthermore, Z/Z ′ is an irreducible Sp(V )module with highest weight ω2 , so by Proposition 3.1 (see Example 3.2) the module Z/Z ′ is orthogonal for Sp(V ). Therefore Z/Z ′ is an orthogonal G-module with no trivial composition factors, so by [GW95, Lemma 1.3] any composition factor of Z/Z ′ with odd multiplicity is an orthogonal Gmodule. Proposition 6.3. Let G = F4 and let V be a non-trivial irreducible representation of G. Then V is orthogonal. 32 Proof. Let τ : G → G be the exceptional isogeny of G as given in [Ste68, Theorem 28]. Then LG (a1 ω1 + a2 ω2 + a3 ω3 + a4 ω4 )τ ∼ = LG (a4 ω1 + a3 ω2 + 2a2 ω3 + 2a1 ω4 ), and by Steinberg’s tensor product theorem this is isomorphic to LG (a4 ω1 + a3 ω2 ) ⊗ L(a2 ω3 + a1 ω4 )F . Thus by lemmas 2.1 and 2.4, it is enough to prove the claim in the case where V = LG (λ) with λ = a3 ω3 + a4 ω4 a 2-restricted dominant weight. Now for λ = ω4 and λ = ω3 + ω4 , we have VG (λ) = LG (λ) = V and so V is orthogonal by Proposition 2.2 (iv)4 . What remains is to show that LG (ω3 ) is orthogonal. Let W = LG (ω4 ). Now dim W = 26, so by Lemma 6.2 (i), it will be enough to prove that ∧2 (W ) has exactly one trivial composition factor and that LG (ω3 ) occurs in ∧2 (W ) with odd multiplicity. We have VG (ω4 ) = LG (ω4 ) and then by a computation with Magma [BCP97] (or [Don85, 7.4.3, pg. 98]) the G-character of ∧2 (W ) is given by ch ∧2 (W ) = ch VG (ω1 )+ch VG (ω3 ). Furthermore, from the data in [Lüb17], we can deduce VG (ω1 ) = LG (ω1 )/LG (ω4 ) and VG (ω3 ) = LG (ω3 )/LG (ω4 )/LG (0). Therefore as a G-module ∧2 (W ) has composition factors LG (ω1 ), LG (ω4 ), LG (ω4 ), LG (ω3 ), and LG (0). G E6 E6 E6 E7 E7 E7 E7 E7 E8 E8 E8 λ ω2 ω4 ω1 + ω6 ω1 ω2 ω5 ω6 ω7 ω1 ω7 ω8 LG (λ) orthogonal? yes yes yes no yes yes yes yes yes yes yes Table 6.1: Orthogonality of some LG (λ) for G of type E6 , E7 and E8 . We finish this section by verifying the information given in Table 6.1. Suppose that G is of type E6 . We have VG (ω2 ) = LG (ω2 ) and VG (ω1 + ω6 ) = L(ω1 +ω6 )/L(ω2 ) by [Lüb17], so LG (ω2 ) and LG (ω1 +ω6 ) are orthogonal by Proposition 2.2 (iv). We show next that L(ω4 ) is orthogonal. Now W = One can also construct a non-degenerate G-invariant quadratic form on LG (ω4 ) explicitly by realizing it as the space of trace zero elements in the Albert algebra. The details of this construction can be found in [Wil09, 4.8.4, pg. 151-152]. 4 33 LG (ω2 ) is self-dual and dim W = 78, so by Lemma 6.2 (ii) it will be enough to prove that ∧2 (W ) has exactly one trivial composition factor and that LG (ω4 ) occurs in ∧2 (W ) with odd multiplicity. Now VG (ω2 ) = LG (ω2 ), and then a computation with Magma [BCP97] (or [Don85, 8.12, pg. 136]) shows that ch ∧2 (W ) = ch VG (ω2 ) + ch VG (ω4 ). From [Lüb17], we can deduce that the composition factors of VG (ω4 ) are LG (ω4 ), LG (ω1 + ω6 ), LG (ω1 + ω6 ), LG (ω2 ) and LG (0). Thus LG (0) and LG (ω4 ) both occur exactly once as a composition factor of ∧2 (W ). Consider next G of type E7 . We can assume that G is simply connected. Then the Weyl module VG (ω1 ) is the Lie algebra of G, and LG (ω1 ) is not orthogonal by [GW95, Theorem 3.4 (a)]. We have VG (ω2 ) = LG (ω2 ), VG (ω5 ) = LG (ω5 )/LG (ω1 + ω7 ) and VG (ω7 ) = LG (ω7 ) by the data in [Lüb17]. Therefore LG (ω2 ), LG (ω5 ) and LG (ω7 ) are orthogonal by Proposition 2.2 (iv). We show that LG (ω6 ) is orthogonal. Now for W = LG (ω7 ) we have dim W = 56, so by Lemma 6.2 (ii), it will be enough to prove that ∧2 (W ) has exactly two trivial composition factors and that LG (ω6 ) occurs in ∧2 (W ) with odd multiplicity. Now VG (ω7 ) = LG (ω7 ), so by a computation with Magma [BCP97] we see ch ∧2 (W ) = ch VG (ω6 ) + ch VG (0). From [Lüb17], we see that VG (ω6 ) has composition factors LG (ω6 ), LG (ω1 ), LG (ω1 ) and LG (0). Therefore ∧2 (W ) has exactly two trivial composition factors and LG (ω6 ) occurs exactly once as a composition factor. For G of type E8 , we have VG (ω8 ) = LG (ω8 ) and so LG (ω8 ) is orthogonal by 2.2 (iv). Finally, we show that LG (ω1 ) and LG (ω7 ) are orthogonal. For W = LG (ω8 ) we have dim W = 248, so by Lemma 6.2 (ii), it will be enough to prove that ∧2 (W ) has exactly two trivial composition factors and that LG (ω1 ) and LG (ω7 ) occur in ∧2 (W ) with odd multiplicity. By a computation with Magma [BCP97] we see that ch ∧2 (W ) = ch VG (ω8 ) + ch VG (ω7 ). From [Lüb17], we see that VG (ω7 ) has composition factors LG (ω7 ), LG (ω1 ), LG (ω8 ), LG (0), and LG (0). Therefore ∧2 (W ) has exactly two trivial composition factors and both LG (ω1 ) and LG (ω7 ) occur with multiplicity one. 7 Applications and further work In this section, we describe consequences of some of our findings and propose some questions motivated by Problem 1.1. Unless otherwise mentioned, we let G be a simply connected algebraic group over K and we assume that char K = 2. 34 7.1 Connection with representations of the symmetric group Denote the symmetric group on n letters by Σn . We will describe a connection between orthogonality of certain irreducible K[Σn ]-representations and the irreducible representations L(ωr ) of Sp2l (K). This is done by an application of Proposition 3.1 and various results from the literature. The result is not too surprising, since the representation theory of the symmetric group plays a key role in the representation theory of the modules L(ωr ) of Sp2l (K). For example, many of the results that we applied in the proof of Proposition 3.1 above are based on studying certain K[Σn ]-representations associated with V (ωr ). It is well known that there exists an embedding Σ2l+1 < Sp2l (K) = G for all l ≥ 2 (see e.g. [GK99] or [Tay92, Theorem 8.9]). Therefore if a representation V of G is orthogonal, it is clear that the same is true for the restriction V ↓ Σ2l+1 . We will proceed to show that the converse is also true when V = L(ωr ) for 2 ≤ r ≤ l, which does not seem to be a priori obvious. First of all, the following result due to Gow and Kleshchev [GK99, Theorem 1.11] gives the structure of L(ωr ) ↓ Σ2l+1 . Theorem 7.1. Let 1 ≤ r ≤ l. Then the restriction L(ωr ) ↓ Σ2l+1 is irreducible, and it is isomorphic to the irreducible K[Σn ]-module D (2l+1−r,r) labeled by the partition (2l + 1 − r, r) of 2l + 1. Now Gow and Quill have determined in [GQ04] when the irreducible K[Σn ]modules D (n−r,r) are orthogonal. Their result is the following. Theorem 7.2. Let 0 ≤ r ≤ n. Then the K[Σn ]-module D (n−r,r) is not orthogonal if and only if r = 2j , j ≥ 0 and n ≡ k mod 2j+2 for some 2j+1 + 2j − 1 ≤ k ≤ 2j+2 − 2. In the case where n = 2l + 1, one can express the result in the following way. Corollary 7.3. Let n = 2l + 1 and 1 ≤ r ≤ l. Then K[Σn ]-module D (n−r,r) is not orthogonal if and only if r = 2i+1 , i ≥ 0 and l + 1 ≡ 2i+1 + 2i + t mod 2i+2 for some 0 ≤ t < 2i . Proof. By Theorem 7.2, the module D (n−1,1) is not orthogonal if and only if 2l + 1 ≡ 2 mod 4, which never happens. Therefore D (n−1,1) is always orthogonal, as desired. Consider then r > 1. According to Theorem 7.2, if D (n−r,r) is not orthogonal, then r = 2j for some j > 0. In this case D (n−r,r) is not orthogonal if and only if 2l + 1 ≡ k mod 2j+2 for some 2j+1 + 2j − 1 ≤ k ≤ 2j+2 − 2. This 35 is equivalent to saying that 2(l + 1) ≡ k mod 2j+2 for some 2j+1 + 2j ≤ k ≤ 2j+2 − 1. Now this condition is equivalent to l + 1 ≡ k mod 2j+1 for some 2j + 2j−1 ≤ k ≤ 2j+1 − 2, giving the claim. Finally combining Theorem 7.1, Proposition 3.1 and Corollary 7.3 gives the following result. Proposition 7.4. Let 2 ≤ r ≤ l. Then L(ωr ) is orthogonal for Sp2l (k) if and only if L(ωr ) ↓ Σ2l+1 is orthogonal for Σ2l+1 . 7.2 Reduction for Problem 1.1 To determine which irreducible G-modules are orthogonal, it is enough to consider LG (λ) with λ ∈ X(T )+ a 2-restricted dominant weight (Remark 2.5). For groups of exceptional type, this leaves finitely many λ to consider. For groups of classical type, we can further reduce the question to G of type Al and type Cl . This follows from the next two lemmas. Note that in Lemma 7.5, we identify the fundamental dominant weights of Bl and Cl by abuse of notation. P Lemma 7.5. Let λ = li=1 ai ωi , where l ≥ 2 and ai ∈ {0, 1} for all 1 ≤ i ≤ l. (i) Let G be of type Bl or Cl . If al = 1, then V = LG (λ) is orthogonal, except when l = 2 and λ = ω2 . (ii) The irreducible Bl -representation LBl (λ) is orthogonal if and only if the irreducible Cl -representation LCl (λ) is orthogonal. Proof. (i) Let ϕ : Bl → Cl be the usual exceptional isogeny between simply connected groups of type Bl and Cl [Ste68, Theorem 28]. Then LCl (λ) ∼ = LBl ( ϕ l−1 X ai ωi + 2al ωl ) ∼ = LBl ( l−1 X ai ωi ) ⊗ LBl (al ωl )F i=1 i=1 where the last equality follows by Steinberg’s tensor product theorem. Assume that al = 1. Note that a Cl -module V is orthogonal if and only if V ϕ is an orthogonal Bl -module. Thus it follows from Lemma 2.4 that LCl (λ) is orthogonal, except possibly when λ = ωl . Finally, we know that LCl (ωl ) is orthogonal if and only if l ≥ 3 by Example 3.4. This proves the claim for G of type Cl . 36 For G of type Bl , let τ : Cl → Bl be the usual exceptional isogeny between simply connected groups of type Cl and Bl . Then LBl (λ) ∼ = LCl ( τ l−1 X 2ai ωi + al ωl ) ∼ = LCl ( l−1 X ai ωi )F ⊗ LCl (al ωl ), i=1 i=1 and now the claim follows as in the type Cl case. (ii) If al = 1, the claim follows from (i). If al = 0, then LCl (λ)ϕ ∼ = LBl (λ) and the claim follows since a Cl -module V is orthogonal if and only if V ϕ is an orthogonal Bl -module. Lemma 7.6. Let λ = Pl i=1 ai ωi , where l ≥ 4 and ai ∈ {0, 1} for all 1 ≤ i ≤ l. (i) If al−1 6= al , then LDl (λ) is orthogonal if it is self-dual. (ii) If al−1 = al , then LDl (λ) is orthogonal if and only if LCl ( orthogonal, except when λ = ω1 . Pl−1 i=1 ai ωi ) is Proof. (i) If al−1 6= al , then for example by [Hum72, Table 13.1] we see that for G of type Dl , the weight λ is not a sum of roots. Therefore LG (0) cannot be a composition factor of VG (λ), and thus LG (λ) is orthogonal by Proposition 2.2 (iv). (ii) Suppose that al−1 = al . Considering Dl <PCl as the subsystem subgroup ∼ generated by long roots, we have LCl ( l−1 i=1 ai ωi ) ↓ Dl = LDl (λ) by [Sei87, Theorem 4.1]. Now the claim follows from Lemma 4.1. 7.3 Application to maximal subgroups of classical groups In this subsection only, we allow char K to be arbitrary. As mentioned in the introduction, one motivation for Problem 1.1 is in the study of maximal closed connected subgroups of classical groups. Let Cl(V ) be a classical simple algebraic group, that is, Cl(V ) = SL(V ), Cl(V ) = Sp(V ), or Cl(V ) = SO(V ). Finding maximal closed connected subgroups of Cl(V ) can be reduced to the representation theory of simple algebraic groups. We proceed to explain how this is done. For more details, see [LS98] and [Sei87]. In [LS98], certain collections C1 , . . ., C6 of geometric subgroups were defined in terms of the natural module V and its geometry. A reduction theorem due to Liebeck and Seitz [LS98, Theorem 1] implies that for a positivedimensional maximal closed subgroup X of Cl(V ) one of the following holds: 37 (i) X belongs to some Ci , (ii) The connected component X ◦ is simple, and V ↓ X ◦ is irreducible and tensor-indecomposable. In particular, the reduction theorem implies the following. Theorem 7.7. Let X < Cl(V ) be a subgroup maximal among the closed connected subgroups of Cl(V ). Then one of the following holds: (i) X is contained in a member of some Ci , (ii) X is simple, and V ↓ X is irreducible and tensor-indecomposable. The maximal closed connected subgroups in case (i) of Theorem 7.7 are well understood [Sei87, Theorem 3]. Furthermore, the maximal closed connected subgroups occurring in case (ii) of Theorem 7.7 can also be described. These were essentially determined by Seitz [Sei87] and Testerman [Tes88]. The result can be stated in the following theorem, which tells when an irreducible tensor-indecomposable subgroup is not maximal. Theorem 7.8. Let Y be a simple algebraic group and let V be a non-trivial irreducible tensor-indecomposable p-restricted and rational Y -module. If X is a closed proper connected subgroup of Y such that X is simple and V ↓ X is irreducible, then (X, Y, V ) occurs in [Sei87, Table 1]. To refine the characterization of maximal closed connected subgroups of Cl(V ) given in [Sei87, Theorem 3], one should determine which of SL(V ), Sp(V ) and SO(V ) contain X and Y in Theorem 7.8. For example, let Y be simple of type D5 and let X < Y be simple of type B4 embedded in the usual way. Then for V = LY (ω5 ) we have V ↓ X = LX (ω4 ); this situation corresponds to entry IV1 in [Sei87, Table 1]. Here V is not self-dual as a Y -module, so Y < SL(V ) only. However, V ↓ X is self-dual and X < SO(V ) if p 6= 2, and X < SO(V ) < Sp(V ) if p = 2 (see Table 2.1 and Theorem 4.2). In this situation Y is maximal in SL(V ), while X is maximal in SO(V ). In fact, the results we have presented in this text allow one to determine for almost all (X, Y, V ) occurring in [Sei87, Table 1] whether V ↓ X and V ↓ Y are orthogonal, symplectic, both, or neither. If p 6= 2, then this is easily done using Table 2.1. For p = 2, we list this information in Table 7.1, which is deduced as follows. Entry IV1 is a consequence of Lemma 7.5, Example 3.4 and Lemma 7.6. In entry S3 , we have V ↓ Y = LC3 (ω2 ) which is orthogonal by Example 38 3.2, and thus V ↓ X is also orthogonal. In entry S4 , we have V ↓ Y = LC3 (ω1 + ω2 ), which is orthogonal by Proposition 2.2 (iv) since VC3 (ω1 + ω2 ) is irreducible. Entry S6 follows from Lemma 2.4 and Lemma 7.6. In entries S7 , S8 , and S9 , we have V ↓ X = LX (λ) and VX (λ) is irreducible, so V ↓ X is orthogonal by Proposition 2.2 (iv). Entries MR2 and MR3 follow from Proposition 6.3, which show that V ↓ Y is orthogonal. Entry MR5 is a consequence of Lemma 7.5 (i). No. X <Y V ↓X IV1 Bl < Dl+1 orthogonal S3 S4 G2 < C 3 G2 < C 3 orthogonal orthogonal P S6 Bn1 · · · Bnk < D1+ S7 S8 S9 MR2 MR3 MR4 MR5 A3 < D7 D4 < D13 C4 < D13 D4 < F4 C4 < F4 Dl < C l Bn1 · · · Bnk < Bn1 +···+nk ni orthogonal orthogonal orthogonal orthogonal orthogonal orthogonal ? orthogonal V ↓Y l + 1 even: orthogonal l + 1 odd: not self-dual orthogonal orthogonal P 1 + P ni even: orthogonal 1 + ni odd: not self-dual not self-dual not self-dual not self-dual orthogonal orthogonal ? orthogonal Table 7.1: Invariant forms on V ↓ X and V ↓ Y for (X, Y, V ) occurring in [Sei87, Table 1] in the case p = 2. What remains is the entry MR4 from [Sei87, Table 1]. Here X = Dl (l ≥ 4) embeddedP in Y = Cl as the subsystem subgroup of longProots, and l−2 we have V = LY ( l−1 i=1 ai ωi ) with ai ∈ {0, 1}, and V ↓ X = LX ( i=1 ai ωi + al−1 (ωl−1 + ωl )). In this situation we do not know in general whether V ↓ Y and Pl−1V ↓ X are orthogonal, but we do know that except in the case where i=1 ai ωi = ω1 , it is true that V ↓ Y is orthogonal if and only if V ↓ X is orthogonal (Lemma 7.6). Using this fact and the information in Table 7.1, we can deduce the following result. Theorem 7.9. Let Y be a simple algebraic group and let V be a non-trivial irreducible tensor-indecomposable p-restricted Y -module. If X is a closed proper connected subgroup of Y such that X is simple and V ↓ X is irreducible, then one of the following holds. 39 (i) V ↓ Y is not self-dual. (ii) Both V ↓ Y and V ↓ X are orthogonal. (iii) Neither of V ↓ Y or V ↓ X is orthogonal. (iv) p = 2, X is of type Dl , Y is of type Cl and V is the natural module of Y. 7.4 Fundamental self-dual irreducible representations Among the irreducible self-dual G-modules that are not orthogonal, so far the only ones that we know of are in some sense minimal among the self-dual irreducible modules of G. We make this more precise in what follows, and pose the question whether any other examples can be found. Recall that LG (λ) is self-dual if and only if λ = −w0 (λ), where w0 is the longest element in the Weyl group. We know that any dominant weight λ ∈ X(T )+ can be written uniquely as a sum of fundamental dominant Pl weights, that is, λ = i=1 ai ωi for unique integers ai ≥ 0. Now similarly, there exists a collection µ1 , . . . , µt ∈ X(T )+ such that µi = −w0 (µi ) for all i, and such that any λ ∈ X(T )+ with λ = −w0 (λ) can be written uniquely as P t i=1 ai µi with ai ≥ 0. For each simple type, these µi are listed below. • Type Al (l odd): µi = ωi + ωl+1−i for 1 ≤ i ≤ l−1 , 2 and µ l+1 = ω l+1 . 2 2 • Type Al (l even): µi = ωi + ωl+1−i for 1 ≤ i ≤ 2l . • Types Bl , Cl , Dl (l even), G2 , F4 , E7 , and E8 : µi = ωi for 1 ≤ i ≤ rank G. • Type Dl (l odd): µi = ωi for 1 ≤ i ≤ l − 2, and µl−1 = ωl−1 + ωl . • Type E6 : µ1 = ω1 + ω6 , µ2 = ω2 , µ3 = ω3 + ω5 , and µ4 = ω4 . Currently the only known examples of non-trivial irreducible modules LG (λ) that are self-dual and not orthogonal are of the form LG (µi ). Are there any others? Problem 7.10. Let λ ∈ X(T )+ be 2-restricted and suppose that λ = λ1 + λ2 , where λi ∈ X(T )+ are nonzero and −w0 (λi ) = λi . Is LG (λ) orthogonal? If the answer to Problem 7.10 is yes, then our results would settle Problem 1.1 almost completely. Indeed, a positive answer to Problem 7.10 would show that any non-orthogonal self-dual irreducible representation of G must be 40 equal to a Frobenius twist of LG (µi ) for some i. Our results determine the orthogonality of LG (µi ) when G is of classical type. The non-orthogonal ones for type Al are the LAl (ωi + ωl+1−i ) described in Theorem 5.1. For G of type Bl , Cl , or Dl , the non-orthogonal ones are LG (ωi ) as described in Theorem 4.2, with the unique exception of LG (ω4 + ω5 ) for G of type D5 (arising from restriction of LC5 (ω4 ) to G). Then a handful of µi still remain for exceptional types. For G simple of exceptional type, the irreducibles LG (µi ) whose orthogonality was not decided in Section 6 are as follows. • LG (ω3 + ω5 ) for G of type E6 , • LG (ω3 ) and LG (ω4 ) for G of type E7 , • LG (ωi ) for 2 ≤ i ≤ 6 for G of type E8 . In any case, a natural next step towards solving Problem 1.1 should be determining an answer to Problem 7.10. The methods we have used in this paper to solve Problem 1.1 for certain families of LG (λ) rely heavily on detailed information about the structure of the Weyl module VG (λ), which is not known in general. For small-dimensional representations the composition factors of VG (λ) can be found using the results of Lübeck given in [Lüb01] and [Lüb17]. However, in general this sort of information is not available, and in characteristic 2 the composition factors of VG (λ) are known only in a relatively few cases. For example, for G of type E8 we do not even know the dimension of LG (ωi ) for all i in characteristic 2. 7.5 Fixed point spaces of unipotent elements We finish by a question about a possible orthogonality criterion for irreducible representations. Let ϕ : G → SL(V ) be a non-trivial irreducible representation of G. Assume that V is self-dual, so that ϕ(G) < Sp(V ) (Lemma 2.1). If V is an orthogonal G-module, then ϕ(G) < O(V ) and so ϕ(G) < SO(V ) since G is connected. Then for any unipotent element u ∈ G, the number of Jordan blocks of ϕ(u) is even [LS12, Proposition 6.22]. In other words, for all u ∈ G we have that dim V u is even, where V u is the subspace of fixed points for u. Does the converse hold? Problem 7.11. Let V be a non-trivial irreducible self-dual representation of G. If V is not orthogonal, does there exist a unipotent element u ∈ G such that dim V u is odd? 41 In Table 7.2, we have listed examples (without proof) of some nonorthogonal representations V of G for which the answer to Problem 7.11 is yes. If the answer to Problem 7.11 turns out to be yes, we would have an interesting criterion for an irreducible representation V of G to be orthogonal. A positive answer would show that the orthogonality of an irreducible representation can be decided from the properties of individual elements of G. Type of G Al , l + 1 ≡ 2 mod 4 Cl Cl , l ≡ 2 mod 4 A4 A5 C5 C6 G2 E7 V LG (ω1 + ωl ) LG (ω1 ) LG (ω2 ) LG (ω2 + ω3 ) LG (ω2 + ω4 ) LG (ω4 ) LG (ω4 ) LG (ω1 ) LG (ω1 ) Conjugacy class of u regular regular regular A3 regular regular in D5 regular regular regular dim V u 2l + 1 1 l−1 19 21 21 25 1 7 Table 7.2: Non-orthogonal irreducible representations V of G with examples of dim V u odd for some unipotent element u ∈ G. References References [Ada84] A. M. Adamovich. An analogue of the space of primitive forms over a field of positive characteristic. Vestnik Moskov. Univ. Ser. I Mat. Mekh., (1):64–66, 1984. [Ada86] A. M. Adamovich. The submodule lattices of Weyl modules of symplectic groups with fundamental highest weights. Vestnik Moskov. Univ. Ser. I Mat. Mekh., (2):7–10, 112, 1986. [Ada92] A. M. Adamovich. Composition factors and submodules of Weyl Modules. PhD thesis, Moscow State University, 1992. in Russian. [AJ84] Henning Haahr Andersen and Jens Carsten Jantzen. Cohomology of induced representations for algebraic groups. Math. Ann., 269(4):487–525, 1984. 42 [BCP97] Wieb Bosma, John Cannon, and Catherine Playoust. The Magma algebra system. I. The user language. J. Symbolic Comput., 24(34):235–265, 1997. Computational algebra and number theory (London, 1993). [Bor70] Armand Borel. Properties and linear representations of Chevalley groups. In Seminar on Algebraic Groups and Related Finite Groups (The Institute for Advanced Study, Princeton, N.J., 1968/69), Lecture Notes in Mathematics, Vol. 131, pages 1–55. Springer, Berlin, 1970. [Bou59] N. Bourbaki. Éléments de mathématique. Première partie: Les structures fondamentales de l’analyse. Livre II: Algèbre. Chapitre 9: Formes sesquilinéaires et formes quadratiques. Actualités Sci. Ind. no. 1272. Hermann, Paris, 1959. [Bou75] N. Bourbaki. Éléments de mathématique. Fasc. XXXVIII: Groupes et algèbres de Lie. Chapitre VII: Sous-algèbres de Cartan, éléments réguliers. Chapitre VIII: Algèbres de Lie semi-simples déployées. Actualités Scientifiques et Industrielles, No. 1364. Hermann, Paris, 1975. [Bro92] A.E. Brouwer. The composition factors of the Weyl modules with fundamental weights for the symplectic group. Unpublished, http://www.win.tue.nl/~aeb/preprints.html, 1992. [BS00] A. A. Baranov and I. D. Suprunenko. Branching rules for modular fundamental representations of symplectic groups. Bull. London Math. Soc., 32(4):409–420, 2000. [BS05] A. A. Baranov and I. D. Suprunenko. Modular branching rules for 2column diagram representations of general linear groups. J. Algebra Appl., 4(5):489–515, 2005. [BW89] Joel G. Broida and S. Gill Williamson. A comprehensive introduction to linear algebra. Addison-Wesley Publishing Company, Advanced Book Program, Redwood City, CA, 1989. [DB09] Bart De Bruyn. Some subspaces of the kth exterior power of a symplectic vector space. Linear Algebra Appl., 430(11-12):3095–3104, 2009. [DB10] Bart De Bruyn. On the Grassmann modules for the symplectic groups. J. Algebra, 324(2):218–230, 2010. 43 [Don85] Stephen Donkin. Rational representations of algebraic groups, volume 1140 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1985. Tensor products and filtration. [FH91] William Fulton and Joe Harris. Representation theory, volume 129 of Graduate Texts in Mathematics. Springer-Verlag, New York, 1991. A first course, Readings in Mathematics. [Fon74] Paul Fong. On decomposition numbers of J1 and R(q). In Symposia Mathematica, Vol. XIII (Convegno di Gruppi e loro Rappresentazioni, INDAM, Rome, 1972), pages 415–422. Academic Press, London, 1974. [GK99] Roderick Gow and Alexander Kleshchev. Connections between the representations of the symmetric group and the symplectic group in characteristic 2. J. Algebra, 221(1):60–89, 1999. [GN16] Skip Garibaldi and Daniel K. Nakano. Bilinear and Quadratic Forms on Rational Modules of Split Reductive Groups. Canad. J. Math., 68(2):395–421, 2016. [Gow97] Rod Gow. Contraction of exterior powers in characteristic 2 and the spin module. Geom. Dedicata, 64(3):283–295, 1997. [GQ04] Rod Gow and Patrick Quill. The quadratic type of certain irreducible modules for the symmetric group in characteristic two. J. Algebra, 277(2):533–541, 2004. [GW95] Roderick Gow and Wolfgang Willems. Methods to decide if simple self-dual modules over fields of characteristic 2 are of quadratic type. J. Algebra, 175(3):1067–1081, 1995. [GW97] Rod Gow and Wolfgang Willems. On the quadratic type of some simple self-dual modules over fields of characteristic two. J. Algebra, 195(2):634–649, 1997. [Hum72] James E. Humphreys. Introduction to Lie algebras and representation theory. Springer-Verlag, New York-Berlin, 1972. Graduate Texts in Mathematics, Vol. 9. [Jan73] Jens Carsten Jantzen. Darstellungen halbeinfacher algebraischer Gruppen und zugeordnete kontravariante Formen. Bonn. Math. Schr., (67):v+124, 1973. 44 [Jan03] Jens Carsten Jantzen. Representations of algebraic groups, volume 107 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, second edition, 2003. [KL90] Peter Kleidman and Martin Liebeck. The subgroup structure of the finite classical groups, volume 129 of London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, 1990. [KS99] A. S. Kleshchev and J. Sheth. On extensions of simple modules over symmetric and algebraic groups. J. Algebra, 221(2):705–722, 1999. [KS01] A. S. Kleshchev and J. Sheth. Corrigendum: “On extensions of simple modules over symmetric and algebraic groups” [J. Algebra 221 (1999), no. 2, 705–722; MR1728406 (2001f:20091)]. J. Algebra, 238(2):843–844, 2001. [LMS79] V. Lakshmibai, C. Musili, and C. S. Seshadri. Geometry of G/P . Bull. Amer. Math. Soc. (N.S.), 1(2):432–435, 1979. [LS96] Martin W. Liebeck and Gary M. Seitz. Reductive subgroups of exceptional algebraic groups. Mem. Amer. Math. Soc., 121(580):vi+111, 1996. [LS98] Martin W. Liebeck and Gary M. Seitz. On the subgroup structure of classical groups. Invent. Math., 134(2):427–453, 1998. [LS12] Martin W. Liebeck and Gary M. Seitz. Unipotent and nilpotent classes in simple algebraic groups and Lie algebras, volume 180 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2012. [Lüb01] Frank Lübeck. Small degree representations of finite Chevalley groups in defining characteristic. LMS J. Comput. Math., 4:135– 169 (electronic), 2001. [Lüb17] Frank Lübeck. Tables of weight multiplicities. http://www.math.rwth-aachen.de/~Frank.Luebeck/chev/WMSmall/index.html, 2017. Accessed: 2017. [Mat90] Olivier Mathieu. Filtrations of G-modules. Ann. Sci. École Norm. Sup. (4), 23(4):625–644, 1990. [McN98] George J. McNinch. Dimensional criteria for semisimplicity of representations. Proc. London Math. Soc. (3), 76(1):95–149, 1998. 45 [PS83] A. A. Premet and I. D. Suprunenko. The Weyl modules and the irreducible representations of the symplectic group with the fundamental highest weights. Comm. Algebra, 11(12):1309–1342, 1983. [Ree57] Rimhak Ree. On some simple groups defined by C. Chevalley. Trans. Amer. Math. Soc., 84:392–400, 1957. [Sei87] Gary M. Seitz. The maximal subgroups of classical algebraic groups. Mem. Amer. Math. Soc., 67(365):iv+286, 1987. [Ste68] Robert Steinberg. Lectures on Chevalley groups. Yale University, New Haven, Conn., 1968. Notes prepared by John Faulkner and Robert Wilson. [Sup95] Irina D. Suprunenko. Irreducible representations of simple algebraic groups containing matrices with big Jordan blocks. Proc. London Math. Soc. (3), 71(2):281–332, 1995. [SW91] Peter Sin and Wolfgang Willems. G-invariant quadratic forms. J. Reine Angew. Math., 420:45–59, 1991. [Tay92] Donald E. Taylor. The geometry of the classical groups, volume 9 of Sigma Series in Pure Mathematics. Heldermann Verlag, Berlin, 1992. [Tes88] Donna M. Testerman. Irreducible subgroups of exceptional algebraic groups. Mem. Amer. Math. Soc., 75(390):iv+190, 1988. [Wan82] Jian Pan Wang. Sheaf cohomology on G/B and tensor products of Weyl modules. J. Algebra, 77(1):162–185, 1982. [Wil77] Wolfgang Willems. Metrische G-Moduln über Körpern der Charakteristik 2. Math. Z., 157(2):131–139, 1977. [Wil09] Robert A. Wilson. The finite simple groups, volume 251 of Graduate Texts in Mathematics. Springer-Verlag London, Ltd., London, 2009. 46
4math.GR
Solving Hard Stable Matching Problems Involving Groups of Similar Agents arXiv:1708.04109v1 [cs.GT] 14 Aug 2017 Kitty Meeks School of Computing Science University of Glasgow Glasgow, UK [email protected] Baharak Rastegari Department of Computer Science University of Bristol Bristol, UK [email protected] Abstract Many important stable matching problems are known to be NP-hard, even when strong restrictions are placed on the input. In this paper we seek to identify simple structural properties of instances of stable matching problems which will allow the design of efficient algorithms. We focus on the setting in which all agents involved in some matching problem can be partitioned into k different types, where the type of an agent determines his or her preferences, and agents have preferences over types (which may be refined by more detailed preferences within a single type). This situation could arise in practice if agents form preferences based on some small collection of agents’ attributes. The notion of types could also be used if we are interested in a relaxation of stability, in which agents will only form a private arrangement if it allows them to be matched with a partner who differs from the current partner in some particularly important characteristic. We show that in this setting several well-studied NP-hard stable matching problems (such as MAX SMTI, MAX SRTI, and MAX SIZE MIN BP SMTI) belong to the parameterised complexity class FPT when parameterised by the number of different types of agents, and so admit efficient algorithms when this number of types is small. 1 Introduction Matching problems occur in various applications and scenarios such as the assignment of children to schools, college students to dorm rooms, junior doctors to hospitals, and so on. In all the aforementioned, and similar, problems, is it understood that the participants (which we will refer to as agents) have preferences over other agents, or subsets of agents. The majority of the literature assumes that these preferences are ordinal, and that is the assumption we make in this work as well. Moreover, it is widely accepted that a “good” and “reasonable” solution to a matching problem must be stable, where stability is defined according to the context of the problem at hand. Intuitively speaking, a stable solution guarantees that no subset of agents find it in their best interest to leave the prescribed 1 solution and seek an assignment amongst themselves. Unfortunately, many interesting and important stable matching problems are known to be NP-hard even for highly restricted cases. In this work, we consider a setting where agents can be grouped into k different “types”, where the type of an agent specifies what the agent’s preferences are and how s/he is compared against other agents. Such a setting could arise in practice if, for example, agents derive their preferences by considering some small collection of attributes of other agents. The notion of types is also useful if we are interested in a relaxation of stability, where agents are only willing to form a private arrangement with a partner who is distinctly superior to their current partner with respect to an important characteristic. We show that we can solve some of the most important hard stable matching problems efficiently from the point of view of parameterized complexity when the number of types is taken as the parameter; many of these results rely on the fixed parameter tractability of Integer Linear Programming. We also demonstrate that, by imposing further restrictions, some of these problems become polynomial-time solvable. Perhaps the most widely studied matching problem is the Stable Marriage problem (SM). In an instance of SM we have two disjoint set of agents, men and women, each having a preference ordering over the individuals of the opposite sex (candidates). A solution to this problem is a matching, that is an mapping from men to women where each man is matched to at most one woman and vice versa. A matching is stable if there are no two agents a and b who prefer each other to their assigned partners. If such a pair exists, we say that (a, b) is a blocking pair and both a and b are blocking agents. In their seminal work, Gale and Shapley [20] showed that every stable marriage problem admits a stable matching that can be found in polynomial time by their proposed algorithm (GS). Simple extensions of GS can be used to identify stable matchings in domains where agents are permitted to declare some candidates unacceptable (Stable Marriage with Incomplete lists (SMI)), are allowed to express preferences in which they are indifferent between two or more agents (Stable Marriage with Ties (SMT)), or both (Stable Marriage with Ties and Incomplete lists (SMTI)). It is known that an instance of SMTI might admit stable matchings of different sizes. In many practical applications, it is important to match as many agents as possible, and thus finding a maximum cardinality stable matching (i.e., a stable matching with the largest size amongst all stable matchings) is a crucial issue. Definition 1. Max SMTI is the problem of determining the maximum cardinality stable matching in an instance of SMTI. Depending on the application, one might be willing to tolerate a small degree of instability if that leads to larger matchings. Two different measurements for the degree of instability have been introduced in the literature: the number of blocking pairs and the number of blocking agents. Definition 2. Max Size Min BP SMI (respectively Max Size Min BA SMI) is the problem of finding a matching, out of all maximum cardinality matchings, which has the minimum number of blocking pairs (respectively, minimum number of blocking agents) in an instance of SMI. 2 The Stable Roommate problem (SR) is a non-bipartite generalization of SM. Extensions allowing for incomplete lists and indifference in preference lists are defined the same way as for SM. An instance of SR need not admit a stable matching. There exists a polynomial time algorithm that finds a stable matching in an instance of SR, or reports that none exists [23]. Definition 3. Max SRT is the problem of identifying a maximum cardinality stable matching in an instance of SRT, or reporting that none exists. Since an instance of SR may not admit a stable matching, it is of interest to find a matching with minimum number of blocking pairs. Definition 4. Min BP SR is the problem of identifying a matching which has the minimum number of blocking pairs in an instance of SR. Hospitals/Residents problem with Ties (HRT) is a famous extension of SMTI that models many practical applications, including the assignment of junior doctors to hospitals, by allowing agents on one side of the market to be assigned to multiple agents on the other side of the market. Mechanisms very similar to that proposed in [20] have been used to compute a stable assignment of residents to hospitals. The concern of computing a maximum size stable matching extends to instances of HRT. Definition 5. Max HRT is the problem of determining the maximum cardinality stable matching in an instance of HRT. All the problems defined above are known to be NP-complete [33, 25, 3, 37, 7, 24, 1]. Note that we have provided Definition 2 for instances of SMI. We chose to do so because these two problems are hard even when there are no ties in preference lists. Likewise, Definition 3 assumes that all agents find all other agents acceptable, as the hardness result holds even under this assumption. Definition 4 is defined for instances of SR, as this problem is hard even if all agents rank all the other agents in strict order of preference. In fact, the aforementioned problems are hard even when the input is heavily restricted. For example, Max SMTI is NP-complete even if each man’s preference list is strictly ordered, and each woman’s preference list is either strictly ordered or is a tie of length 2 [33]. Structure of the rest of the paper. In Section 2 we provide definitions for the settings we study, as well as a brief introduction to parameterised complexity and existing results on Integer Linear Programming. In Section 3 we formally define what we mean by agents having types, discuss some related work, and summarise our results. In subsequent sections we present our results on the complexity of computing a maximum cardinality stable matching in an instance of SMTI or HRT (Section 4), a maximum cardinality stable matching in an instance of SRTI (Section 5), and a maximum cardinality matching with minimum number of blocking pairs/agents (Section 6) in our restricted setting. In Section 7 we provide a definition for the Hospitals/Residents problem with Couples (where a pair of residents can send in preferences over pairs of hospitals) and present our result on the complexity of computing maximum cardinality stable matchings in such scenarios. We motivate our setting by discussing how it may arise in practice in Section 8, along with a brief description on how to extract types from ordinal preference lists. We conclude in Section 9 and provide some directions for future work. 3 2 Preliminaries In this section we introduce the main concepts we use in the paper; we begin with some definitions, then provide a brief introduction to parameterised complexity, before describing existing results on Integer Programming. 2.1 Definitions In this section we provide the key definitions for the stable matching settings we study; for further background and terminology we refer the reader to [31]. Let N denote a set of n agents, which in a bipartite matching setting (i.e. SMTI or HRT) is composed of two disjoint sets. Each hospital h in an instance of HRT is associated with a capacity q(h) that denotes the number of posts it offers. When in a bipartite matching setting, we use the term candidates to refer to the agents on the opposite side of the market to that of an agent under consideration. In a non-bipartite settings, candidates refer to all the other agents except the one under consideration. Each agent finds a subset of candidates acceptable and ranks them in order of preference. Preferences orderings need not to be strict, so it is possible for an agent to be indifferent between two or more candidates. We write b ≻a c , or equivalently c ≺a b, to denote that agent a prefers candidate b to candidate c, and b ≃a c to denote that a is indifferent between b and c. We write b a c to denote that a either prefers b to c or is indifferent between them, and say that a weakly prefers b to c. In an instance of SMTI, a matching M is a pairing of men and women such that no one is paired with an unacceptable partner, each man is paired with at most one woman, and each woman is paired with at most one man. In an SRTI instance, a matching is a pairing of agents such that each agent is matched with at most one other agent whom s/he additionally finds acceptable. We write (a, b) ∈ M to say that a and b are matched in M. In an instance of HRT, a matching is a pairing of hospitals and residents such that no agent is paired with an unacceptable candidate, each resident is matched with at most one hospital, and each hospital h is matched with at most q(h) residents. We use M(a) to denote the agent (or the set of agents in the case of hospitals) matched to a in M. We write M(a) = ∅ if agent a is unmatched in M. We assume that every agent prefers being matched to an acceptable candidate to remaining unmatched. Given an instance of SMTI or SRTI, a matching M is (weakly) stable if there is no pair (a, b) ∈ / M where a prefers b to his current partner in M, i.e., b ≻a M(a), and vice versa. Given an instance of HRT, a matching M is stable if there is no acceptable (resident,hospital) pair (r, h) such that (i) r prefers h to M(r), and (ii) either |M(h)| < q(h) or h prefers r to its worst assigned resident in M. 2.2 Parameterised Complexity In this paper we are concerned with the parameterised complexity of computational problems that are intractable in the classical sense. Parameterised complexity provides a multivariate framework for the analysis of hard problems: if a problem is known to be NP-hard, so that we expect the running-time of any algorithm to depend exponentially on some aspect of 4 the input, we can seek to restrict this combinatorial explosion to one or more parameters of the problem rather than the total input size. This has the potential to provide an efficient solution to the problem if the parameter(s) in question are much smaller than the total input size. A parameterised problem with total input size n and parameter k is considered to be tractable if it can be solved by a so-called fpt-algorithm, an algorithm whose running time is bounded by f (k) · nO(1) , where f can be any computable function. Such problems are said to be fixed parameter tractable, and belong to the complexity class FPT. It should be emphasised that, for a problem to be in FPT, the exponent of the polynomial must be independent of the parameter value; problems which satisfy the weaker condition that the running time is polynomial for any constant value of the parameter(s) (so that the degree of the polynomial may depend on the parameters) belong to the class XP. For further background on the theory of parameterised complexity, we refer the reader to [15, 18]. 2.3 The complexity of Integer Programming Most of the algorithms we present in this paper make use of an algorithm for Integer Linear Programming in some way. This problem is formally stated as follows: given an m × k matrix A and two m-dimensional vectors b and c (all with coefficients in Z), find a m-dimensional vector x ∈ Z which minimizes the scalar product cT · x, subject to the m linear constraints given by Ax ≤ b, or else report that no vector satisfying the constraints exists. Note that we can easily translate problems in which we wish to maximise rather than minimise the objective function into this form, and also we can express constraints based on linear equalities as a combination of linear inequalities; for simplicity of presentation we will use both of these generalisations when expressing problems as instances of Integer Linear Programming. While Integer Linear Programming is NP-hard in general, one of the most celebrated results in parameterised complexity is that this problem belongs to FPT when parameterised by the number of variables. Theorem 1 ([14], based on [19, 26, 29]). An Integer Linear Programming instance of size L with k variables can be solved using O(k 2.5k+o(k) · (L + log Mx ) log(Mx Mc )) arithmetic operations and space polynomial in L + log Mx , where Mx is an upper bound on the absolute value a variable can take in a solution, and Mc is the largest absolute value of a coefficient in the vector c. In Section 6 we also need to solve instances of Integer Quadratic Programming, a variant of Integer Linear Programming in which the objective function is quadratic. Formally, given a k × k integer matrix Q, an m × k integer matrix A and an m-dimensional integer vector b, our goal is to find a vector x ∈ Zk which minimises xT Qx, subject to the m linear constraints Ax ≤ b, or else report that no vector satisfying the constraints exists. As before, we note that we can easily generalise this definition to deal with maximisation problems and constraints in the form of linear equalities. Lokshtanov recently gave an fptalgorithm for this problem. 5 Theorem 2 ([30]). Integer Quadratic Programming is in FPT parameterised by k+α, where α is the maximum absolute value of any entry in the matrices A and Q. 3 Our Contribution In this section we describe our approach in more detail, and provide some context for our results. We begin in Section 3.1 by giving formal definitions of the problems we consider, before discussing some related work in Section 3.2. Finally, in Section 3.3, we summarise our results. 3.1 A typed approach Most hardness results in the study of stable matching problems are based on the premise that agents may have arbitrary preference lists. In practice, however, agents’ preferences are likely to be more structured and correlated. In this work, we consider a setting where each agent is associated with a “type” that specifies his or her preferences, as well as how s/he is perceived by the other agents. We discuss how such settings may arise in practice, and how we can efficiently extract types from ordinal preference lists, in Section 8. 3.1.1 Our first model: agents of the same type are indistinguishable The simplest model is to assume that the agents of the same type are completely indistinguishable. That is, they have the same preference lists, and every other agent that finds their type acceptable is indifferent between them. Assume that there are k types available for agents. Let Ni denote the set of agents that are of type i. Each type i has a preference ordering over types of the candidates, which need not be complete or strict. We assume, without loss of generality, that |Ni | > 0 for all i ∈ [k], and that each type finds at least one other type acceptable. We write j ≻i ℓ, or equivalently ℓ ≺i j, if agents of type i strictly prefer agents of type j to agents of type ℓ. We write j ≃i ℓ to denote that agents of type i are indifferent between agents of types j and ℓ, and j i ℓ if agents of type i prefer agents of type j to those of type ℓ or are indifferent between the two. We assume that given every two agents x and y of the same type: 1. x and y have identical preference lists when restricted to N \ {x, y}, and 2. all other agents are indifferent between x and y. These requirements imply that any agent either finds all agents of a given type acceptable (and is indifferent between them) or finds none of them acceptable. We say that an instance of a stable matching problem satisfying these requirements is typed, and refer to the standard problems with input of this form as Typed Max SMTI etc. We illustrate our definition of typed instances with an example. Example 1. Assume we have 6 types for the agents, {1, . . . , 6}, and that we are in a stable marriage setting where types 1, 2 and 3 correspond to men and the rest of the types correspond 6 to women. Let the preferences over types be as follows, where the preference lists are ordered from left to right in decreasing order of preference, and the types in round brackets are tied. 1:45 2 : (4 5) 6 3:564 4 : 3 (2 1) 5:23 6:32 Assume that there seven men, {m1 , . . . , m7 }, and seven women , {w1 , . . . , w7 }, where men m1 and m2 are of type 1, men m3 , m4 and m5 are of type 2, and the last two men are of type 3. Assume that w1 is of type 4, w2 , w3 and w4 are of type 5 and the last three women are of type 6. Therefore, the preferences of the agents under typed model are as follows. m1 , m2 : w1 (w2 w3 w4 ) m3 , m4 , m5 : (w1 w2 w3 w4 ) (w5 w6 w7 ) m6 , m7 : (w2 w3 ) (w5 w6 w7 ) (w1 ) 3.1.2 w1 : (m6 m7 ) (m1 m2 m3 m4 m5 ) w2 , w3 , w4 : (m3 m4 m5 ) (m6 m7 ) w5 , w6 , w7 : (m6 m7 ) (m3 m4 m5 ) A generalisation: agents of the same type refine their preferences in the same way We also consider a generalisation of typed instances in which agents are no longer necessarily indifferent between two agents of the same type, however agents of the same type occur consecutively in preference lists. Moreover, we assume that being indifferent between two types essentially means being indifferent between the agents of either types. This means that for any two agents x and y of the same type i: 1. x and y have identical preference lists when restricted to N \ {x, y}, 2. no agent of a different type appears between x and y in any preference list, and 3. if type i is indifferent between types j and ℓ then x and y are indifferent between all agents in (Nj ∪ Nℓ ) \ {x, y}. If an instance of a stable matching problem satisfies these slightly weaker requirements, we say that the instance is refined-typed, and refer to the standard problems with input of this form as Refined-Typed Max SMTI etc. We illustrate this definition with two short examples. Example 2. Consider the same stable marriage setting with types as in Example 1. While Example 1 already meets the requirements of refined-typed model, there are many more permitted preference profiles in this setting, as agents are allowed to break ties within a particular type. For example, m4 , m5 and m6 can have the following preference list: m3 , m4 , m5 : (w1 w2 w3 w4 ) w6 (w5 w7 ) 7 However, the following preference list is not allowed for these men, as a tie between t4 and t5 requires these three men to be indifferent between all women who are of either of these two types. m3 , m4 , m5 : (w1 w2 ) (w3 w4 ) (w5 w6 w7 ) Example 3. Assume that we are in a stable roommates setting with six agents a, b, . . . , f of type 1. Assume that type 1 finds type 1 acceptable, and let the refined preferences within type 1 be as follows: (a b c) (d e f ) The agents’ preference lists will then be as listed below. a : (b c) (d e f ) d : (a b c) (e f ) 3.2 b : (a c) (d e f ) e : (a b c) (d f ) c : (a b) (d e f ) f : (a b c) (d e) Related Work NP-hard matching problems. As stated earlier, all the problems defined in Section 1 are NP-complete, even for highly restricted cases. The NP-completeness of Max SMTI has been shown for a variety of restricted settings, for example: (1) even if each man’s list is strictly ordered, and each woman’s list is either strictly ordered or is a tie of length 2 [33], (2) even if each mans preference list is derived from a strictly-ordered master list of women, and each woman’s preference list is derived from a master list of men that contains only one tie [25], and (3) even if the SMTI instance has symmetric preferences; that is, for any acceptable (man, woman) pair (mi , wj ), rank(mi , wj ) = rank(wj , mi ) [3, 37], where rank(a, b) is defined to be one plus the number of candidates that a prefers to b. As SMTI is a special case of HRT, the NP-hardness of finding a maximum stable matching in the latter follows directly from the NP-hardness of this problem in the former. Weak SRT, the problem of deciding weather a stable matching exists in an instance of SRT, is NP-complete [24]. The NP-completeness holds even if each preference list is either strictly ordered or contains a tie of length 2 at the head. It therefore follows that Max SRT is NP-complete. Max Size Min BP SMI and Max Size Min BA SMTI are NP-hard and very hard to approximate [7] even if each agent’s preference list is of length at most 3 [8, 22], but polynomial-time solvable if agents on one side of the market have preference lists of length at most 2 [8]. Parameterized complexity. There are a limited number of works addressing fixed-parameter tractability in the stable matching problems we study in this paper. Marx and Schlotter [34] study the parameterized complexity of Max SMTI. They show that the problem is in FPT when parameterised by the total length of the ties, but is W[1]-hard when parameterised by the number of ties in the instance, even if all the men have strictly ordered preference lists. 8 Attributes and types. Settings in which agents are partitioned into different types, or derive their preferences based on a set of attributes assigned to each candidate, have been considered for the problems of sampling and counting stable matchings in instances of SM or SR (see, e.g., [5, 11, 12]). In [17], the authors study the problem of characterizing matchings that are rationalisable as stable matchings when agents’ preferences are unobserved. They focus on a restricted setting that translates into assigning each agent a type based on several attributes, and assuming that agents of the same type are identical and have identical preferences. They remark that empirical studies on marriage typically make such an assumption [13]. Bounded agent types have been considered in [40, 4] to derive polynomial-time results for the coalition structure generation problem, an important issue in cooperative games when the goal is to partition the participants into exhaustive and disjoint coalitions in order to maximize the social welfare. 3.3 Our Results Observe that the hardness results discussed in the previous section also hold even in the settings of typed and refined-typed models: if we let k = n (so each type contains just one agent) then the definitions of our models do not place any restrictions on the preference lists. Thus we can deduce that (Refined-)Typed Max SMTI, (Refined-)Typed Max HRT, (Refined-)Typed Max SRT, (Refined-)Typed Max Size Min BP SMI and (Refined-)Typed Max Size Min BA SMI are all NP-complete when k is part of the input. In contrast with these hardness results, we are able to provided positive results in the parameterised setting when the number of types is taken to be the parameter. Theorem. (Refined-)Typed Max SMTI, (Refined-)Typed Max HRT, (Refined-)Typed Max SRTI, (Refined-)Typed Max Size Min BP SMTI and (Refined-)Typed Max Size Min BA SMTI all belong to FPT when parameterised by the number of types. Note that in some cases our FPT result applies to a more general version than that which is known to be NP-hard. These results are presented in Sections 4, 5 and 6. We also consider a setting where preferences over types are all strict. Certain known hardness results also carry over to this setting if we allow k = n: (Refined-)Typed Max Size Min BP SMI and (Refined-) Typed Max Size Min BA SMI remain NP-hard even if all preferences over types are strict. However, this additional restriction on the input allows us to give polynomial-time algorithms (where k is taken to be part of the input) for (Refined-)Typed Max SMTI, (Refined-)Typed Max HRT and (Refined-)Typed Max SRTI. Theorem. (Refined-)Typed Max SMTI, (Refined-)Typed Max HRT and (Refined-)Typed Max SRTI are polynomial-time solvable if all preferences over types are strict. We prove these results in Sections 4 and 5. 9 4 Efficient Algorithms for Max SMTI and Max HRT In this section we first present our fpt-algorithm for the NP-complete problem Typed Max SMTI, and then show how to extend it to solve Refined-Typed Max SMTI and to give a polynomial-time algorithm for the special case in which preferences over types are all strict. We conclude with presenting a straightforward reduction from Max HRT to Max SMTI that implies we can use our ftp-algorithms for the latter to solve the former. 4.1 An fpt-algorithm for Typed Max SMTI Let I be a typed instance of SMTI, and let M be a matching in I. We may assume without loss of generality that every agent is matched, by creating sufficiently many dummy agents of type k + 1 which are inserted at the end of each men’s and women’s (possibly incomplete) preference list. We define worstM (i) to be the type of the least desirable agent with which any agent of type i is matched in M. Note that worstM (i) would be a dummy type if an agent of type i is unmatched (i.e. matched to a dummy agent) in M. Let type(a) denote the type of a given agent a. We claim that, in order to determine whether or not M is stable, it suffices to examine the values of worstM (i) for each i ∈ [k]. Lemma 3. Let I be a typed instance of SMTI. Then a matching M in I is stable if and only if there is no pair (i, j) ∈ [k](2) such that j ≻i worstM (i) and i ≻j worstM (j). Proof. Suppose first that M is not stable. In this case, by definition, there exists some pair of agents (a, b) such that a and b are not matched together but each prefers the other over their current partner. Suppose without loss of generality that a is of type i and b is of type j. Then we know that j ≻i type(M(a)) i worstM (i), and similarly i ≻j type(M(b)) j worstM (j). Conversely, suppose that M is stable, and suppose for a contradiction that j ≻i worstM (i) and i ≻j worstM (j). Then there is some agent a of type i which is matched with an agent of type worstM (i), so in particular a is matched with a partner less desirable than any agent of type j. Similarly, there is some agent b of type j which is matched with an agent of type worstM (j) and hence is matched with a partner less desirable than any agent of type i. Thus a and b both prefer each other to their current partner, and so form a blocking pair. This contradicts the assumption that M is stable. We now use the above observation to prove our main result of this section. Theorem 4. Typed Max SMTI is in FPT parameterised by the number k of different types in the instance. Proof. Lemma 3 implies that every stable matching must pass the test of Lemma 3, and any matching that passes its test is stable. Note that there are at most (k + 1)k possibilities for the function worst : [k] → [k + 1]. We say that a given function worst is feasible if for each type i, worst(i) is either a type acceptable to type i or the dummy type. We say that a matching M realises a given feasible function worst if, for each i ∈ [k], the least desirable partner any agent of type i has in M is of type worst(i). 10 Our strategy is to consider each of the feasible possibilities for worst in turn. First we determine (using Lemma 3) whether a matching M realising worst will be stable; this can be done for any candidate function worst in time O(k 2 ). For any feasible function worst which does give rise to a stable matching, we now want to determine what is the maximum number of agents which can be matched in any matching which realises worst. We do this by solving a suitable instance of an Integer Linear Programming formulation. For each unordered pair of distinct values {i, j} ∈ [k + 1](2) , the variable n{i,j} represents the number of pairs in the matching consisting of one agent of type i and another of type j. Recall that Ni is the set of agents of type i. We then have the following integer linear program: maximize X n{i,j} 1≤i,j≤k subject to X n{i,j} = |Ni |, ∀i ∈ [k] j∈[k+1] X n{i,j} = |Ni |, ∀i ∈ [k] ji worst(i) n{i,worst(i)} > 0, and n{i,j} ≥ 0. ∀i ∈ [k] ∀i, j ∈ [k + 1] The first constraint ensures that every agent is involved in exactly one pair, perhaps with a dummy agent. The next two conditions ensure that, for each i, worst(i) is indeed the type of the least desirable partner assigned to any agent of type i. The objective function seeks to maximise the total number of pairs that do not involve dummy agents. The above integer linear program has 3k + (k + 1)2 constraints (where each constraint has constant length) and (k+1)2 variables. The upper bound on the absolute value a variable can take is n. Therefore, by Theorem 1, this maximisation problem for any candidate function 2 worst can be solved in time k O(k ) log3 n. 4.2 An fpt-algorithm for Refined-Typed Max SMTI To extend the result for Typed Max SMTI to Refined-Typed Max SMTI, we need the following result. Lemma 5. Let I be a refined-typed instance of SMTI and suppose that M is a matching in I such that there is no pair (i, j) ∈ k (2) where j ≻i worstM (i) and i ≻j worstM (j). Then there is a stable matching M ′ such that, for every (i, j) ∈ k (2) , both M and M ′ contain the same number of pairs that consist of one agent of type i and another of type j. Moreover, given M, we can compute M ′ in polynomial time. Proof. Given matching M, let n{i,j} (M) denote the number of pairs in M consisting of an agent of type i and an agent of type j. We construct a stable matching M ′ such that n{i,j} (M ′ ) =Sn{i,j} (M). We first present the construction of M ′ and then prove that it is ′ ′ stable. Let i,j∈[k],i<j Mi,j = M ′ be a decomposition of M ′ where Mi,j is the projection of ′ ′ M onto agent types i and j. The construction of M takes place in two steps: 11 • Step 1: For all (i, j) ∈ k 2 , compute Ai,j where Ai,j denotes the set of agents of type i that are to be matched to agents of type j. ′ • Step 2: For all (i, j) ∈ k 2 , i < j, generate Mi,j given Ai,j and Aj,i computed in Step 1. We next elaborate on how each of the above steps are executed. Step 1: To start with, all agents are available. For each type i, take the candidate types in type i’s decreasing order of preference with ties broken arbitrarily, hi1 , i2 , . . . iki i, where is denotes the type that is ranked s’th by type i. Starting with j = i1 , take the topmost (from the perspective of an agent of type j) n{i,j} (M) available agents of type i, and put them in Ai,j ; these agents become unavailable from now on. Note that as the preference lists of agents of type j over agents of type i may include ties, it may not be possible to determine exactly who are the topmost available n{i,j} (M) agents in type i. To be more precise, when going down the preference list of agents of type j over available agents of type i, we may reach a tie τ including z > 1 available agent where we need to pick x < z number of them. If this happens, arbitrarily pick x agents from τ . ′ Step 2: We now show how to generate Mi,j given Ai,j and Aj,i computed in Step 1. We do so by computing a stable matching amongst the agents in Ai,j ∪ Aj,i . Since agents of the same type have the same preference ordering over the candidates of the same type, our problem reduces to the problem of computing a stable matching in an instance of SMTI where all men have identical preferences, as well as all women. The latter problem is solvable in polynomial time using a straightforward greedy procedure. By our construction of M ′ , n{i,j} (M ′ ) = n{i,j} (M). So it only remains to prove that M ′ is weakly stable. Assume for a contradiction that M ′ admits a blocking pair (a, b) where a , b, M ′ (a) and M ′ (b) are of types i, j, j ′ and i′ respectively. Three “kinds” of blocking pairs are possible, depending on how a and b compare each others’ types against the types of their partners. We examine each of them and show that M ′ can admit none. • If a prefers type j to type j ′ , and b prefers type i to type i′ . In the assumption of the lemma we have that given M there is no pair (i, j) ∈ k (2) where j ≻i worstM (i) and i ≻j worstM (j). By construction of M ′ , worstM (i) remains unchanged under M ′ for all types i. It thus directly follows that M ′ cannot admit such a blocking pair. ′ con• If i = i′ and j = j ′ . The existence of such a blocking pair implies that Mi,j structed in Step 2 is not weakly stable w.r.t. the preferences of agents in Ai,j ∪ Aj,i , a contradiction. • Either i = i′ or j = j ′ , but not both. Without loss of generality assume that j = j ′ and b prefers type i to type i′ . Since a prefers b to M ′ (a), and M ′ (a) ∈ Aj,i , therefore it follows the construction in Step 1 that, since b is not in Aj,i and is in Aj,i′ , therefore any agent of type j (including b) prefers type i′ to type i, a contradiction. Notice that we have not considered a scenario where j ≃i j ′ (j 6= j ′ ) or i ≃j i′ (i 6= i′ ). Assume that j ≃i j ′ , j 6= j ′ . Then, by our assumption that the instance is refined-typed, 12 any agent of type i, and hence a, is indifferent between all agents who are of type j or type j ′ . Therefore, M(a) ≃a b and (a, b) cannot be a blocking pair. A similar argument holds for when i ≃j i′ , i 6= i′ . Let I be a refined-typed instance of SMTI and let I ′ be a typed instance of SMTI that is obtained from I by ignoring the refined preferences within each type (i.e. every agent is indifferent between the candidates of the same type). It follows from the definition of stability that every matching that is stable in I is also stable in I ′ . Lemma 5 implies that for any stable matching M in I ′ , there exists a stable matching M ′ in I of the same cardinality as M. Thus, in order to find a maximum cardinality matching in a refined-typed instance of SMTI, it suffices to (1) solve the typed problem (i.e. ignore the refined preferences within each type) and then (2) use the algorithm provided in the proof of Lemma 5 to convert the solution to a matching of the same cardinality that is stable in the instance I. Theorem 6. Refined-Typed Max SMTI is in FPT parameterised by the number k of different types in the instance. 4.3 An fpt-algorithm for Max HRT A typed instance of Max HRT can be reduced to an instance of Max SMTI as follows. For each type i of hospitals, let Ni denote the set of posts offered by the hospitals of type i, and hence |Ni | the total number of posts offered by the hospitals of type i (i.e. the sum of the capacities of the hospitals of type i). To reduce an instance of Refined-Typed Max HRT to an instance of Refined-Typed Max SMTI, additionally let Ai,j denote the set of posts of type i that are matched to residents of type j (if i is a hospital type) or the set of residents of type i that are matched to a post of type j (if i is a resident type). The next result then straightforwardly follows from Theorems 4 and 6. Corollary 7. Typed Max HRT and Refined-Typed Max HRT are in FPT parameterised by the number k of different types in the instance. 4.4 Strict preferences over types In the preceding sections we have assumed that agents can be indifferent between the agents of two or more types. Here we investigate the implication of requiring strict preferences over types on the complexity of the stable matching problems we have studied. We show that the problems (Refined) Typed Max SMTI and (Refined) Typed Max HRT are polynomial-time solvable if agents have strict preferences over types; the argument is based on a private communication with David Manlove. Theorem 8. When preferences over types are strict, both Typed Max SMTI and RefinedTyped Max SMTI are polynomial-time solvable. Furthermore, all stable matchings are of the same size. Proof. Let I be a typed instance of SMTI, and assume that preferences over types are strict. Let I ′ be a refined-typed instance of SMTI that is obtained from I by breaking 13 all ties arbitrarily in a consistent way (recall that agents of the same type have identical preferences). In fact I ′ will be an instance of SMI, as preferences over types are strict. By Lemma 5, if M is a stable matching in I, then there exists a stable matching M ′ in ′ I of the same size. This implies that the size of a maximum cardinality matching in I is the same as the size of some stable matching in I ′ . However I ′ is an instance of SMI and all stable matchings of I ′ are of the same size. Therefore, to solve Max SMTI for I, it is enough to find a stable matching in I ′ , which can be done easily in polynomial time by a simple extension of GS [20]. Additionally, all stable matchings of I are of the same size. Now let I be a refined-typed instance of SMTI and assume that preferences over types are strict. Let I ′ be a typed instance of SMTI that is obtained from I by ignoring the refined preferences within each type (i.e. every agent is indifferent between the candidates of the same type). We have proved, in the preceding paragraph, that all stable matchings of I ′ are of the same size. It thus follows Lemma 5, using a similar argument as in the preceding paragraph, that all stable matchings in I are also of the same size, and thus the maximum cardinality stable matching of I can be found in polynomial time by breaking ties arbitrarily and applying a simple extension of GS [20]. The above result, combined with Corollary 7 gives us the following result. Corollary 9. When preferences over types are strict, Typed Max HRT and RefinedTyped Max HRT is polynomial-time solvable. Furthermore, all stable matchings are of the same size. 5 Efficient Algorithms for Max SRTI Notice that, in the proofs of Lemma 3, Lemma 5, and Theorems 4 and 6, we have not made use of the fact that SMTI is a bipartite matching problem, except that we have implicitly assumed that no type finds itself acceptable. Therefore these proofs will go through immediately for MAX SRTI if we impose such an assumption. To allow for a type to find itself acceptable, we need to make some modifications, which are explained next. 5.1 An fpt-algorithm for Typed Max SRTI Let I be a typed instance of MAX SRTI, and let M be a matching in I. We may assume without loss of generality that every agent is matched, using the same argument as in Section 4.1. We define worstM (i) as in Section 4.1, and additionally define second worstM (i) to be the type of the second least desirable agent with which any agent of type i is matched in M. If there is only one agent of type i, then second worstM (i) is undefined, in which case let second worstM (i) = ∅. Note that it is possible to have second worstM (i) = worstM (i). We claim that, in order to determine whether or not M is stable, it suffices to examine the values of worstM (i) and second worstM (i) for each i ∈ [k]. Lemma 10. Let I be a typed instance of SRTI. Then a matching M is stable if and only if (1) there is no pair (i, j) ∈ [k](2) , i 6= j, such that j ≻i worstM (i) and i ≻j worst(j)M , and (2) there is no pair (i, i), i ∈ [k], such that there is at least two agents of type i and i ≻i second worstM (i). 14 Proof. The proof is similar to the proof of Lemma 3 with slight modification concerning the second case in the statement of the current lemma. Suppose first that M is not stable. In this case, by definition, there exists some pair of agents (a, b) such that a and b are not matched together but each prefers the other over their current partner. If the agents are of different types then the argument is the same as in Lemma 3. Assume that a and b are both of the same type i. Then we know that type i likes type type(M(a)) or type type(M(b)) at least as well as type second worstM (i). Without loss of generality assume that type(M(a)) i second worstM (i). Since (a, b) is a blocking pair, type(b) ≻i type(M(a)), and therefore i ≻i second worstM (i). Conversely, suppose that M is stable. The first case where i 6= j is the same as in the proof of Lemma 3. Suppose, for a contradiction, that there exists a type i corresponding to at least two agents where i ≻i second worstM (i). Then there are two agents a and b of type i who are matched to an agent of type worstM (i) and an agent of type second worstM (i), respectively. Thus a and b both prefer each other to their current partner, and so form a blocking pair. This contradicts the assumption that M is weakly stable. Theorem 11. Typed Max SRTI is in FPT parameterised by the number k of different types in the instance. Proof. The proof is similar to the proof of Theorem 4, with some modifications. Lemma 10 implies that every stable matching must pass the test of Lemma 10, and any matching that )k possibilities passes the test of this lemma is stable. Note that there are at most ( (k+1)(k+2) 2 for the pair of functions worst : [k] → [k + 1] and second worst : [k] → [k + 2]. We say that a given pair of functions worst and second worst are feasible if for each type i, (1) worst(i) is either a type acceptable to type i or the dummy type, (2) second worst(i) is either a type acceptable to type i or a dummy type, or ∅ (only if there is only one agent of type i), and (3) if worst(i) = second worst(i) then there exists at least two agents of type worst(i). We say that a matching M realises functions worst and second worst if, for each i ∈ [k], the least and the second least desirable partner any agent of type i has in M are of types worst(i) and second worst(i) respectively. Our strategy is to consider each of the feasible possibilities for worst and second worst in turn. First we determine (using Lemma 10) whether a matching realising worst and second worst will be stable; this can be done for any candidate functions worst and second worst in time O(k 2 ). Note that it is possible that none of the candidates functions passes the test of Lemma 10, as not all instances of SRTI necessarily admit a stable matching. For any pair of feasible functions worst and second worst which does give rise to a stable matching (if any), we now want to determine what is the maximum number of agents which can be matched in any matching which realises worst and second worst. We do this by solving a suitable instance of an Integer Linear Programming. The formulation is exactly the same as the one provided in the proof of Theorem 4, with the exception of the following two sets of constraints introduced to ensure that second worst(i) is indeed the type of the second least desirable partner assigned to any agent of type i. 15 X n{i,j} = 0, ∀i ∈ [k] such that |Ni | > 1 second worst(i)≻i j≻i worst(i) n{i,worst(i)} + n{i,second worst(i)} > 1, n{i,worst(i)} > 1 ∀i ∈ [k] such that |Ni | > 1 ∀i ∈ [k] such that worst(i) = second worst(i) The above integer linear program has 5k +(k +1)2 constraints and (k +1)2 variables. The upper bound on the absolute value a variable can take is n. Therefore, by Theorem 1, this 2 maximisation problem for any candidate function worst can be solved in time 2O(k ) log3 n. 5.2 An fpt-algorithm for Refined-Typed Max SRTI Using a similar argument as in Section 4.2, and with the aid of the following lemma, we can employ the machinery used to solve Typed Max SRTI to solve Refined-Typed Max SRTI. Lemma 12. Let I be a refined-typed instance of SRTI and suppose that M is a matching in I such that (1) there is no pair (i, j) ∈ [k](2) , i 6= j, such that j ≻i worstM (i) and i ≻j worstM (j), and (2) there is no pair (i, i), i ∈ [k], such that there is at least two agents of type i and i ≻i second worst(i). Then there is a stable matching M ′ such that, for every (i, j) ∈ [k](2) , both M and M ′ contain the same number of pairs that consist of one agent of type i and another of type j. Moreover, given M, we can compute M ′ in polynomial time. Proof. TheSproof is similar to that of Lemma 5 with the following modifications. ′ ′ We let i,j∈[k],i≤j Mi,j = M ′ , to allow for the inclusion of Mi,i . Then in Step 1, if j = i, we take the topmost 2 · n{i,i} (M) available agents of type i, and put them in Ai,i . In Step ′ 2, to generate Mi,i given Ai,i we create a complete stable matching amongst the agents in Ai,i using the following greedy procedure: Until all agents are matched, take the topmost available tie in type i and match the agents in it together. If the number of agents in the topmost tie is odd, match the remaining agent to one of the agents in the second topmost available tie. Label all the matched agents as unavailable. To prove that M ′ is stable, we only additionally need to take care of one type of blocking pair in which two agents a and b of the same type i prefer each other to their partners. In the assumption of the lemma we have that given M there is no pair (i, i) ∈ [k](2) where i ≻i second worstM (i). By the construction of M ′ , second worstM (i) remains unchanged under M ′ for all types i. It thus directly follows that M ′ cannot admit such a blocking pair. Theorem 13. Refined-Typed Max SRTI is in FPT parameterised by the number k of different types in the instance. 5.3 Strict preferences over types A similar argument as in the proof of Theorem 8, with small modifications, shows that the non-bipartite versions are also polynomial-time solvable if preferences over types are strict. 16 Theorem 14. When preferences over types are strict, Typed Max SRTI and RefinedTyped Max SRTI are polynomial-time solvable. Furthermore, all stable matchings (if any exists) are of the same size. Proof. Let I be a typed instance of SRTI, and assume that preferences over types are strict. Let I ′ be a refined-typed instance of SRTI that is obtained from I by breaking all ties arbitrarily in a consistent way (recall that agents of the same type have identical preferences). In fact I ′ will be an instance of SRI, as preferences over types are strict. Let MAX SIZE(I) and MAX SIZE(I ′ ) denote the size of the maximum cardinality stable matching in I and I ′ respectively (which is zero if no stable matching exists). It follows from the definition of stability that every matching that is stable in I ′ (if any) is also stable in I, implying that MAX SIZE(I) ≥ MAX SIZE(I ′ ). Lemma 12 implies that for any stable matching M (if any exists) in I, there exists a stable matching M ′ in I ′ of the same cardinality as M, implying that MAX SIZE(I ′ ) ≥ MAX SIZE(I). Therefore, MAX SIZE(I ′ ) = MAX SIZE(I), and in order to find a maximum cardinality matching in I, or reporting that none exits, it is enough to solve Max SRI for I ′ . The latter problem is polynomial-time solvable [21]. Furthermore, all stable matchings of I ′ (if I ′ admits any) are of the same size [21], and hence all stable matchings of I (if I admits any) are also of the same size. 6 Maximum Cardinality Matchings with Minimum Instability As stated in Section 1, in some settings the size of the matching takes priority over the stability criteria. That is, the mechanism designers are willing to tolerate a small degree of instability if that leads to a matching of larger size. We can extend the methods from previous sections to prove the following results. Theorem 15. Typed Max Size Min BP SMTI belongs to FPT when parameterised by the number k of different types in the given instance. Theorem 16. Typed Max Size Min BA SMTI belongs to FPT when parameterised by the number k of different types in the given instance. We begin by considering the problem of minimising the total number of blocking pairs. For the rest of this section, we assume that types 1 to k ′ are types of women, and types k ′ + 1 to k are types of men. We further assume that all preference lists are extended to include the dummy type as their least desirable acceptable type. If an agent a is unmatched in M, we say that M(a) is of type k + 1, the dummy type. As a first step, we translate the definition of a blocking pair into the setting of typed instances. Proposition 17. Let x and y be agents of type i and type j respectively, and suppose that M(x) is of type j ′ and M(y) is of type i′ . Then (x, y) is a blocking pair in M if and only if j ≻i j ′ and i ≻j i′ . Given this observation, we can obtain an expression for the total number of blocking pairs in M that are comprised of an agent of type i and an agent of type j. 17 Lemma 18. The number of blocking pairs (x, y) in M such that x is a woman of type i and y is a man of type j is given by   ! X X X   n{i′ ,j}  n{i,j ′ } = n{i′ ,j}n{i,j ′ } 1i′ ≺j i 1j ′ ≺i j , i′ ≺ j i j ′ ≺i j 1≤i′ ,j ′ ≤k+1 where 1a≺b c is an indicator function that returns one if a ≺b c and zero otherwise. Proof. It is easy to verify that the left-hand side and the right-hand side of the equation are equal. For the remainder of the proof, we focus on the left-hand side of the equation. By Proposition 17, we know that the set of blocking pairs consisting of one agent of type i and another of type j is precisely {(x, y) : type(M(x)) ≺i tj and type(M(y)) ≺j ti }. The cardinality of this set is thus equal P to the number of agents of type i that are matched to an agent of type inferior to type j ( j ′ ≺i j n{i,j ′ } ) multiplied byPthe number of agents of type j that are matched to an agent of type inferior to type i ( i′ ≺j i n{i′ ,j}). The result follows immediately. Summing over all possibilities for i and j gives the following result. Lemma 19. The total number of blocking pairs in M is given by X X  n{i′ ,j}n{i,j ′ } 1i′ ≺j i 1j ′ ≺i j . 1≤i,j≤k 1≤i′ ,j ′ ≤k+1 We can now prove Theorem 15. Proof of Theorem 15. We begin by computing, in polynomial time, the cardinality Cmax of a maximum matching in our instance. Our strategy then is to formulate Max Size Min BP SMTI as an instance of Integer Quadratic Programming. Our goal is to minimise the following objective function X X  n{i′ ,j}n{i,j ′ } 1i′ ≺j i 1j ′ ≺i j 1≤i,j≤k 1≤i′ ,j ′ ≤k+1 = X 1≤i,i′ ≤k ′ k ′ <j,j ′ ≤k +  1i′ ≺j i 1j ′ ≺i j + 1i≺j′ i′ 1j≺i′ j ′ X n{k+1,j}n{i,j ′ } 1j ′ ≺i j X n{i′ ,j}n{i,k+1} 1i′ ≺j i 1≤i≤k ′ k ′ ≤j,j ′ ≤k + 1≤i,i′ ≤k ′ k ′ ≤j≤k + X 1≤i≤k ′ k ′ <j≤k 18 n{k+1,j} + n{i,k+1} ,  subject to the constraints ∀i ∈ [k] : X n{i,j} = |Ni | j∈[k+1] X n{i,j} = Cmax . 1≤i,j≤k To see that the right-hand side of the objective function equation is equal to the left-hand side of the equation, notice that every pair ({i, j}, {i′, j ′ }) ∈ [k](2) appears twice in the summation on the left-hand side: once with {i, j} coming from the first sum and {i′ , j ′ } coming from the second sum, and once the other way around. The former counts the number of blocking pairs (x, y) where x is of type i, y is of type j, and M(x) and M(y) are of types j ′ and i′ respectively. The latter counts the number of blocking pairs (x, y) where x is of type i′ , y is of type j ′ , M(x) is of type j, and M(y) is of type i. Pairs ({i, j}, {i′ , j ′ }) where at least one of i, i′ , j, j ′ is equal to the dummy type k + 1 are dealt with separately (in this case there can be at most one blocking pair). The linear constraints enforce that every agent is involved in exactly one pair (perhaps with a dummy agent), and that the number of pairs that do not involve dummy agents is equal to the maximum possible cardinality of a matching. We can write our objective function in the form xT Qx where x is the vector (n{1,1} , n{1,2} , . . . , n{k,k} )T and the entry of Q corresponding to n{i,j ′ } and n{i′ ,j} is equal to either 0, 1 or 2 depending on how many of the following conditions hold: 1. j ≻i j ′ and i ≻j i′ , and 2. j ′ ≻i′ j and i′ ≻j ′ i. Thus, by Theorem 2, we have an fpt-algorithm to solve our instance of Max Size Min BP SMTI. We now consider the problem of minimising the number of agents which are involved in at least one blocking pair. We start by characterising the conditions under which an agent of a particular type can belong to one or more blocking pairs; this characterisation follows immediately from the definition of a blocking pair. Lemma 20. Let x be an agent of type i and assume that M(x) is of type j (which would be a dummy type if x is unmatched). Then x belongs to a blocking pair if and only if there is some agent y of type j ′ who is paired with an agent of type i′ (which would be a dummy type if y is unmatched) such that i ≻j ′ i′ and j ′ ≻i j. Using this characterisation, we can now prove Theorem 16. Proof of Theorem 16. For any matching M, we can define a collection of at most k(k + 1)/2 boolean variables vi,j (for 1 ≤ i < j ≤ k + 1, where type k + 1 is a dummy type), so that vi,j is true if and only if the matching contains at least one pair involving an agent of type i and an agent of type j (and unmatched agents are considered to be matched with agents of the dummy type k + 1). For a given matching M, this collection of variables defines a vector vM in {0, 1}k(k+1)/2 , which we call the type-signature of the matching M. 19 2 Note that there are at most 2k(k+1)/2 = O(2k ) possible type-signatures for a matching; we will consider each possible type-signature v in turn and determine the minimum number of agents which can be involved in blocking pairs in a maximum matching which has typesignature v (if a maximum matching with this type-signature exists). Minimising over this set of optimal solutions will give the desired answer. We now describe how to compute the minimum number of agents involved in blocking pairs in a maximum matching with type-signature v or else to report that no such maximum matching exists. Our strategy is to encode the problem as an instance of Integer Linear Programming. First we define the constraints. As usual, we need to ensure that every agent is involved in exactly one pair (potentially involving a dummy agent), and as in the proof of Theorem 15 we need to enforce that the number of pairs that do not involve dummy agents is equal to the maximum cardinality of any matching in our instance. Moreover, we need to make sure that our matching does indeed have type-signature equal to v. This gives rise to the following linear constraints. ∀i ∈ [k] : X n{i,j} = |Ni | j∈[k+1] X n{i,j} = Cmax 1≤i<j≤k ∀1 ≤ i < j ≤ k + 1 with vi,j = 1 : n{i,j} > 0 ∀1 ≤ i < j ≤ k + 1 with vi,j = 0 : n{i,j} = 0. Finally, we define our objective function, which captures the number of agents which are involved in at least one blocking pair. By Lemma 20, we know that an agent of type i matched with an agent of type j belongs to a blocking pair if and only if there exist i ≻j ′ i′ and j ′ ≻i j such that vi′ ,j ′ = 1. Thus, for a given type-signature v, we can compute for each 1 ≤ i, j ≤ k + 1 the indicator variable bi,j which takes the value 1 if an agent of type i matched with an agent of type j in a matching with type-signature v will belong to a blocking pair, and takes the value 0 otherwise. It is now clear that the total number of agents that are involved in at least one blocking pair in the matching is X n{i,j} (bi,j + bj,i ). 1≤i<j≤k+1 This is our linear objective function. We further observe that Theorems 15 and 16 can easily be extended to refined typed instances. Corollary 21. Refined-Typed Max Size Min BP SMTI and Refined-Typed Max Size Min BA SMTI belong to FPT when parameterised by the number of types. 20 Proof. We use the same strategy for both problems. First, we solve the problem for the corresponding typed instance (ignoring the more refined preferences within types). Note that the number of blocking pairs or blocking agents achieved in this simplification of the problem clearly gives a lower bound on the minimum number that can be achieved if we take into account all information; we will argue that in fact we can always obtain a matching which does not increase either quantity when we take into account the full preference lists. To do this, we follow the method described in Lemma 5. Given the number of pairs n{i,j} of type i and j for each 1 ≤ i < j ≤ k, this method allows us to construct a matching M where, for each 1 ≤ i < j ≤ k we have exactly n{i,j} pairs involving one agent of type i and one of type j, and there is no blocking pair (x, y) such that x is currently matched to an agent of the same type as y. Thus the only blocking pairs in M are those of the form (x, y) where x is of type i and y of type j, and j ≻i type(M(x)) and i ≻j type(M(y)). But these are precisely the blocking pairs that occur in the relaxation to a typed instance. Thus we can indeed obtain a solution to Max Size Min BP SMTI or Max Size Min BA SMTI by applying the appropriate algorithm to find the number of pairs of each type under the relaxation to a typed instance, and then use the method of Lemma 5 to extend this to a matching which does not introduce any additional blocking pairs when the full preference lists are taken into consideration. Finally, we consider extending these results to other matching problems. We obtain one immediate corollary. Corollary 22. (Refined-)Typed Max Size Min BP HRT and (Refined-)Typed Max Size Min BA HRT belong to FPT when parameterised by the total number of types. To generalise to the non-bipartite case takes just slightly more care: if two agents of type i which are both matched to agents of type j form a blocking pair, then the total number of blocking pairs that results is n{i,j} (n{i,j} − 1)/2 rather than n2{i,j} . Otherwise, exactly the same method works. Thus we obtain the following corollary. Corollary 23. (Refined-)Typed Max Size Min BP SRTI and (Refined-)Typed Max Size Min BA SRTI belong to FPT when parameterised by the total number of types. It is also fairly straightforward to modify the IQP and ILP in the proofs of Theorem 15 and 16 to solve MIN BP SRTI and MIN BA SRTI respectively. We only need to remove the constraint that enforces the matching to be of size Cmax . Corollary 24. (Refined-)Typed Min BP SRTI and (Refined-)Typed Min BA SRTI belong to FPT when parameterised by the total number of types. A related problem to MIN BP SR is EXACT BP SR which given an instance I of SR and an integer Z, decides whether I admits a matching with exactly Z blocking pairs. Even this problem is NP-hard [1] in general. If I is a (Refined-)Typed instance, to solve EXACT BP SRTI we only have to move the objective function in the IQP of Theorem 15 to the set of constraints, and enforce it to be equal to Z. Corollary 25. (Refined-)Typed EXACT BP SRTI belongs to FPT when parameterised by the total number of types. 21 7 Hospitals/Residents problem with Couples The job market for medical residents underwent a change since mid 1970’s, due to married couples seeking posts in nearby hospitals. Subsequently, the central matching systems had to be adapted to take into account couple’s preferences, as otherwise couples would seek to arrange their own matches outside the centralized clearinghouse. The Hospitals/Residents problem with Couples (HRC) models such settings, where couples can send in preferences over pairs of hospitals. The size of a given matching in an instance of HRC is equal to the number of residents matched in the matching. Definition 6. Stable HRC is the problem of deciding whether an instance of HRC admits a stable matching. Definition 7. Max HRC is the problem of identifying a maximum cardinality stable matching in an HRC instance, or reporting that none exists. Stable HRC is NP-complete, and as a corollary so is Max HRC. The result holds even if each hospital has capacity 1 and there are no single residents [38]. The parameterized complexity of Stable HRC has been studied in [35, 6], where [35] also studies Max HRC. Stable HRC is W[1]-hard when the problem is parametrized by the number of couples [35]. The result holds even if each hospital has capacity 1. However, Stable HRC belongs to FPT when the problem is parametrized by the number of couples in the instance and the hospitals’ list are derived from a strictly-ordered master list of residents [6]. In the remainder of this section, we first provide a stability definition for instances of HRC, then extend our Typed and Refined-typed models to such instances, and conclude with presenting an fpt-algorithm for Typed Max HRC. 7.1 Stability in HRC In an instance of HRC, the set of residents R includes an even size subset R′ consisting of those residents who belong to couples, where every resident in R′ belongs to exactly one couple. Let C denote the set of ordered pairs (ri , rj ) where ri and rj form a couple. Single residents and hospitals find a subsets of candidates acceptable and rank them in strict order of preference. Every couple (ri , rj ) ∈ C has a joint strict preference ordering over an acceptable subset of ordered hospital pairs. Different stability definitions for an instance of HRC have been provided in the literature, most of them distinct from one another (see, e.g., [6, 28, 35, 36, 9, 16, 27, 39]). In this paper we adopt the definition of [36] and provide it as it has been given in [31]. We say that a hospital hj is undersubscribed if |M(hj )| < q(hj ) . Given an instance of HRC, a matching M is stable if it admits no blocking pair, where a blocking pair satisfies one of the following properties: 1. it involves a single resident ri and a hospital hj where (a) ri prefers hj to M(ri ) and (b) hj is undersubscribed or prefers ri to its worst assigned resident in M; 2. it involves a couple (ri , rj ) ∈ C and a hospital hk such that either 22 (a) (rj , rj ) prefers (hk , M(rj )) to (M(rj ), M(rk )), and hk is undersubscribed or prefers ri to a resident in M(hk ) \ {rj }; or (b) (rj , rj ) prefers (M(ri ), hk ) to (M(rj ), M(rk )), and hk is undersubscribed or prefers rj to a resident in M(hk ) \ {ri }; 3. it involves a couple (ri , rj ) ∈ C and a pair of (not necessarily distinct) hospitals hk , hℓ such that hk 6= M(ri ), hℓ 6= M(rj ), (rj , rj ) prefers (hk , hℓ ) to (M(rj ), M(rk )), and either (a) hk 6= hℓ , and hk (respectively hℓ ) is either undersubscribed or prefers ri (respectively rj ) to at least one of its assigned residents; or (b) hk = hℓ and q(hk ) − |M(hk )| ≥ 2; or (c) hk = hℓ and q(hk ) − |M(hk )| ≥ 1 and hk prefers at least one of ri , rj to one of its assigned residents; or (d) hk = hℓ , q(hk ) = |M(hk )|, hk prefers ri to some resident rs ∈ M(hk ), and prefers rj to some resident in M(hk ) distinct from rs . 7.2 Typed Instances In an instance of HRC, each type i that is associated with a single resident or a hospital has a preference ordering over types of the candidates, similarly to that defined earlier for instances of HRT. We assume, without loss of generality, that each pair of residents that form a couple are ordered such that the type of the first one is no larger than the type of the second one. That is, for each (x, y) ∈ C where x is of type i and y is of type j, it is the case that i ≤ j. Each pair of types (i, j) (with i ≤ j) associated with a couple has a joint preference ordering over ordered pairs of “hospital” types. The same conditions outlined in Section 3.1.1 apply to single residents and hospitals. That is, (1) every two single residents or hospitals of the same type have identical preference lists, and (2) all single residents are indifferent between hospitals of the same type, and all hospitals are indifferent between residents of the same type. These notions extend naturally to the preference lists of couples: 1. if (x, y) and (x′ , y ′) are couples of the same type (i, j) then (x, y) and (x′ , y ′) have identical preference lists, and 2. all couples are indifferent between (hs , ht ) and (hk , hℓ ) if hs and hk are hospitals of the same type, say i, and ht and hℓ are hospitals of the same type, say j. 7.3 An fpt-algorithm for Typed MAX HRC We adapt the method used to prove Theorems 4 and 11 in order to show that Typed Max HRC is in FPT when parameterised by the number of different types in the instance. The idea, once again, is to (1) consider a number of possibilities for the matching, where this number of possibilities is bounded by a function of k, (2) determine for each such possibility whether a matching which meets the conditions will be stable, and then (3) 23 express the maximisation problem associated with each set of stable conditions as an instance of Integer Linear Programming. As in previous sections, we assume without loss of generality that there is at least one agent of each type; however we cannot be sure that there will always be a single resident of a given type, or a couple of a specific pair of types. Also, as in previous sections, we assume that every agent is matched by creating sufficiently many dummy agents of type k + 1 which are inserted at the end of each single resident’s and hospital’s (possibly incomplete) preference list; we insert (k + 1, k + 1) at the end of each couple’s preference list. We start by providing conditions that are necessary and sufficient for a matching to be stable in a given typed instance of HRC. Given a matching M (in which every agent is matched, perhaps to a dummy agent), we define three functions worstM , second worstM and assignedM . The function worstM is defined as follows: • for any resident type i of which there is at least one single agent, worstM (i) is the type of the least desirable hospital (with respect to the preference list for i) to which any single resident of type i is assigned in M; • for any pair of resident types i and j (with i ≤ j) such that at least one couple has type (i, j), worstM (i, j) is the least desirable pair of hospital types (with respect to the joint preference list for (i, j)) to which any couple of types i and j are assigned in M; • for any hospital type p, worstM (p) is the type of the least desirable resident (with respect to the preference list for p) assigned to a hospital of type p in M. The function second worstM is only defined for hospital types where the total capacity of hospitals of the type is at least two: second worstM (p) is the type of the second least desirable resident assigned to any hospital of type p in M. The boolean function assignedM is defined for each combination of a pair of resident types and a pair of hospital types: assignedM ((i, j), (ℓ, p)) = 1 if and only if there is at least one couple involving residents of type i and j who are assigned to hospitals of types ℓ and p respectively. For the remainder of this section, we assume that types 1 to k ′ are types of residents, and types k ′ + 1 to k are types of hospitals. Let (k ′ : k] denote the set {k ′ + 1, . . . k}. Let |Ni | (with i ∈ [k ′ ]) denote the number of single residents of type i, |Np | (with p ∈ (k ′ : k]) denote the total capacity of hospitals of type p, and |Ni,j | (with i, j ∈ [k ′ ] where i ≤ j) the number of couples where the first resident is of type i and the second one is of type j. Lemma 26. Let I be a typed instance of HRC. Then a matching M is stable if and only if none of the following holds: 1. There is a pair of types (i, p), i ∈ [k ′ ], p ∈ (k ′ : k] such that |Ni | > 0, p ≻i worstM (i) and i ≻p worstM (p) . 2. There is a couple-type (i, j) with |Ni,j | > 0 and a hospital type ℓ such that either (a) i ≻ℓ worstM (ℓ) and there is a pair of hospital types (p, q) such that assignedM ((i, j), (p, q)) = 1 and (ℓ, q) ≻i,j (p, q), or (b) j ≻ℓ worstM (ℓ) and there is a pair of hospital types (p, q) such that assignedM ((i, j), (p, q)) = 1 and (p, ℓ) ≻i,j (p, q). 24 3. There is a couple-type (i, j) with |Ni,j | > 0 and two hospital types ℓ and p (where ℓ 6= p) such that (i) (ℓ, p) ≻i,j worstM (i, j), (ii) i ≻ℓ worstM (ℓ), and (iii) j ≻p worstM (p); 4. There is a couple-type (i, j) with |Ni,j | > 0 and a hospital type ℓ such that (i) (ℓ, ℓ) ≻i,j worstM (i, j), (ii) i, j ≻ℓ worstM (ℓ), and either (a) i ≻ℓ second worst(ℓ), or (b) j ≻ℓ second worst(ℓ). Proof sketch. It is straightforward to prove this claim using a similar approach as in the proof of Lemma 3 and the definition of a stable matching for an instance of HRC. Theorem 27. Typed Max HRC is in FPT when parameterised by the number k of different types in the instance. Proof. Lemma 26 gives a necessary and sufficient condition for a matching realising the functions worst, second worst and assigned to be stable. We consider each of the feasible possibilities for the functions worst, second worst and assigned in turn and determine, using Lemma 26, whether a matching realising them will be stable; this can be done for any set of candidate functions in time O(k 3 ). For any trio of feasible functions worst, second worst and assigned that give rise to a stable matching (if any), we need to determine the maximum number of residents that can be matched in any matching that realises the trio, and we do this (as in the rest of the paper) by solving an instance of Integer Linear Programming. We define variables ni,p and n(i,j),(p,q) as follows. For each pair of types (i, p), i ∈ [k ′ ], p ∈ (k ′ : k], we have a variable ni,p which denotes the number of single residents of type i matched to hospitals of type p. Additionally, for each combination of a pair of resident types (i, j) (with i ≤ j) and a pair of hospital types (p, q) (where p and q may or may not be distinct), we have a variable n(i,j),(p,q) which denotes the number of couples consisting of residents of types i and j respectively such that the resident of type i is assigned to a hospital of type p and the resident of type j is assigned to a hospital of type q. Our objective is to maximise the size of the matching hence to maximise: X X ni,p + 2 · n(i,j),(p,q) i≤k ′ ,k ′ <p≤k i,j≤k ′ ,k ′ <p,q≤k Our first set of constraints ensures that every agent is involved in exactly one pair. The first constraint handles single agents, the second one the hospitals, and the last one couples. X ni,p = |Ni |, ∀i ∈ [k ′ ] k ′ <p≤k X ni,p + i≤k ′ X X (n(i,j),(p,q) + n(i,j),(q,p) ) = |Np |, ∀p ∈ (k ′ : k] i,j≤k ′ ,k ′ <q≤k ∀i, j : i ≤ j ≤ k ′ n(i,j),(p,q) = |Ni,j |, k ′ <p,q≤k 25 The next set of constraints ensures that function worst complies with its definition. X ni,p = |Ni |, ∀i ∈ [k ′ ] : |Ni | > 0 pi worst(i) ∀i ∈ [k ′ ] : |Ni | > 0 ni,worst(i) > 0, X X ni,p + ip worst(p) (n(i,j),(p,q) + n(i,j),(q,p)) = |Np |, i,jp worst(p) k ′ <q≤k ∀p ∈ (k ′ : k] nworst(p),p > 0, X ∀p ∈ (k ′ : k] ∀i, j : i ≤ j ≤ k ′ : |Ni,j | > 0 n(i,j),(p,q) = |Ni,j |, (p,q)(i,j) worst(i,j) ∀i, j : i ≤ j ≤ k ′ : |Ni,j | > 0 n(i,j),worst(i,j) > 0, The following set of constraints ensures that function second worst for the hospitals complies with its definition. X ni,p = 0, ∀p ∈ (k ′ : k] : |Np | > 1 second worst(p)≻p i≻p worst(p) nworst(p),p + nsecond worst(p),p > 1, nworst(p),p > 1 ∀p ∈ (k ′ : k] : |Np | > 1 ∀p ∈ (k ′ : k] : |Np | > 1 and second worst(p) = worst(p) And finally, the last set of constraints ensure that boolean variables assigned are set correctly. n(i,j),(p,q) > 0, ∀i, j ∈ [k ′ ]∀p, q ∈ (k ′ , k] such that assigned((i, j)(p, q)) = 1 The above integer linear program has O(k 4 ) variables and O(k 2 ) constraints. The upper bound on the absolute value a variable can take is n. Therefore, by Theorem 1, this maximisation problem for any trio of candidate functions worst, second worst and assigned can 2 be solved in time 2O(k ) log3 n. 8 Types in practice In this section we discuss some issues relating to the applicability of our models to real world matching problems. We begin in Section 8.1 by discussing how an instance with relatively few types could arise if agents derive their preference lists from a small set of attributes of other agents. In Section 8.2 we discuss how the same ideas can be applied to consider a relaxation of the stability requirement in a more realistic setting in which agents only feel strongly about a small collection of attributes, but may refine these preferences arbitrarily. Finally, in Section 8.3, we explain how we can efficiently identify agents of the same type if we only have access to the agents’ ordinal preference lists. 26 8.1 Attributes It seems believable in many settings (particularly those involving very many agents) that agents will derive their preferences by considering some collection of attributes of the other agents, where the number of attributes considered is likely to be much smaller than the total number of agents. For example, in a centralised job allocation scheme, employers might rank applicants based on a number of criteria from the application form e.g. exam grade, interview score etc. In the setting of hospitals/residents allocation, a junior doctor might rank hospitals based on the programs they offer, their reputation on the quality of their programs, their geographic location etc. A similar observation has been made in [5, 11, 12, 17]. We assume that there is a set B = {B1 , . . . , Br } of attributes and each attribute Bi can take one of ui distinct possible values (perhaps including a “not applicable” value if different subsets of attributes are relevant for different agents); we write u for max1≤i≤r ui . The attribute profile of an agent is an r-tuple which lists its value Q for each of the r attributes that is relevant to the agent. Note that there are at most ri=1 ui ≤ ur possible attribute profiles for an agent. Since agents form their preferences solely on the basis of the attributes of other agents, every agent is necessarily indifferent between two agents which have the same attribute profile. Thus we can express each agent’s preferences as an ordering of the possible attribute profiles, allowing for some attribute profiles to be declared unacceptable. This implies that (up to the reordering of agents involved in ties) there are at most ur ·(ur )! possible preference lists. As we can capture all relevant information about an agent by specifying its attribute profile and its ordering of the possible attribute profiles, we can therefore obtain a typed instance with at most u2r · (ur )! types, by grouping together agents who have the same attribute profile and preferences over other attribute profiles. In certain models, it might be reasonable to make a stronger assumption, namely that an agents’ preferences are also determined by their own attributes: for example, in a stable roommates problem, agents might be looking for partners who are similar to themselves in certain ways. If agents rank others based solely on their own attributes, this then means that the number of types is bounded by ur rather than ur · (ur )!. 8.2 A relaxation of the stability requirement While attributes may well play an important role in determining agents’ preferences, the situation described in Section 8.1, where only a very small number of attributes (taking few distinct values) are relevant, might be too simplified to reflect reality in many situations. In this section we discuss how the same ideas may be applied in a much wider range of settings if we relax our notion of stability. It is reasonable to assume that a certain amount of effort is required by both agents in a blocking pair to make a private arrangement outside the matching, and so agents are unlikely to make this effort for a very small improvement in their utility. We say that a blocking pair is dangerous if both agents can make a significant improvement in their utility (where we will define what it means to make a significant improvement later). The goal is then to find a maximum cardinality matching which contains no dangerous blocking pairs. In our attribute-based model (with potentially many different attributes), there are two 27 natural ways to quantify the relative utility of two partners: one is to consider which attributes can be improved by switching partners, and another is to consider by how much these attributes can be improved. It seems reasonable to assume that: 1. each agent considers some (small) subset of attributes to be important, and will only make a private arrangement to improve the value of an important attribute, and 2. the possible values for each attribute can split into a small number of groups of similar values, and an agent will only make a private arrangement to obtain a partner whose attribute value comes from a superior group. If we further assume that there are relatively few attributes that are important to one or more agents (in particular we do not have a large number of agents with disjoint sets of important attributes), we can define a simpler problem instance which captures coarse information about the original instance, and crucially contains enough information to determine whether there will be any dangerous blocking pairs in a matching. This idea can be formalised as follows. We denote by AI the set of t attributes which are important to at least one agent, and for each Ai ∈ AI we split the values into groups Gi1 , . . . , Gisi . We write s = maxAi ∈AI si . We now define the coarse attribute profile of an agent to be a t-tuple which lists the group to which the agent’s value belongs for each attribute in AI . Based on our assumptions above, we know that an agent will only make a private arrangement with a new partner whose coarse attribute profile is preferable to that of their current partner. Note that there are at most st possible coarse attribute profiles, and that there are at most (st )! possible rankings of coarse attribute profiles. We now group agents into types so that two agents are assigned to the same type if and only if they have the same coarse attribute profile and the same preferences over coarse attribute profiles; the number of types is at most st (st )!. If we simplify the instance I by assuming that all agents are in fact indifferent between any two agents with the same attribute profile, we obtain a typed instance I ′ . Of course, a matching which is stable in this simplified instance I ′ need not be stable in the original instance I; however, it is easy to check that a matching M is weakly stable in I ′ if and only if M contains no dangerous blocking pair in I. Thus, if we are willing to accept matchings which are not stable but which contain no dangerous blocking pairs, we can use the machinery developed to deal with typed instances in this more general setting. 8.3 Identifying types from preference lists In order to make use of the algorithms we have developed for typed and refined-typed instances, we need to know the type of each agent, but in practice we may only be presented with lists of ordinal preferences for each agent. In this section we describe how we can compute the coarsest partition into types which meet the conditions of either of the models. 8.4 Finding types in a typed instance We identify the type of each agent in two stages. 28 In the first step, we sort the list of agents by their preference lists; this makes it easy to identify which agents have identical preference lits, which is one of the requirements for agents having the same type. We write x ∼pref y if agents x and y have the same preference list. Next, we consider each pair of agents (x, y) such that x ∼pref y (that is, pairs (x, y) where we have not yet found evidence that x and y cannot have the same type). We then consider the preference lists of all other agents, and for each list check whether x and y belong to the same tie. If there is any agent which strictly prefers x to y (or vice versa) then x and y cannot have the same type; otherwise we write x ∼ y. The types are then taken to be the equivalence classes with respect to the equivalence relation ∼. 8.5 Finding types in a refined-typed instance The process of identifying types for refined-typed instances is only slightly more complicated. Again, we require that all agents of the same type have identical preferences, so we can once again begin by sorting preference lists to determine the pairs for which x ∼pref y. We call each equivalence class with respect to ∼pref a group. We now refine the partition as follows by making a number of passes through the list of preference lists. On every pass, we consider each preference list in turn, and check for each group whether its members occur consecutively on the preference list. For every group where this is not the case, we split the group into the maximal subsets which occur consecutively in the preference list being considered, then move onto the next preference list with this new, refined partition. We continue this process until we complete a pass through the preference lists in which we do not split any group; when this occurs we stop with the current partition. At this point, it is clear that we have obtained a partition that satisfies the requirements that agents of the same type have identical preferences and appear consecutively in all other agents’ lists. It remains to observe that we do not have to complete too many passes to find such a partition. Note that, on each pass in which we do split some group, we must increase the number of parts in the partition by at least one. Therefore, if the instance can be split into a set of k types satisfying the conditions for a refined-typed instance, we cannot possibly make more than k passes. 9 Summary and Future Work We considered a setting in which agents are partitioned into k different types, and the type of an agent determines his or her preferences, as well as how s/he is compared against other agents. Agents have preferences over types, which may be refined by more detailed preferences within a single type, and agents of the same type have identical preferences. We showed that in this setting several important NP-hard stable matching, namely Max SMTI, Max HRT, Max SRTI, Max Size Min BP/BA SMTI, Max Size Min Bp/BA SRTI, and Min BP SRTI belong to the parameterised complexity class FPT when parameterised by the number of different types of agents, and so admit efficient algorithms when this number of types is small. We were further able to prove that Max SMTI, Max HRT and 29 Max SRTI are polynomial-time solvable when agents have strict preferences over types. Additionally, we were able to show Max HRC is in FPT parameterised by the number of different types of agents, when agents are indifferent between the agents of the same type. It would be interesting to investigate whether our approach might yield fpt-algorithms for other NP-hard stable matching problems. so that agents of the same type have “similar” preference lists, but not necessarily identical. Another intriguing question would be to understand how the complexity of the stable matching problems we have studied changes when agents on only one side of the market are associated with types. Acknowledgements. The first author is supported by a Personal Research Fellowship from the Royal Society of Edinburgh (funded by the Scottish Government). Both authors are extremely grateful to David Manlove for his insightful comments on a preliminary version of this manuscript. References [1] D.J. Abraham, P. Biró, and D.F. Manlove. “Almost stable” matchings in the Roommates problem. In Proceedings of WAOA ’05: the 3rd Workshop on Approximation and Online Algorithms, volume 3879 of Lecture Notes in Computer Science, pages 1–14. Springer, 2006. [2] D.J. Abraham, A. Levavi, D.F. Manlove, and G. O’Malley. The Stable Roommates Problem with Globally-Ranked Pairs. In Proceedings of WINE ’07: the 3rd International Workshop on Internet and Network Economics, volume 4858 of Lecture Notes in Computer Science, pages 431–444. Springer, 2007. [3] D.J. Abraham, A. Levavi, D.F. Manlove, and G. O’Malley. The Stable Roommates Problem with Globally-Ranked Pairs. Internet Mathematics, 5(4):493–515, 2008. Preliminary version appeared as [2]. [4] H. Aziz and B. de Keijzer. Complexity of coalition structure generation. In The 10th International Conference on Autonomous Agents and Multiagent Systems, AAMAS ’11, pages 191–198, 2011. [5] N. Bhatnagar, S. Greenberg, and D. Randall. Sampling stable marriages: why spouseswapping won’t work. In Proceedings of SODA ’08: the 19th ACM/SIAM Symposium on Discrete Algorithms, pages 1223–1232. ACM-SIAM, 2008. [6] P. Biró, R.W. Irving, and I. Schlotter. Stable matching with couples: an empirical study. ACM Journal of Experimental Algorithmics, 16, 2011. Section 1, article 2, 27 pages. [7] P. Biró, D.F. Manlove, and S. Mittal. Size versus stability in the marriage problem. Technical Report TR-2008-283, University of Glasgow, Department of Computing Science, 2008. 30 [8] P. Biró, D.F. Manlove, and S. Mittal. Size versus stability in the marriage problem. Theoretical Computer Science, 411:1828–1841, 2010. [9] D. Cantala. Matching markets: the particular case of couples. Economics Bulletin, 3(45):1–11, 2004. [10] P. Chebolu, L.A. Goldberg, and R. Martin. The complexity of approximately counting stable matchings. In Proceedings of APPROX ’10: the 13th International Workshop on Approximation Algorithms for Combinatorial Optimization Problems, volume 6302 of Lecture Notes in Computer Science, pages 81–94. Springer, 2010. Full version available at http://arxiv.org/abs/1004.1836. [11] P. Chebolu, L.A. Goldberg, and R. Martin. The complexity of approximately counting stable matchings. Theoretical Computer Science, 437:35–68, 2012. Preliminary version appeared as [10]. [12] P. Chebolu, L.A. Goldberg, and R. Martin. The complexity of approximately counting stable roommate assignments. Journal of Computer and System Sciences, 78(5):1579– 1605, 2012. [13] E. Choo and A. Siow. Who marries whom and why. Journal of Political Economy, 114(1):175–201, 2006. [14] M. Cygan, F.V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk, M. Pilipczuk, and S. Saurabh. Parameterized Algorithms. Springer International Publishing, 2015. [15] R. G. Downey and M. R. Fellows. Fundamentals of Parameterized Complexity. Springer London, 2013. [16] B. Dutta and J. Massó. Stability of matchings when individuals have preferences over colleagues. Journal of Economic Theory, 75:464–475, 1997. [17] F. Echenique, S. Lee, M. Shum, and M. B. Yenmez. The revealed preference theory of stable and extremal stable matchings. Econometrica, 81(1):153–171, 2013. [18] J. Flum and M. Grohe. Parameterized Complexity Theory. Springer, 2006. [19] A. Frank and É Tardos. An application of simultaneous Diophantine approximation in combinatorial optimization. Combinatorica, 7(1):49–65, 1987. [20] D. Gale and L.S. Shapley. College admissions and the stability of marriage. American Mathematical Monthly, 69:9–15, 1962. [21] D. Gusfield and R.W. Irving. The Stable Marriage Problem: Structure and Algorithms. MIT Press, 1989. [22] K. Hamada, K. Iwama, and S. Miyazaki. An improved approximation lower bound for finding almost stable maximum matchings. Information Processing Letters, 109(18):1036–1040, 2009. 31 [23] R.W. Irving. On the stable room-mates problem. Technical Report CSC/86/R5, University of Glasgow, Department of Computing Science, 1986. [24] R.W. Irving and D.F. Manlove. The Stable Roommates Problem with Ties. Journal of Algorithms, 43:85–105, 2002. [25] R.W. Irving, D.F. Manlove, and S. Scott. The stable marriage problem with master preference lists. Discrete Applied Mathematics, 156(15):2959–2977, 2008. [26] R. Kannan. Minkowskis convex body theorem and integer programming. Mathematics of Operations Research, 12(3):415–440, 1987. [27] B. Klaus, F. Klijn, and T. Nakamura. Corrigendum: Stable matchings and preferences of couples. Journal of Economic Theory, 144(5):2227–2233, 2009. [28] F. Kojima, P.A. Pathak, and A.E. Roth. Matching with couples: Stability and incentives in large matching markets. Quarterly Journal of Economics, 128(4):1585–1632, 2013. [29] H.W. Lenstra Jr. Integer programming with a fixed number of variables. Mathematics of Operations Research, 8(4):538–548, 1983. [30] D. Lokshtanov. Parameterized integer quadratic programming: Variables and coefficients. arXiv:1511.00310 [cs.DS], 2015. [31] D.F. Manlove. Algorithmics of Matching Under Preferences. World Scientific, 2013. [32] D.F. Manlove, R.W. Irving, K. Iwama, S. Miyazaki, and Y. Morita. Hard variants of stable marriage. Technical Report TR-1999-43, University of Glasgow, School of Computing Science, 1999. [33] D.F. Manlove, R.W. Irving, K. Iwama, S. Miyazaki, and Y. Morita. Hard variants of stable marriage. Theoretical Computer Science, 276(1-2):261–279, 2002. Full version available as [32]. [34] D. Marx and I. Schlotter. Parameterized complexity and local search approaches for the stable marriage problem with ties. Algorithmica, 58(1):170–187, 2010. [35] D. Marx and I. Schlotter. Stable assignment with couples: parameterized complexity and local search. Discrete Optimization, 8:25–40, 2011. [36] E.J. McDermid and D.F. Manlove. Keeping partners together: Algorithmic results for the hospitals / residents problem with couples. Journal of Combinatorial Optimization, 19(3):279–303, 2010. [37] G. O’Malley. Algorithmic Aspects of Stable Matching Problems. PhD thesis, University of Glasgow, Department of Computing Science, 2007. [38] E. Ronn. NP-complete stable matching problems. Journal of Algorithms, 11:285–304, 1990. 32 [39] A.E. Roth. The evolution of the labor market for medical interns and residents: a case study in game theory. Journal of Political Economy, 92(6):991–1016, 1984. [40] T. Shrot, Y. Aumann, and S. Kraus. On agent types in coalition formation problems. In Proceedings of the 9th International Conference on Autonomous Agents and Multiagent Systems, AAMAS ’10, pages 757–764, 2010. 33
8cs.DS
Control Synthesis for Multi-Agent Systems under Metric Interval Temporal Logic Specifications ? Sofie Andersson ∗ Alexandros Nikou ∗ Dimos V. Dimarogonas ∗ arXiv:1703.02780v1 [cs.SY] 8 Mar 2017 ∗ ACCESS Linnaeus Center, School of Electrical Engineering and KTH Center for Autonomous Systems, KTH Royal Institute of Technology, SE-100 44, Stockholm, Sweden. E-mail: {sofa, anikou, dimos}@kth.se Abstract: This paper presents a framework for automatic synthesis of a control sequence for multi-agent systems governed by continuous linear dynamics under timed constraints. First, the motion of the agents in the workspace is abstracted into individual Transition Systems (TS). Second, each agent is assigned with an individual formula given in Metric Interval Temporal Logic (MITL) and in parallel, the team of agents is assigned with a collaborative team formula. The proposed method is based on a correct-by-construction control synthesis method, and hence guarantees that the resulting closed-loop system will satisfy the specifications. The specifications considers boolean-valued properties under real-time. Extended simulations has been performed in order to demonstrate the efficiency of the proposed controllers. Keywords: Reachability analysis, verification and abstraction of hybrid systems, Multi-agent systems, Control design for hybrid systems, Modelling and control of hybrid and discrete event systems, Temporal Logic 1. INTRODUCTION Multi-agent systems are composed by N ≥ 2 number of agents which interact in an environment. Cooperative control for multi-agent systems allows the agents to collaborate on tasks and plan more efficiently. In this paper, the former is considered by regarding collaborative team specifications which requires more than one agent to satisfy some property at the same time. The aim is to construct a framework that will start from an environment and a set of tasks, both local (i.e. specific to an individual agent) and global (i.e. requires collaboration between multiple agents), and yield the closed-loop system that will achieve satisfaction of the specifications, by control synthesis. The specification language that has been introduced to express such tasks is Linear Temporal Logic (LTL) (see e.g., [Loizou and Kyriakopoulos 2004]). The general framework that is used is based on a three-steps procedure ([Kloetzer and Belta 2008, Kress-Gazit et al. 2007]): First the agent dynamics is abstracted into a Transition System. Second a discrete plan that meets the high level task is synthesized. Third, this plan is translated into a sequence of continuous controllers for the original system. Control synthesis for multi-agent systems under LTL specifications has been addressed in [Kloetzer et al. 2011, Guo and Dimarogonas 2015, Kantaros and Zavlanos 2016]. Due ? This work was supported by the H2020 ERC Starting Grand BUCOPHSYS, the Swedish Research Council (VR), the Swedish Foundation for Strategic Research (SSF) and the Knut och Alice Wallenberg Foundation. to the fact that we are interested in imposing timed constraints to the system, the aforementioned works cannot be directly utilized. Timed constraints have been introduced for the single agent case in [Gol and Belta 2013, Raman et al. 2015, Fu and Topcu 2015, Zhou et al. 2016] and for the multi-agent case in [Karaman and Frazzoli 2008, Nikou et al. 2016b]. Authors in [Karaman and Frazzoli 2008] addressed the vehicle routing problem, under Metric Temporal Logic (MTL) specifications. The corresponding approach does not rely on automata-based verification, as it is based on a construction of linear inequalities and the solution of a resulting Mixed-Integer Linear Programming (MILP) problem. In our previous work [Nikou et al. 2016b], we proposed an automatic framework for multiagent systems such that each agent satisfies an individual formula and the team of agents one global formula. The approach to solution suggested in this paper follows similar principles as in [Nikou et al. 2016b]. Here however, we start from the continuous linear system itself rather than assuming an abstraction, by adding a way to abstract the environment in a suitable manner such that the transition time is taken explicitly into account. The suggested abstraction is based on the work presented in [Gol and Belta 2013], which considered time bounds on facet reachability for a continuous-time multi-affine single agent system. Here, we consider multi-agent systems and suggest an alternative time estimation and provide a proof for its validity. Furthermore, we present alternative definitions of the local BWTS, the product BWTS and the global BWTS, compared to the work presented in [Nikou et al. 2016b]. The definitions suggested here requires a smaller number of states and hence, a lower computational demand. The drawback of the suggested definitions is an increased risk of a false negative result and a required modification to the applied graph-searchalgorithm. However, this will have no effect on the fact that the method is correct-by-construction. The method, in its entirety, has been implemented in simulations, demonstrating the satisfaction of the specifications through the resulting controller. • d :→→ R+ is a positive weight assignment map; the expression d(ri , σj , rk ) is used to express the weight σj assigned to the transition ri → rk . Definition 3. A timed run rt = (r(0), τ0 )(r(1), τ1 )... of a WTS T is an infinite sequence where r(0) ∈ Πinit , and σ r(j) ∈ Π, r(j) →i r(j + 1) ∀j ≥ 1 s.t. The contribution of this paper is summarized in four parts; (1) it extends the method suggested in [Nikou et al. 2016b] with the ability to define the environment directly as a continuous linear system rather than treating the abstraction as a given, (2) it provides for a less computationally demanding alternative, (3) simulation results which support the claims are included, (4) it considers linear dynamics in contrast to the already investigated (in [Nikou et al. 2016b]) single integrator. for some σi ∈ Σ. Definition 4. A timed word produced by a timed run is an infinite sequence of pairs w(rt ) = (L(r(0)), τ0 )(L(r(1)), τ1 )..., where rt = (r(0), τ0 )(r(1), τ1 )... is the timed run. Definition 5. The syntax of MITL over a set of atomic propositions AP is defined by the grammar φ := > | ap | ¬ φ | φ ∨ ψ | φ U[a,b] ψ (1) where ap ∈ AP and φ, ψ are formulas over AP . The operators are Negation (¬), Conjunction (∨) and Until (U) respectively. The extended operators Eventually (♦) and Always () are defined as: ♦[a,b] φ := >U[a,b] φ, (2a) [a,b] φ := ¬♦[a,b] ¬φ. (2b) This paper is structured as follows. Section 2 introduces some preliminaries and notations that will be applied throughout the paper, Section 3 defines the considered problem and Section 4 presents the main result, namely the solution framework. Finally, simulation result is presented in Section 5, illustrating the framework when applied to a simple example, and conclusions are made in section 6. 2. PRELIMINARIES AND NOTATION In this section, the mathematical notation and preliminary definitions from formal methods that are required for this paper are introduced. Given a set S, we denote by |S|, 2S its cardinality and the set of all its subsets respectively. Let A ∈ Rn×m , B ∈ Rn be a matrix and a vector respectively. Denote by [A]ij the element in the i-th row and j-th column of matrix A. Similarly, denote by [B]i the i-th element of vector B. Given a set of nonnegative rational numbers T ⊂ Q+ a time sequence is defined as: Definition 1. [Alur and Dill 1994] A time sequence τ = τ0 τ1 ... is an infinite sequence of time values which satisfies all the following: • τi ∈ T ⊂ Q+ , • τi < τi+1 , ∀i ≥ 0 and • ∃i ≥ 1, s.t. τi > t, ∀t ∈ T. An atomic proposition ap is a statement over the system variables that is either true (>) or false (⊥). Definition 2. A Weighted Transition System (WTS) is a tuple T = (Π, Πinit , Σ, →, AP, L, d) where Π = {ri : i = 0, ..., M } is a set of states, Πinit ⊂ Π is a set of initial states, Σ = {σi : i = 0, ..., l} is a set of inputs, →⊂ Π × Σ × Π is a transition map; the expression σj ri → rk is used to express transition from ri to rk under the action σj , • AP is a set of observations (atomic propositions), • L : Π → 2AP is an observation map and • • • • • τ0 = 0, • τj+1 = τj + d(r(j), σi , r(j + 1)), ∀j ≥ 1, Given a timed run rt = (r(0), τ0 )(r(1), τ1 ), ..., (r(n), τn ) of a WTS, the semantics of the satisfaction relation is then defined as: (rt , i) |= ap ⇔ ap ∈ L(r(i)), (rt , i) |= ¬φ ⇔ (rt , i) 2 φ, (rt , i) |= φ ∧ ψ ⇔ (rt , i) |= φ and (rt , i) |= ψ, (rt , i) |= φ U[a,b] ψ ⇔ ∃j ∈ [a, b], s.t. (rt , j) |= ψ and ∀i ≤ j, (rt , i)  φ. Definition 6. A clock constraint Φx is a conjunctive formula on the form x ./ a, where ./∈ {<, >, ≤, ≥}, x is a clock and a is some constant. Let ΦX denote the set of clock constraints. The TBA was first introduced in [Alur and Dill 1994] and is defined as Definition 7. A Timed Büchi Automaton (TBA) is a tuple A = (S, S0 , X , I, E, F, AP, L) where • • • • • • • • S = {si : i = 0, 1, ..., M } is a finite set of locations, S0 ∈ S is the set of initial locations, X is a finite set of clocks, I : S → ΦX is a map labelling each state si with some clock constraints ΦX , E ⊆ S × ΦX × 2X × S is a set of transitions and F ⊂ S is a set of accepting locations, AP is a finite set of atomic propositions, L is a labelling function, labelling every state with a subset of atomic propositions. A state of A is a pair (s, v) where s ∈ S is a location and v is a clock valuation that satisfies the clock constraint I(s). The initial state of A is a pair (s0 , (0, 0, ..., 0)), where s0 ∈ S0 and the null-vector (0, 0, ..., 0) is a vector of |X | number of valuations vi = 0. For the semantics and examples of the above TBA definition we refer the reader to [Nikou et al. 2016a]. It has been shown in previous work [Alur et al. 1996] that any MITL formula can be algorithmically translated to a TBA such that the language that satisfies the MITL formula is also the language that produces accepting runs by the TBA. The TBA expresses all possibilities, both satisfaction and violation of the MITL formula. All timed runs which result in the satisfaction of the MITL formula are called accepting: Definition 8. An accepting run is a run for which there are infinitely many j ≥ 0 s.t. qj ∈ F , i.e. a run which consists of infinitely many accepting states. In motion-planning, the movement of an agent can be described by a timed run. For the multi-agent case, the movement of all agents can be collectively described by a collective run. The definition is Definition 9. [Nikou et al. 2016b] The collective timed run rG = (rG (0), τG (0))(rG (1), τG (1))... of N agents, is defined as follows • (rG (0), τG (0)) = (r1 (0), ..., rN (0), τG (0)) • (rG (i + 1), τG (i + 1)) = (r1 (j1 ), ..., rN (jN ), τG (i + 1)), for i ≥ 0 where (rG (i), τG (i)) = (r1 (i), ..., rN (i), τG (i)) and · l = argmin{τk (ik + 1)}, k∈I · τG (i + 1)= τl (il + 1), rl (il + 1), if k = l · rk (jk ) = rk (il ), otherwise. 3. PROBLEM DEFINITION since by taking into account the counterexample in [Nikou et al. 2016b, Section III], the following holds: ^ t t rG |= ϕk < r1t |= ϕ1 ∧ . . . ∧ rN |= ϕN . (5) k∈I 4. PROPOSED SOLUTION The solution approach involves the following steps: (1) For each agent, we abstract the continuous-time linear system (3) into a WTS which describes the possible movements of the agent considering the dynamics and limitations of the state space (section 4.1). (2) For each agent, we construct a local BWTS out of its WTS and a TBA representing the local MITL specification. The accepting timed runs of the local BWTS satisfy the local specification (section 4.2). (3) Next, we construct a product BWTS out of the local BWTSs. The accepting timed runs of the product BWTS satisfy all local specifications (section 4.3). (4) Next, we construct a global BWTS out of the product BWTS and the TBA representing the global MITL specification. The accepting runs of the global BWTS satisfy both the global specification and all local specifications (section 4.4). (5) Finally, we determine the control input by applying a graph-search algorithm to find an accepting run of the global BWTS and projecting this accepting run onto the individual WTSs (section 4.5). The computational complexity of the proposed approach is discussed in Section 4.6. 3.1 System Model Consider N agents performing in a bounded workspace X ⊂ Rn and governed by the dynamics ẋm = Am xm + Bm um , m ∈ I, xm (0) = x0m , xm ∈ X (3) where I = {1, ..., N } is a set containing a label for each agent. 3.2 Problem Statement The problem considered in this paper consists in synthesizing a control input sequence, um , m ∈ I, such that each agent satisfies a local individual MITL formula φm over the set of atomic propositions APm . At the same time, the team of agents should satisfy a team specification MITL formula φG over the set of atomic propositions S APG = APm . m∈I Following the terminology presented in Section 2, the problem becomes: Problem 10. Synthesize a sequence of individual timed t runs r1t , ..., rN such that the following holds:   t rG |= φG ∧ r1t |= φ1 ∧ ... ∧ rN |= φN (4) where the collective run rG was defined in 3 9. Remark 11. Initially it might seem that if a run rG that satisfies the conjunction of the local formulas i.e., rG |= t r1t ∧ . . . ∧ rN can be found, then the Problem 10 is solved in a straightforward centralized way. This does not hold 4.1 Constructing a WTS In this section we consider the abstraction of the environment into a WTS. The definition of a WTS was given in Section 2. The abstraction is performed for each agent m ∈ I, resulting in N number of WTSs. Following the idea of [Gol and Belta 2013], we begin by dividing the state space Xm into p-dimensional rectangles, defined as in Definition 12 Definition 12. A p-dimensional rectangle Rp (a, b) ⊂ Rp is characterized by two vectors a, b, where a = (a1 , a2 , ..., ap ), b = (b1 , b2 , ..., bp ) and ai < bi , ∀i = 1, 2, ..., p. The rectangle is then given by Rp (a, b) = {x ∈ Rp : ai ≤ xi ≤ bi , ∀i ∈ {1, 2, .., p}} (6) such that formula (7) is satisfied for each rectangle, i.e, such that each atomic proposition in the set APm is either true at all points within a rectangle Rp (a, b) or false at all points within the rectangle, i.e. api = (>, ∀x ∈ Rp (a, b)) or api = (⊥, ∀x ∈ Rp (a, b)), ∀api ∈ APm . (7) The set of states Π = {r0 , r1 , ..., rM } of the WTS is then defined as the set of rectangles R = {Rp (a0 , b0 ), Rp (a1 , b1 ), ..., Rp (aM , bM )}. From this, the definition of the initial state Πinit , transitions → and labelling L follows directly: Πinit = {ri ∈ Π|x0m ∈ Rp (ai , bi )} (8) ri → rj iff Rp (ai , bi ) and Rp (aj , bj ) (9) xi have a common edge, L(ri ) = {api ∈ APm |api = T rue ∀x ∈ Rp (ai , bi )} (10) The set Σ is given as the set of control inputs which induce transitions. In particular, a control input must be defined for each possible transition such that it guarantees the transition, that is no other transition can be allowed to occur and the edge of which the transition goes through must be reachable. This conditions on control inputs are required both to ensure that the synthesized path is followed and to guarantee that the following time estimation holds. A suggested low-level controller for a transition rk → rl in direction i, based on [Gol and Belta 2013], is given by max [ẋm ]i rl = Rp (al , bl ) ekl x1 rk = Rp (ak , bk ) ekl x0 [xm ]min := akj j [xm ]max := bkj j um ∈Um s. t. [ẋm ]i ≥  > 0, [ẋm ]j ≤ [ẋm ]j ≥ xj −, ∀j 6= i, j = {1, ..., p}, if [xm ]j = bkj , , ∀j 6= i, j = {1, ..., p}, if [xm ]j = akj . (11) where Um = [−umax , umax ] is some bound on um and  is a robustness parameter. The idea is to maximize the transition speed, under the conditions that the speed in direction j is negative at the edge with norm direction j, where j is not the transition direction. Finally, the weights d are assigned as the maximum transition times. These times are given according to Theorem 13 below. The theorem depends on the assumption Bm um = Bm1 xm +Bm2 , where Bm1 and Bm2 are matrices of dimension N ×N and N ×1 respectively. The assumption corresponds to um being affine. Theorem 13. The maximum time T max (rk , rl ) required for the transition rk → rl to occur, where Rp (ak , bk ) and Rp (al , bl ) share the edge ekl , ekl is the edge located opposite to ekl in Rp (ak , bk ), i is the direction of the transition, and assuming that ekl is reachable from all points within rk , is defined as: ! ∗ ∗ ([A∗m ]ii x0 + Cm + [Bm ]i ) 1 max T (rk , rl ) = ln ∗ + [B ∗ ] ) ∗ ] ([A∗m ]ii x1 + Cm [A i m m ii (12) where ∗ Cm = n X j=1 min  [xm ]j ∈rk  [A∗m ]ij [xm ]j , j6=i ( min [xm ]j ∈rk [A∗m ]ij [xm ]j  = [A∗m ]ij akj if [A∗m ]ij > 0 , (13) [A∗m ]ij bkj if [A∗m ]ij < 0 and x0 ∈ ekl , x1 ∈ ekl (note that x0 , x1 are the ith coordinate of the initial and final positions of the transition), ∗ A∗m = Am + Bm1 and Bm = Bm2 , where ẋm = Am xm + Bm um = Am xm + Bm1 xm + Bm2 . See figure 1 for illustration of the variables of Theorem 13 in 2 dimensions. Proof. of Theorem 1 T max - the maximum transition time for rk → rl in a Fig. 1. Illustration of the variables in Theorem 13 in 2 dimensions. system following the linear dynamics (3) is determined by considering the minimum transition speed. Consider the dynamics of agent m projected onto the direction of the transition i, i.e [ẋm ]i = [Am xm + Bm um ]i , (14)   0 0 xm (0) i = [xm ]i = x ,   xm (t1 ) i = [x1m ]i = x1 , where x0 is the ith coordinate of some point on the edge ekl , and x1 is the ith coordinate of some point on the edge ekl . Since Bm um = Bm1 xm + Bm2 , system (14) can be rewritten to (15), by introducing A∗m = Am + Bm1 and ∗ Bm = Bm2 . n X ∗ ∗ [ẋm ]i = [Am ]ii [xm ]i + [A∗m ]ij [xm ]j + [Bm ]i (15) j=1 j6=i  xm (0) i = x0   xm (t1 ) i = x1  The maximum transition time is determined by solving (15) for t1 . The equation can be solved by separating [xm ]i from t, if and only if [A∗m ]ij [xm ]j is a constant ∀j. Since [A∗m ]ij is a constant this holds if and only if [ẋm ]j = 0 or [A∗m ]ij = 0. Otherwise, the maximum transition time can be overestimated by considering the minimum transition speed [ẋm ]min , at every point in rk , i which can be determined by considering the limits of [xm ]j in rk , namely akj and bkj (  [A∗m ]ij akj if [A∗m ]ij > 0 ∗ min [Am ]ij [xm ]j = (16) [A∗m ]ij bkj if [A∗m ]ij < 0 [xm ]j ∈rk The maximum transition time, denoted T max , can then be overestimated as the solution to ∗ ∗ ẏ = [A∗m ]ii y + Cm + [Bm ]i y(0) = x y(T max 0 ) = x1 (17) where ∗ Cm = n n X X min [A∗m ]ij [xm ]j = min [A∗m ]ij [xm ]j [xm ]j ∈rk j=1 j=1 j6=i j6=i [xm ]j ∈rk . Which can be solved as: dy ∗ ∗ = [A∗m ]ii y + Cm + [Bm ]i =⇒ dt  Z Z  1 dt = dy =⇒ ∗ + [B ∗ ] [A∗m ]ii y + Cm m i ∗ ∗ ln([A∗m ]ii y + Cm + [Bm ]i ) t + tc = ∗ [Am ]ii Now, y(0) = x0 yields ∗ ∗ ln([A∗m ]ii x0 + Cm + [Bm ]i ) tc = [A∗m ]ii and y(T max ) = x1 yields ! ∗ ∗ [A∗m ]ii x1 + Cm + [Bm ]i 1 max T = ln ∗ 0 ∗ ∗ ∗ [Am ]ii x + Cm + [Bm ]i [Am ]ii (18) (19) (20) ∗ Remark 14. If Cm = 0 or [ẋm ]j = 0 ∀j, then T max is the maximal time required for the transition to occur. Otherwise T max is an over-approximation. Finally, the weights of the WTS are defined as d(ri , rj ) = T max (ri , rj ) where (ri , σ, rj ) ∈→ . 4.3 Constructing a Product BWTS Now, a product BWTS should be constructed from the local BWTSs. The definition is given as follows: Definition 17. Given N local BWTSs T1p , ..., TNp , defined as in Definition 15, and Mk = |Xk | for k ∈ {1, .., N }, the product BWTS TG = T1p ⊗ ... ⊗ TNp = p p p p = (Q1 , Qinit 1 , ;1 , d1 , F1 , AP1 , L1 , I1 , C1 , M1 ) ⊗ ... p p p p ... ⊗(QN , Qinit N , ;N , dN , FN , APN , LN , IN , CN , MN ) (21) for σ = um (ri , rj ). 4.2 Constructing a Local BWTS Next, a local BWTS is constructed out of the WTS and a TBA representing the local MITL specification for each agent. As stated in Section 2 any MITL formula can be represented by a TBA [Alur et al. 1996]. Approaches for the translation were suggested in [Maler et al. 2006], [Brihaye et al. 2013] and [Ničković and Piterman 2010]. Note that the time-intervals considered by the MITL formulas must be on the form ≤ a due to the overapproximation of time in the abstraction. The local BWTS is defined as: Definition 15. Given a weighted transition system T = (Π, Πinit , Σ, →, AP, L, d) and a timed Büchi automaton A = (S, Sinit , X , I, E, F, AP, L) their local BWTS is defined as T p = T ⊗ A = (Q, Qinit , ;, dp , F p , AP, Lp , I p , C) with: • Q ⊆ {(r, s) ∈ Π × S : L(r) = L(s)}, • Qinit = Πinit × Sinit • q ; q 0 iff · q = (r, s), q 0 = (r0 , s0 ) ∈ Q · (r, r0 ) ∈→ and · ∃ γ, R, s.t. (s, γ, R, s0 ) ∈ E, • dp (q, q 0 ) = d(r, r0 ) if (q, q 0 ) ∈;, • F p = {(r, s) ∈ Q : s ∈ F } and • Lp (r, s) = L(r) • I p (q) = I(s) • C = {c1 , ..., cM } ci = {(q, q 0 ) | ∃ R s.t. (s, R, s0 ) ∈ E and xi ∈ R} where M = |X |. It follows from the construction and automata-based LTL model checking theory [Baier and Katoen 2007] that all possible runs of the local BWTS correspond to a possible run of the WTS. Furthermore, all accepting states of the local BWTS corresponds to accepting states in the TBA. This is formalized in Lemma 16. Lemma 16. An accepting timed run rkt = (qk (0), τk (0))(qk (1), τk (1))... of the local BWTS projects onto the timed run rt = (r(0), τ (0))(r(1), τ (1))... of the WTS that produces the timed word w(rt ) = (Lk (r(0)), τ (0))(Lk (r(1)), τ (1))... accepted by the TBA. Also, if there is a timed run that produces an accepting timed word of the TBA, then there is an accepting timed run of the local BWTS. = (QG , Qinit G , →G , dG , FG , APG , LG , IG , CG , M ) is defined as: • QG ⊆ Q1 × ... × QN = Qinit × ... × Qinit • Qinit 1 N G 0 ) ∈→G iff • (qG , qG · qG = (q1 , ..., qN ) ∈ QG , 0 0 ) ∈ QG , = (q10 , ..., qN · qG 0 · ∃qk ∈ Qk s.t. (qk , qk0 ) ∈;k , ∀ k ∈ I, 0 0 ) ∈→G , ) = dmax = max(dpi ), if (qG , qG • dG (qG , qG i=I p • FG = {(qS 1 , ..., qN ) ∈ QG s.t. qk ∈ Fk , ∀k ∈ I}, • APG = APk , k∈I S p • LG (q1 , .., qN ) = Lk (qk ), S p k∈I • IG (qG ) = Ik (qk ), k∈I • CG = {C 1 , ..., C N }, C i = {ci1 , ..., ciMi } 0 0 0 {(qG , qG ), qG = (q1 , ..., qN ), qG = (q10 , ..., qN ) cik = 0 s.t (qi , qi ) ∈ ck , ck ∈ Ci } • M = {M1 , .., MN } It follows from the construction that an accepting collective run of the product BWTS corresponds to accepting runs of each local BWTS. Formally Lemma 18. An accepting collective run rG of the product BWTS projects onto an accepting timed run rkt of a local BWTS, for each k ∈ I. Moreover, if there exists an accepting timed run for every local BWTS, then there exists an accepting collective run. Remark 19. Note that the definition does not allow for the agents to start transitions at different times. This causes an overestimation of required time which increases the risk for false negative result. An alternative definition which allows the mentioned behaviour was suggested in [Nikou et al. 2016b]. However, the definition suggested here requires less number of states and hence less computational time. 4.4 Constructing a Global BWTS Finally, a global BWTS is constructed from the product BWTS and a TBA representing the global MITL specification. Definition 20. Given a product BWTS TG = (QG , Qinit G , →G , dG , FG , APG , LG , IG , CG , M ) and a init global TBA AG = (SG , SG , XG , IG , EG , FG , LG ), with MG = |XG |, their global BWTS T̂G = TG ⊗ AG = ˆ (Q̂G , Q̂init G , ;G , dG , F̂G , APG , L̂G ) is defined as: • Q̂G ⊆ {(q, s) ∈ QG × SG s.t. LG (q) = LG (s)} × i Z0 × ... × ZN × {1, 2}, where Zi = {z1i , ..., zM } for i 0 0 i = 1, ..., N and Z0 = {z1 , ..., zMG } init • Q̂init = Qinit G G ×SG ×{1, .., 1}×...×{1, ..., 1}×{1, 2}, where {1, ..., 1} × ... × {1, ..., 1} consists of N + 1 sets, where the first set contains MG ones, and the remaining sets contains Mi ones each, 0 • (qG , qG ) ∈;G iff · qG = (q, s, Z0 , ..., ZN , l) ∈ Q̂G , 0 0 · qG = (q 0 , s0 , Z00 , ..., ZN , l0 ) ∈ Q̂G , 0 · (q, q ) ∈→G , · ∃γ, R s.t. (s, γ, R, s0 ) ∈ EG s.t, For all i ∈ {1, ..., N }, Zi and Zi0 are such that 0 zki = 0 and zki = 1, if (q, q 0 ) ∈ cik 0 zki = zki , otherwise Z0 and Z00 are such that  0 if xk ∈ R 0 zk = 1 otherwise 0 zk0  = 1 if xk ∈ R zk0 otherwise   1, if l = 1 and q ∈ FG or l = 2 and s ∈ FG · l0 =  2, otherwise • • • • 0 0 dˆG (qG , qG ) = dG (q, q 0 ) if (qG , qG ) ∈;G , F̂G = {(q, s, Z0 , ..., ZN , 1) ∈ Q̂G s.t. q ∈ FG } and L̂G (q, s, Z0 , ..., ZN , l) = LG (r). I(qG ) = IG (q) ∪ I(s) It follows from the construction that an accepting run of the global BWTS corresponds to an accepting run of the product BWTS as well as an accepting run of the TBA representing the global specification. Formally t Lemma 21. An accepting timed run rG of the global BWTS projects onto an accepting collective run rG of the product BWTS that produces a timed word w(rG ) which is accepted by the TBA representing the global specification. Also, if there exists an accepting collective run that produces a timed word accepted by the TBA, t then there is an accepting timed run rG of the global BWTS. 4.5 Control Synthesis The controller can now be designed by applying a modified graph-search algorithm (such as a modified Dijkstra) to find an accepting run of the global product. The modification of the algorithm includes a clock valuation when considering a transition. A sketch of the modification is given in Algorithm 1. The idea is to calculate the clock valuation for each clock given the predecessors of the current state, if a valuation does not satisfy the clock constraint the transition is not valid. When the algorithm is complete the accepting run is projected onto the WTSs following Lemma 16, Lemma 18 and Lemma 21. Finally, the set of controllers are given as the sequences of control t inputs which induces the timed runs (r1t , r2t , ...rN ) which in turn produce accepted timed words of all local TBAs as well as of the global TBA. Algorithm 1: Modification to search-algorithm to evaluate clocks Result: Clock Valuation M =Total number of clocks; q=current state; q 0 =possible successor of q; for i = 1 : M do vi = d(q, q 0 ); k = q; if zi (k) == 1 then while zi (P red(k)) == 1 do vi =vi + d(P red(k), k); k = P red(k); if P red(k) isempty then break; end end end end if v1 , ..., vM 2 I(q) then Transition is illegal - don’t add q as a successor to P red(q). end 4.6 Complexity The framework proposed in this paper requires at most |T̂G | = N Y (|Ti | × |Ai | × 2Mi ) × |AG | × 2MG × 2 (22) i=1 number of states. The method suggested in [Nikou et al. 2016b] requires N   Y |T̂G | = |Ti | × |Ai | × (Cmax,i + 1)Mi × i=1 × |AG | × 2 × (Cmax,G + 1)MG × 2 number of states, where all possible clock values are integers in the set [0, Cmax,i ] and [0, Cmax,G ] for the local and global TBA’s respectively. Hence the number of states required in the proposed framework is a factor QN Mi × (Cmax,G + 1)MG i=1 (Cmax,i + 1) PN 2 i=1 (Mi )+MG less. Room 1 Room 2 1 Corridor 2 Room 4 Room 5 Table 1. The worst case estimation of the transition times as well as the actual required time. The actual transition times are defined as the maximum of the times the individual agents require to complete the transition. *These transitions require agent 2 to stay in place, hence the actual time is here defined as the time agent 1 requires to complete the transition.?Numbered in order of transitions, see figure 4. Room 3 Room 6 Position? Agent 1 Agent 2 Worst Case Time Estimation Actual Time 0 1 2 3 4 5 6 7 8 9 5 6 6 5 8 7 8 5 6 6 0 0.0589 0.04 0.0771 0.0645 0.0668 0.0465 0.2027 0.1438 0.04 0 0.0368 0.026* 0.0212 0.0403 0.0551 0.0151 0.1115 0.1366 0.027* Fig. 2. Draft of the problem described in section 5. Fig. 3. Partition constructed by the MATLAB script. The circles represents the initial states of each agent. 5. SIMULATION RESULT Consider N = 2 agents with dynamics in the form:     21 10 ẋ = x+ u (23a) 02 01     10 01 ẋ = x+ u (23b) 01 10 evolving in a bounded workspace X consisting of 6 rooms and a hallway as can be seen in Figure 2. Each agent is assigned with the local MITL formula φL = ♦0.1 r2 ∧ r2 → ♦0.3 r6 (’Eventually, within 0.1 time units, the agent must be in room 2, and if the agent enters room 2 it must then enter room 6 within 0.3 time units.’). Furthermore, they are assigned with the global MITL formula φG = ♦1 (a1 = r1 ∧ a2 = r2 ) (’Eventually, within 1 time units, agent 1 must be in room 1 and agent 2 must be in room 2, at the same time.’). The initial positions of each agent is indicated by the encircled 1 and 2 in Figure 2. Remark 22. As can be seen in figure 2, some walls have been added to the environment. Transitions through these are forbidden. This is handled by the abstraction since the edges on which the walls are placed aren’t reachable. The suggested environment can be abstracted to a WTS of 9 states (see figure 3), while the local MITL formula can be represented by a TBA of 4 states. This results in a local BWTS of 36 states. Notable is that the local BWTSs for each agent will be identical if and only if the dynamics are identical. Furthermore, if the problem at hand only considers local MITL formulas - that is, if no global tasks are considered - the five step procedure described earlier can stop here. In that case, the control 2 5 6 5 8 7 8 5 2 3 design can be performed based on accepting runs of each local BWTS. Since a global task is considered in this case, the product BWTS and the global BWTS must be constructed. The product BWTS will consist of (|Q1 | · |Q2 |) = 1296 states while the global BWTS will consist of 2 · (|QpBT W S | × |QgT BA | × 2M1 × 2M2 × 2MG ) = 248832 states. MATLAB was used to simulate the problem by constructing all transition systems and applying a modified Dijkstra algorithm to find an accepting path as well as a control sequence that satisfies the specifications. The projection of the found accepting run onto each WTS, yielded [2, 5, 6, 5, 8, 7, 8, 5, 2, 3] and [5, 6, 6, 5, 8, 7, 8, 5, 6, 6], for the respective agent. The result is visualized in Figure 4, which shows the evolution of each closed-loop system for the given initial positions. The figure was constructed by implementing the built-in function ode45 for the determined closed-loop system in each state with the initial position equal to the last position of the former transition. The switching between controllers is performed based on the position of the agent; namely the switching from controller uij to ujk is performed when the agent has entered far enough into state j, where ”far enough” was defined as 5 iterations of ode45 upon exiting the previous state. The estimated time distances for each joined transition are given in table 1. That is, the worst case transition times yields; 0 Agent 1 and Agent 2 begins at their respective initial position in the corridor 1 Agent 2 enters room 2 within 0.0589 time units from start 2 Agent 1 enters room 2 within 0.0989 time units from start 5 Agent 1 and Agent 2 enters room 6 within 0.2084 and 0.2484 time units respectively from entering room 2 9 Agent 1 is in room 1 while Agent 2 is in room 2 within 0.7404 time units from start. From this, it is clear that the given path will satisfy the MITL formulas. (a) Agent 1 (b) Agent 2 Fig. 4. Illustration of the paths of each agent in the example. The numbers 0-9 represent the end of each joined transition. The actual arrival time at each location as well as the time the agent is required to wait till the worst case transition time has been reached (and it is guaranteed that all other agents have transitioned), is noted to the right of the figure. The time the agent has to wait till corresponds to the worst case estimation of the required transition time and is due to the requirement that the agents make transitions simultaneously. It is notable that both agents finish all transitions on less time than the worst case estimation. Hence, the waiting time can be further cut by allowing the agents to communicate to each other when a transition is done. The simulation presented in this section was run in MATLAB on a laptop with a Core i7-6600U 2.80 GHz processor, the runtime was approximately 30min. 6. CONCLUSIONS AND FUTURE WORK A correct-by-construction framework to synthesize a controller for a multi-agent system following continuous linear dynamics such that some local MITL formulas as well as a global MITL formula are satisfied, has been presented. The method is supported by result of the simulations in the MATLAB environment. Future work includes communication constraints between the agents. REFERENCES Alur, R. and Dill, D.L. (1994). A theory of timed automata. Theoretical computer science, 126(2), 183– 235. Alur, R., Feder, T., and Henzinger, T.A. (1996). The benefits of relaxing punctuality. Journal of the ACM (JACM), 43(1), 116–146. Baier, C. and Katoen, J.P. (2007). Principles of model checking. MIT press. Brihaye, T., Estiévenart, M., and Geeraerts, G. (2013). On mitl and alternating timed automata. In Formal Modeling and Analysis of Timed Systems, 47–61. Springer. Fu, J. and Topcu, U. (2015). Computational methods for stochastic control with metric interval temporal logic specifications. CoRR, abs/1503.07193. Gol, E.A. and Belta, C. (2013). Time-constrained temporal logic control of multi-affine systems. Nonlinear Analysis: Hybrid Systems, 10, 21–33. Guo, M. and Dimarogonas, D. (2015). Multi-Agent Plan Reconfiguration Under Local LTL Specifications. The International Journal of Robotics Research, 34(2), 218– 235. Kantaros, Y. and Zavlanos, M. (2016). A Distributed LTLBased Approach for Intermittent Communication in Mobile Robot Networks. American Control Conference (ACC), 2016, 5557–5562. Karaman, S. and Frazzoli, E. (2008). Vehicle Routing Problem with Metric Temporal Logic Specifications. 47th IEEE Conference on Decision and Control (CDC 2008), 3953–3958. Kloetzer, M., Ding, X.C., and Belta, C. (2011). MultiRobot Deployment from LTL Specifications with Reduced Communication. 50th IEEE Conference on Decision and Control (CDC 2011), 4867–4872. Kloetzer, M. and Belta, C. (2008). A fully automated framework for control of linear systems from temporal logic specifications. Automatic Control, IEEE Transactions on, 53(1), 287–297. Kress-Gazit, H., Fainekos, G.E., and Pappas, G.J. (2007). Where is waldo? sensor based temporal logic motion planning. mag. Loizou, S. and Kyriakopoulos, K. (2004). Automatic Synthesis of Multi-Agent Motion Tasks Based on LTL Specifications. 43rd IEEE Conference on Decision and Control (CDC 2004), 1, 153–158. Maler, O., Nickovic, D., and Pnueli, A. (2006). From mitl to timed automata. In Formal Modeling and Analysis of Timed Systems, 274–289. Springer. Ničković, D. and Piterman, N. (2010). From MTL to deterministic timed automata. Springer. Nikou, A., Boskos, D., Tumova, J., and Dimarogonas, D.V. (2016a). Cooperative Planning for Coupled MultiAgent Systems under Timed Temporal Specifications. http://arxiv.org/pdf/1603.05097v2.pdf. Nikou, A., Tumova, J., and Dimarogonas, D.V. (2016b). Cooperative task planning of multi-agent systems under timed temporal specifications. Raman, V., Donzé, A., Sadigh, D., Murray, R., and Seshia, S. (2015). Reactive Synthesis from Signal Temporal Logic Specifications. 18th International Conference on Hybrid Systems: Computation and Control (HSCC 2015), 239–248. Zhou, Y., Maity, D., and Baras, J.S. (2016). Timed Automata Approach for Motion Planning Using Metric Interval Temporal Logic. European Control Conference (ECC 2016).
3cs.SY
THEORY OF DISGUISED-GROUPS arXiv:1607.07345v2 [math.GR] 22 Aug 2016 EDUARDO BLANCO GÓMEZ Abstract. In this paper we define a new algebraic object: the disguised-groups. We show the main properties of the disguisedgroups and, as a consequence, we will see that disguised-groups coincide with regular semigroups. We prove many of the results of the theory of groups adapted to this case and some unknown results in the theory of groups and regular semigroups. 1. Introduction In this paper it is defined and developed a new algebraic structure, disguised-groups. This new object lives between semigroups and groups and it is called like that because there are many traditional properties of groups that turn disguised-groups on groups. This is a new algebraic structure because its definition does not coincide with the monoid and, in fact, a disguised-group is never just a monoid because, as we will see later, a disguised-group that is simultaneously a monoid is, in fact, a group. The most similar known algebraic structure to the disguisedgroups are regular semigroups; we will see that one can not deduce the equality of both algebraic objects just from the definition of disguisedgroups. In some sense, disguised-groups are groupoids for which the binary operation can be done with every couple of elements. From the topological point of view, the use of the theory of groups is very extended into the algebraic topology but the use of groupoids not. One of the reasons is that the groupoids are more complicated algebraic structures and is more difficult to obtain results (see [2]). From this point of view, we propose the disguised-groups like a good tool to attack problems into algebraic topology as it is shown in [1]. For information about the theory of groups we suggest references [4], [5] and [7]. 2. Disguised-groups and disguised-subgroups Definition 2.1. Let G be a set and ∗ a binary operation in G. An element id ∈ G is an identity if for some g ∈ G we have id ∗ g = g or g ∗ id = g; if id ∗ g = g we will say that the element id is a left identity for g and if g ∗ id = g we will say that the element id is a right identity for g. An element g2 ∈ G is an inverse for an element g1 ∈ G if g1 ∗ g2 = id or g2 ∗ g1 = id for some id an identity of g1 . 1 2 EDUARDO BLANCO GÓMEZ Definition 2.2. Let G be a set with ∗ a binary operation in G. We will call (G, ∗) a disguised-group if the next four conditions hold, (a) ∗ is a closed operation in G, i.e., for every g1 , g2 ∈ G we have that g1 ∗ g2 ∈ G. (b) ∗ is an associative operation in G, i.e., for every g1 , g2 , g3 ∈ G, we have that, (g1 ∗ g2 ) ∗ g3 = g1 ∗ (g2 ∗ g3 ). (c) For every g ∈ G there exist some id a right identity for g and some id0 a left identity for g. (d) Every element g ∈ G has, at least, one inverse and every inverse −1 g ∈ G of an element g ∈ G is such that, g −1 ∗ g = id g∗g −1 = id for some id a right identity for g 0 for some id0 a left identity for g. Looking at the definition of disguised-group one can realize some interesting things. The first one is that the difference between a group and a disguised-group is that, taking one element of the disguisedgroup, the identities on the right and on the left can be different between them and among the identities of the other elements of the disguised-group; from the definition one deduces that the category (later we will describe the morphisms) of disguised-groups include the category of groups so the first one is not empty. A natural question is if one can find a disguised-group that is not a group; in paper [1] we are going to define a disguised-group that does not seem to be a group (though in fact it is). The second interesting thing one can realize from the definition of disguised-groups is their relation with monoids. In a monoid there is a unique identity on the right and on the left and the identity is unique for all the elements of the monoid. This fact is what we refer before: in a monoid every element has a unique right identity and a unique left identity and they are equal, so a disguised-group that is a monoid is a group. The most similar known algebraic structure to the disguised-groups are regular semigroups; from definition 2.2 one can not deduce that disguised-groups are exactly regular semigroups. The difference lies in (d) of definition 2.2; from definition of regular semigroups, see [3] or [6], identities should hold (c) and (d) of definition 2.2 both at the same time; that is not a condition for the identities of an element g of a disguised-group. Let’s denote I(G) the set of the identities in G. Let g ∈ G; we will denote IR (g) the set of the right identities of g and IL (g) the set of the left identities of g. We are going to prove a lemma that will help us to state the important inverse identity relations. Remark 2.3. Looking at definition 2.2 from the point of view of an inverse g −1 ∈ G of an element g ∈ G we have that every id ∈ IR (g) for THEORY OF DISGUISED-GROUPS 3 which g −1 ∗ g = id is such that id ∈ IL (g −1 ) and every id0 ∈ IL (g) for which g ∗ g −1 = id0 is such that id0 ∈ IR (g −1 ). We require more work to prove the inverse statement. Lemma 2.4. Let (G, ∗) be a disguised-group. Let g ∈ G, id1 ∈ IR (g) be such that g −1 ∗ g = id1 . Then g −1 ∗ g = id2 for every id2 ∈ IR (g) (for IL (g) the lemma is analogue). Proof. As g −1 ∗ g = id1 , by the associative property (b) of definition 2.2, ∗id g −1 ∗ g = id1 =⇒2 g −1 ∗ g = id1 ∗ id2 . Suppose now that g −1 ∗ g 6= id2 . Using associative property (b) of definition 2.2 and remark 2.3, id ∗ 1 id1 ∗ (g −1 ∗ g) 6= id1 ∗ id2 ⇒ g −1 ∗ g 6= id1 ∗ id2 , g −1 ∗ g 6= id2 =⇒ and this is clearly a contradiction with the beginning of the proof.  From the definition of disguised-group and the last lemma we have the inverse identity relations, ( IR (g) = IL (g −1 ) (1) ∀g ∈ G and ∀g −1 an inverse of g. −1 IL (g) = IR (g ) Definition 2.5. Let (G, ∗) be a disguised-group. For every g ∈ G and every n ∈ N we define, g n = g ∗ . . . ∗ g. n−times A positive integer n ∈ N is the order of the element g if it is the minimum n ∈ N for which g n = id for some id ∈ I(G). If it does not exist such positive integer then we will say that the element g ∈ G is of infinite order. A disguised-group G is said to be cyclic if there exists g ∈ G such that, G = {g n : n ∈ N}. A disguised-group (G, ∗) is abelian if every two elements conmute, i.e., g1 ∗ g2 = g2 ∗ g1 ∀g1 , g2 ∈ G. The most of the results of this section 2 are known in the literature of regular semigroups (see [3] and [6]) excepting proposition 2.11 that seems to be new. Into the next proposition we will prove some properties of the disguised-groups that arise from the definition. One of them very surprising is the last one, the cancellative property. We say ”surprising” because, as in other algebraic structures (see, e.g., Grothendiek’s group -see page 40 of [7]-), the cancellative property improve the algebraic object. In this case, if the cancellative property holds, a disguised-group turns into a group. 4 EDUARDO BLANCO GÓMEZ Proposition 2.6. Let (G, ∗) be a disguised-group. Then, the following properties hold, (a) Let g ∈ G. Then its right identity and its left one are unique. Moreover, if there exists an element g ∈ G such that IR (g) = I(G) (or IL (g) = I(G)) then G is a group. (b) id = idn , ∀id ∈ I(G) and ∀n ∈ N. Particularly, an identity has itself as a right and left inverse and identity. (c) For every g ∈ G its inverse is unique. Moreover, (g −1 )−1 = g with (g −1 )−1 the inverse element of g −1 and we can extend definition 2.5 and case (c) to every n ∈ Z+ . (d) Let g ∈ G be an element of finite order n ∈ Z+ . Then its right and left identities are equal and its inverse is g n−1 . (e) ∀g1 , g2 ∈ G, IR (g1 ∗ g2 ) = IR (g2 ) and IL (g1 ∗ g2 ) = IL (g1 ). (f ) If there is a unique right identity or a unique left identity for all g ∈ G, then G is a group. (g) Suppose it is true that, If g1 ∗ g2 = g1 ∗ g3 ⇒ g2 = g3 ∀g1 , g2 , g3 ∈ G, or, If g2 ∗ g1 = g3 ∗ g1 ∗ g3 ⇒ g2 = g3 ∀g1 , g2 , g3 ∈ G, then G is a group. Proof. (a) Let id1 , id2 ∈ IL (g) then by lemma 2.4, id1 = g ∗ g −1 = id2 . For the right identity is analogue. By uniqueness of left and right identities we have that |IR (g)| = 1 and |IL (g)| = 1. So, looking at definition 2.2, if IR (g) = I(G) or IL (g) = I(G), then the identity element in G would be unique so G would be a group. (b) Let id ∈ I(G). Without loss of generality (using the inverse identity relations (1)), it exists g ∈ G such that id ∈ IR (g). Then g ∗ id = g. By induction, suppose that the statement (b) is true for all natural numbers lower than n. Then, using the associative property of definition 2.2 and by the hypothesis of induction, g ∗ idn = (g ∗ idn−1 ) ∗ id = g ∗ id = g. Therefore, by (a) we have id = idn ∀n ∈ N. (c) Let g1 ∈ G and g2 , g3 ∈ G two inverses of g1 . Let id ∈ IR (g1 ). Then by lemma 2.4, ∗g2 g2 ∗ g1 = id = g3 ∗ g1 =⇒ (g2 ∗ g1 ) ∗ g2 = (g3 ∗ g1 ) ∗ g2 . Let id0 ∈ IL (g1 ). Using associative property of definition 2.2, the definition of identity and lemma 2.4, g2 ∗ id0 = g3 ∗ id0 . THEORY OF DISGUISED-GROUPS 5 Now the inverse identity relations (1) lead us to, g2 = g3 . To prove (g −1 )−1 = g it is just necessary to know that (g −1 )−1 is an inverse of g −1 (this is by hypotheses), because in that case, as g is an inverse of g −1 too, and we have proved before that the inverse is unique, the equality holds. (d) First of all it is important to say that in every disguised-group G there are elements with finite order because the identities have order 1 by (b) of this proposition 2.6. Then, using the definition 2.5, proposition 2.6 (c) and the equalities g n−1 ∗ g = g n = id and g ∗ g n−1 = g n = id for some id ∈ I(G) we conclude that the inverse of g is g n−1 . Using the same equalities we obtain that id ∈ I(G) is a right and a left identity of g so by (a) of this proposition 2.6 the proof of (d) is finished. (e) We will prove the equality for the right identity set and the other is analogue. Let id ∈ IR (g2 ), then, from the associative property of definition 2.2, (g1 ∗ g2 ) ∗ id = g1 ∗ (g2 ∗ id) = g1 ∗ g2 . Now we just need to apply property (a) of this proposition 2.6. (f) Let id ∈ I(G) be the unique right identity of the disguised-group. Take g ∈ G. Using the hypotheses, id ∈ IR (g)∩IR (g −1 ). By the inverse identity relations id ∈ IL (g −1 ) ∩ IL (g). Like the last deduction is valid ∀g ∈ G then id = id0 ∀id0 ∈ I(G) a left identity, by (a), so we conclude that G is a group. (g) We will prove the first case because the second one is analogue. Suppose that, If g1 ∗ g2 = g1 ∗ g3 ⇒ g2 = g3 ∀g1 , g2 , g3 ∈ G. As (G, ∗) is a disguised-group, for every triad g1 , g2 , g3 ∈ G with g2 6= g3 , there exist g, g 0 ∈ G such that, ( g = g1 ∗ g2 hypotheses of (g) =⇒=⇒ g 6= g 0 . 0 g = g1 ∗ g3 Then the disguised-group G stays invariant by the left action of g1 , for every g1 ∈ G, i.e., for every g ∈ G it exists g 0 ∈ G such that g = g1 ∗ g 0 . Let fix a g1 ∈ G with id ∈ IL (g1 ). Then looking at (e) of this proposition 2.6, id = IL (G). But now, using (f) of this proposition 2.6 we conclude that G is a group.  Remark 2.7. After proposition 2.6 we have the next equalities, ( From (a) proposition 2.6 : Disguised-groups = Regular semigroups From (c) proposition 2.6 : Disguised-groups = Inverse semigroups ⇒ Regular semigroups = Disguised-groups = Inverse semigroups, 6 EDUARDO BLANCO GÓMEZ just looking in references [3] and [6] the definition of regular and inverse semigroups. We will mantain the name of disguised-groups for the shake of clearness and having account that the equality proved in this remark does not come directly from definition 2.2. From now til the end we will use the following notation: let g ∈ G, we will denote idR (g) and idL (g) for its right and left identities respectively (they are unique by (a) of proposition 2.6) and g −1 for its inverse (it is unique by (c) of proposition 2.6). One important difference of disguised-groups comparing them with groups is that the set of identities is not necessarily closed by the operation ∗. This fact generates a lot of problems when we want to prove the results of the theory of groups for disguised-groups. However, the most of the concepts and results of groups theory can be obtained without a big effort. For some of the traditional results we need several surprising theorems for disguised-groups. Let’s define the concept of disguised-subgroup. Definition 2.8. Let (G, ∗) be a disguised-group. A subset Q ⊂ G is a disguised-subgroup if the following three conditions hold, (a) ∗ is a closed operation in Q. (b) For every g ∈ Q, idR (g), idL (g) ∈ Q. (c) For every g ∈ Q, g −1 ∈ Q. If the identity is unique ∀g ∈ Q then we will say that Q is a subgroup of the disguised-group (G, ∗). It is not necessary to demand the associative property of ∗ to a disguised-subgroup because this property holds just with the condition Q ⊂ G. There exist trivial disguised-subgroups into every group. Take g ∈ G, then the subsets {idR (g)} and {idL (g)} are disguised-subgroups (in fact subgroups) because of (b) and (c) of proposition 2.6. It is not true that, in general, I(G) (or some of its subsets with more than one element) is a disguised-subgroup because it is not possible to prove that the operation ∗ is closed into a subset of identities of G. Let’s see some examples of disguised-subgroups into a disguisedgroup. Let g ∈ G and let, Qg = {g 0 ∈ G : {idR (g 0 ), idL (g 0 )} = {idR (g), idL (g)}}. It is easy to see that Q is a disguised-subgroup of G just using properties (d) and (b) of proposition 2.6 and the inverse identity relations (1). If G is a disguised-group that is not a group, then it exists, at least, one element g ∈ G for which idR (g) 6= idL (g); the set Qg for that g, is a disguised-subgroup that is not a subgroup of G. Another example is the following one. Let g ∈ G be an element of finite order n. Let Qg = {g k : k ∈ {1, . . . , n}} ∪ {id(g)}. Let’s see Qg is a subgroup of G. Let’s prove first the condition (b) in the definition 2.8. By (b) of proposition 2.6 the right and left identities THEORY OF DISGUISED-GROUPS 7 of g are equal. Let k ∈ {1, . . . , n}; then the element g k ∈ Qg has the same right and left identities of g just using (a) of proposition 2.6, the equalities g k = g ∗ g k−1 and g k = g k−1 ∗ g and (e) of proposition 2.6. The operation ∗ is closed in Qg because, for every k1 , k2 ∈ Z+ , using associative property of ∗, g k1 ∗ g k2 = (g ∗ ... k1 −times = (g ∗ ∗ g) ∗ (g ∗ ... ... k2 −times ∗ g) ∗ g) = g k1 +k2 . (k1 +k2 )−times Moreover, if k1 + k2 > n we have that, g k1 +k2 = g k1 +k2 −n ∗ g n = g k1 +k2 −n ∗ id = g k1 +k2 −n ∈ G, because we have proved before that the identity for all the elements of the set Qg is the same and k1 +k2 ≤ 2n. Let’s prove condition (c) in the definition 2.8 of disguised-subgroup; let k ∈ {1, . . . , n} and g k ∈ Qg . Then, as g has finite order n, g k ∗ g n−k = g n = id(g), and , g n−k ∗ g k = g n = id(g), so using (c) of 2.6, g n−k is the inverse of g k , for all k ∈ {1, . . . , n}. Then Qg is a subgroup of G. In the next proposition we give a criteria (shorter than definition) for a subset Q of a disguised-group G to be a disguised-subgroup. Proposition 2.9. Let (G, ∗) be a disguised-group and Q ⊂ G. Then, Q is a disguised-subgroup ⇔ q1 ∗ q2−1 ∈ Q ∀q1 , q2 ∈ Q. Proof. (⇒) Suppose Q is a disguised-subgroup of G. If q1 , q2 ∈ Q then by condition (c) in the definition 2.8, q2−1 ∈ Q, so by condition (a) in the definition 2.8, q1 ∗ q2−1 ∈ Q. (⇐) Suppose q1 ∗ q2−1 ∈ Q ∀q1 , q2 ∈ Q. Let q ∈ Q, then q ∗ q −1 ∈ Q so idL (q) ∈ Q. As idL (q), q ∈ Q then q −1 ∗ idL (q) ∈ Q so using inverse identity relations (1) q −1 ∈ Q. Like q −1 ∈ Q using (c) of proposition 2.6, q −1 ∗ (q −1 )−1 ∈ Q so idR (q) = q −1 ∗ q = q −1 ∗ (q −1 )−1 ∈ Q. Finally, ∗ is a closed operation in Q because if q1 , q2 ∈ Q we have proven before that q2−1 ∈ Q, and remembering (c) of proposition 2.6 q1 ∗ q2 = q1 ∗ (q2−1 )−1 ∈ Q.  Remark 2.10. This remark aims to show an important fact we are going to use frequently later. Most of the properties and results of groups can be extended to disguised-groups with a little bit of work. Here, we prove an elementary fact that is obvious for groups but not so obvious for disguised-groups (however quite easy to see). Let (G, ∗) be a disguised-group and Q a disguised-subgroup. Let id ∈ I(G) ∩ Q. We are going to prove the invariance of a disguised-subgroup under 8 EDUARDO BLANCO GÓMEZ its own identities, i.e., ( Q = {q ∗ id : q ∈ Q} (2) Q = {id ∗ q : q ∈ Q}. Let’s prove the first equality (the second one is analogue). It is obvious that {q ∗ id : q ∈ Q} ⊂ Q because id ∈ Q. Suppose there is a q 0 ∈ Q such that q 0 6= q ∗ id for all q ∈ Q; then operating on the right by ∗id and using associativity and (b) of proposition 2.6, q 0 ∗ id 6= (q ∗ id) ∗ id = q ∗ (id ∗ id) = q ∗ id, so q 0 ∗ id 6= q ∗ id for all q ∈ Q. As this is clearly false, we conclude that q 0 ∈ {q ∗ id : q ∈ Q} so Q ⊂ {q ∗ id : q ∈ Q}. The next proposition is an specific result of the disguised-groups theory because it does not exist in groups theory. The main idea is that, due to the multiplicity of identities, despite the fact that a disguised-group has a disguised-subgroup, there can be identities of the disguised-group which are not in the disguised-subgroup. Let (G, ∗) be a disguised-group and Q, S ⊂ G. We define, Q ∗ S = {q ∗ s : q ∈ Q and s ∈ S}. Proposition 2.11. Let (G, ∗) be a disguised-group, let Q ⊂ G be a disguised-subgroup and let’s define, S = {g ∈ G \ Q : idR (g), idL (g) ∈ G \ Q}. Then, the following statements hold, (1) If I(G) ⊆ Q then S = ∅. (2) If I(G) * Q we have: (i) S 6= ∅ and S is a disguised-subgroup of G. (ii) Q∗S and S∗Q are subsets of G\(Q∪S) and all g ∈ G\(Q∪S) has its inverse into G \ (Q ∪ S). (iii) Q ∗ (G \ Q) and (G \ Q) ∗ Q are subsets of G \ S. Proof. Suppose that I(G) * Q. Let’s prove (i). Let id ∈ I(G) but id ∈ / Q. We have id ∈ S because id ∈ G \ Q and (a) of proposition 2.6. Let’s see S is a disguisedsubgroup. The operation ∗ is closed in S just using its definition and (e) proposition 2.6. Let s ∈ S, then s has its identities into S by definition and so it does s−1 ∈ S thanks to inverse identity relations (1). Let’s prove (ii). Let q ∈ Q and s ∈ S. Then both elements q ∗ s and s ∗ q has exactly an identity in Q and an identity in S using (e) from proposition 2.6. So by definition of S none of them can be into it and, as Q is a disguised-group, all its elements has their identities into Q, so, like the identities are unique by (a) of proposition 2.6, and, S ∩ Q = ∅, then Q ∗ S and S ∗ Q are subsets of G \ (Q ∪ S). Let g ∈ G\(Q∪S). If g −1 ∈ Q or g −1 ∈ S, as Q and S are disguised-groups, THEORY OF DISGUISED-GROUPS 9 remembering (c) from proposition 2.6 we would have g = (g −1 )−1 ∈ Q or g = (g −1 )−1 ∈ S and this is a contradiction. Let’s prove (iii). Let q ∈ Q and g ∈ G \ Q. Then both elements q ∗ g and g ∗ q has an identity in Q by (e) of proposition 2.6. Then those elements can not be in S by definition so they have to be in G \ S.  3. Normal and disguised-normal disguised-subgroups The concepts and results exposed in this section are new in the theory of regular semigroups. At this moment we are going to define a very useful concept in groups theory that is going to be very important too in disguised-groups theory. Definition 3.1. A disguised-subgroup Q of a disguised-group (G, ∗) is normal if ∀g1 , g2 ∈ G and ∀q1 , q2 ∈ Q we have that, (g1 ∗ q1 ) ∗ (g2 ∗ q2 ) = q3 ∗ (g1 ∗ g2 ), for some q3 ∈ Q. It is not possible, in general, to find normal disguised-subgroups because the trivial example in groups theory here does not work due to the multiplicity of the identities. However, the existence of normal disguised-subgroups into a disguised-group produces some effects on the disguised-group so, for the future, we are going to prove a criteria (shorter and more useful than definition) for a subset Q of a disguisedgroup G to be a normal disguised-subgroup. Lemma 3.2. Let (G, ∗) be a disguised-group and Q a disguised-subgroup. If Q is normal, then I(G) ⊂ Q. Proof. Let id ∈ I(G), then id is the right or the left identity for some element g ∈ G; without loss of generality, by the inverse identity relations (1), we can suppose it exists g ∈ G such that id = idR (g). By definition 3.1, (g1 ∗ q1 ) ∗ (g ∗ q2 ) = q3 ∗ (g1 ∗ g) ∀g1 ∈ G and ∀q1 , q2 ∈ Q. But the element of the right hand side has right identity id by (e) of proposition 2.6 and the element of the left hand side has right identity in Q by (e) of proposition 2.6. As both sides are equal and the right identity is unique for every element of a disguised-subgroup (by (a) of proposition 2.6) then id ∈ Q.  Proposition 3.3. Let (G, ∗) be a disguised-group and Q a disguisedsubgroup of G. Then the following conditions are equivalent, (i) Q is a normal disguised-subgroup of G. (ii) g ∗ q ∗ g −1 ∈ Q ∀g ∈ G and ∀q ∈ Q. (iii) For every q1 ∈ Q it exists q2 ∈ Q such that, g ∗ q1 = q2 ∗ g ∀g ∈ G, 10 EDUARDO BLANCO GÓMEZ and for every q2 ∈ Q it exists q1 ∈ Q such that, q2 ∗ g = g ∗ q1 ∀g ∈ G. Proof. We are going to prove the implications (i)⇒(ii)⇒(iii)⇒(i). Suppose Q is a normal disguised-subgroup of G. Let g ∈ G and −1 g be its inverse. By lemma 3.2 I(G) ⊂ Q; as Q is normal then for every g ∈ G and for every q ∈ Q, there exist q 0 ∈ Q such that g ∗ q ∗ g −1 = (g ∗ q) ∗ (g −1 ∗ idR (g −1 )) = q 0 ∗ (g ∗ g −1 ) = q 0 ∗ idL (g) ∈ Q. Suppose g ∗ q ∗ g −1 ∈ Q ∀g ∈ G and ∀q ∈ Q. First of all, let’s prove with that hypotheses and Q being a disguised-subgroup, we have I(G) ⊂ Q. Suppose there is an identity id ∈ I(G) \ Q; using (b) and (e) of proposition 2.6 the element id ∗ q ∗ id−1 = id ∗ q ∗ id has left and right identity id ∈ / Q so id ∗ q ∗ id−1 ∈ / Q, as Q is a disguisedsubgroup, and this is a contradiction with the hypotheses of this case. Then I(G) ⊂ Q. Now ∀g ∈ G and ∀q ∈ Q, it exists q 0 such that g ∗ q ∗ g −1 = q 0 . But by associativity, ∗g g ∗ q ∗ g −1 = q 0 ⇒ g ∗ q ∗ idR (g) = q 0 ∗ g. The element q ∗ idR (g) takes all the values of Q when q does, by the invariance of a disguised-subgroup under its own identities 2.10, so taking q1 = q ∗ idR (g), we have that ∀g ∈ G and ∀q1 ∈ G it exists q2 ∈ Q such that q ∗ q1 = q2 ∗ g just taking q2 = q 0 . Suppose that for every q1 ∈ Q it exists q2 ∈ Q such that, g ∗ q1 = q2 ∗ g ∀g ∈ G. Take g1 , g2 ∈ G and q1 , q2 ∈ Q. Then by hypothesis, there exist q10 , q20 ∈ Q such that, (g1 ∗ q1 ) ∗ (g2 ∗ q2 ) = (q10 ∗ g1 ) ∗ (g2 ∗ q2 ) = q10 ∗ (g1 ∗ g2 ) ∗ q2 = q10 ∗ q20 ∗ (g1 ∗ g2 ) = (q10 ∗ q20 ) ∗ (g1 ∗ g2 ). With this we have proven that ∀g1 , g2 ∈ G and ∀q1 , q2 ∈ Q we have that, (g1 ∗ q1 ) ∗ (g2 ∗ q2 ) = q3 ∗ (g1 ∗ g2 ), for some q3 ∈ Q. Now take g1 , g2 ∈ G and q ∈ Q. Then by hypothesis there exist q10 , q20 ∈ Q such that, q ∗ (g1 ∗ g2 ) = idL (q) ∗ q ∗ (g1 ∗ g2 ) = idL (q) ∗ (g1 ∗ g2 ) ∗ q20 = (idL (q) ∗ g1 ) ∗ (g2 ∗ q20 ) = (g1 ∗ q10 ) ∗ (g2 ∗ q20 ). so we have proven that ∀g1 , g2 ∈ G and ∀q3 ∈ Q we have that, q3 ∗ (g1 ∗ g2 ) = (g1 ∗ q1 ) ∗ (g2 ∗ q2 ), for some q1 , q2 ∈ Q.  The next theorem explains very good the name of disguised-groups given to this new algebraic object. THEORY OF DISGUISED-GROUPS 11 Theorem 3.4. Let (G, ∗) be a disguised-group. Then, (a) If it exists g ∈ G such that g conmutes with every element of the disguised-group then G is a group. Particularly, if G is abelian, then G is a group. (b) If G is cyclic then G is a group. (c) If G contains a normal disguised-subgroup then G is a group. Proof. (a) Suppose that it exists g ∈ G such that g conmutes with every element of the disguised-group. It is just necessary to prove that there is a unique identity for all the elements of the disguised-group. Particularly, g conmute with its inverse so the right and left identities of g are equal, i.e., idR (g) = idL (g) = id. Let g 0 ∈ G different from g and with right and left identities idR (g 0 ) and idL (g 0 ) respectively. From (e) of proposition 2.6 we have that the element g ∗ g 0 has right identity idR (g 0 ) and left identity id and that the element g 0 ∗ g has right identity id and left identity idL (g 0 ). As, for every element of the disguisedgroup, its right and left identities are unique by (a) proposition 2.6, and g ∗ g 0 = g 0 ∗ g as g conmutes with all the elements of the disguisedgroup, then idR (g 0 ) = idL (g 0 ) = id. Like this can be proved ∀g 0 ∈ G then G is a group. Suppose G is cyclic. Then it exists g ∈ G for which G = {g n : n ∈ Z \ {0}}, so for every g 0 ∈ G it exists k ∈ Z such that g 0 = g k . Take k 6= 1. Writing g 0 = g k−1 ∗ g and g 0 = g ∗ g k−1 and using (a) of proposition 2.6 we conclude that every g 0 ∈ G has the same right identity as g and the same left identity as g, so remembering (f) of proposition 2.6, G is a group. Suppose G contains a normal disguised-subgroup. Let Q be that disguised-subgroup. By (iii) of proposition 3.3 for every q1 ∈ Q it exists q2 ∈ Q such that, g ∗ q 1 = q2 ∗ g ∀g ∈ G. Fixed q1 so q2 is fixed too. By an analogue deduction as in the last case we conclude, idR (g) = idR (q1 ) , and, idL (g) = idL (q2 ), for all g ∈ G so using (f) of proposition 2.6, G is a group.  With the last theorem one realizes clearly the reason for the name disguised-groups. More surprising is the theorem after the next definitions, and crucial for our purposes in the future. Definition 3.5. Let (G, ∗) be a disguised-group and Q be a disguisedsubgroup. We say Q is disguised-normal if I(G) ⊂ Q. Definition 3.6. Let (G, ∗) be a disguised-group and Q be a disguisedsubgroup. Let g ∈ G and define the subsets of G, [g] = {g ∗ q : q ∈ Q}. 12 EDUARDO BLANCO GÓMEZ Define the operation ~ as, [g1 ] ~ [g2 ] = [g1 ∗ g2 ] ∀g1 , g2 ∈ G. We define the quotient set G/Q with the binary operation ~ as the set which contains all the subsets [g] ⊂ G with g ∈ G. For the kind of disguised-normal disguised-subgroups we have a surprising theorem that we are going to call fundamental theorem of disguised-groups. Theorem 3.7. Let (G, ∗) be a disguised-group and Q be a disguisedsubgroup. If Q is disguised-normal then, (G/Q, ~) is a group. Proof. From definition 3.6 it is obvious that the operation ~ is closed and associative in G/Q (because so it is ∗ in G). Let’s prove that there is a unique identity in G/Q. Define e = [id] for some id ∈ I(G) ⊂ Q (because Q is disguised-normal). Then by remark 2.10, [id] = Q. Take now id1 , id2 ∈ I(G) ⊂ Q. Then using twice remark 2.10, [id1 ∗ id2 ] = {(id1 ∗ id2 ) ∗ q : q ∈ Q} = {id1 ∗ (id2 ∗ q) : q ∈ Q} = {id1 ∗ q : q ∈ Q} = Q. So, using the same reasoning for all n ∈ N and for all {idj }nj=1 ⊂ I(G) ⊂ Q we have that [id1 ∗ · · · ∗ idn ] = Q, then we conclude that e = Q is the unique identity element in G/Q. The last step requires the proof that every [g] ∈ G/Q has an inverse in G/Q; but taking one g ∈ [g] it is enough to see that [g]−1 = [g −1 ].  Corolary 3.8. Let (G, ∗) be a disguised-group and Q be a disguisedsubgroup. If G \ Q contains no disguised-subgroup then (G/Q, ~) is a group. Proof. If G \ Q contains no disguised-subgroup, particularly, G \ Q contains no id ∈ I(G) because {id} is a disguised-subgroup (a subgroup in fact) for all id ∈ I(G). So I(G) ⊂ Q, then Q is disguised-normal and applying the fundamental theorem of disguised-groups 3.7 we finish.  Corolary 3.9. Let (G, ∗) be a disguised-group and Q be a normal disguised-subgroup. Let ∼ be the following equivalence relation for every g1 , g2 ∈ G, g1 ∼ g2 ⇔ g1 ∗ g2−1 ∈ Q In that case we will say that g2 ∈ [g1 ]. Let G/ ∼ be the set of all subsets [g] ⊂ G for every g ∈ G, with the binary operation  defined by, [g1 ][g2 ] = [g1 ∗ g2 ] Then (G/Q, ~) = (G/ ∼, ). ∀g1 , g2 ∈ G. THEORY OF DISGUISED-GROUPS 13 Proof. As Q is a normal disguised-subgroup then G is a group by (c) theorem 3.4, so (G/Q, ~) is a group by the traditional group theory. Knowing that G is a group, it is just an exercise to prove that ∼ is an equivalence relation and (G/ ∼, ) is a group. For g ∈ G let’s call [g]1 = {g ∗ q : q ∈ Q} and [g]2 = {g 0 ∈ G : g ∗ g 0−1 ∈ Q}. We are going to prove that [g]1 = [g]2 . Let g 0 ∈ [g]2 . Then g ∗ g 0−1 ∈ Q, so g ∗ g 0−1 = q for some q ∈ Q. As G is a group then q −1 ∗ g = g 0 . As Q is normal then it exists q 0 such that g ∗ q 0 = g 0−1 , so g 0 ∈ [g]1 . Let now g 0 ∈ [g]1 ; then it exists q ∈ Q such that g 0 = g ∗ q. As Q is normal it exists q 0 ∈ Q such that g 0 = q 0 ∗ g; as G is a group g ∗ g 0−1 = q 0−1 so g 0 ∈ [g]2 . All in all (G/Q, ~) = (G/ ∼, ).  4. Disguised-homomorphisms and isomorphy theorems for disguised-groups At this moment in which we have developed and proved all the basic properties and results of disguised-groups, we are going to define the concept of disguised-homomorphism. We will consider a different category from the natural one associate with disguised groups, i.e., we are going to consider the category of disguised-groups with the morphisms between disguised-groups and groups. This fact happens for a good reason, isomorphy theorems can be proved for this category but we would not be able to prove them considering the morphisms between disguised-groups. The main reason is that the binary operation on a disguised-group is not close for the set of the identities. Definition 4.1. Let (G1 , ∗) be a disguised-group and (G2 , ) be a group. A morphism h : G1 → G2 is said to be a disguised-homomorphism if, h(g1 ∗ g2 ) = h(g1 )  h(g2 ) ∀g1 , g2 ∈ G1 . Definition 4.2. Let (G1 , ∗) be a disguised-group and (G2 , ) be a group with identity element e. A morphism h : G1 → G2 is said to be disguised-injective if, ∀g, g 0 ∈ G2 \ {e} with g 6= g 0 : h(g) 6= h(g 0 ). A homomorphism h : G1 → G2 is said to be a disguised-monomorphism if it is disguised-injective, it is said to be a disguised-epimorphism if it is suprajective and it is said to be a disguised-isomorphism if it is suprajective and disguised-injective. Two disguised-groups (G, ∗) and (G0 , M) are said to be isomorphic e ) and two disguised-isomorphisms h, h0 such if there exists a group (G, 14 EDUARDO BLANCO GÓMEZ that, G h > G. e h0 0 G Proposition 4.3. Let (G1 , ∗) be a disguised-group and (G2 , ) be a group with identity e. Let h : G1 → G2 be a disguised-homomorphism. Then, (a) h(id) = e for all id ∈ I(G1 ). (b) h(g1 )−1 = h(g1−1 ) for all g1 ∈ G1 . Proof. (a) Let g ∈ G. Using the definition of disguised-homomorphism, h(g)  h(idR (g)) = h(g ∗ idR (g)) = h(g) = h(g)  e. Using now the cancellative property of the group G2 we conclude that h(idR (g)) = e. As the last proceeding can be done for every g ∈ G1 then h(idR ) = e for every idR a right identity in G1 . By the inverse identity relations h(id) = e for every id ∈ I(G1 ). (b) Let g1 ∈ G. By the definition of disguised-homomorphism and (a), h(g1 )  h(g1−1 ) = h(g1 ∗ g1−1 ) = h(idL (g1 )) = e. An analogue deduction can be done for idR (g1 ) operating h(g1−1 )h(g1 ). As the inverse is unique for every disguised-group (by (c) of proposition 2.6) and also for every group then (b) is true.  Definition 4.4. Let (G1 , ∗) be a disguised-group and (G2 , ) be a group with identity e. Let h : G1 → G2 be a disguised-homomorphism. We define the kernel of h as the set, Ker(h) = {g1 ∈ G1 : h(g1 ) = e} ⊂ G1 . We define the image of h as the set, Im(h) = {g2 ∈ G2 : ∃g1 ∈ G1 with h(g1 ) = g2 } ⊂ G2 . Proposition 4.5. Let (G1 , ∗) be a disguised-group and (G2 , ) be a group with identity e. Let h : G1 → G2 be a disguised-homomorphism. Then Ker(h) is a normal disguised-subgroup of G1 . Proof. Let’s prove Ker(h) is a disguised-group. Take g1 , ge1 ∈ Ker(h). Using proposition 4.3, ge1−1 ∈ Ker(h) because like h(g1 ) = e then h(g1−1 ) = h(g1 )−1 = e−1 = e. Then, h(g1 ∗ ge1−1 ) = h(g1 )  h(e g1−1 ) = e  e = e. THEORY OF DISGUISED-GROUPS 15 Using proposition 2.9, Ker(h) is a disguised-group. Let’s see Ker(h) is normal. Take g ∈ G1 and g1 ∈ Ker(h). Using that h is a disguisedhomomorphism and proposition 4.3 we have that, h(g ∗ g1 ∗ g −1 ) = h(g)  h(g1 )  h(g)−1 = h(g)  h(g)−1 = e, so g ∗ g1 ∗ g −1 ∈ Ker(h) for all g ∈ G1 and for all g1 ∈ Ker(h). Remembering (b) of proposition 3.3 the proof is finished.  Corolary 4.6. Let (G1 , ∗) be a disguised-group and (G2 , ) be a group with identity e. If it exists h : G1 → G2 a disguised-homomorphism, then (G1 , ∗) is a group. Proof. It is just necessary to use propositions 4.5 and 3.4 (c).  From now til the end of the section we are going to declare some results, concerning disguised-groups, without proof. This is because in all of them appear the condition that exists a disguised-homomorphism so, using this last corollary 4.6, the result turn into that one from traditional groups theory. For more information about the theory of groups see [4], [5] or the great and complete book [7]. Proposition 4.7. Let (G1 , ∗) be a disguised-group and (G2 , ) be a group with identity e. Let h : G1 → G2 be a disguised-homomorphism. Then, (a) If Q1 is a disguised-subgroup of G1 then Q2 = h(Q1 ) is a subgroup of G2 . (b) If Q2 is a subgroup of G2 then Q1 = h−1 (Q2 ) = {q1 ∈ G1 : h(q1 ) ∈ Q2 } is a subgroup of G1 . (c) If Q2 is a normal subgroup of G2 then Q1 = h−1 (Q2 ) is a normal subgroup of G2 . (d) If Q1 is a normal disguised-subgroup of G1 and h is a disguisedepimorphism then Q2 = h(Q1 ) is a normal subgroup of G2 . In the next proposition we obtain a consequence of the definition of isomorphic disguised-groups. Proposition 4.8. Let (G1 , ∗) and (G2 , M) be disguised-groups. If G1 and G2 are isomorphic then for all Q1 disguised-normal disguisedsubgroup of G1 exists Q2 disguised-normal disguised-subgroup of G2 such that the quotient groups G1 /Q1 and G2 /Q2 are isomorphic. Now we are going to declare the statements called, in the theory of groups, isomorphy theorems. Theorem 4.9. (First isomorphy theorem for disguised-groups) Let (G1 , ∗) be a disguised-group, (G2 , ) be a group and h : G1 → G2 be a disguised-homomorphism. Then G1 /Ker(h) is a group isomorphic to G2 . Proof. This is a direct consequence of corollary 4.6 and the first isomorphy theorem for groups.  16 EDUARDO BLANCO GÓMEZ Let’s use the notation ≈ to express isomorphy. Here we have the second isomorphy theorem for disguised-groups. Theorem 4.10. (Second isomorphy theorem for disguised-groups) Let (G, ∗) be a disguised-group and Q1 ⊂ Q2 be normal disguised-subgroups of G. Then Q1 /Q2 is a normal subgroup of the group G/Q1 and, (G/Q1 )/(Q1 /Q2 ) ≈ G/Q2 . Proof. This is a direct consequence of proposition 3.4 (c) and the second isomorphy theorem for groups.  Theorem 4.11. (Third isomorphy theorem for disguised-groups) Let (G, ∗) be a disguised-group and Q1 and Q2 be disguised-subgroups of G, with Q2 normal into G. Then Q1 ∗Q2 is a normal subgroup of the group G, Q2 is a normal subgroup of Q1 and Q1 ∩ Q2 is a normal subgroup of Q1 . Furthermore, (Q1 ∗ Q2 )/Q2 ≈ Q1 /(Q1 ∩ Q2 ) Proof. This is a direct consequence of proposition 3.4 (c) and the third isomorphy theorem for groups.  References [1] E. BLANCO, Homotopy groups of symmetric products, preprint. [2] R. BROWN, Topology and groupoids, Booksurge LLC, 2006. [3] A.H. CLIFFORD and G.B. PRESTON, The algebraic theory of semigroups, volume I, Providence: AMS, 1964. [4] J. DORRONSORO and E. HERNÁNDEZ, Números, grupos y anillos, Addison-Wesley, 1996. [5] P. DUBREIL, Teorı́a de grupos, Reverté, 1975. [6] J.M. HOWIE, Fundamentals of semigroup theory, Clarendon Press, 1995. [7] S. LANG, Algebra, Springer, 2002.
4math.GR
arXiv:1204.4106v3 [cs.DS] 25 Dec 2016 Coalescing Random Walks and Voting on Connected Graphs∗ Colin Cooper† Robert Elsässer‡ Tomasz Radzik¶ Hirotaka Ono§ December 28, 2016 Abstract In a coalescing random walk, a set of particles make independent discrete-time random walks on a graph. Whenever one or more particles meet at a vertex, they unite to form a single particle, which then continues a random walk through the graph. Let G = (V, E), be an undirected and connected graph, with n vertices and m edges. The coalescence time, C(n), is the expected time for all particles to coalesce, when initially one particle is located at each vertex. We study the problem of bounding the coalescence time for general connected graphs, and prove that   n 1  4 . log n + C(n) = O 1 − λ2 ν Here λ2 is the second eigenvalue of the transition matrix of the random walk. To avoid problems arising from e.g. lack of coalescence on bipartite graphs, we assume the random walk can be made lazy if ∗ Partially supported by the Royal Society International Joint Project grant JP090592 ”Random Walks, Interacting Particles and Faster Network Exploration,” and the EPSRC grant EP/J006300/1 ”Random walks on computer networks.” A Preliminary version of the results in this paper was presented in the Proceedings of PODC 2012 [4]. † Department of Informatics, King’s College London, UK ‡ Department of Computer Science, University of Salzburg, Austria § Department of Economic Engineering, University of Kyushu, Fukuoka, Japan ¶ Department of Informatics, King’s College London, UK 1 P required. The value of ν is given by ν = v∈V d2 (v)/(d2 n), where d(v) is the degree of vertex v, and d = 2m/n is the average degree. The parameter ν is an indicator of the variability of vertex degrees: 1 ≤ ν = O(n), with ν = 1 for regular graphs. Our general bound on C(n) holds for all connected graphs. This implies, for example, that C(n) = O(n/(1 − λ2 )) for d-regular graphs with expansion parameterized by the eigenvalue gap 1 − λ2 . The bound on C(n) given above is sub-linear for some classes of graphs with skewed degree distributions. In the voter model, initially each vertex has a distinct opinion, and at each step each vertex changes its opinion to that of a random neighbour. Let E(Cv ) be the expected time for voting to complete, that is, for a unique opinion to emerge. A system of coalescing particles, where initially one particle is located at each vertex, corresponds to the voter model in that E(Cv ) = C(n). Thus our result stated above for C(n) also gives general bounds for E(Cv ). 1 Introduction In a coalescing random walk, a set of particles make independent discretetime random walks in an undirected connected graph. Whenever two or more particles meet at a vertex, then they unite to form a single particle which then continues to make a random walk through the graph. Let G = (V, E) be an undirected connected graph with n vertices and m edges. The coalescence time is the expected time for all particles to coalesce, when initially one particle is located at each vertex of the graph. We study the problem of bounding the coalescence time for general connected graphs. For a given graph G we denote the coalescence time of an n particle system by C(n) = CG (n). In order to bound C(n), we study the coalescence time C(k) = CG (k) of a system of k particles for any 2 ≤ k ≤ n. The expected time for the k particles to coalesce to a single particle depends on their initial positions. Let Ck (i1 , ..., ik ), be the coalescence time when the particles start from distinct vertices i1 , ..., ik . The worst case expected coalescence time for k particles is C(k) = max E(Ck (i1 , ..., ik )). i1 ,...,ik In the special case of two particles, C(2) is more naturally referred to as the (worst case expected) meeting time of two random walks. 2 A system of n coalescing particles where initially one particle is located at each vertex, corresponds to another classical problem, the voter model, which is defined as follows. Initially each vertex has a distinct opinion, and at each step each vertex changes its opinion to that of a random neighbour. Let Cv be the number of steps for voting to be completed, i.e., for a unique opinion to emerge. The expected completion time of voting, E(Cv ), is called the voting time. The random variable Cv has the same distribution, and hence the same expected value, as the coalescence time Cn of n coalescing particles, where one particle is initially located at each vertex (see [2]). Thus C(n) ≡ E(Cn ) = E(Cv ), and any bound on coalescence time C(n) applies equally to the voting time E(Cv ). As the coalescence time is easier to estimate, we focus on this quantity henceforth. The coalescing random walk is the key ingredient in the self-stabilizing mutual exclusion algorithm of Israeli and Jalfon [8]. Initially each vertex emits a token which makes a random walk on G. On meeting at a vertex, tokens coalesce. Provided the graph is connected, and not bipartite, eventually only one token will remain, and the vertex with the token has exclusive access to some resource. The token makes a random walk on G, so in the long run it will visit all vertices of G in proportion to their stationary distribution. Previous work on coalescing random walks We summarize some known results for coalescing random walks. There are two distinct models for the transition times of random walks on finite graphs. In the discrete-time model, all walks make transitions synchronously at steps t = 1, 2, ... In the continuous-time model, each walk W waits for a random time tW independently of other walks, and then makes a transition. The wait time tW is an independent exponential random variable with rate 1. Let Hu,v denote the hitting time of vertex v starting from vertex u, that is, the random variable which gives the time taken for a random walk starting from vertex u to reach vertex v; and let Hmax = maxu,v E(Hu,v ). Aldous [1] considers C(2), the meeting time of two random walks, in the continuoustime model, and shows that C(2) = Ω(m/∆) and C(2) = O(Hmax ), where ∆ is the maximum degree of G. These upper and lower bounds can be far apart, e.g. for a star graph (with loops), C(2) = Θ(1) whereas Hmax = Θ(n). 3 The O(Hmax) bound on C(2) implies that C(n) = O(Hmax log n), since the number of particles halves in O(Hmax) time. Aldous [1] conjectured that C(n) is actually O(Hmax ). Earlier results by Cox [5] for the continuous-time model, imply C(n) = O(Hmax) for constant dimension tori and grids. For regular graphs, in the continuous-time model, Aldous and Fill [2] show that, C(n) ≤ e(log n + 2)Hmax , C(n) ≤ rn2 /(4s) for r-regular s-edge connected graphs, and C(n) ∼ n for complete graphs. Cooper et al. [3] confirmed that the conjecture C(n) = O(Hmax ) holds for discrete-time random walks on random regular graphs. This follows from their result that for r-regular random graphs C(n) ∼ 2((r − 1)/(r − 2))n, with high probability. We use the notation with high probability (whp), to mean with probability tending to 1 as n → ∞. The notation f (n) ∼ g(n) means that f (n) = (1 ± o(1))g(n). Simple bounds on Hmax can be obtained from the commute time between any pair of vertices (see e.g. Corollary 3.3 of Lovasz [10]). For a graph G with n vertices, m edges and minimum degree δ(G), we have m 4m ≤ Hmax ≤ . 2δ(G) (1 − λ2 )δ(G) (1) As δ(G) ≤ d the average degree, it follows that Hmax ≥ n/4 for any graph. An upper bound, for connected graphs, of Hmax ≤ 4m/(1 − λ2 ) follows from δ(G) ≥ 1. General results for coalescing walks on graphs In this article, we study the problem of bounding the coalescence time C(n) of any connected graph. We assume that the graphs G we consider are not bipartite, or that if G is bipartite, then the random walks are lazy and pause with probability 1/2 at each step. Equivalently, for the voting process, we assume that vertices may choose their own opinion with this probability. Our main result, stated formally below, is given in terms of the second eigenvalue of the transition matrix of the random walk, λ2 , and a parameter ν which measures the variability of the degree sequence. Let d(v) be the degree of vertex v, and d = 2m/n the average degree. The parameter ν, the ratio of the average squared degree to the average degree squared, is defined as P (d(v))2 . (2) ν = v∈V 2 dn 4  P 2 2 This can also be written as ν = n v∈V d (v) /(2m) . The parameter ν ranges from 1 for regular graphs to Θ(n) for a star graph. We prove the following general theorem. Theorem G be a connected graph with n vertices, m edges, and let P 1 Let 2 2 ν= n v∈V d (v) /(2m) . Let C(n) be the expected coalescence time for a system of n particles making a lazy random walk on G, where originally one particle starts at each vertex. Then   1  4 n C(n) = O log n + . (3) 1 − λ2 ν By the equivalence between coalescence and voting, the expected time E(Cv ) to complete voting on G has the same upper bound as C(n). Although Theorem 1 is a general statement of our results, the bound (3) can be improved in extremal cases. It is established in (29) of Section 5 that  2  1 m C(n) = O log n . (4) 1 − λ2 ∆ This bound is better than (3), if ∆ = ω(m/ log n). For example, for a star, (3) gives C(n) = O(log4 n) and (4) gives C(n) = O(log2 n), whereas the correct value is C(n) = Θ(log n) (since a star is a bipartite graph, we consider the lazy walk). Hassin and Peleg [7] showed that voting (hence also coalescence) is completed in expected O(n3 log n) time on any connected graph. The bound (3) is parameterized by the eigenvalue gap, and can offer a refinement of Hassin and Peleg’s bound. As 1 − λ2 = Ω(1/n2 ) for any connected regular graph, coalescence for these graphs is completed in O(n3 ) expected time. An example of a (non-regular) graph with coalescence time Θ(n3 ) is given by two cliques of size n/4 joined by a path of length n/2. On the other hand 1 − λ2 = Θ(1/n3 ) for lollypop graphs, indicating that the bound (3) is not tight. The parameter ν is related to the second moment of the degree distribution and measures the variability of the degree sequence. If ∆ is the maximum degree of G, then 1 ≤ ν ≤ ∆/d ≤ n. For near regular graphs, when the ratio of the largest to the smallest vertex degree is bounded by a constant, we have ν ≤ ∆/d = O(1), so the bound (3) becomes   n C(n) = O . 1 − λ2 5 In particular, if G is an expander in the classic sense that it is regular and its eigenvalue gap (1 − λ2 ) is constant, then C(n) = O(n). In parallel with our work, Oliveira [11] recently proved the conjecture C(n) = O(Hmax) for continuous-time random walks. The result of Oliveira implies an analogous linear bound C(n) = O(n) for continuous-time random walks on expanders. We note that the bound (3) is qualitatively different from O(Hmax ), as the graph structure is made explicit through the parameter ν. As Hmax = Ω(n) for any graph (see (1)), the bound (3) can improve on C(n) = O(Hmax). This can occur for example if ν = ω(1), but also when ν = Θ(1), since there are graphs with Hmax = ω(n/(1 − λ2 )). Some examples follow. For graphs with a power law (heavy tailed) degree distribution, Theorem 1 can give sublinear bounds on the coalescence and voting times as the following example shows. Mihail et al. [6] prove that for 2 < α < 3, the random Θ(n)-vertex graph with ⌈n/dα ⌉ vertices of degree d, for d = 3, 4, . . . , n1/2 , −2 has an Ω(log n) eigenvalue gap. For this class of power law graphs, ν =  (3−α)/2 Θ n , so Theorem 1 implies a sublinear O(n(α−1)/2 log2 n) voting time, whereas for any graph, Hmax = Ω(n). There are also examples of graphs with ν = Θ(1) for which our bound is asymptotically better than O(Hmax). Consider the graph consisting of (log n)-degree expander (1 − λ2 ≤ c < 1) with an additional vertex attached to one of the vertices of the expander. For this graph ν = Θ(1) and 1 − λ2 is a positive constant, so C(n) = O(n), but Hmax = Θ(n log n). It would be interesting to have a general lower bound on C(n) which incorporates the graph structure in a similar way to the upper bounds (3) and (4), but it is not clear what form such a bound might take. A weaker conjecture is C(n) = Ω(1/(1 − λ2 )). This bound is tight for a path on n vertices, where 1/(1 − λ2 ) = Θ(n2 ) and C(n) = C(2) = Θ(n2 ). Indeed C(n) ≤ n2 , the cover time of the graph by a particle starting from the left most vertex; and C(2) ≥ n2 /4, the expected hitting time of the central vertex by particles starting from the left most and right most vertices. Structure of the paper The analysis of the coalescence process (that is, the proof of Theorem 1) is divided into two phases. During the first phase the number of particles decreases from the initial n to a threshold value k ∗ . This phase is analysed by showing that for a suitably chosen number of steps t∗ = t∗ (k ∗ ), the probability 6 that there exist k ∗ particles which do not have a single meeting between them within the first t∗ steps is at most 1/2. This implies that with probability at least 1/2, the number of particles at step t∗ is less than k ∗ . The second phase, when the number of particles decreases from k ∗ to 1, is analysed by bounding the expected time we have to wait until the first meeting between any of k particles, where 2 ≤ k ≤ k ∗ . At the time of this first meeting, the number of particles decreases from k to k − 1 (with some relatively small probability, the first meeting could involve more than 2 particles, reducing the number of particles to fewer than k − 1). The analysis of the second phase is based on the following theorem bounding the expected time to first meeting between any of k particles. Theorem 2 Let k ∗ be given by     n 1/2 m ∗ k = max 2, min , log n , , ν 2∆ (5) where ∆ is the maximum degree, and ν given by (2). For 2 ≤ k ≤ k ∗ particles starting from arbitrary vertices in G, let Mk be the time to first meeting. Then   n  1  k log n + 2 . (6) E(Mk ) = O 1 − λ2 νk The expression (5) for the threshold value k ∗ is not very transparent, but seems to be necessary to deal with the generality of degree sequences of connected graphs. Provided the maximum degree of the graph satisfies ∆ ≤ 2m/ log2 n, then k ∗ = log n. The other terms are there to cover extremal cases such as star graphs. The condition that k ∗ ≥ 2 ensures there are at least 2 particles to coalesce. Section 2 gives background material on random walks. Section 3 replaces multiple random walks by a single walk on a suitably defined product graph. Theorem 2 is proven in Section 4 and the proof of Theorem 1 is concluded in Section 5. 2 Random walk properties Let G = (V, E) denote a connected undirected graph, |V | = n, |E| = m, and let d(v) be the degree of a vertex v. A simple random walk Wu , u ∈ V , 7 on graph G is a Markov chain modeled by a particle moving from vertex to vertex according to the following rule. The probability of transition from vertex v to vertex w is equal to 1/d(v), if w is a neighbour of v, and 0 otherwise. The walk Wu starts from vertex u at t = 0. Denote by Wu (t) the vertex reached at step t; Wu (0) = u. We assume G is connected, and the random walk Wu on G is ergodic with stationary distribution π, where πv = d(v)/(2m). If this is not the case, e.g. G is bipartite, then the walk can be made ergodic, by making it lazy. A random walk is lazy, if it moves from v to one of its neighbours w with probability 1/(2d(v)), and stays at vertex v with probability 1/2. Let P = P (G) be the matrix of transition probabilities of the walk and let Put (v) = Pr(Wu (t) = v). The eigenvalues of P (G) are real, and can be ordered λ1 = 1 > λ2 ≥ · · · ≥ λn , where λn > −1 as the walk is ergodic. Let λ = max(λ2 , |λn |) < 1. The rate of convergence of the walk is given by |Put (x) − πx | ≤ (πx /πu )1/2 λt , (7) where |r| is the absolute value of the real number r. For a proof see for example, Lovasz [10] Theorem 5.1. We assume henceforth that λ = λ2 . If not, the standard way to ensure that λ = λ2 = λ2 (G), is to make the chain lazy. We use the following definition of mixing time TG , for a graph G. For all vertices u and x in G and any t ≥ TG ,   1 (t) |Pu (x) − πx | ≤ o . (8) n2 For convenience we assume that TG = Ω(log n), even if this is not necessary. Let Eπ (Hw ) denote the expected hitting time of a vertex w from the stationary distribution π. The quantity Eπ (Hw ) can be expressed as (see e.g. [2], Chapter 2) Eπ (Hv ) = Zvv /πv , (9) where Zvv = ∞ X (Pv(t) (v) − πv ). (10) t=0 Let Av (t; u) denote the event that Wu does not visit vertex v in steps 0, ..., t. The following lemma gives a bound on the probability of this event in terms of Eπ (Hv ) and the mixing time of the walk. 8 Lemma 1 Let T = TG be a mixing time of a random walk Wu on G satisfying (8). Then Pr(Av (t; u)) ≤ e−⌊t/(T +3Eπ (Hv ))⌋ . (T ) Proof. Let ρ ≡ Pu be the distribution of Wu on G after T steps. Then (8) and the fact that πx ≥ 1/n2 for any connected graph imply Eρ(Hv ) = (1 + o(1))Eπ (Hv ). (11) Let Hv (ρ) be the time to hit v starting from ρ, and let τ = T + 3Eπ (Hv ). Then, noting that Eρ (Hv ) ≡ E(Hv (ρ)), Pr(Av (τ ; u)) = Pr( Av (T ; u) and Hv (ρ) ≥ 3Eπ (Hv ) ) ≤ Pr ( Hv (ρ) ≥ 3Eπ (Hv ) ) ≤ Pr ( Hv (ρ) ≥ e · E(Hv (ρ)) ) 1 ≤ . e By restarting the process Wu at Wu (0) = u, Wu (τ ), Wu (2τ ), . . . , Wu ((⌊t/τ ⌋− 1)τ ), we obtain Pr(Av (t; u)) ≤ e−⌊t/τ ⌋ . ✷ 3 Multiple random walks We consider the coalescence of k ≥ 2 independent random walks on a graph G = (VG , EG ). To do this we replace the k walks by a single walk as follows. Let graph Q = Qk = (VQ , EQ ) have vertex set VQ = V k . Thus a vertex v of Qk is a k-tuple v = (v1 , v2 , ..., vk ) of vertices vi ∈ VG , i = 1, ..., k, with repeats allowed. Two vertices v, w ∈ VQ are adjacent if {v1 , w1 }, ..., {vk , wk } are edges of G. There is a direct equivalence between k random walks Wui (t) on G with starting positions ui and a single random walk Wu (t) on Qk with starting position u = (u1 , u2 , ..., uk ). For any starting positions u = (u1 , ..., uk ) of the walks, let Mk (u) be the time until the first meeting in G. Let Sk ⊆ V (Qk ), the diagonal set of vertices, be defined by S = Sk = {(v1 , ..., vk ) : vi = vj some 1 ≤ i < j ≤ k}. 9 If the random walk on Qk visits this set, two particles occupy the same vertex in the underlying graph G and a (coalescing) meeting occurs. The number of visits to the set Sk by a random walk is not a readily manipulated quantity. An easier approach is to contract Sk to a single vertex γ = γk = γ(Sk ), thus replacing Qk by a graph Γ = Γk . On contraction, all edges, including loops, are retained. Thus dΓ (γ) = dQ (S), where dF denotes vertex degree in graph F , and the degree dF (X) of a set X is the sum of the degrees of the vertices in X. Moreover Γ and Q have the same total degree, and the degree of any vertex of Γ other than γ is the same as in graph Q. Let π and π̂ be the stationary distributions of a random walk on Q and Γ, P respectively. If v 6∈ S then π̂v = πv , and π̂γ = πS ≡ x∈S πx . It follows that, if TΓ is a mixing time satisfying (8) in Γ, then E(Mk (u)) ≤ TΓ + (1 + o(1))Eπb (Hγk ), (12) where Eπb (Hγk ) is the hitting time of γk in Γ from stationarity. Since we have replaced k individual walks on G by a single walk on Qk , and then on Γ, we need to relate mixing times on TQ and TΓ directly to a given mixing time TG of a single random walk on the underlying graph G. (We will need TΓ in two places: in the bound (12) and when applying Lemma 1 to graph Γ.) Lemma 2 For random walks in graphs G, Q and Γ, there are mixing times   log n TG = O , TQ = O(kTG ), TΓ = O(kTG ), (13) 1 − λ2 (G) such that max |Put (x) − πx | = o(1/n2F ), for any t ≥ TF , u,x∈VF where F is any of the graphs G, Q or Γ, and nF = |VF |. Proof. The bound on TG is well known (see for example, Sinclair [12]): 1/(1−λ2 ) use (7), observing that πx /πu = O(n) and λ2 has a constant c < 1 upper bound. To use (7) also to derive bounds on TQ and TΓ , we need to know the eigenvalues of Qk and Γ in terms of the eigenvalues of G. We have λ2 (Γ) ≤ λ2 (Qk ) and λ2 (Qk ) = λ2 (G). This follows from established results, as we next explain. In the notation of Markov processes, the random walk on Qk is known as the tensor product chain, and its eigenvalues are the k-wise products of the 10 Graph G Qk Γk vertices nG = n nQ = nk nΓ ≤ nk Stationary distribution π πv = d(v)/2m πv = d(v1 ) · · · d(vk )/(2m)k πγ ≥ k 2 ν/(8n) Mixing time TG = O(log n/(1 − λ2 )) TQ = O(kTG ) TΓ = O(TQ ) Table 1: The main parameters of the random walks on graphs G, Qk and Γk . eigenvalues of G. Thus, assuming λ2 (G) ≥ λn (G), it follows that λ2 (Qk ) = λ2 (G). See [9] page 168 for more details. In the notation of [2, Ch. 3], the random walk on Γ is the random walk on Qk with S collapsed to γ(S). It is proved in [2, Ch. 3], Corollary 27, that if a subset A of vertices is collapsed to a single vertex, then the second eigenvalue of the transition matrix cannot increase (in that corollary the variable τ2 = 1/(1 − λ2 )). Thus λ2 (Q) ≥ λ2 (Γ). We get the factor k in the bounds (13) on the mixing times TQ and TΓ , because πx /πu = O(n2k ) and we need |PuT (x)−πx | = o(1/n2k ), as the number of vertices in graphs Q and Γ is O(nk ). ✷ For reference, we record the salient facts for the graphs G, Q, Γ in Table 1. The bound on πγ will be established in Lemma 4. 4 Hitting time from stationarity – Proof of Theorem 2 The proof of Theorem 2 is based on Inequality (12) and on a good upper bound on the expected hitting time of vertex γ by a random walk in Γ which starts from the stationary distribution. We obtain such a bound using (9) by deriving an upper bound on Zγγ (Lemma 3) and a lower bound on the stationary probability πγ = π bγ (Lemma 4). Lemma 3 Let F be a graph with the eigenvalue gap 1 − λ2 , then Zvv ≤ 1 . 1 − λ2 In particular, for any vertex v of G, Q or Γ, Zvv ≤ 1/(1 − λ2 (G)). 11 (14) Proof. Let λ2 = λ2 (F ). Using (7) with x = u = v gives |Pvt(v) − πv | ≤ λt2 , and thus Zvv = X (Pvt (v) − πv ) ≤ t≥0 X λt2 = t≥0 1 . 1 − λ2 The proof of Lemma 2 establishes that (1 − λ2 (Γ)) ≥ 1 − λ2 (Q) = 1 − λ2 (G). ✷ Lemma 4 Let G be a connected graph with n vertices and m edges. Let     n 1/2 m ∗ , log n , (15) , k = max 2, min ν 2∆ P where ∆ is the maximum degree of G and ν = (n/(2m)2 ) v∈V d2 (v). Let k be integer, 2 ≤ k ≤ k ∗ . Let γ = γk in Γ be the contraction of S = Sk in Q. Then k2ν d(γ) ≥ . (16) πγ = (2m)k 8n Proof. By definition, d(γ) = d(S). If k = 2, then, X ν d(S) = d2 (v) = (2m)2 . n v∈V If 3 ≤ k ≤ k ∗ , for 1 ≤ x < y ≤ k, define the following subsets of S: S(x,y) = {(v1 , . . . , vk ) : vx = vy }. We have S= [ S(x,y) , 1≤x<y≤k and d S(x,y)  = (2m)k−2  X v∈V ν d2 (v) = (2m)k . n For {x, y} = 6 {p, q}, d S(x,y) ∩ S(p,q) equals to P (2m)k−4 v,u∈V d2 (v)d2(u), if {x, y} ∩ {p, q} = ∅, or P (2m)k−3 v∈V d3 (v), if |{x, y} ∩ {p, q}| = 1. 12 Therefore, from the inclusion-exclusion principle, X X   d(S) ≥ d S(x,y) − d S(x,y) ∩ S(p,q) {x,y} {x,y}6={p,q}       2 ∆ν k k k kν kν (2m) 2 − 3 (2m)k (2m) − 3 ≥ n 4 n 3 2mn 2     2 ν k k ν k∆ (2m)k ≥ 1− − 2 n 4n 2m   ν k (2m)k . ≥ 2n 2 (17) (18) (19) The factor 3 in (17) occurs as the number of ways to partition 4 objects into disjoint sets of size 2, and partition 3 objects into sets of size 2 with single intersection, respectively. The bound (19) follows from (18), by noting the upper bound on k in (15). ✷ Proof of Theorem 2. Let Mk be the time of the first meeting among k ≤ k ∗ particles in G, and let γ = γk be the contraction of the diagonal set S = Sk . Using (9) for graph Γ and with v = γ, and Lemmas 3 and 4 we have, that the hitting time Hγ of γ from stationarity has expected value 1 1 π(γ) 1 − λ2 8 n 1 ≤ 2 . k ν 1 − λ2 Eπ (Hγ ) ≤ (20) (21) Since TΓ = O(kTG ), and referring to (12) and Table 1, E(Mk ) ≤ O(kTG ) + (1 + o(1))Eπ (Hγ )   1  n  = O k log n + 2 . 1 − λ2 νk (22) (23) ✷ Let Ck be the time for k ≤ k ∗ particles to coalesce. For use in the proof of Theorem 1 in the next section, we state an upper bound on E(C P k ) which follows directly from Theorem 2. Using (23) and noting that s (1/s2 ) ≤ π 2 /6 is constant, we have E(Ck ) ≤ k X E(Ms ) = O s=2 13   n 1  2 k log n + . 1 − λ2 ν (24) 5 Coalescence time: Proof of Theorem 1 We consider the case of n coalescing particles, where each particle is initially located at a distinct vertex of the graph. The purpose of this section is to conclude the proof that for any connected graph   1  4 n C(n) = O log n + . (25) 1 − λ2 ν To establish this result, we first prove that the probability that there exist k ∗ particles which do not have a single meeting between them within the first t∗ steps is at most 1/2, if t∗ = k ∗ log n (TΓ + 3Eπ (Hγ )) , where Γ = Γk∗ , γ = γk∗ and the value of k ∗ is given in (15). An upper bound on the expected time E(Ck ) for k ≤ k ∗ particles to coalesce is given in (24) above, and we can deal with that part separately. Let P = P(v) be the set of k ∗ particles starting from vertices v = (v1 , ..., vk∗ ). The probability that the particles in P do not meet by time t is the same as the probability that the random walk in Γ starting from v does not visit γ by time t. We apply Lemma 1 to graph Γ, vertex γ and t = t∗ , and obtained that Pr(no meeting among particles in P before t∗ ) ∗ ∗ ≤ e−k log n = n−k . In the coalescence process, we can assume that if two or more particles meet at the same vertex, then the lowest index particle survives (and continues its random walk) while the other particles die. Thus if there are k ∗ or more particles after t∗ steps, then there is a set P of k ∗ particles which do not meet within t∗ steps. Therefore, Pr(at least k ∗ particles after t∗ steps) ≤ Pr(exists a set P of k ∗ particles with no meeting before t∗ )   1 n −k∗ n ≤ . ≤ (26) ∗ k 2  The last inequality holds because nk ≤ nk /k! and k ∗ ≥ 2. The bound (26) implies that the expected number of steps until fewer than k ∗ particles remain is at most t∗ + 12 (2t∗ ) + 41 (3t∗ ) + · · · = 4t∗ . Therefore, using TΓ = 14 O(k ∗ log n/(1 − λ2)) from Lemma 2, the bound on Eπ (Hγ ) given in (21), and the bound on E(Ck∗ ) given in (24), we obtain the bound (25): C(n) ≤ 4t∗ + E(Ck∗ )  ∗  (k log n)2 n 1 log n n 1  ∗ 2 = O (k ) log n + + + 1 − λ2 1 − λ2 k ∗ ν 1 − λ2 ν    1 log n n = O (k ∗ log n)2 + ∗ (27) 1 − λ2 k ν   n 1  4 log n + . = O 1 − λ2 ν The last bound above is obvious if k ∗ = log n. If k ∗ < log n, then the last bound holds because the second term in the sum in (27), that is (n/ν) log n/k ∗ , is O(log3 n). Indeed, if k ∗ < log n, then from the definition of k ∗ , either  n 1/2 m < log n or 2∆ < log n. If the former, then the second term in the sum ν in (27) is clearly O(log3 n). Observe that n n2 d2 ≤ = ν ∆2  2m ∆ 2 . (28) m < log n, then n/ν = O(log2 n), and the second term in the sum Thus if 2∆ in (27) is again O(log3 n). m We conclude by noting that since k ∗ ≤ 2∆ and (28), then (27) implies  2  1 m C(n) = O log n . (29) 1 − λ2 ∆ The above bound is better than (25), if ∆ = ω(m/ log n). References [1] D. Aldous. Meeting times for independent Markov chains. Stochastic Processes and their Applications 38(2):185-193, August 1991. [2] D. Aldous and J. Fill. Reversible Markov Chains and Random Walks on Graphs, http://stat-www.berkeley.edu/pub/users/aldous/RWG/book.html. 15 [3] C. Cooper, A. M. Frieze, and T. Radzik. Multiple Random Walks in Random Regular Graphs. SIAM J. Discrete Math. 23(4):1738-1761, 2009. [4] C. Cooper, R. Elsässer, H. Ono, T. Radzik. Coalescing random walks and voting on graphs. In PODC 2012: Proceedings of the 2012 ACM Symposium on Principles of Distributed Computing, pages 47-56, July 2012. [5] J. T. Cox. Coalescing random walks and voter model consensus times on the torus in Zd . The Annals of Probability 17(4):1333-1366, October 1989. [6] C. Gkantsidis, M. Mihail, and A. Saberi. Conductance and congestion in power law graphs. In SIGMETRICS 2003: Proceedings of 2003 ACM SIGMETRICS Intl. Conf. on Measurement and Modeling of Computer Systems, New York, NY, USA, pages 148-159, 2003. [7] Y. Hassin and D. Peleg. Distributed probabilistic polling and applications to proportionate agreement. Information & Computation 171(2):248-268, December 2001. [8] A. Israeli and M. Jalfon. Token management schemes and random walks yeild self stabilizing mutual exclusion. In PODC 1990: Proceedings of the 9th Annual ACM Symposium on Principles of Distributed Computing, Quebec City, Quebec, Canada, pages 119-131, August 1990. [9] D. Levin, Y. Peres, and E. Wilmer. Markov Chains and Mixing Times. American Mathematical Society, 2009. [10] L. Lovász. Random walks on graphs: a survey. Bolyai Society Mathematical Studies. Combinatorics, Paul Erdős is Eighty 2:1-46, Keszthely, Hungary, 1993. [11] R. Oliveira. On the coalescence time of reversible random walks. Trans. Amer. Math. Soc. 364(4): 2109-2128, 2012. [12] A. Sinclair. Improved bounds for mixing rates of Markov chains and multicommodity flow. Combinatorics, Probability and Computing 1(4):351370, December 1992. 16
8cs.DS
arXiv:1611.00532v1 [cs.DS] 2 Nov 2016 An asymptotically optimal, online algorithm for weighted random sampling with replacement MICHAŁ PIOTR STARTEK University of Warsaw, Faculty of Mathematics, Informatics, and Mechanics November 3, 2016 Abstract This paper presents a novel algorithm solving the classic problem of generating a random sample of size s from population of size n with nonuniform probabilities. The sampling is done with replacement. The algorithm requires constant additional memory, and works in O(n) time (even when s >> n, in which case the algorithm produces a list containing, for every population member, the number of times it has been selected for sample). The algorithm works online, and as such is well-suited to processing streams. In addition, a novel method of mass-sampling from any discrete distribution using the algorithm is presented. 1 Introduction n−1 Assume that we are given a population of elements P = {ei }i=0 , n ∈ N (at least at first, the problem of infinite populations will be elaborated on later), along with a sequence of probabilities of each element of P , denoted {pi }n−1 i=0 , Pn−1 such that i=0 pi = 1, and a single number s, the sample size, which might be greater or lower than n. The task is to compute a random sample of size s from the population P , such that each element Xi from the sample is one of the elements of P , each with its corresponding probability. Note that without loss of generality we can (and will) assume that P = {0, ..., n − 1} The algorithm assumes a non-naive (constant-time) implementation of procedures for sampling single random numbers from the beta (in the easy case, where α and β parameters are integer and ≥ 1), and binomial distributions, as Author’s address: M. Startek Wydział Matematyki, Informatyki i Mechaniki Uniwersytetu Warszawskiego ul. Banacha 2 02-097 Warsaw Poland 1 ALGORITHM 1: The naive sampling algorithm 1 2 3 4 5 6 7 8 n−1 Input: The sequence of probabilities pi i=0 , desired sample size s Output: A multiset R containing the random sample Pi−1 Compute array {Si }n−1 i=0 = j=0 pj ; R = new empty multiset; repeat s times Randomize X ∼ U (0, 1) ; Find greatest k s.t. S[k] < X using binary search ; Add k to R until; R contains the result ; well as lack of numerical errors. Some consideration to mitigating the effects of numerical inaccuracies will be given in later sections. The algorithm is best presented (as the author feels) by starting from the naive algorithm, and iteratively refining it, until the desired time and memory complexity are reached. 2 The naive algorithm The naive algorithm (which, despite its non-optimal costs, in practice is reasonably efficient, and is used, in its second variant, for example by the numpy numeric library for Python) is based on a cruicial idea, which will be used also in the novel version presented here. The idea is based on a geomertical intuition: if an interval [0, 1] is divided into parts with lengths pi , then sampling a random number X from the uniform distribution U(0, 1) and picking the subinterval of [0, 1] into which it falls (and the corresponding element of P ) results in a choice of a single element of P with the desired probability distribution. Efficient finding of the selected subinterval is faciliated by precomputing an array of cumulative sums of probabilities, then performing a binary search on it. The algorithm consumes O(n) time for initialization (lines 1 and 2), then O(s log(n)) time for the actual sampling, and O(n) memory space for additional data structures (not counting the O(s) for the result). 3 Omitting the computation of cumulative sums table The first modification of the algorithm makes it possible to skip the necessity to precompute the array of cumulative sums in line one. Instead it samples all the necessary random numbers Xi ∼ U(0, 1), sorts the X array, and then processes iid the pi sequence at the same time as X, in fasion similar to the merge step of the mergesort algorithm. 2 ALGORITHM 2: The sampling algorithm without cumulative sums table 1 2 3 4 5 6 7 8 9 10 11 12 13 n−1 Input: The sequence of probabilities pi i=0 , desired sample size s Output: A multiset R containing the random sample R = new empty multiset; idx = 0 ; cumulativeP robSum = 0.0 ; Randomize X ∼ U (0, 1)s ; Sort X in ascending order; foreach x ∈ X do while cumulativeProbSum < x do cumulativeP robSum+ = p[idx] ; idx + + end Add idx − 1 to R; end R contains the result ; The algorithm runs in O(s log(s) + n) time, (which is not an improvement over the previous version): lines 4 and 5 take a total of O(s log(s)), the outer loop runs s times, while the inner loop runs a total of n times (as the variable idx is bounded by n). The algorithm uses O(s) memory. 4 Omitting the sorting The algorithm might be further improved if the table X could be generated in an already sorted order. This is, in fact, possible: it is a well-known fact that if X0 , ..., Xn−1 ∼ U(0, 1), then min(X0 , ..., Xn−1 ) ∼ Beta(1, n), and is the first iid element of the sought table [1]. Since the variables are independent, then, after sampling the minumum using this method, it is easy to see that the remaining variables (under condition that they have to be not less than the minimum) are distributed according to U(M, 1), where M is the minimum. The second-lowest variable might be sampled with the same method after rescaling U(M, 1) to U(0, 1), and so on. In fact, this allows us to drop the step of precomputing the table X altogether, and to just compute the consecutive variables "on the go", making the algorithm capable of online operation, as well as improving the runtime. The algorithm runs in O(n + s) time, and requires constant additional memory if working online: in that case, every intermediate result is immediately provided to the calling procedure for consumption (and possibly, immediately discarded), instead of being explicitly stored in R. 3 ALGORITHM 3: The online algorithm 1 2 3 4 5 6 7 8 9 10 11 12 13 n−1 Input: The sequence of probabilities pi i=0 , desired sample size s Output: A multiset R containing the random sample R = new empty multiset; idx = 0 ; cumulativeP robSum = 0.0 ; currentX = 0.0 ; for i in s, ..., 1 do currentX+ = Beta(1, i) ∗ (1.0 − currentX) ; while cumulativeP robSum < currentX do cumulativeP robSum+ = p[idx] ; idx + + ; end Add idx − 1 to R ; end R contains the result ; 5 The case of s >> n The practical speed of the above algorithm is constrained by the speed of the sampling from the beta distribution, the remaining operations being trivial in comparison. This provides an opportunity for optimization: if the population in small with respect to the numer of samples required, then the algorithm will have to sample from the beta distribution many times for any population member. This can be avoided by changing the reasoning: insead of asking "where will the next Xi be?" we can ask "how many Xes we will encounter while going through the current pi ?". The answer to that, for p1 is the binomial distribution: Binom(s, p1 ). For further pi s the answer is the same distribution, only conditioned on the number of Xes and the population probability already Pi−1 consumed: Binom(s − |Xused |, pi /(1.0 − j=0 pj )).This is in fact a standard algorithm for sampling from multinomial distribution (which is exactly the same problem as random sample with replacement: only the former terminology is most often used in contexts where s >> n, and the latter otherwise). The provided algorithm runs in O(n) time (assuming that R behaves like a counter, and increasing the count of a given item is done in constant time), and in constant memory. It is capable of working online. It should be noted, that although it achieves the optimal theoretical asymptotic runtime, its practical implementation will be very inefficient when s << n: even if a sample consisting only of one element is desired, it will perform n expensive operations of sampling from binomial distribution. 6 A practical algorithm The previous two algorithms are opposites in terms of their practical pessimistic case: the one using beta distribution has to randomize once per each requested 4 ALGORITHM 4: The online algorithm for s > n 1 2 3 4 5 6 7 8 9 10 n−1 Input: The sequence of probabilities pi i=0 , desired sample size s Output: A multiset R containing the random sample R = new empty multiset; idx = 0 ; cumulativeP robSum = 0.0 ; for i in 0, ..., n − 1 do Randomize N ∼ Binom(s, p[i]/(1.0 − cumulativeP robSum)) ; Add i to R with multiplicity N ; s− = N ; cumulativeP robSum+ = p[i] ; end R contains the result ; sample member, and so, runs fast if s << n, and slowly if n << s, while the one using the binomial distibution has to randomize once for every member of the population, and as such is efficient in practice only for large values of s and small n. It turns out that it is actually possible to create a hybrid algorithm which combines the strengths of both of them. What’s more, the algorithm doesn’t work by first examining the data, and then choosing one of the previous versions and runnig it, instead it adapts "on the fly", is capable of switching back and forth between modes during runtime as needed, and does not need to examine the data in advance, which keeps it compatible with online operation. Recall the metaphor of a segment divided into fragments corresponding to the population members, with lengths equal to their probabilities. The algorithm may be imagined as if walking along the segment, picking its sample along the way. It can make two kinds of steps: first is the "beta" step, with constant average length, which may pass over multiple small population elements, and results in adding to the sample the population member in which it ends (with multiplicity of one). The disadvantage is that if a large population member is encountered then it may take multiple beta steps to pass it. The other kind of step, the binomial step immediately travels forward to the end of the current population member, adding it to the sample with multiplicity according to the result of randomization. Obviously it makes sense to use this type of step while traversing population members with large probabilities. This is achieved through the condition in line 7: the expected number of samples randomized from the current member of the population is compared to a constant (1.0 here). The result of this comparison is used to determine whether to proceed in beta or in binomial mode. Any positive constant here is good enough to achieve the theoretical bounds, however, in practice it is best to choose it based on the relative costs of sampling from beta and from binomial distributions. The algorithm is still online (although presented in non-online form for readability), works in constant memory if online (results are immediately consumed by caller, instead of being stored in R). A careful reader might notice that the algoritm as presented runs in pessimistic O(n + s) time. The pessimistic 5 ALGORITHM 5: The final algorithm 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 n−1 Input: The sequence of probabilities pi i=0 , desired sample size s Output: A multiset R containing the random sample R = new empty multiset; idx = 0 ; cumulativeP robSum = 0.0 ; currentP osition = 0.0 ; while s > 0 do cumulativeP robSum+ = p[idx] ; while (cumulativeP robSum − currentP osition) ∗ s/(1.0 − currentP osition) < 1.0 do currentP osition+ = Beta(1.0, s) ∗ (1.0 − currentP osition) ; while cumulativeP robSum < currentP osition do idx + + ; cumulativeP robSum+ = p[idx] ; end Increase multiplicity of idx in R by 1 ; s− = 1 ; if s == 0 then Terminate algorithm, R contains the result end end Randomize N ∼ Binom(s, (cumulativeP robSum − currentP osition)/(1.0 − currentP osition)) ; Increase multiplicity of idx in R by N ; s− = N ; idx+ = 1 ; currentP osition = cumulativeP robSum ; end R contains the result ; time is achieved if the algorithm encounters an element of the population with probability small enough that it decides to use the beta mode, but, due to bad luck, proceeds to draw O(s) infinitesimal samples from the beta distribution before leaving the element and proceeding forward. This may be easily avoided by adding a hard condition that would force a binomial mode after a constant number of consecutive beta samples. This was omitted from the main code presented here for readability, and also because it is not a concern for any practical application. However, with the hard limit, the algorithm can perform at most n binomial samples (as each binomial sample increases the idx variable - which is bounded by n - and with the hard limit it is possible to perform at most O(n) beta samples) - therefore its runtime is bounded by O(n). 6 7 Practical notes The algorithms presented here depend heavily on good implementations of functions for sampling from binomial and from beta distribution. In particular, anyone undertaking the implementation of the algorithm is advised to write a custom version of function for sampling from β(α, β) distribution: the algorithm always samples with α = 1, and in this case, the distribution has an explicit, invertible CDF - and so, a custom sampler using the inverse CDF method will always be faster than a sampler from any scientific library which has to handle the general case. Regarding the binomial distribution, the C++11 function for sampling as implemented by the GNU project’s libstdc++ (standard C++ library on most Unix systems) is inadequate for the task as it seems to have non-constant complexity with regard to its parameters. In any tests performed I have used the function as implemented by the Boost library which seems to work not only faster than libstdc++’s, but also runs in constant time. Regarding the numerical stability of the algorithm, the algorithm computes a cumulative sum of all encountered probabilities, which is of course tricky. If precise numerical correctness is required, then the summation should be done using Kahan’s [4] or even Shewchuk’s [5] summation algorithms. However, for most if not all practical purposes imaginable, this is not necessary. Special care must be taken, however, as sometimes, due to numerical errors, the last sampling from binomial distribution might be performed with probability of success slightly greater than 1 - the programmer must ensure that the sampling function just assumes 1 instead of crashing. As a side note: the algorithm does benefit slightly from having its input data sorted (in either ascending or descending order, doesn’t matter): it casues the algorithm to perform less switches between beta and binomial mode, and so, it minimises the number of elements which are partially dealt with in binomial mode, and partially in beta mode. The very slight speed benefit does not justify spending the computational time (and especially the loss of asymptotic optimality and the ability to work online) needed to sort the data. However, this means that in order to avoid any bias in runtime tests described further, the input data for all tests was randomly shuffled. 8 Comparison with other sampling methods and runtime tests The algorithms presented in this paper have been implemented in C++11 programming language for purposes of testing and speed comparison. These are compared with the implementation of standard Walker’s alias method [6] as implemented by the R programming language (later referred to under name "Walker-R"), the full implementation of R’s sampling function (which examines the data, heuristically chooses between Walker’s algorithm or a naive algorithm, and the runs it), referred to as "full-R", an alternative, standalone implemen7 Comparison of runtimes Alg5-Array Alg5-Pop Alg5-HT Alg3 Alg4-Array Alg4-Pop Walker-R full-R ransampl numpy "Gaussian" population Population size fixed at 1000000 101 100 10-1 Elapsed runtime (sec) Elapsed runtime (sec) 100 10-2 10-3 10-4 10-5 0 10 Sample size fixed at 1000000 101 10-1 10-2 10-3 10-4 10-5 101 102 103 104 105 Sample size (s) 106 107 10-6 0 10 108 101 104 105 103 Population size (n) 102 106 107 108 "Uniform" population Population size fixed at 1000000 101 100 Elapsed runtime (sec) Elapsed runtime (sec) 100 10-1 10-2 10-3 10-4 0 10 Sample size fixed at 1000000 101 10-1 10-2 10-3 10-4 10-5 101 102 103 104 105 Sample size (s) 106 107 10-6 0 10 108 101 102 104 105 103 Population size (n) 106 107 108 107 108 "Geometric" population Population size fixed at 1000000 100 10-1 10-1 10-2 10-3 10-4 10-5 10-6 0 10 Sample size fixed at 1000000 101 100 Elapsed runtime (sec) Elapsed runtime (sec) 101 10-2 10-3 10-4 10-5 101 102 103 104 105 Sample size (s) 106 107 10-6 0 10 108 101 102 104 105 103 Population size (n) 106 Figure 1: Comparison of runtimes of various random choice algorithms. The various versions of the novel algorithm proposed in this work are marked with large dots, some example algorithms are marked with small dots, competing algorithms used by various scientific libraries are marked with crosses. 8 Algorithm Algorithm 1 (numpy) Algorithm 2 Algorithm 3 Algorithm 4 Algorithm 5 Walker’s algorithm Pessimistic runtime Calls to RNG Can work online O(n + s log n) Additional memory used O(n) O(s) No O(n + s log s) O(n + s) O(n) O(n) O(n + s) O(s) O(1) O(1) O(1) O(n) O(s) O(s) O(n) O(min(n, s)) O(s) Yes Yes Yes Yes No Table 1: A comparison of properties of sampling algorithms tation of Walker’s algorithm in C by ransampl library, and with numpy’s implementation which follows Algorithm 1. The novel algorithm proposed in this work (Algorithm 5) has been tested in two versions, one which produces an array of size s (the sample with repetitions), referred to as the "Alg5-array" algorithm, and one which produces the multiset (with integer counts instead of the repetitions). The multiset was implemented either as a hashtable (using standard C++ unordered_map data structure) or trivially as an array of size n. The implementation using the former is referred to as "Alg5-HT" in plots, the one using the latter as "Alg5-Pop". Similarly, Algorithm 4 has been implemented and tested both outputting an array with repetitions and multiset (based on an array, hashtable implementation was produced but skipped in effort to avoid complicating the plots further). A summary of theoretical properties of each of the algorithms is presented in Table 8 The R functions are written in C code, and the code for these functions is reused, slightly modified to remove dependencies on R’s internals. Numpy, although written in Python, is compiled to native code using Cython, and as such, should run at near-native speeds, like the rest of the tested algorithms, without suffering overhead due to Python being an interpreted, high-level language. The algorithms have been tested on three different random populations, one drawn from uniform distribution U([0, 1])n , representing population with mostly equal probabilities, one drawn from a geometric sequence starting with 1.0 and ending at 10−100 , representing a population with skewed probabilities. The third type of population is generated by applying a Gaussian PDF function to n points evenly spaced between 0.0 and 10 times the stdev of the Gaussian function. This is meant to simulate the usual application of sampling function in modelling in population genetics (which in fact was the inspiration for this research): in population genetics models, selection and reproduction of modelled organisms is often done precisely by randomly sampling with replacement of n organisms (that reproduce and pass their offspring to a next generation) from a population of n. The probability of a given organism being chosen to reproduce is proportional to its fitness function - which is often Gaussian. Each population type (uniform, geometric and Gaussian) is rescaled so that it sums to 1.0, and randomly shuffled. The results of the tests are presented on 9 Figure 1. It is evident from the results of the tests that not only is the proposed algorithm asymptotically optimal, but, unlike Algorithm 4 it is also efficient in practice, outperforming the competing methods in most scenarios, by as much as several orders of magnitude in some cases. In the single pessimistic case, where the distribution of probabilities in the population is close to uniform and s ≈ n, although it runs slightly slower, it still remains competetive, moreover, the difference in runtime grows smaller as s = n → ∞, and it overtakes the Walker’s method at s = n ≈ 108 (data not shown). The algorithm is able to adapt to the input data and use any skew from uniform distribution to its advantage, to increase its runtime, as evidenced by the tests on Gaussian and especially geometric populations. Unlike the popular algorithms it works in constant additional memory, and is capable of online operation. An implementation of this algorithm in a few programming languages may be downloaded from http://bioputer.mimuw.edu.pl/~mist/stats 9 Application: mass sampling from any discrete distribution As the proposed algorithm is online, it may accept an infinite sequence of states as its population, and can still be expected to produce a sample in finite time, without exhausting the whole sequence. As such, one application is immediately obvoius: mass sampling of iid variates from any discrete distribution. All one needs ito do is to exhaustively walk through the configuration space of the distribution, preferably (though not necessarily) in order of decreasing probability mass function (PMF), and feed the resulting sequence into the proposed algorithm. The result is a sample from the input distribution of any desired size. The advantage of the proposed solution is that the input distribution does not need to have an easily invertible CDF, only a computable PMF. The runtime is usually sublinear, wrt. to the sample size, but that depends on the exact properties of the distribution being sampled, distributions with light tails being faster to sample from than heavy-tailed ones. As an example: generating a sample of size 109 from Poisson distribution with λ = 10000.0 using R programming language’s rpois function takes about 90 seconds, while using the scheme proposed above elapses 0.7 seconds. Such an algorithm itself consumes constant memory plus any memory needed for datastructures needed to walk through the configuration space (trivially constant in case of distributions with integer support, at worst a linear "visited" hashtable plus a linear priority queue when the configuration space is complicated, and needs to be traversed in a Dijkstra-like fashion [2]). The algorithm works online, in the meaning that the generated part of the sample is immediately available for consumption, before computations proceed to generate the rest of it. 10 This could be used to provide an alternative implementation of sampling functions in many programming languages, most of which accept an argument denoting sample size, but then proceed to generate even a large sample in naive, iterative fashion. One point worth noting, however, is that the algorithm, as presented, returns the sample sorted in the order in which the configuration space was traversed. If this is undesirable, a Fisher-Yates shuffle [3] may be performed on the resulting stream, at the cost of loss of online property. 10 Acknowledgments I would like to thank prof. Anna Gambin, Błażej Miasojedow PhD, and Mateusz Łącki MSc for their helpful comments. This research was funded by grant no. 2012/06/M/ST6/00438 by Polish National Science Centre, and grant POLONIUM „Matematyczne i obliczeniowe modelowanie ewolucji ruchomych elementów genetycznych”. References [1] David, H. A., and Nagaraja, H. N. Order statistics. Wiley Online Library, 1981. [2] Dijkstra, E. W. A note on two problems in connexion with graphs. Numerische mathematik 1, 1 (1959), 269–271. [3] Fisher, R. A., and Yates, F. Statistical tables for biological, agricultural and medical research. Longman, 1938. [4] Kahan, W. Pracniques: Further remarks on reducing truncation errors. Commun. ACM 8, 1 (Jan. 1965), 40–. [5] Shewchuk, J. R. Adaptive precision floating-point arithmetic and fast robust geometric predicates. Discrete & Computational Geometry 18, 3 (1997), 305–363. [6] Walker, A. J. An efficient method for generating discrete random variables with general distributions. ACM Trans. Math. Softw. 3, 3 (Sept. 1977), 253– 256. 11
8cs.DS
Time-Staging Enhancement of Hybrid System Falsification arXiv:1803.03866v1 [cs.SY] 10 Mar 2018 Gidon Ernst Ichiro Hasuo Zhenya Zhang Sean Sedwards National Institute of Informatics, Tokyo, Japan University of Waterloo, Waterloo, Canada {gidon,hasuo,zhangzy}@nii.ac.jp [email protected] Optimization-based falsification employs stochastic optimization algorithms to search for error input of hybrid systems. In this paper we introduce a simple idea to enhance falsification, namely time staging, that allows the time-causal structure of time-dependent signals to be exploited by the optimizers. Time staging consists of running a falsification solver multiple times, from one interval to another, incrementally constructing an input signal candidate. Our experiments show that time staging can dramatically increase performance in some realistic examples. We also present theoretical results that suggest the kinds of models and specifications for which time staging is likely to be effective. 1 Introduction Hybrid Systems Quality assurance of cyber-physical systems (CPS) has been recognized as an important challenge, where many CPS are hybrid systems that combine the discrete dynamics of computers and the continuous dynamics of physical components. Unfortunately, analysis of hybrid systems poses unique challenges, such as the limited applicability of formal verification. In formal verification one aims to give a mathematical proof for a system’s correctness. This is much harder for hybrid systems than for computer software/hardware, where the presence of continuous dynamics makes many problems more complex or even undecidable (e.g. reachability in hybrid automata). Optimization-Based Falsification Bemore robustly This way to ?? cause of these difficulties, an increasclimb down :) true ing number of researchers are turning true input input to optimization-based falsification as a less so quality assurance measure. It is a testing false quantitative method rather than that of formal verifiBoolean semantics robust semantics cation; the problem is formalized as folFigure 1: From Boolean to robust semantics lows. • Given: a model M (that takes an input signal u and yields an output signal M (u)), and a specification ϕ (a temporal M (u) u / formula) / M 6|= ϕ ? • Answer: error input, that is, an input signal u such that the corresponding output M (u) violates ϕ In the optimization-based falsification approach, the above falsification problem is turned into an optimization problem. This is possible thanks to robust semantics of temporal formulas [14]. Instead of the Boolean satisfaction relation v |= ϕ, robust semantics assigns a quantity Jv, ϕK ∈ R ∪ {∞, −∞} that tells 2 Time-Staging Enhancement of Hybrid System Falsification us, not only whether ϕ is true or not (by the sign), but also how robustly the formula is true or false. This allows one to employ hill-climbing optimization (see Fig. 1): we iteratively generate input signals, in the direction of decreasing robustness, hoping that eventually we hit negative robustness. Optimization-based falsification is a subclass of search-based testing: it adaptively chooses test cases (input signals u) based on previous observations. One can use stochastic algorithms for optimization (such as simulated annealing), which turn out to be much more scalable than many model checking algorithms that rely on exhaustive search. Note also that the system model M can be black box: it is enough to know the correspondence between input u and output M (u). An error input u is concrete evidence for the system’s need for improvement, and thus has great appeal to practitioners. The approach of optimization-based falsification was initiated in [14] and has been actively pursued ever since [3, 5, 7, 8, 10, 11, 20]. There are now mature tools, such as Breach [8] and S-Taliro [5], which work with industry-standard Simulink models. Contribution We introduce a simple idea of time staging for enhancement of optimization-based falsification. Time staging consists of running a falsification solver repeatedly, from one input segment to another, incrementally constructing an input signal. In general, in solving a concrete problem C by a metaheuristic H (such as stochastic optimization, evolutionary computation, etc.), a key to success is to communicate as much information as possible in the translation from C to H —that is, to let H exploit structures unique to C . Our idea of time staging follows this philosophy. More specifically, via time staging we communicate the time-causal structure of time-dependent signals—a structure that is present in some instances of the falsification problem but not in optimization problems in general—to stochastic optimization solvers. Our implementation of time-staged falsification is based on Breach [8]. We show that this simple idea can dramatically enhance its performance in some examples. We also present some theoretical considerations on the kinds of problem instances where time staging is likely to work, and some results that aid implementation of time staging. Structure of the Paper In §2 we informally outline optimization-based falsification and illustrate the idea of time staging. We then turn to formal developments: in §3 we review existing falsification works and in §4 we present our algorithm, augmented by some theoretical results that aid its implementation. §5 is devoted to the theoretical consideration of two specific settings in which time staging is guaranteed to work well. These settings will serve as useful “rules of thumb” for practical applications. In §6 we discuss our implementation and experimental results. We conclude in §8. 2 Schematic Overview: Falsification and Time Staging We illustrate falsification and time staging, informally with an example. Example Setting We take as a system model M a simple automotive powertrain whose input signal is 1-dimensional (the throttle u) and whose output signal is the vehicle speed v. We assume that M exhibits the following natural behavior: the larger u is, the quicker v grows. Let our specification be  ϕ ≡ 2(v ≤ 120) , where ≡ denotes the syntactic equality. To falsify ϕ the vehicle speed v must exceed 120. From the assumption about M ’s behavior, we expect u to be large in a falsifying input signal. Note that this is a simplified version of one of our experiments in §6. G. Ernst, I. Hasuo, Z. Zhang & S. Sedwards 3 (i + 1)-th sampling i-th sampling throttle throttle (i) (i+1) u4 (i+1) u1 u2 (i+1) (i) (i+1) u3 (i) u2 (i) u4 time · · · −→ Choosing (i+1) (i+1) , . . . , uK u1 by optimization −−−−−−−−−−−−−−−→ vehicle speed 120 u3 u1 time −→ · · · vehicle speed 120 v(i+1) v(i) time time Figure 2: Conventional optimization-based falsification (without time staging). first stage second stage throttle throttle (1) u1 ... time (1) u2 time vehicle speed (n2 ) v1 (2) v1 (1) v1 (n1 ) v1 optimization (n1 ) .. . (2) u2 .. . v2 (1) v2 −→ · · · optimization vehicle speed 120 Choosing the best prefix (n ) u 1 −−1−−−−−−→ optimization (2) u1 (n2 ) u2 optimization ... (n1 ) u1 (n1 ) u1 time time Figure 3: Falsification with time staging ⇒ ⇒ ng bi m cli llhi robustness JM(u), ϕK ··· u(2) u(1) u(i) u(i+1) ? ⇒ ··· ⇒ input signal u Figure 5: Nelder-Mead optimization. Here the inFigure 4: Hill-climbing optimization in falsification put space is the two-dimensional square, and the (unknown) score function is depicted by contour lines. Figures are from Wikipedia 4 Time-Staging Enhancement of Hybrid System Falsification Optimization-Based Falsification Fig. 2 illustrates how a conventional optimization-based falsification procedure works. In the i-th sampling one tries an input signal u(i) . Following the falsification litera(i) (i) ture we focus on piecewise constant signals. Thus a signal u(i) is represented by a sequence (u1 , . . . , uK ) of real numbers. See the top left of Fig. 2. The corresponding output signal v(i) = M (u(i) ) is shown below it. Since v(i) does not reach the threshold 120, we move on to the (i + 1)-th sampling and try a new (i+1) (i+1) input signal u(i+1) = (u1 , . . . , uK ). The choice of u(i+1) is made by an optimization algorithm. Specifically, the optimization algorithm observes the results of the previously sampled input signals u(1) , . . . , u(i) —especially the robustness value JM (u(i) ), ϕK that each input u(i) achieves. In the current  setting where ϕ ≡ 2(v ≤ 120) , the robustness value is simply the difference between 120 and the peak vehicle speed. The optimization algorithm tries to derive some general tendency, which it then uses to increase the probability that the next input signal u(i+1) will make the robustness smaller (i.e. the peak vehicle speed higher). Hill climbing is a prototype of such optimization algorithms. Its use in falsification is illustrated in Fig. 4, where u = (u1 , . . . , uK ) is depicted as one-dimensional for clarity. The actual curve for the robustness value JM (u), ϕK (gray and dashed) is unknown. Still the previous observations under input u(1) , . . . , u(i) suggest that to the right is the climbing down direction. The next candidate u(i+1) is picked accordingly, towards negative robustness. Another well-known optimization algorithm is the NelderMead algorithm. See Fig. 5, where the input space is two-dimensional and the (unknown) robustness function is depicted by contour lines. (i+1) (i+1) We see in the right of Fig. 2 that the new input signal u(i+1) = (u1 , . . . , uK ) leads to a corresponding output signal v(i+1) that reduces the robustness value by achieving a higher peak speed. We continue this way, u(i+2) , u(i+3) , . . . , hoping to eventually reach a falsifying input signal. Absence of the Time-Causal Information A closer look at Fig. 2 reveals room for improvement. In Fig. 2, the new input signal u(i+1) indeed achieves a smaller overall robustness JM (u), ϕK than u(i) . (i+1) (i) However, its initial segment u1 is smaller than u1 ; consequently the vehicle speed v(i+1) is smaller (i) than v(i) in the first few seconds. Keeping the initial segment u1 would have achieved an even greater peak speed. The problem here is that the time-causal structure inherent in the problem is not explicitly communicated to the optimization algorithm. The relevant structure is more specifically time monotonicity: an input prefix that achieves smaller robustness (i.e. a greater peak speed) is more likely to extend to a full falsifying input signal. Although it is possible that a stochastic optimization algorithm somehow “learns” time monotonicity, it is not guaranteed, because the structure of input spaces (the horizontal axis in Fig. 4 and the squares in Fig. 5) does not explicitly reflect time-causal structures. While the time monotonicity is not shared by all instances of the falsification problems, we find many realistic instances that approximately satisfy the property. We discuss time monotonicity in §5, as well as in the context of our experiments in §6. Falsification with Time Staging Our proposal of time staging consists of incrementally synthesizing a candidate input signal. We illustrate this in Fig. 3. In the first stage (left), we run a falsification algorithm and try to find an initial input segment that achieves low robustness (i.e. high peak speed). This first stage comprises running n1 samplings, as illustrated in Fig. 2. This process will gradually improve candidates for the initial input segment, in the way the arrows ↑ on the left in Fig. 3 designate. Let us assume that (n ) the last candidate u1 1 is the (tentative) best, achieving the smallest robustness. G. Ernst, I. Hasuo, Z. Zhang & S. Sedwards 5 (n ) In the second stage (on the right in Fig. 3) we continue u1 1 and synthesize the second input segment. This is again by running a falsification algorithm, as depicted. Note that, in each stage (a box in Fig. 3), the whole iterated process in Fig. 2 is conducted. In this way we continue to the K-th stage, always starting with the input segment that performed the best in the previous stage, thus exploiting the timecausal structure. While time staging is not difficult to implement, there is a challenge in using it effectively. An immediate question is whether choosing the single best input segment in each stage is the optimal approach. Our current strategy favors exploitation over exploration: it might miss a falsifying signal whose robustness must decrease slowly in the earlier segments and only quickly in the latter segments. Indeed we are working on an evolutionary variant of the above time-staged algorithm, where multiple segments are passed over from one stage to another, in order to maintain diversity and conduct exploration. That said, even under the current simple strategy of picking the best one, we observe significant performance enhancement in some falsification problems. See §6. We can summarize this trade-off in terms of the size of search spaces. Let U be the set of candidates for input segments, and K be the number of stages. Then the size of the set of whole input signals is |U|K , choosing one input segment for each stage. In our staged algorithm, in contrast, the search space for each stage is U and overall our search space is K · |U|. This reduction comes with the risk of missing some falsifying input signals. The experimental results in §6 suggest this risk is worth taking. Moreover, in §5 we present some theoretical conditions for the absence of such risk. They help users decide in practical applications when time staging will be effective. 3 Optimization-Based Falsification From this section on we turn to the formal description and analysis of our algorithm. This section presents a review of existing works on optimization-based falsification. System Models Let us formalize our system models. Definition 3.1 (time-bounded signal). Let T ∈ R>0 be a positive real. A (time-bounded) m-dimensional signal with a time horizon T is a function w : [0, T ] → Rm . Let w : [0, T ] → Rm and w0 : [0, T 0 ] → Rm be (time-bounded) signals. Their concatenation w · w0 : [0, T + T 0 ] → Rm is defined by (w · w0 )(t) := w(t) if t ∈ [0, T ], and w0 (t − T ) if t ∈ (T, T + T 0 ]. Let T1 , T2 ∈ (0, T ] such that T1 < T2 . The restriction w|[T1 ,T2 ] : [0, T2 − T1 ] → Rm of w : [0, T ] → Rm to the interval [T1 , T2 ] is defined by (w|[T1 ,T2 ] )(t) := w(T1 + t). Definition 3.2 (system model M ). A system model, with M-dimensional input, is a function M that takes an input signal u : [0, T ] → RM and returns M (u) : [0, T ] → RN . Here the common time horizon T ∈ R>0 is arbitrary. Furthermore, we impose the following causality condition on M . For any time-bounded signals u : [0, T ] → RM and u0 , we require that M (u · u0 ) [0,T ] = M (u). Note that M (u · u0 ) = M (u) · M (u0 ) does not hold in general: feeding u can change the internal state of M . This motivates the following definition. Definition 3.3 (continuation Mu ). Let M be a system model and u : [0, T ] → RM be a signal. The continuation of M after u, denoted by Mu , is defined as follows. For an input signal u0 : [0, T 0 ] → RM : Mu (u0 )(t) := M (u · u0 )(T + t). 6 Time-Staging Enhancement of Hybrid System Falsification Signal Temporal Logic and Robust Semantics We review signal temporal logic (STL) [21] and its robust semantics [10, 14]. Var is the set of variables, and let N := |Var|. Variables stand for physical quantities, control modes, etc. ≡ denotes syntactic equality. Definition 3.4 (syntax). In STL, atomic propositions and formulas are defined as follows, respectively: α ::≡ f (x1 , . . . , xn ) > 0, and ϕ ::≡ α | ⊥ | ¬ϕ | ϕ ∧ ϕ | ϕ UI ϕ. Here f is an n-ary function f : Rn → R, x1 , . . . , xn ∈ Var, and I is a closed non-singular interval in R≥0 , i.e. I = [a, b] or [a, ∞) where a, b ∈ R and a < b. We omit subscripts I for temporal operators if I = [0, ∞). Other common connectives like ∨, →, >, 2I (always) and 3I (eventually), are introduced as abbreviations: 3I ϕ ≡ > UI ϕ and 2I ϕ ≡ ¬3I ¬ϕ. Atomic formulas like f (~x) ≤ c, where c ∈ R is a constant, are also accommodated by using negation and the function f 0 (~x) := f (~x) − c. Definition 3.5 (robust semantics [9, 10]). For an unbounded n-dimensional signal w : R≥0 → Rn and t ∈ R≥0 , wt denotes the t-shift of w, that is, wt (t 0 ) := w(t + t 0 ). Let w : R≥0 → RN be a signal (recall N = |Var|), and ϕ be an STL formula. We define the robustness Jw, ϕK ∈ R ∪ {∞, −∞} as follows, by induction. Here u and t denote infimums and supremums of real numbers, respectively.  Jw, f (x1 , · · · , xn ) > 0K := f w(0)(x1 ), · · · , w(0)(xn ) Jw, ⊥K := −∞ Jw, ¬ϕK := −Jw, ϕK Jw, ϕ1 ∧ ϕ2 K := Jw, ϕ1 K u Jw, ϕ2 K d F 0 Jw, ϕ1 UI ϕ2 K := t∈I Jwt , ϕ2 K u t 0 ∈[0,t) Jwt , ϕ1 K Here are some intuitions and consequences. The robustness Jw, f (~x) > cK stands for the vertical margin f (~x) − c for the signal w at time 0. A negative robustness value indicates how far the formula is from being true. The robustness for the eventually modality is computed by Jw, 3[a,b] (x > 0)K = F t∈[a,b] w(t)(x). The original semantics of STL is Boolean, given by a binary relation |= between signals and formulas. The robust semantics refines the Boolean one, in the sense that: Jw, ϕK > 0 implies w |= ϕ, and Jw, ϕK < 0 implies w 6|= ϕ. Optimization-based falsification via robust semantics [14] hinges on this refinement. See [10]. Although the definitions so far are for unbounded signals only, we note that the robust semantics Jw, ϕK, as well as the Boolean satisfaction w |= ϕ, allows straightforward adaptation to time-bounded signals (Def. 3.1). See Appendix A. Falsification Solvers In the next definition, a prototype of a score function ρ is given by the robustness ρϕ of a given STL specification ϕ. The generality of allowing other ρ is needed later in §4. ρϕ (v) := Jv, ϕK (1) Definition 3.6 (falsification solver). A falsification solver is a stochastic algorithm Falsify that takes, as input: 1) a system model M (Def. 3.2) with M-dimensional input; 2) a score function ρ that takes an output signal v of M and returns a score ρ(v) ∈ R ∪ {−∞, ∞}; and 3) a time horizon T ∈ R>0 . The algorithm Falsify returns an M-dimensional signal u : [0, T ] → RM . Each invocation Falsify(M , ρ, T ) of the solver is called a falsification trial. It is successful if the returned signal u satisfies ρ(M (u)) < 0. Note that the returned signal u can differ in every trial, since Falsify is a stochastic algorithm. G. Ernst, I. Hasuo, Z. Zhang & S. Sedwards 7 Algorithm 1 Internal Structure of a Falsification Solver Falsify(M , ρ, T ) Require: a system model M , a score function ρ, and T ∈ R>0 1: U ← () . the list U collects all the candidates u : [0, T ] → RM 2: while ¬InitialSamplingDone(T, U) do 3: u ← InitialSampling(T ) . u : [0, T ] → RM is sampled following some recipe 4: U ← cons(U, u) 5: while ¬OptimizationSamplingDone(M , ρ, T, U) do 6: u ← OptimizationSampling(M , ρ, T, U) 7: . u is sampled, so that ρ(M (u)) becomes small, based on previous samples in U 8: U ← cons(U, u) 9: u ← arg minu∈U ρ(M (u)) 10: return u . a trial is successful if ρ(M (u)) < 0 We further assume the internal structure of the solver Falsify follows the scheme in Algorithm 1. It consists of two phases. The first initial sampling phase collects some candidates for u : [0, T ] → RM regardless of the system model M or the score function ρ. In the second optimization sampling phase, a stochastic optimization algorithm is employed to sample a candidate u that is likely to make the score ρ(M (u)) small. Implementation of Falsification Solvers Both Breach [8] and S-Taliro [5] take industry-standard Simulink models as system models. For input signal candidates the tools focus on piecewise constant signals; they are represented by sequences (u1 , . . . , uK ) of real numbers, much like in §2. Here K is the number of control points; in our staged algorithm we use the same K for the number of stages. The tools offer multiple stochastic optimization algorithms for the optimization sampling phase, including CMA-ES [6], global Nelder-Mead and simulated annealing. The initial sampling phase is mostly by random sampling. Additionally, in Breach with global Nelder-Mead, so-called corner samples are added to the list U. The number of corner samples grows exponentially as K grows, i.e. as we have more control points. 4 Time Staging in Optimization-Based Falsification Definition 4.1 (time-staged deployment of falsification solver). Let M be a system model, ϕ be an STL formula, and T ∈ R>0 be a time horizon. Let K ∈ N be a parameter; it is the number of time stages. The time-staged deployment of a falsification solver Falsify is the procedure in Algorithm 2. On the line 3, the model Mu is the continuation of M after u (Def. 3.3); the score function ∂v ρϕ is defined by (1) (∂v ρϕ )(v0 ) := ρϕ (v · v0 ) = Jv · v0 , ϕK . (2) The whole procedure is stochastic (since Falsify is); an invocation is called a time-staged falsification trial. It is successful if the returned signal u satisfies JM (u), ϕK < 0. A falsification trial (i.e. an invocation of Algorithm 1) is an iterative process: the more we sample, the more likely we obtain a falsifying input signal. Since we run multiple falsification trials in Algorithm 2 (one trial for each of the K stages), an important question is how we distribute available time to different stages. 8 Time-Staging Enhancement of Hybrid System Falsification Algorithm 2 Time-Staged Deployment of a Falsification Solver Require: a falsification solver Falsify, a system model M , an STL formula ϕ, T ∈ R>0 and K ∈ N 1: u ← () . the input prefix obtained so far. We start with the empty signal () 2: for j ∈ {1, . . . , K} do 3: . synthesizing the j-th input segment u0 ← Falsify(Mu , ∂M (u) ρϕ , KT ) 4: u ← u · u0 . concatenate u0 , after which the length of u is jT K 5: return u . a time-staged falsification trial is successful if JM (u), ϕK < 0 A simple strategy is to fix the number of samples in each phase of Algorithm 1. Then the predicates init and InitialSamplingDone(T, U) and OptimizationSamplingDone(M , ρ, T, U) are given by |U| > Nmax opt opt init |U| > Nmax , where Nmax , Nmax are constants. An adaptive strategy, that we also implemented for the optimization sampling phase, is to continue stuck , and we stop after N stuck consampling until we stop seeing progress. Here we fix a parameter Nmax max secutive samplings without reducing robustness. A similar strategy of adaptively choosing the number of samples can be introduced for random sampling in the initial sampling phase (the lines 2–4 of Algorithm 1). 4.1 Towards Efficient Implementation A key to speedup of Algorithm 2 is in the line 3; more specifically, how we handle the previous input prefix u. Here we discuss two directions, one on the model Mu and the other on the score function ∂v ρϕ . (We note that the suggested enhancements are not currently used in our implementation, because of performance reasons. See below.) Continuation of Models Optimization-based falsification has a very wide application domain. Since it only requires a black-box model M , the concrete form of M can vary from a program to a Simulink model and even a system with hardware components (HILS). These models can be very big, and usually the bottleneck in falsification lies in simulation, that is, to compute M (u) given an input signal u. In the line 3 of Algorithm 2, therefore, using the definition M (u · u0 )(T +t) in Def. 3.3 is in principle not a good strategy: it requires simulation of M for the whole prefix u·u0 , which can be avoided if we can directly simulate the continuation Mu . In Simulink this is possible by saving the snapshot of the model after a simulation, via the SaveFinalState model configuration parameter. In our implementation we do not do so, though, because the overhead of saving and loading snapshots is currently greater than the cost of simulating. This balance can become different, if we figure out a less expensive way to use snapshots, or if we study more complex models. Derivative of Formulas The situation is similar with the score function ∂M (u) ρϕ in the line 3 of Algorithm 2. Using the presentation ρϕ (M (u) · v0 ) in (2) requires scanning the same prefix M (u) repeatedly. Desired here is a syntactic presentation of ∂M (u) ρϕ , that will be given as an STL formula ∂M (u) ϕ such that ∂M (u) ρϕ = ρ(∂M (u) ϕ) . This would allow one to utilize available algorithms for computing robustness values Jv, ∂M (u) ϕK. Definition 4.2 (derivative of flat STL formulas). Let T ∈ R>0 , and v : [0, T ] → RN be a signal. Given an STL formula ϕ that is flat in the sense that it does not have nested temporal operators, the derivative G. Ernst, I. Hasuo, Z. Zhang & S. Sedwards 9 ∂v ϕ by v is defined inductively as follows.  ∂v f (~x) > 0 :≡ cJv, f (~x)>0K ∂v ⊥ :≡ ⊥ ∂v (¬ϕ) :≡ ¬∂v ϕ ∂v (ϕ1 ∧ ϕ2 ) :≡ (∂v ϕ1 ) ∧ (∂v ϕ2 ) ∂v (ϕ1 UI ϕ2 ) :≡ cJv,ϕ1 UI ϕ2 K ∨ (cJv,2ϕ1 K ∧ ϕ1 ) UI−T ϕ2 Here the interval I − T is obtained by shifting endpoints, such as [a, b] − T = [a − T, b − T ]. For each r ∈ R, the notation cr abbreviates the atomic formula r > 0, where r is thought of as a constant function. We use the fact that Jw, cr K = r. Until formulas ϕ1 UI ϕ2 are split into the evaluation on the signal prefix v (first disjunct), and a continuation (second disjunct). The constant cJv,2ϕ1 K injects the robustness of ϕ1 seen so far into the residual formula (recall  that both of 2 and ∧ take the infimum). It follows that ∂v 2I ϕ ≡ cJv,2ϕK ∧ 2I−T ϕ and ∂v 3 ϕ ≡ cJv,3ϕK ∨ 3I−T ϕ. Proposition 4.3. Let T ∈ R>0 , v : [0, T ] → RN be a signal, and ϕ be a flat STL formula. We have, for any v0 : [0, T 0 ] → RN , Jv0 , ∂v ϕK = Jv · v0 , ϕK. A proof is in Appendix B. Use of derivatives for timed specifications is also found e.g. in [22]. The settings are different, though: Boolean semantics in [22] while our semantics is quantitative. Our restriction to flat formulas comes mainly from this difference, and lifting the flatness restriction seems hard. 5 Sufficient Conditions for Time Staging We present some theoretical analyses of the performance of time staging that indicate to which class of systems the time-staged approach can apply. We give some sufficient conditions under which the approach is guaranteed to work. However, it should be noted that it is not necessary that a concrete system satisfies these conditions strictly as these are rather restrictive. Nevertheless, we believe that users with expert domain knowledge can judge whether their models satisfy these conditions approximately. This way our results provide those users with “rules of thumb.” As we discussed in the last paragraph of §2, the potential performance advantage by time staging comes from the reduction of search spaces from |U|K to K · |U|. Here U is the set of potential input segments for each stage, and K is the number of stages. This advantage comes at the risk of missing out some error input signals. The following basic condition (3), that we call incremental falsification, ensures that there is no such risk. More precisely, we can decompose the “best” input signal u into a first stage u1 and its remainder u2 such that the entire falsification problem (left hand side) is solved by greedy optimization of the initial segment (inner arg minu1 ), and subsequent optimization of the continuation (outer minu2 ). For all choices of T1 , T2 with ranges u : [0, T1 + T2 ] → Rm and ui : [0, Ti ] → Rm :  z r   minJM (u), ϕK = min M arg minJM (u1 ), ϕK · u2 , ϕ u u2 (3) u1 Algorithm 2 repeatedly unfolds (3) by picking constant T1 = T /K where T is the time horizon and K is the number of stages. The rest of this section is devoted to the search for concrete sufficient conditions for (3). 10 Time-Staging Enhancement of Hybrid System Falsification Monotone Systems and Ceiling Specifications We formalize the time monotonicity property in §2. That it implies incremental falsification (3) can be easily proved. Definition 5.1 (time-monotone falsification problem). A system model M and an STL formula ϕ are said to constitute a time-monotone falsification problem if, for any input signals u1 , u01 : [0, T1 ] → Rm and u2 : [0, T2 ] → Rm , JM (u1 ), ϕK ≤ JM (u01 ), ϕK implies JM (u1 · u2 ), ϕK ≤ JM (u01 · u2 ), ϕK. We investigate yet more concrete conditions that ensures time monotonicity. The following condition on system models is assumed in the example of §2. Definition 5.2 (monotone system, ceiling specification). Let x be a variable (for output). A system model M is said to be monotone in x if, for each u1 , u01 : [0, T1 ] → RM and u2 : [0, T2 ] → RM , M (u1 )(T1 )(x) ≤ M (u01 )(T1 )(x) implies M (u1 · u2 )(T1 + T2 )(x) ≤ M (u01 · u2 )(T1 + T2 )(x). An STL formula of the form 2(x < c), where x is a variable and c ∈ R is a constant, is called a ceiling formula. One can speculate that a monotone system and a ceiling specification 2(x < c), like those in §2, constitute a time-monotone falsification problem. The speculation is not true, unfortunately; a counterexample is easily constructed using a model M whose output signal is not increasing. We can instead show the following weaker property. Definition 5.3 (truncated time monotonicity). A system model M and an STL formula ϕ constitute a truncated time-monotone falsification problem if, for any input u1 , u01 : [0, T1 ] → Rm and u2 : [0, T2 ] → Rm , JM (u1 ), ϕK ≤ JM (u01 ), ϕK implies existence of T ∈ (0, T1 ] such that JM ((u1 |[0,T ] ) · u2 ), ϕK ≤ JM ((u01 |[0,T ] ) · u2 ), ϕK .  Proposition 5.4. Let M be a model monotone in x, and ϕ ≡ 2(x < c) . Then M and ϕ constitute a truncated time-monotone falsification problem. The proof, in Appendix C.1, constructs a concrete choice of T in Def. 5.3. Specifically it is the instant T ∈ [0, T1 ] in which the robustness JM (u1 |[0,T ] ), ϕK is minimum. In the scenario of §2 this is the instant that the vehicle speed is in its peak. Note that truncated time monotonicity does not guarantee (K−1)T incremental falsification as per (3), but it implies that the current rigid time staging at 0, KT , 2T K ,..., K is not optimal. These theoretical considerations suggest potential improvement of the staged procedure in Def. 4.1 with adaptive choice of stages, which is left for future work. Stateless Systems and Reachability Specifications Here is another sufficient condition. Definition 5.5 (stateless system, reachability formula). A system model M is said to be stateless if, for any input signals u1 , u01 : [0, T1 ] → Rm and u2 : [0, T2 ] → Rm , we have M (u1 · u2 )|(T1 ,T2 ] = M (u01 · u2 )|(T1 ,T2 ] . An STL formula 3ψ, where ψ is modality-free, is called a reachability formula. Note that being stateless is a sufficient but not necessary condition for M (u1 · u2 ) = M (u1 ) · M (u2 ). Statelessness requires insensitivity to previous input prefixes, but a stateless system can still be sensitive to time. Proposition 5.6. Let M be a stateless system and ϕ be a reachability specification ϕ ≡ (3ψ). Then M and ϕ satisfy the incremental falsification property (3). A proof is easy. A typical situation in which we would appeal to Prop. 5.6 is when: the specification is 3(x < c1 ∨ x > c2 ) (which can be hard to falsify if c1 < c2 are close); and the system is already in its stable state (so that its behavior does not depend much on what happened during the transient phase). Our experiments in §6 demonstrate that time staging can drastically improve performance in such settings. G. Ernst, I. Hasuo, Z. Zhang & S. Sedwards 11 Table 1: Experimental results. Each column shows how many falsification trials succeeded (out of 20), and the average runtime. S1: 2[0,30] (v < 120). S2: 2[0,30] (g = 3 → v ≥ 30). S3: 3[10,30] (v ≤ vmin ∨ v ≥ vmax ), where: vmin = 50, vmax = 60 (easy); vmin = 53, vmax = 57 (hard). S4: 2[0,10] (v < v) ∨ 3[0,30] (ω > ωmax ), where: vmin = 80, ωmax = 4500 (easy); vmin = 50, ωmax = 2700 (mid); vmin = 50, ωmax = 2520 (hard). The specification S for the Abstract Fuel Control model is ¬(3[t1 ,t2 ] [0,t 0 ] (AF − AF ref > δ ∗ 14.7)), where: t1 = 0, t2 = 6, t 0 = 3, δ = 0.07 (init); t1 = 6, t2 = 26, t 0 = 4, δ = 0.01 (stable). Starred numbers 0* or 20* indicate that GNM is deterministic so all trials yield the same result. model Automatic Transmission spec. S1 S2 S3 easy S3 hard S4 easy algorithm time #/20 time #/20 time #/20 time #/20 time #/20 CMA-ES 27s 20 5s 20 39s 14 57s 0 32s 16 +TS 52s 15 15s 16 9s 19 23s 11 15s 14 +A-TS 41s 18 15s 17 9s 16 21s 10 26s 14 SA 50s 5 43s 7 37s 9 55s 0 35s 6 +TS 37s 20 33s 16 11s 19 33s 8 21s 14 +A-TS 34s 20 18s 17 9s 18 26s 4 16s 18 GNM 6s 20* 61s 0* 56s 0* 55s 0* 43s 0* +TS 42s 20* 15s 20* 13s 20* 25s 20* 11s 20* +A-TS 20s 20* 16s 20* 10s 20* 26s 20* 13s 20* S4 mid time #/20 37s 9 14s 14 22s 14 36s 9 25s 13 21s 11 46s 0* 45s 0* 45s 0* Abst. Fuel Ctrl. S4 hard S init S stable time #/20 time #/20 time #/20 59s 0 49s 0 82s 1 24s 3 30s 0 42s 1 20s 5 26s 0 41s 0 47s 5 51s 0 76s 2 51s 0 47s 1 54s 7 30s 2 34s 0 42s 5 53s 0* 50s 0* 86s 0* 52s 0* 30s 20* 20s 20* 43s 0* 37s 0* 19s 20* Corner Samples for Global Nelder-Mead The reduction of search spaces from |U|K to K · |U| has its analogue in the number of corner samples in Breach with global Nelder-Mead (lines 2–4 of Algorithm 1, see the last paragraph of §3). Originally the number of corner samples is 2K·M , where K is the number of control points and M is the number of input values. By introducing K time stages, the total number of corner samples is reduced to K · 2M . 6 Experiments We compare the success rate and time consumption of the proposed method. The benchmarks here use automotive Simulink models that are commonly used in the falsification literature. Specifications are chosen taking the deliberations of §5 into account, namely with ceiling specifications (Def. 5.2, including the example of §2), a reachability specification (Def. 5.5) and a combination thereof. The base line is Algorithm 1 implemented by Breach [8]. The methods proposed in §4 are implemented on top of Breach: the time-staged Algorithm 2 (TS), and the adaptive strategy (A-TS, the one described after Def. 4.1). All three algorithms (plain, TS, A-TS) are combined with different optimization solvers: CMA-ES, simulated annealing (SA), global Nelder-Mead (GNM), obtaining a total of nine configurations. The results in Table 1 indicate that both success rate and runtime performance are significantly improved by time staging, often finding counterexamples when non-staged Breach fails (e.g. columns S3 hard and S init). Furthermore, we see that while the adaptive algorithm (A-TS) does not necessarily lead to a higher success rate in comparison to the time-staged one (TS), it gives yet another runtime performance improvement. However, as discussed in detail in §6, there is no overall best algorithm, and time staging affects the optimization algorithms differently depending on the problem. Benchmarks Automatic Transmission is a Simulink model that was proposed as a benchmark for falsification in [15]. It has input values throttle ∈ [0, 100] and brake ∈ [0, 325], and outputs the car’s speed v, 12 Time-Staging Enhancement of Hybrid System Falsification the engine rotation ω, and the selected gear g. With this model we consider five specifications S1–5. The first two are ceiling ones. Specification S1 2[0,30] (v < 120) (cf. the example in §2) states the speed be always below a threshold. This property is easily falsified with throttle = 100. Specification S2 2[0,30] (g = 3 → v ≥ 30) states that it is not possible to drive slowly in a high gear. A falsifying trajectory first has to speed up to reach this gear and subsequently roll out until speed falls below the threshold. This latter part of the trajectory can again be seen as a ceiling specification. Note that this property is interesting because the robustnes does not provide any guidance unless gear 3 has been entered by the system. Specification S3 is a reachability problem, 3[10,30] (v ≤ vmin ∨ v ≥ vmax ), that encodes the search for a trajectory that keeps the speed between a lower and upper bound. The falsification problem consists of two sub-challenges: 1) hitting this speed interval precisely after an initial acceleration up to 10s simulated time; and 2) maintaining a correct speed till the time horizon. This suggests that a natural decomposition of the problem can indeed be achieved by separating these two aspects in time. Specification S4 2[0,10] (vmin < v) ∨ 3[0,30] (ω > ωmax ) expresses that speed vmin can only be reached with an engine rotation exceeding a threshold ωmax . This specification is mentioned in [15] and evaluated in e.g., [3,4], too. To falsify, a trajectory must be found that reaches speed v early with an engine rotation lower than ωmax . The difficulty increases with higher vmin and lower ωmax , respectively. The formula represents the mixture of ceiling and reachability specifications. The second system model is Abstract Fuel Control from [17]. It takes two input values, pedal angle (from [0, 61.1]) and engine speed (from [0, 1100]); it outputs air-fuel ratio AF, which influences fuel efficiency and performance of the car. The value of AF is expected to be close to a reference value AF ref . According to [17], this setting corresponds to the so-called normal mode, where AF ref = 14.7 is constant. We used the specification ¬(3[t1 ,t2 ] [0,t 0 ] (AF − AF ref > δ ∗ 14.7)): the air-fuel ratio does not deviate from an acceptable range for more than t 0 seconds. We evaluated this specification with two parameter sets: the initial period with a larger error margin, and the stable period with a smaller margin. See Table 1 for parameter values. Experimental Setup and Results For the experiments with the Automatic Transmission model, the input signals were piecewise constant with 5 control points. The time horizon was T = 30. The parameters outlined in §4 were as follows: the maximum number of samplings for each plain (non-staged) falsification trial was 150 (initial and optimization samplings combined). In the time-staged (TS) trials, we make the number of stages coincide with that of control points. Analogously, the sampling budget per stage was set to 30 for K = 5 stages, resulting in overall 150 samplings. The adaptive algorithm (A-TS) stall = 30/2 = 15 per each of five stages. The experiments with the Abstract Fuel ran with the threshold Nmax Control model were run up to the time horizon T = t2 + t 0 where t2 and t 0 are as in Table 1. We used three and five stages, respectively, for the initial and stable specifications. These again coincide with the number of control points. The TS algorithms conducted 30 samplings in each stage. The experiments ran Breach version 1.2.9 and MATLAB R2017b on an Amazon EC2 c4.8xlarge instance with a 36 core Intel(R) Xeon(R) CPU (2.90GHz) and 58G of main memory. However, we did not use the opportunity to parallelize, and the time reported is in the same order of magnitude as that of a modern desktop workstation. The results are shown in Table 1. They are grouped by the underlying stochastic optimization algorithm: CMA-ES, simulated annealing (SA) and global Nelder-Mead (GNM). In each group, we compare plain (unstaged) Breach to the time-staged (TS) and the adaptive time-staged (A-TS) ones. We compare average runtimes (lower is better) and the success rate (higher is better), aggregated over 20 falsifica- G. Ernst, I. Hasuo, Z. Zhang & S. Sedwards 13 tion trials for each configuration. Those good results which deserve attention are highlighted in bold. Note that the implementation of the global Nelder-Mead algorithm in Breach uses a deterministic source of quasi-randomness (Halton sequences), which implies that whether GNM finds a counterexample is consistent across all trials (marked with an asterisk ∗). Discussion Focusing on the Automatic Transmission model first, we see that CMA-ES works well for S1, although GNM performs even better (6s, supposedly because it uses corner samples, see §3). Time staging introduces overhead to CMA-ES and GNM, because each stage is optimized individually. In contrast, simulated annealing (SA) benefits from time staging for the two ceiling specifications S1–2. We presume that since SA emphasizes exploration, it benefits from exploitation added by time staging (cf. §2). The second specification S2 is slightly more complex: before gear 3 is reached, there is no guidance from the robustness semantics, because J , g = 3K = −∞ masks any quantitative information on v. Hence, falsifying this property needs some luck during the collection of initial samples in Algorithm 1. CMAES apparently exploits this, see top of column S2 of Table 1. Considering the other algorithms, SA and GNM, both benefit from time staging: exploitation of time causality prevents these good trajectory prefixes from being discarded accidentally once the required gear is reached (cf. Fig 2). The results for S3 are evaluated with two different choices of parameters. The harder instance was falsified by the time-staged algorithms only, which can likely be attributed to the flattening of the search space from size |U|K to K · |U| (§5). S4 is evidently harder than the previous ones. Time staging improves performance in a general tendency but not in all cases (SA for S4 hard). The results for the Abstract Fuel Control model (the last two columns in Table 1) show that the time-staged algorithms boost the ability to falsify some rare events. The specification for the initial stage where AF is still unstable (S init) can be considered a rare event since all the three non-staged algorithms failed to falsify it. Time-staged SA and time-staged GNM managed to find error inputs. In the last column (stable period) is remarkable, too, where success rate and run time of SA and GNM significantly improved. Overall, while the performance of the non-staged algorithms suffers from tightening the bounds, the time-staged versions are able to find falsifying trajectories with good success rates while at the same time exhibiting significantly shorter runtimes. 7 Related Work Falsification is a special case of search-based testing, so considerable research efforts have been made towards coverage [3, 7, 20]. The benefits of coverage in falsification guarantees are twofold. Firstly, they indicate confidence for correctness in case no counterexamples are found. Paired with sound robustness estimates for simulations, one can cover a an infinite parameter space with finitely many simulations. C2CE [13] is a recent tool that computes approximations of reachable states using such an approach. Secondly, coverage can be utilized for a better balance between exploration and exploitation: stochastic optimization algorithms can be called in an interleaved manner, in which coverage guides further exploration. The approach based on Rapidly-Exploring Random Trees [11] puts an emphasis on exploration by achieving high coverage of the state space. In their algorithm, robustness-guided hill-climbing optimization plays a supplementary role. Compared to these works, our current results go in an orthogonal direction, by utilizing time causality to enhance exploitation. The so-called multiple-shooting approach 14 Time-Staging Enhancement of Hybrid System Falsification to falsification [24] can be seen of a generalization of RRTs. It consists of: an upper layer that searches for an abstract error trace given by a succession of cells; and a lower layer where an abstract error trace is concretized to an actual error trace by picking points from cells. The approach can discover falsifying traces by backwards search from a goal region, but needs to concatenate partial traces with potential gaps, which can fail. Furthermore it is unclear how to extend it to general STL specifications. A survey of simulation based approaches has been done by Kapinski et al. [18]. Monotonicity has been exploited in different ways for falsification. Robust Neighborhood Descent [1, 2] (RED) searches for trajectories incrementally, restarting the search from points of low robustness. Descent computation of RED assumes explicit derivatives of the dynamics to guarantee convergence to (local) minima. It is the same principle underlying Prop. 5.4) and our experiments indicate that this principle is useful for black-box optimization, too. In [2], RED is paired with simulated annealing to combine local and global search and to account for more exploration. Doing so for our present work remains to be done in the future. In [16], Hoxha et al. mine parameters θ under which specifications φ [θ ] are satisfied or falsified by the system. They show that the robust semantics of formulas is monotone in θ and use that fact to tighten such parameters. This is orthogonal to this work as it does not use monotonicity of the system itself. Kim et al. [19] use an idea similar to [16] to partition specifications into upper bounds and lower ceilings. However, instead of robustness-guided optimization they use exhaustive exploration of the input space in a way that in turn requires that the system dynamics is monotone in the choice of each input at each time point. This is different from our Def 5.1 of timemonotonicity that aims at incrementally composing good partial choices. The recent work [12] introduces a compositional falsification framework, focusing on those systems which include machine-learning (ML) components that perform tasks such as image recognition. While the current work aims at the orthogonal direction of finding rare counterexamples, we are interested in its combination with the results in [12], given the increasingly important roles of ML algorithms in CPS. 8 Conclusions and Future Work We have introduced and evaluated the idea of time staging to enhance falsification for hybrid systems. The proposed method emphasizes exploitation over exploration as part of stochastic optimization. As there is no single algorithm that fits every problem (as a consequence of having no free lunch [23]), having a variety of methods at disposal permits the user of a system to choose the one suitable for the problem at hand. We have shown that the proposed approach is a good fit for problems that suitable exhibit time-causal structures, where it significantly outperforms non-staged algorithms. Two obvious directions for future work have been pointed out already. Instead of just picking the best trajectory for each stage, it might be beneficial to retain a few, potentially diverse ones in the spirit of evolutionary algorithms (§2). For example, it would be interesting to explore the space between this work on one hand and coverage-driven rapidly-exploring random trees. Another idea is to discover time stages adaptively (§5, the discussion after Prop. 5.4). For the experiments presented here, we chose to set uniformly fixed stages, which runs the risk of either being too coarse grained (missing some falsifying input), or being too fine grained (wasting analysis time). Finally, another future direction is to explore variations of robust semantics to mitigate discrete propositions like g = 3 (§6), for example using averaging modalities [4]. Other t-norms than min/max for the semantics of conjunction/disjunction could preserve more information from different subformulas. G. Ernst, I. Hasuo, Z. Zhang & S. Sedwards 15 Acknowledgement. This work is supported by ERATO HASUO Metamathematics for Systems Design Project (No. JPMJER1603), Japan Science and Technology Agency. References [1] H. Abbas, A. Winn, G. Fainekos, and A. A. Julius. Functional gradient descent method for metric temporal logic specifications. In American Control Conference (ACC), 2014, pages 2312–2317. IEEE, 2014. [2] H. Y. Abbas. Test-based falsification and conformance testing for cyber-physical systems. Arizona State University, 2015. [3] A. S. Adimoolam, T. Dang, A. Donzé, J. Kapinski, and X. Jin. Classification and coverage-based falsification for embedded control systems. In R. Majumdar and V. Kuncak, editors, Computer Aided Verification - 29th Int. Conf., CAV 2017, volume 10426 of LNCS, pages 483–503. Springer, 2017. [4] T. Akazaki and I. Hasuo. Time robustness in MTL and expressivity in hybrid system falsification. In D. Kroening and C. S. Pasareanu, editors, Computer Aided Verification - 27th Int. Conf., CAV 2015, volume 9207 of LNCS, pages 356–374. Springer, 2015. [5] Y. Annpureddy, C. Liu, G. E. Fainekos, and S. Sankaranarayanan. S-taliro: A tool for temporal logic falsification for hybrid systems. In P. A. Abdulla and K. R. M. Leino, editors, Tools and Algorithms for the Construction and Analysis of Systems - 17th Int. Conf., TACAS 2011, volume 6605 of LNCS, pages 254–257. Springer, 2011. [6] A. Auger and N. Hansen. A restart CMA evolution strategy with increasing population size. In Proceedings of the IEEE Congress on Evolutionary Computation, CEC 2005, pages 1769–1776. IEEE, 2005. [7] J. V. Deshmukh, X. Jin, J. Kapinski, and O. Maler. Stochastic local search for falsification of hybrid systems. In B. Finkbeiner, G. Pu, and L. Zhang, editors, Automated Technology for Verification and Analysis - 13th Int. Symp., ATVA 2015, volume 9364 of LNCS, pages 500–517. Springer, 2015. [8] A. Donzé. Breach, A toolbox for verification and parameter synthesis of hybrid systems. In T. Touili, B. Cook, and P. B. Jackson, editors, Computer Aided Verification, 22nd Int. Conf., CAV 2010, volume 6174 of LNCS, pages 167–170. Springer, 2010. [9] A. Donzé, T. Ferrère, and O. Maler. Efficient robust monitoring for STL. In N. Sharygina and H. Veith, editors, Computer Aided Verification - 25th Int. Conf., CAV 2013, volume 8044 of LNCS, pages 264–279. Springer, 2013. [10] A. Donzé and O. Maler. Robust satisfaction of temporal logic over real-valued signals. In K. Chatterjee and T. A. Henzinger, editors, Formal Modeling and Analysis of Timed Systems - 8th Int. Conf., FORMATS 2010, volume 6246 of LNCS, pages 92–106. Springer, 2010. [11] T. Dreossi, T. Dang, A. Donzé, J. Kapinski, X. Jin, and J. V. Deshmukh. Efficient guiding strategies for testing of temporal properties of hybrid systems. In K. Havelund, G. J. Holzmann, and R. Joshi, editors, NASA Formal Methods - 7th Int. Symp., NFM 2015, volume 9058 of LNCS, pages 127–142. Springer, 2015. [12] T. Dreossi, A. Donzé, and S. A. Seshia. Compositional falsification of cyber-physical systems with machine learning components. In C. Barrett, M. Davies, and T. Kahsai, editors, NASA Formal Methods - 9th Int. Symp., NFM 2017, volume 10227 of LNCS, pages 357–372, 2017. [13] P. S. Duggirala, S. Mitra, M. Viswanathan, and M. Potok. C2E2: A verification tool for stateflow models. In C. Baier and C. Tinelli, editors, Tools and Algorithms for the Construction and Analysis of Systems - 21st Int. Conf., TACAS 2015, volume 9035 of LNCS, pages 68–82. Springer, 2015. [14] G. E. Fainekos and G. J. Pappas. Robustness of temporal logic specifications for continuous-time signals. Theor. Comput. Sci., 410(42):4262–4291, 2009. [15] B. Hoxha, H. Abbas, and G. E. Fainekos. Benchmarks for temporal logic requirements for automotive systems. In G. Frehse and M. Althoff, editors, 1st and 2nd Int. Workshops on Applied veRification for 16 Time-Staging Enhancement of Hybrid System Falsification Continuous and Hybrid Systems, ARCH@CPSWeek 2014 and 2105, volume 34 of EPiC Series in Computing, pages 25–30. EasyChair, 2014. [16] B. Hoxha, A. Dokhanchi, and G. E. Fainekos. Mining parametric temporal logic properties in model-based design for cyber-physical systems. STTT, 20(1):79–93, 2018. [17] X. Jin, J. V. Deshmukh, J. Kapinski, K. Ueda, and K. R. Butts. Powertrain control verification benchmark. In M. Fränzle and J. Lygeros, editors, 17th International Conference on Hybrid Systems: Computation and Control (part of CPS Week), HSCC’14, Berlin, Germany, April 15-17, 2014, pages 253–262. ACM, 2014. [18] J. Kapinski, J. V. Deshmukh, X. Jin, H. Ito, and K. Butts. Simulation-based approaches for verification of embedded control systems: an overview of traditional and advanced modeling, testing, and verification techniques. IEEE Control Systems, 36(6):45–64, 2016. [19] E. S. Kim, M. Arcak, and S. A. Seshia. Directed specifications and assumption mining for monotone dynamical systems. In A. Abate and G. E. Fainekos, editors, Proceedings of the 19th International Conference on Hybrid Systems: Computation and Control, HSCC 2016, Vienna, Austria, April 12-14, 2016, pages 21–30. ACM, 2016. [20] J. Kurátko and S. Ratschan. Combined global and local search for the falsification of hybrid systems. In A. Legay and M. Bozga, editors, Formal Modeling and Analysis of Timed Systems - 12th Int. Conf., FORMATS 2014, volume 8711 of LNCS, pages 146–160. Springer, 2014. [21] O. Maler and D. Nickovic. Monitoring temporal properties of continuous signals. In Y. Lakhnech and S. Yovine, editors, Formal Techniques, Modelling and Analysis of Timed and Fault-Tolerant Systems, Joint Int. Confs. on Formal Modelling and Analysis of Timed Systems, FORMATS 2004 and Formal Techniques in Real-Time and Fault-Tolerant Systems, FTRTFT 2004, volume 3253 of LNCS, pages 152–166. Springer, 2004. [22] D. Ulus, T. Ferrère, E. Asarin, and O. Maler. Online timed pattern matching using derivatives. In M. Chechik and J. Raskin, editors, Tools and Algorithms for the Construction and Analysis of Systems - 22nd Int. Conf., TACAS 2016, volume 9636 of LNCS, pages 736–751. Springer, 2016. [23] D. Wolpert and W. G. Macready. No free lunch theorems for optimization. IEEE Trans. Evolutionary Computation, 1(1):67–82, 1997. [24] A. Zutshi, J. V. Deshmukh, S. Sankaranarayanan, and J. Kapinski. Multiple shooting, cegar-based falsification for hybrid systems. In T. Mitra and J. Reineke, editors, 2014 International Conference on Embedded Software, EMSOFT 2014, New Delhi, India, October 12-17, 2014, pages 5:1–5:10. ACM, 2014. G. Ernst, I. Hasuo, Z. Zhang & S. Sedwards A 17 STL Semantics for Time-Bounded Signals Definition A.1 (robust semantics for time-bounded signals). Let T ∈ R>0 , w : [0, T ] → RN be a timebounded signal, and ϕ be an STL formula. We define the robustness Jw, ϕKT ∈ R ∪ {∞, −∞} of w with respect to ϕ as follows, by induction. Here the superscript T is an annotation that designates the time horizon.  Jw, f (x1 , . . . , xn ) > 0KT := f w(0)(x1 ), . . . , w(0)(xn ) Jw, ⊥KT := −∞ Jw, ¬ϕKT := −Jw, ϕKT Jw, ϕ1 ∧ ϕ2 KT := Jw, ϕ1 KT u Jw, ϕ2 KT l G T −t 0  0 T −t Jw, ϕ1 UI ϕ2 KT := Jwt , ϕ2 K u Jwt , ϕ1 K t∈I∩[0,T ] too. B t 0 ∈[0,t) The Boolean semantics |=, found e.g. in [10], allows a similar adaptation to time-bounded signals, Brzozowski Derivative of Flat STL Formulas In the time-staged falsification procedure we often encounter the following situation: an STL formula ϕ and a signal v : [0, T ] → RN are fixed; and we have to compute robustness Jv · v0 , ϕK for a number of different signals v0 : [0, T 0 ] → RN . To aid such computation, a natural idea is to use a syntactic construct ∂v ϕ of (Brzozowski) derivative. It should be compatible with robust semantics in the sense that Jv · v0 , ϕK = Jv0 , ∂v ϕK, reducing the computation of the LHS to that of the RHS. Similar use of derivatives is found e.g. in [22]. The settings are different, though: Boolean semantics is used in [22] while we use quantitative robust semantics. In fact, the definition of derivatives in this section focuses on flat formulas (i.e. free from nested modalities). This restriction is mandated by the quantitative semantics, as our proof later suggests. Anyway, the definitions and results in this section are new to the best of our knowledge. We need the following extension of STL syntax. Definition B.1 (extended STL). We extend the syntax of STL (Def. 3.4) by atomic propositions cr for each r ∈ R. The robust semantics in Def. 3.5 (and that in Def. A.1 in Appendix A) is extended accordingly: Jw, cr K := r. Intuitively cr is an atomic proposition that constantly returns the robustness value r. Definition B.2 (derivative). Let T ∈ R>0 , and v : [0, T ] → RN be a time-bounded signal. For each extended STL formula ϕ, we define its derivative ∂v ϕ by v by the following induction.  ∂v f (~x) > 0 :≡ cJv, f (~x)>0K ∂v cr :≡ cr ∂v ⊥ :≡ ⊥ ∂v (¬ϕ) :≡ ¬∂v ϕ ∂v (ϕ1 ∧ ϕ2 ) :≡ (∂v ϕ1 ) ∧ (∂v ϕ2 ) ∂v (ϕ1 UI ϕ2 ) :≡ cJv,ϕ1 UI ϕ2 K ∨ (cJv,2ϕ1 K ∧ ϕ1 ) UI−T ϕ2 Here the interval I −T is obtained from I by shifting both of its endpoints earlier by T , such as [a, b]−T = [a − T, b − T ]. Definition B.3 (flat STL formula). An STL formula ϕ is flat if it does not have nested temporal modal operators. This means: if ϕ1 UI ϕ2 is a subformula of ϕ, then neither ϕ1 nor ϕ2 contains U. Proposition B.4. Let T ∈ R>0 , v : [0, T ] → RN be a signal, and ϕ be a flat STL formula. We have, for each T 0 ∈ R>0 and v0 : [0, T 0 ] → RN , Jv0 , ∂v ϕK = Jv · v0 , ϕK. 18 Time-Staging Enhancement of Hybrid System Falsification Proof. By induction on the construction of ϕ. Most equalities below follow from the definition of ∂ and that of J K.  Jv0 , ∂v f (~x) > 0 K = Jv0 , cJv, f (~x)>0K K = Jv, f (~x) > 0K = Jv · v0 , f (~x) > 0K Jv0 , ∂v cr K = Jv0 , cr K = r = Jv · v0 , cr K Jv0 , ∂v ⊥K = Jv0 , ⊥K = −∞ = Jv · v0 , ⊥K I.H. Jv0 , ∂v (¬ϕ)K = Jv0 , ¬∂v ϕK = −Jv0 , ∂v ϕK = −Jv · v0 , ϕK = Jv · v0 , ¬ϕK Jv0 , ∂v (ϕ1 ∧ ϕ2 )K = Jv0 , (∂v ϕ1 ) ∧ (∂v ϕ2 )K = Jv0 , ∂v ϕ1 K u Jv0 , ∂v ϕ2 K I.H. = Jv · v0 , ϕ1 K u Jv · v0 , ϕ2 K = Jv · v0 , ϕ1 ∧ ϕ2 K Here is a nontrivial case. Jv0 , ∂v (ϕ1 UI ϕ2 )K  = Jv0 , cJv,ϕ1 UI ϕ2 K K t Jv0 , (cJv,2ϕ1 K ∧ ϕ1 ) UI−T ϕ2 K = Jv, ϕ1 UI ϕ2 K t = G t∈I∩[0,T ] t (∗) = t 00 ∈I∩[T,T +T 0 ] t∈I∩[0,T ] = l t0 t 0 ∈[0,t) t 00 ∈I∩[T,T +T 0 ] G t∈I∩[0,T +T 00 ] 00 J(v · v0 )t , ϕ2 K u l J(v · v0 )t , ϕ2 K u l  t 0 ∈[0,t) l 0 J(v · v0 )t , ϕ1 K  t 0 ∈[T,T +t) t 0 ∈[0,T ] t0 l 0 Jv0t , ϕ1 K l   0 0 J(v · v0 )t , ϕ1 K u J(v · v0 )t , ϕ1 K  J(v · v0 ) , ϕ1 K t 0 ∈[0,t) 00 l  0 J(v · v0 )t , ϕ1 K t 0 ∈[0,t) l  0 Jvt , ϕ1 K u t 0 ∈[0,T ] l J(v · v0 )t , ϕ2 K u = Jv · v0 , ϕ1 UI ϕ2 K l Jv0t , ϕ2 K u J(v · v0 )t , ϕ2 K u G Jv0t , ϕ2 K u Jv, 2ϕ1 K u  Jv , ϕ1 K J(v · v0 )t , ϕ2 K u G G t Jvt , ϕ2 K u t∈(I−T )∩[0,T 0 ] t∈I∩[0,T ] = t∈(I−T )∩[0,T 0 ] G G t G t 0 ∈[T,t 00 ) 0 J(v · v0 )t , ϕ1 K  t 0 ∈[0,t 00 )  0 J(v · v0 )t , ϕ1 K t 0 ∈[0,t) In (∗) we used the following facts. Firstly, for a formula ψ without temporal operators, we have Jv, ψK = Jv · v0 , ψK. Secondly, if v’s domain is [0, T ] and t ∈ [0, T ], then vt · v0 = (v · v0 )t . Note that the flatness assumption on ϕ is crucially used in the proof step. Modifying Def. 4.2 in order to accommodate nested modalities seems hard, after analyzing the proof step (∗). G. Ernst, I. Hasuo, Z. Zhang & S. Sedwards C C.1 19 Omitted Proofs Proof of Prop. 5.4 Proof. By the definitions we have, for each input signal u : [0, T ] → RM , l JM (u), 2(x < c)K = c − M (u)(t)(x) . t∈[0,T ] Therefore the assumption JM (u1 ), ϕK ≤ JM (u01 ), ϕK expands to l l c − M (u1 )(t)(x) ≤ c − M (u01 )(t)(x) . t∈[0,T1 ] (4) t∈[0,T1 ] The first infimum in the above is taken over a compact domain [0, T1 ]; therefore there exists T ∈ [0, T1 ] that achieves the infimum. Let T be such a real number. The following is obvious. l l c − M (u1 )(t)(x) = c − M (u1 )(t)(x) = c − M (u1 )(T )(x) t∈[0,T1 ] ≤ l t∈[0,T ] c − M (u01 )(t)(x) ≤ t∈[0,T1 ] l c − M (u01 )(t)(x) . (5) t∈[0,T ] Another immediate consequence, derived using the causality of M (Def. 3.2), is c − M (u1 |[0,T ] )(T )(x) ≤ c − M (u01 |[0,T ] )(T )(x) . (6) Our goal is to show JM (u1 |[0,T ] · u2 ), ϕK ≤ JM (u01 |[0,T ] · u2 ), ϕK. JM (u01 |[0,T ] · u2 ), 2(x < c)K l = c − M (u01 |[0,T ] · u2 )(t)(x) t∈[0,T +T2 ] = l c − M (u01 )(t)(x) u t∈[0,T ] ≥ ≥ l c − M (u1 )(t)(x) u t∈[0,T ] l c − M (u1 )(t)(x) u t∈[0,T ] = ··· l c − M (u01 |[0,T ] · u2 |[0,t−T ] )(t)(x) (∗) t∈(T,T +T2 ] l c − M (u01 |[0,T ] · u2 |[0,t−T ] )(t)(x) by (4) t∈(T,T +T2 ] l c − M (u1 |[0,T ] · u2 |[0,t−T ] )(t)(x) (†) t∈(T,T +T2 ] = JM (u1 |[0,T ] · u2 ), 2(x < c)K. In the above we heavily used the causality of M (Def. 3.2). For example, in the step (∗) above, causality is used in deriving M (u01 |[0,T ] · u2 )(t) = M (u01 )(t). In the step (†) we applied the monotonicity of M to the signals u1 |[0,T ] , u01 |[0,T ] and u2 |[0,t−T ] . Note that (6) allows to do so.
3cs.SY
Vector Clocks in Coq: An Experience Report Christopher Meiklejohn arXiv:1406.4291v1 [cs.DC] 17 Jun 2014 Basho Technologies, Inc. [email protected] Abstract • Authoring code in a statically-typed language and performing extraction to a dynamically-typed language is problematic due to the required encoding of type information and constructors in the exported objects. This report documents the process of implementing vector clocks in the Coq proof assistant for extraction and use in the distributed Dynamo-inspired data store, Riak. In this report, we focus on the technical challenges of using Core Erlang code extracted from the proof assistant in a production-grade Erlang application, as opposed to verification of the model itself. • Required use of a subset of the source language due to inherent differences in the implementation of the destination language; for example, extracting functions that rely on currying to a language that does not support currying. Categories and Subject Descriptors D.1.3 [Programming Techniques]: Concurrent Programming; D.3.3 [Programming Techniques]: Language Constructs and Features - abstract data types, patterns, control structures General Terms tion Keywords • Required use of an adapter layer to perform translations be- tween the native data structures of the destination language and the exported data structures of the source language. The main contributions of this work are the following: Algorithms, Design, Reliability, Formal Verifica- • A Coq module providing an implementation of vector clocks. Dynamo, Erlang, Riak, vector clocks • An extracted Erlang module providing an implementation of vector clocks, authored in Core Erlang, extracted from the Coq proof assistant. 1. Introduction • An Erlang module used as glue code that provides helpers to In a distributed system, where data structures are replicated or shared between multiple communicating nodes, it is highly desirable to have a method for asserting that certain invariants are preserved after objects are manipulated. In industry, one of the major approaches is using property-based testing utilities such as QuickCheck [1] to verify that certain properties hold true over multiple inter-leavings of operations applied to these data structures. However, given the amount of possible executions as the number of nodes or processes grow, exhausting that state space becomes more difficult. An alternative approach for this type of verification is to use an interactive theorem prover. We explore vvclocks [9], an Erlang library that aims to provide a formally verified implementation of vector clocks for use in Erlang applications. We describe the process of building this implementation, as described in 1988 [4], in the Coq [5] proof assistant and walk through the process of using Coq’s code extraction utilities to generate Core Erlang [14] code. vvclocks leverages the verlang project [3], an extension to Coq providing extraction of Core Erlang, written by Tim Carstens. Given the differences between the source and target languages, we outline the process of adapting both the generated Core Erlang and Coq source to properly execute and compile in the Erlang environment. To verify the applicability of this implementation, we replace the vector clock implementation in the open source Dynamo-based data store, Riak. We explore the process of writing an adapter to translate between the Coq-derived data structures and the base Erlang data structures, adapting the existing test suite to pass, as well as running the Riak data store with the newly extracted implementation of vector clocks. Based on our experience, we have identified a series of issues that makes this approach not immediately viable for use in production systems. The issues are the following: perform type conversion between native Erlang types and Coq’s data structures. • Details of where the extraction process was difficult; e.g. man- ual modifications to the extracted source or non-idiomatic Coq implementations. The following goals are out of scope: • We do not explore verification of the vector clock model, specif- ically because the proofs are tied very closely to the implementation of data structures that have changed extensively during the development of this research. • We do not focus on using the most efficient implementation of the data structures. We have found that a simple representation although inefficient, is much easier to debug. 2. Background The verified vector clock implementation discussed below is authored in the Coq [5] proof assistant. Coq’s code extraction utilities are used to generate Core Erlang [14] code, using a supporting library called verlang [3]. This Core Erlang code is then compiled and used to replace one of the modules in the Erlang-based data store Riak. The following subsections provide background information about Coq, Core Erlang, verlang, and vector clocks. 2.1 Coq Coq is an interactive theorem prover that implements a dependentlytyped functional programming language called Gallina. In addition, Coq provides a vernacular for stating function definitions and theorems and a tactic language for allowing the user to define custom 1 2014/6/18 1 2 3 4 5 6 7 8 9 10 (∗∗ Less than or equal to comparson for natural numbers. ∗) Fixpoint ble nat (n m : nat) {struct n} : bool := match n with | O => true | S n’ => match m with | O => false | S m’ => ble nat n’ m’ end end. 1 2 3 4 (∗∗ Type definitions for actors, counts and timestamps. ∗) Definition actor := nat. Definition count := nat. Definition timestamp := nat. 5 6 7 (∗∗ Model clocks as triples. ∗) Definition clock := prod actor (prod count timestamp). 8 9 10 (∗∗ Model vector clocks as a list of clock triples. ∗) Definition vclock := (list clock)%type. 11 Figure 1. Fixpoint that computes less than or equal to over two natural numbers. Example taken from Benjamin C. Pierce’s ”Software Foundations”. [10] 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 12 13 (∗∗ Create a vector clocks. ∗) Definition fresh : vclock := nil. Figure 3. Types specifications for vector clocks in Coq. ’ble nat’/2 = fun ( n, m) −> case n of ’O’ when ’true’ −> ’True’ {’S’, n@} when ’true’ −> case m of ’O’ when ’true’ −> ’False’ {’S’, m@} when ’true’ −> call ’vvclock’:’ble nat’ ( n@ , m@ ) end end • Currying is not supported in Core Erlang and any code that relies on this produces code that does not execute correctly during extraction. • While Core Erlang differentiates between inter- and intra- mod- ule calls, the extraction code treats all calls to functions as intermodule calls, and fully qualifies them with the module name. • There is currently no known good way to handle the Erlang receive primitive. 2.4 Vector Clocks Vector clocks [7] provide a mechanism for reasoning about the relationship of events in a distributed system, specifically focusing on differentiating between causal and concurrent events. Vector clocks are traditionally modeled as a list of pairs, each pair being composed of an actor and an operation count. As events occur in the system, the entire vector clock is shipped with the event. Each actor in the event increments its own count. This allows for a partial ordering to be established over events in the system by comparing each actor’s count between two vector clocks. We are able to reason about events that happen concurrently or causally by calculating whether the vector clocks descend or diverge from one another. In the Riak system, objects are replicated across a series of nodes, and each object is stored with a vector clock representing each actor’s modifications of the object. When dealing with divergent values across replicas, vector clocks are used to reason about whether the value is a descendent of another, and effectively replaces it, or was written concurrently as another object, in that case both values need to be preserved. Figure 2. Generated Core Erlang function that computes less than or equal to over two natural numbers, modeled using Peano numbers. Note the recursive call to itself at line 10. proof methods. Coq also provides the ability to perform extraction of certified programs as Scheme, Haskell or OCaml code through this vernacular. Figure 1 provides an example of a fixed point computation authored in Coq. 2.2 Core Erlang Core Erlang [2] is an intermediate representation of the Erlang programming language, designed to be operable on by other programs. The major design goals of Core Erlang are providing a simple grammar that can be converted to normal Erlang programs, as well as providing a regular, concise structure to allow for code-walking tools to operate on programs translated to Core Erlang. While the semantics of Core Erlang are not the focus of this paper, there is a sample Core Erlang function in Figure 2. 3. Implementation The following subsections deal with the implementation of the vector clocks in Coq, the problems with the extracted Core Erlang code, and writing the adapter layer in Erlang. 2.3 verlang Verlang [3] is an experimental project by Tim Carstens that adds the Core Erlang extraction target to Coq, in an attempt to enable formally verified Erlang development with the Coq proof assistant. verlang provides a mapping between MiniML, the language that serves as an intermediary translation step during extraction, to Core Erlang. There are a number of interesting problems to address during the translation, which is outlined by Tim Carstens in the verlang code repository. [3] In summary: 3.1 Vector clocks in Coq The package inside of Riak that provides the Erlang vector clock implementation is riak core. In implementing vector clocks in Coq for use in Riak, we attempt to stick as closely as possible to the existing API. For simplicity, we model each clock as a triple of natural numbers representing the actor, current count, and timestamp. This can be seen in Figure 3. The timestamps are used for pruning only, and are not critical to the semantics of the vector clock. These timestamps are modeled as monotonically advancing natural numbers, similar to a UNIX timestamp. We model vector clocks as a list of these clocks. • While Coq supports module nesting, there is no such concept in Core Erlang. In the extracted code, nested modules are supported by name mangling. 2 2014/6/18 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 (∗∗ Increment a vector clock. ∗) Definition increment (actor : actor) (vclock : vclock) := match find (fun clock => match clock with | pair x => beq nat actor x end) vclock with | None => cons (pair actor (pair init count init timestamp)) vclock | Some (pair x (pair count timestamp)) => cons (pair x (pair (incr count count) (incr timestamp timestamp))) (filter (fun clock => match clock with | pair x => negb (beq nat actor x) end) vclock) end. Figure 4. increment function for a vector clock. 1 2 Figure 5. f resh vector clock data constructor 1 2 3 4 1 3 4 5 6 7 8 • f resh call ’vvclock.VVClock’:’ble nat’ ( actor , a ) Figure 6. Incorrectly qualified call with module name. 2 The API we provide mimics the API exposed by riak core. This includes the following functions: ’fresh’/0 = fun () −> [] 9 ’descends’/2 = fun ( vc1, vc2) −> case call ’Coq.Lists.List’:’fold left’ ( ’descends@’ , vc2 , { ’Pair’ , ’True’ , vc1 } ) of Used to generate a new, empty vector clock • increment Figure 7. Missing arity when passing function to the fold left call next to ′ descends@′ . Used to increment the vector clock for a particular actor. • equal Used to test equality between two vector clocks. 3.2 Code extraction to Core Erlang • descends When performing the code extraction, we are required to extract not only the code of the vector clocks implementation itself, but also all of the supporting Coq libraries used in our implementation. In our example, this includes modules supporting the core Coq datatypes: specifically Peano numbers, the List module, and a module providing arithmetic equality over natural numbers. In extracting our vector clock library to Core Erlang, we encounter numerous problems with our implementation, but none with the extraction of the core libraries. The following subsections provide details and an example of each of these issues, with the workaround we identified. Used to determine if one vector clock is an ancestor of another. • merge Used to merge two vector clocks. • get counter Used to extract the count for a particular actor. • get timestamp Used to extract the timestamp for a particular actor. • all nodes Used to extract all actors that have incremented the vector clock. 3.2.1 Missing data constructors • prune The first problem we experience is that the data constructors do not export correctly. Figure 5 shows our addition to the code to resolve the problem of the missing data constructor, f resh. Given a series of timestamps representing the bounds for pruning, prune the vector clock and return the result of pruning. Not included in this paper, but available in the provided source code, is the implementation of all of these functions. Figure 4 is an example showing how the increment function destructures the vector clock, attempts to locate the appropriate actor if it exists, and either adds a new actor or increments an existing actors count. There are a few important things to note about this example: 3.2.2 Incorrectly qualified calls We also run into problems related to calls that are incorrectly qualified that have to be manually modified. Specifically, we run into cases where files containing modules are overqualified, with the filename repeated in the module name (as in Figure 6) and cases where function calls are missing their arity, resulting in a failed function call at runtime (as in Figure 7). The workaround we identified for dealing with these issues was to manually modify the extracted code to fix the locations with missing arities, or incorrectly qualified functions. • We abstract the increment function that operates over times- tamps. This is important because Coq has no notion of generating an Erlang, or UNIX, timestamp. We then patch the generated Core Erlang code to call timestamp aware replacements. • There are two anonymous functions passed as arguments here 3.2.3 Lack of currying — to the find and filter functions respectively — that in our initial version of the library were abstracted into their own function, as they are used by many of the API functions. Because this relies on the use of partial application, these functions had to be inlined to ensure the extracted code compiled correctly. Figure 8 provides an example of using currying in Coq. In this example, the f ind′′ method returns a closure that returns true if a given actor is associated with this clock. In this example, the extracted Core Erlang code does execute, but has turned a arity-1 call that returned a function into a arity-2 call that immediately computes a final result. 3 2014/6/18 1 2 3 4 5 (∗∗ Source Coq function definition. ∗) Definition find’’ (actor : actor) := fun clock : clock => match clock with | pair x => negb (beq nat actor x) end. 6 7 8 9 10 11 12 13 14 15 16 17 18 1 2 3 4 5 6 %% Generated Core Erlang function. ’find@’/2 = fun ( actor, clock) −> case clock of { ’Pair’ , c , x } when ’true’ −> call ’Coq.Arith.EqNat’:’beq nat’ ( actor , c ) end 7 8 9 10 11 12 13 14 15 16 17 2 3 4 5 %% @doc Convert a natural number into a peano. natural to peano(0) −> ’O’; natural to peano(Natural) −> {’S’, natural to peano(Natural − 1)}. 6 7 8 9 10 11 1 2 3 %% @doc Return natural timestamp. timestamp() −> calendar:datetime to gregorian seconds(erlang:universaltime()). 4 5 %% @doc Convert a peano number into a natural. peano to natural(’O’) −> 0; peano to natural({’S’, Peano}) −> 1 + Peano. %% @doc Determine if one vector clock is an ancestor of another. descends(VClock1, VClock2) −> case vvclock:descends(VClock1, VClock2) of ’True’ −> true; ’False’ −> false end. Figure 10. Wrapper functions for converting between Coq’s boolean type and Erlang’s representation of booleans as atoms when dealing with equalities or inequalities. Figure 8. Coq code that generates incorrect Erlang code when extracted. 1 %% @doc Compare equality of two vclocks. equal(VClock1, VClock2) −> case vvclock:equal(VClock1, VClock2) of ’True’ −> true; ’False’ −> false end. 6 7 %% @doc Peanoized timestamp. peano timestamp() −> term to peano(timestamp()). Figure 11. Conversion of Erlang’s representation of UNIX timestamps to Peano numbers. Figure 9. Conversion of Peano numbers to Erlang’s built in numerics. ture: {′ S ′ , {′ S ′ ,′ O′ }}. Figure 9 shows our wrapper function used to convert back and forth between these data types. Next, we look at booleans which are implemented as part of the Coq.Bool.Bool module. In Coq, booleans are modeled as an abstract data type with two constructors T rue and F alse. In Erlang, booleans are modeled using atoms, specifically the true and f alse atom, which are immutable constants. We write a series of small wrapper function, as seen in Figure 10, which call into our exported module, and pattern match on the return value. The workaround we identified for this issue was to manually inline all of these function definitions, which, in addition to being non-idiomatic, leads to a large amount of code duplication in the implementation. 3.3 Erlang adapter between Riak and vvclocks Next, we look at the adapter layer that is required to convert between the Coq-derived data structures exported to Core Erlang and the data structures provided by Erlang. The following subsections look at the conversion of Peano numbers, boolean types, UNIX timestamps, and data located in the application environment. 3.3.2 Timestamps Timestamps are another area where we have to provide wrapper functions for performing type conversions. As Coq has no notion of what a UNIX timestamp is, our original implementation of the prune method models timestamps as monotonically advancing natural numbers. To account for this, we provide helper functions that are used to convert Erlang’s notion of a timestamp to a Peano number that can be used by the functions exported by Coq. Figure 11 provides an example of this conversion. 3.3.1 Type conversion verlang models the abstract data types provided by Coq using nsized tuples with the first position of the tuple representing the constructor name, and the remaining n − 1 slots of the tuple as the constructor arguments. Nested function applications, as used in recursive or inductively defined data types, are modeled as nested tuples. We see two examples of this below, specifically when working with booleans and natural numbers. First, we look at the modeling of naturals in Coq, using Peano numbers as provided by Coq.Init.P eano. Concretely, these numbers are modeled using an inductive data type with two constructors: one for the base case that takes no arguments O, and one for the inductive case that takes an argument of a Peano number, S. For example, to represent the Arabic numeral 2 the following would be used: S S O. This is translated to the following Erlang struc- 3.3.3 Actors The original increment function arguments, as provided by riak core’s vector clock implementation, includes an actor ID, which is typically an atom but can be any Erlang term. In our Coq implementation, we model this as string, which is an inductively defined data type over ASCII characters, which are themselves defined inductively over eight bits. See Figure 12 for an example. 4 2014/6/18 1 on type specifications in the form of annotations, as well as debugging symbols in the compiled byte code; neither are supported by Core Erlang. Definition zero := Ascii false false false false false false false false. Figure 12. Coq representation of ASCII character.[6] 5. Future Work 1 2 3 4 5 6 7 Based on our experience of implementing the vvclocks library, we have identified a series of work items that would improve the viability of this approach for use in production systems authored in Erlang. %% @doc Prune vclocks. prune(VClock, Timestamp, BProps) −> Old = term to peano(get property(old vclock, BProps)), Young = term to peano(get property(young vclock, BProps)), Large = term to peano(get property(large vclock, BProps)), Small = term to peano(get property(small vclock, BProps)), vvclock:prune(VClock, Small, Large, Young, Old). 5.1 Bugs in verlang It is clear that some of the issues we experienced during extraction of the vector clocks implementation are related to bugs in the verlang extraction process. Specifically, we refer to the following issues: Figure 13. Wrapper function for dealing with information that must come from other parts of the system. In this example, get property is retrieving information from a Riak bucket, but could easily also be accessing something stored in the runtime environment. The function term to peano is used to convert the Erlang timestamp format to a Peano number, leveraging the datetime to gregorian seconds function. • Incorrectly qualified function calls. (Section 3.2.2) • Missing data constructors. (Section 3.2.3) It appears that these bugs can be addressed in the extraction process. 3.3.4 Environment variables 5.2 Other Applications The original prune function, as provided by riak core’s vector clock implementation, takes three arguments: the vector clock, a timestamp, and an object that provides a series of settings which should apply to the vector clock, specifically referred to as bucket properties. A simple dictionary-like data structure stores these properties. Given that we cannot operate over this data structure in Coq, we break apart our certif ied prune function as implemented in Coq from our wrapper function, which extracts these values out of the environment and directly passes them as formal arguments. Figure 13 provides an example of this. The original motivation for this work was to take the partially verified implementations [8] of two convergent replicated data types, specifically the G-Counter and PN-Counter as described by Shapiro, et. al [13] and extract them for use in the Riak system. In performing the extraction of this code, we ran into similar problems as discussed in 5.1. However, rather than these issues occurring in the vector clocks implementation we authored, they occurred in the extraction of support libraries for the built-in data structures provided by Coq. This made working with and debugging this implementation more difficult, because of the lack of control over the internal implementation of Coq’s data structures. Beacuse of this, we decided to simplify the model and focus on the implementation of a vector clock library using the basic list data structures provided by Coq, which provides similar semantics to a G-Counter. 4. Evaluation When evaluating the vvclock package with the Riak data store, we immediately ran into problems related to how the inductive data structures have been modeled in Erlang. We will explore one particular example, related to timestamps. For example, to get the current time in Erlang and convert it to a UNIX timestamp takes anywhere between 5 - 8 microseconds. However, a timestamp for the year 2014 to a Peano number takes much longer given the nested tuple encoding and how memory is allocated on the stack for tuples. In a matter of minutes we were able to exhaust just over 8.0 GB of memory in attempting to encode a timestamp like 1390525760, before the process terminated because we ran out of virtual memory. To allow us to continue testing the rest of the library, we modified the timestamp encoding function to store the timestamp as a much smaller value which was easier to encode. Once completed, we were able to successfully validate the existing test suite against our exported module. While we were able to successfully validate that our unit tests passed, we still cannot use the library in a production Riak instance because of the problems of data structure encoding. In addition, even though the vvclock module itself may be verified, the vclock module, which provides the adapter layer between the vvclock module and the rest of the system, must still be fully tested to verify correct behavior. This most likely would be done using Erlang’s existing unit testing utility, eunit, or the aforementioned QuickCheck. Finally, debugging of modules compiled from Core Erlang is still difficult. Tooling such as Dialyzer [11] and Proper [12] all rely 5.3 Adapter Layer It is highly desirable to eliminate the adapter layer for several reasons: • P erf ormance The process of converting Coq data structures into Erlang data structures requires the encoding of constructor calls, in order to retain typing information in a language without abstract data types. The overhead of performing this conversion is high, especially when dealing with large or complex data structures such as timestamps. • T esting Erlang functions must be made to perform conver- sions of data structures to and from their Coq representations. This is problematic in that it requires an additional level of testing where bugs in the creation of objects can be introduced. However, this is problematic. Consider the case of naturals, modeled as Peano numbers: any axioms that are defined, or theorems proven about functions executing over these types is only guaranteed because of the way the data structure has been modeled. Specifically with inductive types, certain functions are only guaranteed to terminate because of properties held by the inductive hypothesis. In the case of naturals, if we modeled this using Erlang’s built in integer type, none of the properties regarding subtraction would hold once we passed zero. Regardless, we feel it is worth exploring an alternate means of extraction that attempts to leverage the built-in Erlang types more 5 2014/6/18 efficiently; sacrificing some safety for performance. For example, a function operating over a Coq string would be exported to Core Erlang code that operated over an Erlang binary instead of the tuple encoded Coq structure. [14] Uppsala University. Core erlang initiative. http://www.it.uu.se/research/group/hipe/cerl/. 5.4 QuickCheck Property Generation Finally, we feel that the ability to generate QuickCheck models instead of generating source code for the data structures might be a more viable approach. Based on a series of of axioms defined regarding the Erlang run-time system, and given the theorems defined in the proof assistant, we could generate a series of properties to guide development of the data structures that preserve invariants. A. Code Availability All of the code discussed is available on GitHub under the Apache 2.0 License at http://github.com/cmeiklejohn/vvclocks. Acknowledgments Thanks to Andy Gross, Andrew Stone, Ryan Zezeski, Scott Lystig Fritchie, and Reid Draper for providing feedback during the research and development of this work. References [1] T. Arts, L. M. Castro, and J. Hughes. Testing erlang data types with quviq quickcheck. In Proceedings of the 7th ACM SIGPLAN Workshop on ERLANG, ERLANG ’08, pages 1–8, New York, NY, USA, 2008. ACM. ISBN 978-1-60558-065-4. . URL http://doi.acm.org/10.1145/1411273.1411275 . [2] R. Carlsson, B. Gustavsson, E. Johansson, T. Lindgren, S.-O. Nyström, M. Pettersson, and R. Virding. Core Erlang 1.0 language specification. Technical Report 2000-030, Department of Information Technology, Uppsala University, Nov. 2000. [3] T. Carstens. verlang source code repository. https://github.com/tcarstens/verlang. [4] C. J. Fidge. Timestamps in message-passing systems that preserve the partial ordering. Proceedings of the 11th Australian Computer Science Conference, 10(1):5666, 1988. URL http://sky.scitech.qut.edu.au/~ fidgec/Publications/fidge88a.pdf . [5] INRIA. The coq proof assistant homepage. http://coq.inria.fr, . [6] INRIA. Library coq.strings.ascii. http://coq.inria.fr/library/Coq.Strings.Ascii.html, . [7] L. Lamport. Time, clocks, and the ordering of events in a distributed system. Commun. ACM, 21(7):558–565, July 1978. ISSN 0001-0782. . URL http://doi.acm.org/10.1145/359545.359563. [8] C. S. Meiklejohn. distributed-data-structures source code repository. https://github.com/cmeiklejohn/distributed-data-structures , . [9] C. S. Meiklejohn. vvclocks source code https://github.com/cmeiklejohn/vvclocks, . repository. [10] B. C. Pierce. Software foundations. http://www.cis.upenn.edu/~ bcpierce/sf/. [11] K. Sagonas. Detecting defects in erlang programs using static analysis. In Proceedings of the 9th ACM SIGPLAN International Conference on Principles and Practice of Declarative Programming, PPDP ’07, pages 37–37, New York, NY, USA, 2007. ACM. ISBN 978-1-59593-769-8. . URL http://doi.acm.org/10.1145/1273920.1273926 . [12] Sagonas, Kostis et. al. Proper source code repository. https://github.com/manopapad/proper . [13] M. Shapiro, N. Preguiça, C. Baquero, and M. Zawirski. A comprehensive study of Convergent and Commutative Replicated Data Types. Rapport de recherche RR-7506, INRIA, Jan. 2011. URL http://hal.inria.fr/inria-00555588 . 6 2014/6/18
6cs.PL